repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
imagefusion-rfn-nest
imagefusion-rfn-nest-main/test_40pairs.py
# -*- coding:utf-8 -*- # @Author: Li Hui, Jiangnan University # @Email: hui_li_jnu@163.com # @File : test_40pairs.py # @Time : 2020/8/14 17:11 # test phase import os import torch from torch.autograd import Variable from net import NestFuse_light2_nodense, Fusion_network, Fusion_strategy import utils from args_fusion import args import numpy as np def load_model(path_auto, path_fusion, fs_type, flag_img): if flag_img is True: nc = 3 else: nc =1 input_nc = nc output_nc = nc nb_filter = [64, 112, 160, 208, 256] nest_model = NestFuse_light2_nodense(nb_filter, input_nc, output_nc, deepsupervision=False) nest_model.load_state_dict(torch.load(path_auto)) fusion_model = Fusion_network(nb_filter, fs_type) fusion_model.load_state_dict(torch.load(path_fusion)) fusion_strategy = Fusion_strategy(fs_type) para = sum([np.prod(list(p.size())) for p in nest_model.parameters()]) type_size = 4 print('Model {} : params: {:4f}M'.format(nest_model._get_name(), para * type_size / 1000 / 1000)) para = sum([np.prod(list(p.size())) for p in fusion_model.parameters()]) type_size = 4 print('Model {} : params: {:4f}M'.format(fusion_model._get_name(), para * type_size / 1000 / 1000)) nest_model.eval() fusion_model.eval() nest_model.cuda() fusion_model.cuda() return nest_model, fusion_model, fusion_strategy def run_demo(nest_model, fusion_model, fusion_strategy, infrared_path, visible_path, output_path_root, name_ir, fs_type, use_strategy, flag_img, alpha): img_ir, h, w, c = utils.get_test_image(infrared_path, flag=flag_img) # True for rgb img_vi, h, w, c = utils.get_test_image(visible_path, flag=flag_img) # dim = img_ir.shape if c is 1: if args.cuda: img_ir = img_ir.cuda() img_vi = img_vi.cuda() img_ir = Variable(img_ir, requires_grad=False) img_vi = Variable(img_vi, requires_grad=False) # encoder en_r = nest_model.encoder(img_ir) en_v = nest_model.encoder(img_vi) # fusion net if use_strategy: f = fusion_strategy(en_r, en_v) else: f = fusion_model(en_r, en_v) # decoder img_fusion_list = nest_model.decoder_eval(f) else: # fusion each block img_fusion_blocks = [] for i in range(c): # encoder img_vi_temp = img_vi[i] img_ir_temp = img_ir[i] if args.cuda: img_vi_temp = img_vi_temp.cuda() img_ir_temp = img_ir_temp.cuda() img_vi_temp = Variable(img_vi_temp, requires_grad=False) img_ir_temp = Variable(img_ir_temp, requires_grad=False) en_r = nest_model.encoder(img_ir_temp) en_v = nest_model.encoder(img_vi_temp) # fusion net if use_strategy: f = fusion_strategy(en_r, en_v) else: f = fusion_model(en_r, en_v) # decoder img_fusion_temp = nest_model.decoder_eval(f) img_fusion_blocks.append(img_fusion_temp) img_fusion_list = utils.recons_fusion_images(img_fusion_blocks, h, w) # ########################### multi-outputs ############################################## output_count = 0 for img_fusion in img_fusion_list: file_name = 'fused_' + str(alpha) + '_' + name_ir output_path = output_path_root + file_name output_count += 1 # save images utils.save_image_test(img_fusion, output_path) print(output_path) def main(): # False - gray flag_img = False # ################# gray scale ######################################## test_path = "images/40_pairs_tno_vot/ir/" path_auto = args.resume_nestfuse output_path_root = "./outputs/alpha_1e4_40/" if os.path.exists(output_path_root) is False: os.mkdir(output_path_root) fs_type = 'res' # res (RFN), add, avg, max, spa, nuclear use_strategy = False # True - static strategy; False - RFN path_fusion_root = args.fusion_model with torch.no_grad(): alpha_list = [700] w_all_list = [[6.0, 3.0]] for alpha in alpha_list: for w_all in w_all_list: w, w2 = w_all temp = 'rfnnest_' + str(alpha) + '_wir_' + str(w) + '_wvi_' + str(w2) output_path_list = 'fused_' + temp + '_40' output_path1 = output_path_root + output_path_list + '/' if os.path.exists(output_path1) is False: os.mkdir(output_path1) output_path = output_path1 # load network path_fusion = path_fusion_root + str(w) + '/' + 'Final_epoch_2_alpha_' + str(alpha) + '_wir_' + str( w) + '_wvi_' + str(w2) + '_ssim_vi.model' model, fusion_model, fusion_strategy = load_model(path_auto, path_fusion, fs_type, flag_img) imgs_paths_ir, names = utils.list_images(test_path) num = len(imgs_paths_ir) for i in range(num): name_ir = names[i] infrared_path = imgs_paths_ir[i] visible_path = infrared_path.replace('ir/', 'vis/') if visible_path.__contains__('IR'): visible_path = visible_path.replace('IR', 'VIS') else: visible_path = visible_path.replace('i.', 'v.') run_demo(model, fusion_model, fusion_strategy, infrared_path, visible_path, output_path, name_ir, fs_type, use_strategy, flag_img, temp) print('Done......') if __name__ == '__main__': main()
4,952
30.75
152
py
imagefusion-rfn-nest
imagefusion-rfn-nest-main/train_fusionnet.py
# Training a NestFuse network # auto-encoder import os os.environ["CUDA_VISIBLE_DEVICES"] = "0" import sys import time from tqdm import tqdm, trange import scipy.io as scio import random import torch from torch.optim import Adam from torch.autograd import Variable import utils from net import NestFuse_light2_nodense, Fusion_network from args_fusion import args import pytorch_msssim EPSILON = 1e-5 def main(): original_imgs_path, _ = utils.list_images(args.dataset_ir) train_num = 80000 original_imgs_path = original_imgs_path[:train_num] random.shuffle(original_imgs_path) # True - RGB , False - gray img_flag = False alpha_list = [700] w_all_list = [[6.0, 3.0]] for w_w in w_all_list: w1, w2 = w_w for alpha in alpha_list: train(original_imgs_path, img_flag, alpha, w1, w2) def train(original_imgs_path, img_flag, alpha, w1, w2): batch_size = args.batch_size # load network model nc = 1 input_nc = nc output_nc = nc nb_filter = [64, 112, 160, 208, 256] f_type = 'res' with torch.no_grad(): deepsupervision = False nest_model = NestFuse_light2_nodense(nb_filter, input_nc, output_nc, deepsupervision) model_path = args.resume_nestfuse # load auto-encoder network print('Resuming, initializing auto-encoder using weight from {}.'.format(model_path)) nest_model.load_state_dict(torch.load(model_path)) nest_model.eval() # fusion network fusion_model = Fusion_network(nb_filter, f_type) if args.resume_fusion_model is not None: print('Resuming, initializing fusion net using weight from {}.'.format(args.resume_fusion_model)) fusion_model.load_state_dict(torch.load(args.resume_fusion_model)) optimizer = Adam(fusion_model.parameters(), args.lr) mse_loss = torch.nn.MSELoss() ssim_loss = pytorch_msssim.msssim if args.cuda: nest_model.cuda() fusion_model.cuda() tbar = trange(args.epochs) print('Start training.....') # creating save path temp_path_model = os.path.join(args.save_fusion_model) temp_path_loss = os.path.join(args.save_loss_dir) if os.path.exists(temp_path_model) is False: os.mkdir(temp_path_model) if os.path.exists(temp_path_loss) is False: os.mkdir(temp_path_loss) temp_path_model_w = os.path.join(args.save_fusion_model, str(w1)) temp_path_loss_w = os.path.join(args.save_loss_dir, str(w1)) if os.path.exists(temp_path_model_w) is False: os.mkdir(temp_path_model_w) if os.path.exists(temp_path_loss_w) is False: os.mkdir(temp_path_loss_w) Loss_feature = [] Loss_ssim = [] Loss_all = [] count_loss = 0 all_ssim_loss = 0. all_fea_loss = 0. for e in tbar: print('Epoch %d.....' % e) # load training database image_set_ir, batches = utils.load_dataset(original_imgs_path, batch_size) fusion_model.train() count = 0 for batch in range(batches): image_paths_ir = image_set_ir[batch * batch_size:(batch * batch_size + batch_size)] img_ir = utils.get_train_images(image_paths_ir, height=args.HEIGHT, width=args.WIDTH, flag=img_flag) image_paths_vi = [x.replace('lwir', 'visible') for x in image_paths_ir] img_vi = utils.get_train_images(image_paths_vi, height=args.HEIGHT, width=args.WIDTH, flag=img_flag) count += 1 optimizer.zero_grad() img_ir = Variable(img_ir, requires_grad=False) img_vi = Variable(img_vi, requires_grad=False) if args.cuda: img_ir = img_ir.cuda() img_vi = img_vi.cuda() # get fusion image # encoder en_ir = nest_model.encoder(img_ir) en_vi = nest_model.encoder(img_vi) # fusion f = fusion_model(en_ir, en_vi) # decoder outputs = nest_model.decoder_eval(f) # resolution loss: between fusion image and visible image x_ir = Variable(img_ir.data.clone(), requires_grad=False) x_vi = Variable(img_vi.data.clone(), requires_grad=False) ######################### LOSS FUNCTION ######################### loss1_value = 0. loss2_value = 0. for output in outputs: output = (output - torch.min(output)) / (torch.max(output) - torch.min(output) + EPSILON) output = output * 255 # ---------------------- LOSS IMAGES ------------------------------------ # detail loss # ssim_loss_temp1 = ssim_loss(output, x_ir, normalize=True) ssim_loss_temp2 = ssim_loss(output, x_vi, normalize=True) loss1_value += alpha * (1 - ssim_loss_temp2) # feature loss g2_ir_fea = en_ir g2_vi_fea = en_vi g2_fuse_fea = f # w_ir = [3.5, 3.5, 3.5, 3.5] w_ir = [w1, w1, w1, w1] w_vi = [w2, w2, w2, w2] w_fea = [1, 10, 100, 1000] for ii in range(4): g2_ir_temp = g2_ir_fea[ii] g2_vi_temp = g2_vi_fea[ii] g2_fuse_temp = g2_fuse_fea[ii] (bt, cht, ht, wt) = g2_ir_temp.size() loss2_value += w_fea[ii]*mse_loss(g2_fuse_temp, w_ir[ii]*g2_ir_temp + w_vi[ii]*g2_vi_temp) loss1_value /= len(outputs) loss2_value /= len(outputs) # total loss total_loss = loss1_value + loss2_value total_loss.backward() optimizer.step() all_fea_loss += loss2_value.item() # all_ssim_loss += loss1_value.item() # if (batch + 1) % args.log_interval == 0: mesg = "{}\t Alpha: {} \tW-IR: {}\tEpoch {}:\t[{}/{}]\t ssim loss: {:.6f}\t fea loss: {:.6f}\t total: {:.6f}".format( time.ctime(), alpha, w1, e + 1, count, batches, all_ssim_loss / args.log_interval, all_fea_loss / args.log_interval, (all_fea_loss + all_ssim_loss) / args.log_interval ) tbar.set_description(mesg) Loss_ssim.append( all_ssim_loss / args.log_interval) Loss_feature.append(all_fea_loss / args.log_interval) Loss_all.append((all_fea_loss + all_ssim_loss) / args.log_interval) count_loss = count_loss + 1 all_ssim_loss = 0. all_fea_loss = 0. if (batch + 1) % (200 * args.log_interval) == 0: # save model fusion_model.eval() fusion_model.cpu() save_model_filename = "Epoch_" + str(e) + "_iters_" + str(count) + "_alpha_" + str(alpha) + "_wir_" + str(w1) + "_wvi_" + str(w2) + ".model" save_model_path = os.path.join(temp_path_model, save_model_filename) torch.save(fusion_model.state_dict(), save_model_path) # save loss data # pixel loss loss_data_ssim = Loss_ssim loss_filename_path = temp_path_loss_w + "/loss_ssim_epoch_" + str(args.epochs) + "_iters_" + str(count) + "_alpha_" + str(alpha) + "_wir_" + str(w1) + "_wvi_" + str(w2) + ".mat" scio.savemat(loss_filename_path, {'loss_ssim': loss_data_ssim}) # SSIM loss loss_data_fea = Loss_feature loss_filename_path = temp_path_loss_w + "/loss_fea_epoch_" + str(args.epochs) + "_iters_" + str(count) + "_alpha_" + str(alpha) + "_wir_" + str(w1) + "_wvi_" + str(w2) + ".mat" scio.savemat(loss_filename_path, {'loss_fea': loss_data_fea}) # all loss loss_data = Loss_all loss_filename_path = temp_path_loss_w + "/loss_all_epoch_" + str(args.epochs) + "_iters_" + str(count) + "_alpha_" + str(alpha) + "_wir_" + str(w1) + "_wvi_" + str(w2) + ".mat" scio.savemat(loss_filename_path, {'loss_all': loss_data}) fusion_model.train() fusion_model.cuda() tbar.set_description("\nCheckpoint, trained model saved at", save_model_path) # ssim loss loss_data_ssim = Loss_ssim loss_filename_path = temp_path_loss_w + "/Final_loss_ssim_epoch_" + str( args.epochs) + "_alpha_" + str(alpha) + "_wir_" + str(w1) + "_wvi_" + str(w2) + ".mat" scio.savemat(loss_filename_path, {'final_loss_ssim': loss_data_ssim}) loss_data_fea = Loss_feature loss_filename_path = temp_path_loss_w + "/Final_loss_2_epoch_" + str( args.epochs) + "_alpha_" + str(alpha) + "_wir_" + str(w1) + "_wvi_" + str(w2) + ".mat" scio.savemat(loss_filename_path, {'final_loss_fea': loss_data_fea}) # SSIM loss loss_data = Loss_all loss_filename_path = temp_path_loss_w + "/Final_loss_all_epoch_" + str( args.epochs) + "_alpha_" + str(alpha) + "_wir_" + str(w1) + "_wvi_" + str(w2) + ".mat" scio.savemat(loss_filename_path, {'final_loss_all': loss_data}) # save model fusion_model.eval() fusion_model.cpu() save_model_filename = "Final_epoch_" + str(args.epochs) + "_alpha_" + str(alpha) + "_wir_" + str( w1) + "_wvi_" + str(w2) + ".model" save_model_path = os.path.join(temp_path_model_w, save_model_filename) torch.save(fusion_model.state_dict(), save_model_path) print("\nDone, trained model saved at", save_model_path) def check_paths(args): try: if not os.path.exists(args.vgg_model_dir): os.makedirs(args.vgg_model_dir) if not os.path.exists(args.save_model_dir): os.makedirs(args.save_model_dir) except OSError as e: print(e) sys.exit(1) if __name__ == "__main__": main()
8,542
33.447581
181
py
imagefusion-rfn-nest
imagefusion-rfn-nest-main/pytorch_msssim/__init__.py
import torch import torch.nn.functional as F from math import exp import numpy as np def gaussian(window_size, sigma): gauss = torch.Tensor([exp(-(x - window_size//2)**2/float(2*sigma**2)) for x in range(window_size)]) return gauss/gauss.sum() def create_window(window_size, channel=1): _1D_window = gaussian(window_size, 1.5).unsqueeze(1) _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0) window = _2D_window.expand(channel, 1, window_size, window_size).contiguous() return window def ssim(img1, img2, window_size=11, window=None, size_average=True, full=False, val_range=None): # Value range can be different from 255. Other common ranges are 1 (sigmoid) and 2 (tanh). if val_range is None: if torch.max(img1) > 128: max_val = 255 else: max_val = 1 if torch.min(img1) < -0.5: min_val = -1 else: min_val = 0 L = max_val - min_val else: L = val_range padd = 0 (_, channel, height, width) = img1.size() if window is None: real_size = min(window_size, height, width) window = create_window(real_size, channel=channel).to(img1.device) mu1 = F.conv2d(img1, window, padding=padd, groups=channel) mu2 = F.conv2d(img2, window, padding=padd, groups=channel) mu1_sq = mu1.pow(2) mu2_sq = mu2.pow(2) mu1_mu2 = mu1 * mu2 sigma1_sq = F.conv2d(img1 * img1, window, padding=padd, groups=channel) - mu1_sq sigma2_sq = F.conv2d(img2 * img2, window, padding=padd, groups=channel) - mu2_sq sigma12 = F.conv2d(img1 * img2, window, padding=padd, groups=channel) - mu1_mu2 C1 = (0.01 * L) ** 2 C2 = (0.03 * L) ** 2 v1 = 2.0 * sigma12 + C2 v2 = sigma1_sq + sigma2_sq + C2 cs = torch.mean(v1 / v2) # contrast sensitivity ssim_map = ((2 * mu1_mu2 + C1) * v1) / ((mu1_sq + mu2_sq + C1) * v2) if size_average: ret = ssim_map.mean() else: ret = ssim_map.mean(1).mean(1).mean(1) if full: return ret, cs return ret def msssim(img1, img2, window_size=11, size_average=True, val_range=None, normalize=False): device = img1.device weights = torch.FloatTensor([0.0448, 0.2856, 0.3001, 0.2363, 0.1333]).to(device) levels = weights.size()[0] mssim = [] mcs = [] for _ in range(levels): sim, cs = ssim(img1, img2, window_size=window_size, size_average=size_average, full=True, val_range=val_range) mssim.append(sim) mcs.append(cs) img1 = F.avg_pool2d(img1, (2, 2)) img2 = F.avg_pool2d(img2, (2, 2)) mssim = torch.stack(mssim) mcs = torch.stack(mcs) # Normalize (to avoid NaNs during training unstable models, not compliant with original definition) if normalize: mssim = (mssim + 1) / 2 mcs = (mcs + 1) / 2 pow1 = mcs ** weights pow2 = mssim ** weights # From Matlab implementation https://ece.uwaterloo.ca/~z70wang/research/iwssim/ output = torch.prod(pow1[:-1] * pow2[-1]) return output # Classes to re-use window class SSIM(torch.nn.Module): def __init__(self, window_size=11, size_average=True, val_range=None): super(SSIM, self).__init__() self.window_size = window_size self.size_average = size_average self.val_range = val_range # Assume 1 channel for SSIM self.channel = 1 self.window = create_window(window_size) def forward(self, img1, img2): (_, channel, _, _) = img1.size() if channel == self.channel and self.window.dtype == img1.dtype: window = self.window else: window = create_window(self.window_size, channel).to(img1.device).type(img1.dtype) self.window = window self.channel = channel return ssim(img1, img2, window=window, window_size=self.window_size, size_average=self.size_average) class MSSSIM(torch.nn.Module): def __init__(self, window_size=11, size_average=True, channel=3): super(MSSSIM, self).__init__() self.window_size = window_size self.size_average = size_average self.channel = channel def forward(self, img1, img2): # TODO: store window between calls if possible return msssim(img1, img2, window_size=self.window_size, size_average=self.size_average)
4,380
31.69403
118
py
tasksource
tasksource-main/src/tasksource/metadata/popularity.py
dataset_rank = {'glue': 0, 'super_glue': 12, 'tweet_eval': 23, 'blimp': 34, 'imdb': 101, 'wikitext': 102, 'squad': 106, 'trec': 107, 'openwebtext': 108, 'rotten_tomatoes': 109, 'anli': 110, 'adversarial_qa': 111, 'ai2_arc': 115, 'xsum': 117, 'amazon_reviews_multi': 118, 'ag_news': 125, 'yelp_review_full': 126, 'wino_bias': 127, 'piqa': 131, 'duorc': 132, 'quail': 134, 'trivia_qa': 135, 'cnn_dailymail': 143, 'common_gen': 146, 'sst': 147, 'conll2003': 150, 'financial_phrasebank': 151, 'babi_qa': 155, 'poem_sentiment': 163, 'dream': 164, 'paws': 165, 'emotion': 168, 'kilt_tasks': 169, 'sciq': 180, 'cos_e': 181, 'dbpedia_14': 183, 'newsgroup': 184, 'cosmos_qa': 244, 'squad_v2': 245, 'samsum': 246, 'amazon_polarity': 247, 'multi_news': 248, 'wiki_hop': 249, 'quartz': 251, 'qasc': 252, 'wiki_qa': 253, 'openbookqa': 254, 'ropes': 256, 'quoref': 257, 'snli': 258, 'app_reviews': 259, 'gigaword': 260, 'wiki_bio': 261, 'amazon_us_reviews': 262, 'scan': 308, 'race': 320, 'swag': 323, 'codah': 325, 'ccdv/arxiv-summarization': 331, 'subjqa': 333, 'universal_morphologies': 339, 'hans': 447, 'sst2': 448, 'guardian_authorship': 449, 'math_qa': 465, 'librispeech_asr': 466, 'hendrycks_test': 469, 'openai_humaneval': 526, 'ptb_text_only': 527, 'pubmed_qa': 528, 'head_qa': 531, 'ought/raft': 533, 'ade_corpus_v2': 544, 'cbt': 547, 'bookcorpus': 552, 'squadshifts': 553, 'story_cloze': 557, 'multi_nli': 559, 'qanta': 560, 'hate_speech18': 564, 'gem': 565, 'lex_glue': 599, 'deepmind/code_contests': 606, 'imagenet-1k': 607, 'blended_skill_talk': 608, 'sms_spam': 609, 'asset': 610, 'fever': 612, 'commonsense_qa': 615, 'scientific_papers': 616, 'evidence_infer_treatment': 618, 'hotpot_qa': 620, 'superb': 622, 'sick': 628, 'humicroedit': 629, 'snips_built_in_intents': 631, 'winograd_wsc': 632, 'bigbench': 634, 'multi_woz_v22': 801, 'lambada': 803, 'banking77': 804, 'hate_speech_offensive': 805, 'yahoo_answers_topics': 806, 'ccdv/cnn_dailymail': 807, 'hyperpartisan_news_detection': 810, 'gsm8k': 812, 'wikisql': 814, 'the_pile': 815, 'health_fact': 825, 'mdd': 826, 'web_questions': 830, 'ethos': 831, 'wnut_17': 833, 'medical_questions_pairs': 834, 'scitldr': 835, 'drop': 838, 'squad_adversarial': 839, 'e2e_nlg_cleaned': 841, 'onestop_english': 842, 'pragmeval': 843, 'relbert/analogy_questions': 863, 'nq_open': 869, 'daily_dialog': 870, 'mc_taco': 871, 'crows_pairs': 872, 'go_emotions': 873, 'ncbi_disease': 875, 'boolq': 876, 'movie_rationales': 877, 'climate_fever': 878, 'discovery': 879, 'lama': 881, 'ecthr_cases': 885, 'jfleg': 887, 'selqa': 888, 'acronym_identification': 892, 'scicite': 893, 'tab_fact': 894, 'wiki_asp': 896, 'enriched_web_nlg': 916, 'svhn': 918, 'docred': 920, 'conllpp': 921, 'liar': 922, 'multi_x_science_sum': 923, 'discofuse': 924, 'competition_math': 926, 'biosses': 927, 'jnlpba': 928, 'web_nlg': 929, 'qa_srl': 937, 'neural_code_search': 938, 'conv_ai_2': 940, 'craigslist_bargains': 941, 'qed': 942, 'conv_ai_3': 943, 'conv_ai': 944, 'turk': 945, 'covid_qa_castorini': 946, 'sem_eval_2014_task_1': 947, 'mwsc': 948, 'gutenberg_time': 949, 'billsum': 950, 'riddle_sense': 951, 'species_800': 952, 'hlgd': 953, 'definite_pronoun_resolution': 954, 'tmu_gfm_dataset': 955, 'relbert/semeval2012_relational_similarity_v4': 956, 'clinc_oos': 957, 'imppres': 960, 'mrqa': 976, 'cc_news': 977, 'lmqg/qag_tweetqa': 978, 'aeslc': 979, 'big_patent': 980, 'eli5': 990, 'scene_parse_150': 991, 'circa': 993, 'aqua_rat': 994, 'nlu_evaluation_data': 996, 'newspop': 997, 'relbert/lexical_relation_classification': 998, 'yahoo_answers_qa': 1003, 'emo': 1004, 'silicone': 1005, 'cord19': 1015, 'tweet_qa': 1018, 'meta_woz': 1019, 'md_gender_bias': 1021, 'art': 1031, 'google_wellformed_query': 1032, 'ambig_qa': 1033, 'taskmaster2': 1035, 'quac': 1042, 'freebase_qa': 1043, 'quora': 1044, 'numer_sense': 1045, 'narrativeqa': 1046, 'ccdv/pubmed-summarization': 1047, 'qa_zre': 1049, 'limit': 1050, 'tweets_hate_speech_detection': 1051, 'mocha': 1052, 'hatexplain': 1053, 'bing_coronavirus_query_set': 1054, 'great_code': 1055, 'medal': 1056, 'sent_comp': 1057, 'kelm': 1058, 'natural_questions': 1059, 'wiki_split': 1061, 'zest': 1062, 'cfq': 1063, 'multi_re_qa': 1071, 'stereoset': 1080, 'coqa': 1082, 'cuad': 1083, 'break_data': 1084, 'mbpp': 1089, 'knkarthick/dialogsum': 1091, 'wiki_auto': 1092, 'pile-of-law/pile-of-law': 1097, 'pg19': 1132, 'DFKI-SLT/few-nerd': 1133, 'wikicorpus': 1136, 'e2e_nlg': 1142, 'anton-l/superb': 1143, 'ghomasHudson/muld': 1144, 'Exr0n/wiki-entity-similarity': 1150, 'BeIR/nfcorpus': 1156, 'ccdv/govreport-summarization': 1158, 'woz_dialogue': 1159, 'reddit': 1164, 'EMBO/sd-nlp': 1165, 'empathetic_dialogues': 1170, 'BeIR/fiqa': 1171, 'generics_kb': 1173, 'swda': 1177, 'wikitablequestions': 1178, 'pubmed': 1183, 'chr_en': 1184, 'sharc': 1185, 'sharc_modified': 1186, 'BeIR/scifact': 1190, 'nell': 1192, 'patriziobellan/PET': 1196, 'EMBO/biolang': 1198, 'dynabench/qa': 1202, 'reddit_tifu': 1206, 'BeIR/scidocs': 1208, 'pec': 1210, 'tner/tweetner7': 1213, 'BeIR/arguana': 1214, 'multidoc2dial': 1216, 'taskmaster1': 1219, 'spider': 1221, 'adv_glue': 1222, 'allenai/mslr2022': 1228, 'conceptnet5': 1230, 'tyqiangz/multilingual-sentiments': 1233, 'newsqa': 1246, 'metashift': 1249, 'so_stacksample': 1250, 'doc2dial': 1253, 'search_qa': 1256, 'yhavinga/mc4_nl_cleaned': 1258, 'hope_edi': 1270, 'proto_qa': 1273, 'tuple_ie': 1276, 'simple_questions_v2': 1279, 'nlpaueb/finer-139': 1282, 'bookcorpusopen': 1283, 'tner/ontonotes5': 1284, 'crd3': 1285, 'ucberkeley-dlab/measuring-hate-speech': 1286, 'gap': 1287, 'recipe_nlg': 1288, 'schema_guided_dstc8': 1289, 'BeIR/beir': 1291, 'sagnikrayc/mctest': 1294, 'eurlex': 1296, 'corypaik/coda': 1297, 'bc2gm_corpus': 1298, 'ascent_kb': 1299, 'curiosity_dialogs': 1301, 'covid_qa_deepset': 1302, 'air_dialogue': 1303, 'taskmaster3': 1305, 'xsum_factuality': 1306, 'medical_dialog': 1308, 'BeIR/trec-covid': 1312, 'lhoestq/test': 1314, 'newsroom': 1315, 'tne': 1316, 'covid_qa_ucsd': 1317, 'fhamborg/news_sentiment_newsmtsc': 1319, 'prachathai67k': 1321, 'cardiffnlp/tweet_topic_multi': 1322, 'datacommons_factcheck': 1323, 'deal_or_no_dialog': 1325, 'ubuntu_dialogs_corpus': 1327, 'eu_regulatory_ir': 1329, 'scifact': 1331, 'wi_locness': 1333, 'relbert/relation_mapping': 1335, 'coastalcph/fairlex': 1336, 'asnq': 1340, 'peer_read': 1341, 'metaeval/linguisticprobing': 1343, 'jigsaw_unintended_bias': 1353, 'totto': 1354, 'irc_disentangle': 1355, 'med_hop': 1357, 'numeric_fused_head': 1359, 'ollie': 1361, 'per_sent': 1363, 'SocialGrep/ten-million-reddit-answers': 1364, 'lmqg/qg_squad': 1366, 's2orc': 1367, 'Hellisotherpeople/DebateSum': 1368, 'SocialGrep/reddit-crypto-aug-2021': 1369, 'jigsaw_toxicity_pred': 1371, 'GroNLP/ik-nlp-22_slp': 1372, 'SocialGrep/reddit-nonewnormal-complete': 1374, 'SocialGrep/reddit-wallstreetbets-aug-2021': 1376, 'SocialGrep/the-reddit-covid-dataset': 1378, 'SocialGrep/top-american-universities-on-reddit': 1380, 'BeIR/beir-corpus': 1382, 'SocialGrep/one-year-of-r-india': 1384, 'BritishLibraryLabs/EThOS-PhD-metadata': 1386, 'librispeech_lm': 1388, 'few_rel': 1389, 'arxiv_dataset': 1390, 'lc_quad': 1391, 'diplomacy_detection': 1392, 'lmqg/qa_squadshifts_pseudo': 1393, 'grail_qa': 1461, 'tner/wnut2017': 1462, 'demo-org/auditor_review': 1463, 'allenai/real-toxicity-prompts': 1464, 'BeIR/nfcorpus-qrels': 1465, 'onestop_qa': 1466, 'demelin/moral_stories': 1467, 'atomic': 1493, 'crawl_domain': 1494, 'BeIR/quora': 1495, 'Abirate/english_quotes': 1497, 'narrativeqa_manual': 1498, 'BeIR/fiqa-qrels': 1499, 'social_bias_frames': 1500, 'pkavumba/balanced-copa': 1501, 'eraser_multi_rc': 1502, 'sled-umich/TRIP': 1503, 'opinosis': 1504, 'PiC/phrase_sense_disambiguation': 1505, 'enwik8': 1506, 'sem_eval_2020_task_11': 1508, 'gooaq': 1509, 'linnaeus': 1510, 'hover': 1511, 'GonzaloA/fake_news': 1512, 'consumer-finance-complaints': 1513, 'ohsumed': 1514, 'casino': 1515, 'gfissore/arxiv-abstracts-2021': 1516, 'conv_questions': 1517, 'hate_offensive': 1518, 'sofc_materials_articles': 1519, 'wanyu/IteraTeR_human_sent': 1520, 'dialog_re': 1521, 'fake_news_english': 1522, 'dart': 1523, 'blog_authorship_corpus': 1524, 'msr_zhen_translation_parity': 1525, 'cryptonite': 1526, 'disfl_qa': 1527, 'olm/olm-CC-MAIN-2022-21-sampling-ratio-0.14775510204': 1528, 'olm/olm-CC-MAIN-2022-33-sampling-ratio-0.20': 1529, 'coarse_discourse': 1530, 'eth_py150_open': 1531, 'event2Mind': 1532, 'Paul/hatecheck': 1533, 'eli5_category': 1534, 'hippocorpus': 1535, 'the_pile_books3': 1536, 'coached_conv_pref': 1537, 'has_part': 1538, 'times_of_india_news_headlines': 1539, 'medmcqa': 1540, 'Babelscape/rebel-dataset': 1541, 'glucose': 1542, 'msr_text_compression': 1543, 'msr_genomics_kbcomp': 1544, 'SpeedOfMagic/ontonotes_english': 1545, 'msr_sqa': 1546, 'wiki_movies': 1547, 'hybrid_qa': 1548, 'metooma': 1549, 'multi_nli_mismatch': 1550, 'text2log': 1551, 'the_pile_stack_exchange': 1552, 're_dial': 1553, 'inquisitive_qg': 1554, 'SocialGrep/one-million-reddit-jokes': 1555, 'time_dial': 1556, 'BeIR/scifact-qrels': 1557, 'sede': 1558, 'mutual_friends': 1559, 'pass': 1560, 'allenai/multi_lexsum': 1561, 'youtube_caption_corrections': 1562, 'NbAiLab/norec_agg': 1563, 'DanL/scientific-challenges-and-directions-dataset': 1564, 'SocialGrep/one-million-reddit-questions': 1565, 'Motahar/github-issues': 1566, 'SocialGrep/the-2022-trucker-strike-on-reddit': 1567, 'allenai/qasper': 1568, 'CyranoB/polarity': 1569, 'SocialGrep/one-million-reddit-confessions': 1570, 'debatelab/deepa2': 1571, 'bhavnicksm/sentihood': 1572, 'debatelab/aaac': 1573, 'jgammack/SAE-door-abstracts': 1574, 'erwanlc/cocktails_recipe': 1575, 'erwanlc/cocktails_recipe_no_brand': 1576, 'BeIR/arguana-qrels': 1577, 'tner/fin': 1578, 'BeIR/scidocs-qrels': 1579, 'tner/bc5cdr': 1580, 'olm/olm-CC-MAIN-2022-27-sampling-ratio-0.16142697881': 1581, 'BeIR/fever': 1582, 'cardiffnlp/tweet_topic_single': 1584, 'speechcolab/gigaspeech': 1585, 'BeIR/webis-touche2020': 1586, 'aquamuse': 1588, 'olm/olm-CC-MAIN-2022-40-sampling-ratio-0.15894621295': 1590, 'tner/btc': 1591, 'truthful_qa': 1592, 'McGill-NLP/FaithDial': 1594, 'ekinakyurek/ftrace': 1595, 'tomasg25/scientific_lay_summarisation': 1597, 'tner/mit_restaurant': 1599, 'bigscience-biomedical/bioasq_task_b': 1600, 'strombergnlp/broad_twitter_corpus': 1619, 'tner/bionlp2004': 1620, 'metaeval/recast': 1621, 'the_pile_openwebtext2': 1629, 'taln-ls2n/inspec': 1630, 'lmqg/qa_squadshifts': 1631, 'BeIR/hotpotqa': 1636, 'jpwahle/machine-paraphrase-dataset': 1638, 'tner/mit_movie_trivia': 1639, 'tner/conll2003': 1640, 'OxAISH-AL-LLM/wiki_toxic': 1641, 'ccdv/WCEP-10': 1642, 'BeIR/trec-covid-qrels': 1646, 'g8a9/europarl_en-it': 1647, 'carblacac/twitter-sentiment-analysis': 1648, 'usc-isi/WikiConvert': 1649, 'visual_genome': 1650, 'florianbussmann/FUNSD-vu2020revising': 1660, 'Felix-ML/quoteli3': 1661, 'allenai/scico': 1662, 'drAbreu/bc4chemd_ner': 1663, 'tner/tweebank_ner': 1664, 'alisawuffles/WANLI': 1665, 'Team-PIXEL/rendered-bookcorpus': 1666, 'Team-PIXEL/rendered-wikipedia-english': 1667, 'wanyu/IteraTeR_full_sent': 1668, 'EMBO/BLURB': 1669, 'metaeval/crowdflower': 1676, 'AlexaAI/bold': 1685, 'metaeval/ethics': 1686, 'sileod/movie_recommendation': 1691, 'lmqg/qg_subjqa': 1692, 'copenlu/scientific-exaggeration-detection': 1699, 'esb/datasets': 1700, 'BeIR/msmarco': 1701, 'biwi_kinect_head_pose': 1703, 'BeIR/quora-qrels': 1704, 'wardenga/lsoie': 1705, 'nlphuji/vasr': 1707, 'BeIR/nq': 1708, 'BeIR/dbpedia-entity': 1710, 'sadrasabouri/ShahNegar': 1712, 'knkarthick/xsum': 1713, 'ColumbiaNLP/FLUTE': 1714, 'bigscience-biomedical/scitail': 1715, 'lmqg/qg_squadshifts': 1717, 'BeIR/climate-fever': 1722, 'PiC/phrase_retrieval': 1724, 'bdotloh/empathetic-dialogues-contexts': 1726, 'ccdv/mediasum': 1727, 'BeIR/msmarco-qrels': 1735, 'alexfabbri/answersumm': 1736, 'pszemraj/text2image-multi-prompt': 1737, 'shibing624/source_code': 1738, 'kensho/spgispeech': 1741, 'jamescalam/channel-metadata': 1742, 'EMBO/sd-nlp-non-tokenized': 1743, 'facebook/pmd': 1748, 'drt/kqa_pro': 1749, 'BeIR/fever-qrels': 1751, 'TheFusion21/PokemonCards': 1752, 'zeroshot/twitter-financial-news-sentiment': 1753, 'bigscience-biomedical/blurb': 1754, 'mteb/bucc-bitext-mining': 1759, 'pinecone/core-2020-05-10-deduplication': 1763, 'tals/vitaminc': 1764, 'BeIR/hotpotqa-qrels': 1765, 'gigant/ted_descriptions': 1766, 'jpwahle/autoencoder-paraphrase-dataset': 1767, 'beki/privy': 1768, 'Muennighoff/P3': 1770, 'jpwahle/dblp-discovery-dataset': 1771, 'taln-ls2n/kp20k': 1773, 'bigscience-biomedical/biosses': 1774, 'allenai/prosocial-dialog': 1776, 'pacovaldez/stackoverflow-questions': 1777, 'kasnerz/hitab': 1778, 'relbert/semeval2012_relational_similarity': 1779, 'sagnikrayc/snli-cf-kaushik': 1780, 'mwritescode/slither-audited-smart-contracts': 1781, 'BeIR/webis-touche2020-qrels': 1787, 'bigscience-biomedical/mednli': 1788, 'pinecone/movielens-recent-ratings': 1790, 'BeIR/dbpedia-entity-qrels': 1791, 'shanya/crd3': 1792, 'knkarthick/samsum': 1793, 'BeIR/climate-fever-qrels': 1794, 'BeIR/nq-qrels': 1795, 'sanchit-gandhi/librispeech_asr_dummy': 1796, 'taln-ls2n/semeval-2010-pre': 1797, 'Bingsu/openwebtext_20p': 1798, 'PolyAI/banking77': 1799, 'JulesBelveze/tldr_news': 1800, 'Freed-Wu/kodak': 1801, 'biglam/gutenberg-poetry-corpus': 1802, 'SocialGrep/reddit-r-bitcoin-data-for-jun-2022': 1803, 'taln-ls2n/kptimes': 1805, 'biglam/old_bailey_proceedings': 1806, 'launch/gov_report': 1807, 'knkarthick/AMI': 1810, 'voidful/NMSQA': 1811, 'DTU54DL/dmeo': 1812, 'FinanceInc/auditor_sentiment': 1813, 'jamescalam/unsplash-25k-photos': 1814, 'Tidrael/tsl_news': 1815, 'DTU54DL/common3k-train': 1816, 'okite97/news-data': 1817, 'lmqg/qa_squad': 1818, 'ConvLab/woz': 1819, 'ConvLab/camrest': 1820, 'ConvLab/metalwoz': 1821, 'kakaobrain/coyo-700m': 1822, 'taln-ls2n/kpbiomed': 1823, 'abhinavk/openpi_v2': 1826, 'mwong/fever-claim-related': 1831, 'ConvLab/tm1': 1832, 'joey234/nan-nli': 1833, 'ConvLab/tm2': 1834, 'ConvLab/tm3': 1835, 'ConvLab/kvret': 1836, 'ConvLab/sgd': 1837, 'relbert/semeval2012_relational_similarity_v5': 1838, 'cmudrc/wave-energy': 1839, 'llangnickel/long-covid-classification-data': 1840, 'webis/args_me': 1841, 'HuggingFaceM4/something_something_v2': 1844, 'ConvLab/dailydialog': 1845, 'huanggab/reddit_haiku': 1846, 'relbert/semeval2012_relational_similarity_v6': 1847, 'pszemraj/riddlesense_plusplus': 1848, 'rungalileo/20_Newsgroups_Fixed': 1849, 'DTU54DL/common-voice-test16k': 1850, 'lhoestq/custom_squad': 1851, 'merve/poetry': 1852, 'yoshitomo-matsubara/srsd-feynman_easy': 1853, 'nightingal3/fig-qa': 1854, 'matejklemen/vuamc': 1855, 'strombergnlp/twitter_pos': 1856, 'nlphuji/winogavil': 1858, 'DFKI-SLT/tacred': 1859, 'valurank/News_Articles_Categorization': 1861, 'nbroad/mediasum': 1862, 'asapp/slue': 1863, 'zbnsl/emoteModified': 1865, 'adsabs/WIESP2022-NER': 1866, 'arize-ai/ecommerce_reviews_with_language_drift': 1867, 'UCL-DARK/ludwig': 1868, 'Aunsiels/InfantBooks': 1874, 'openclimatefix/uk_pv': 1875, 'copenlu/fever_gold_evidence': 1876, 'rungalileo/mit_movies_fixed_connll_format': 1877, 'jamescalam/youtube-transcriptions': 1878, 'lmqg/qa_harvesting_from_wikipedia': 1879, 'qanastek/Biosses-BLUE': 1880, 'zeronix1020/Strawberry-Disease': 1881, 'dferndz/cSQuAD2': 1882, 'taln-ls2n/pubmed': 1883, 'BeIR/scidocs-generated-queries': 1884, 'jmhessel/newyorker_caption_contest': 1885, 'inverse-scaling/NeQA': 1915, 'DTU54DL/common-voice': 1916, 'turingbench/TuringBench': 1917, 'demelin/understanding_fables': 1937, 'RUCAIBox/Open-Dialogue': 1938, 'allenai/multinews_sparse_max': 1939, 'RamAnanth1/lex-fridman-podcasts': 1940, 'sled-umich/Conversation-Entailment': 1941, 'stevhliu/demo': 1942, 'svakulenk0/qrecc': 1943, 'arize-ai/movie_reviews_with_context_drift': 1944, 'launch/ampere': 1945, 'AnonymousSub/recipe_RL_data_roberta-base': 1946, 'dreamproit/bill_summary_us': 1947, 'bgstud/libri-whisper-raw': 1948, 'jpwahle/etpc': 1949, 'DTU54DL/common-native-proc': 1950, 'mbartolo/synQA': 1951, 'wanyu/IteraTeR_full_doc': 1952, 'wanyu/IteraTeR_human_doc': 1953, 'orieg/elsevier-oa-cc-by': 1954, 'climatebert/environmental_claims': 1955, 'SocialGrep/the-reddit-climate-change-dataset': 1956, 'KGraph/FB15k-237': 1958, 'KheemDH/data': 1959, 'mwong/fever-evidence-related': 1960, 'HuggingFaceM4/TGIF': 1961, 'BeIR/fever-generated-queries': 1962, 'nateraw/ade20k-tiny': 1963, 'BeIR/cqadupstack-qrels': 1964, 'knkarthick/highlightsum': 1965, 'RUCAIBox/Data-to-text-Generation': 1966, 'GateNLP/broad_twitter_corpus': 1967, 'Tidrael/finance-headlines': 1968, 'lmqg/qag_squad': 1969, 'pacovaldez/stackoverflow-questions-2016': 1970, 'BeIR/fiqa-generated-queries': 1971, 'BeIR/signal1m-generated-queries': 1972, 'MicPie/unpredictable_msdn-microsoft-com': 1973, 'zeroshot/twitter-financial-news-topic': 1974, 'inverse-scaling/quote-repetition': 1975, 'esc-bench/esc-diagnostic-backup': 1976, 'lmqg/qg_annotation': 1977, 'sileod/wep-probes': 1978, 'DTU54DL/common-voice-test3k': 1981, 'jakartaresearch/causalqa': 1982, 'copenlu/sufficient_facts': 2002, 'ConvLab/multiwoz21': 2005, 'arka0821/multi_document_summarization': 2006, 'strombergnlp/rumoureval_2019': 2007, 'rongzhangibm/NaturalQuestionsV2': 2008, 'Muennighoff/mbpp': 2009, 'RUCAIBox/Simplification': 2011, 'shubhamg2208/lexicap': 2012, 'olm/olm-wikipedia-20220701': 2013, 'esc-bench/esc-diagnostic-dataset': 2014, 'jpwahle/autoregressive-paraphrase-dataset': 2015, 'GabrielVidal/dead-by-daylight-perks': 2016, 'DTU54DL/common-proc-whisper': 2017, 'valurank/PoliticalBias': 2018, 'McGill-NLP/TopiOCQA': 2019, 'gsarti/magpie': 2020, 'BeIR/cqadupstack-generated-queries': 2021, 'MicPie/unpredictable_mmo-champion-com': 2022, 'RUCAIBox/Question-Generation': 2023, 'allenai/multinews_sparse_mean': 2024, 'demo-org/diabetes': 2025, 'StonyBrookNLP/tellmewhy': 2026, 'bergr7/weakly_supervised_ag_news': 2027, 'din0s/msmarco-nlgen': 2028, 'frankier/cross_domain_reviews': 2029, 'gart-labor/pumpnli': 2030, 'AndyChiang/cloth': 2031, 'olm/olm-CC-MAIN-2017-22-sampling-ratio-0.16178770949': 2032, 'bgstud/libri': 2033, 'DTU54DL/commonvoice_accent_test': 2034, 'lewtun/my-awesome-dataset': 2035, 'peixian/rtGender': 2036, 'pmc/open_access': 2039, 'uva-irlab/trec-cast-2019-multi-turn': 2043, 'DFKI-SLT/scidtb': 2044, 'surrey-nlp/PLOD-filtered': 2045, 'wanyu/IteraTeR_v2': 2046, 'strombergnlp/ipm_nel': 2047, 'HuggingFaceM4/charades': 2048, 'ncats/EpiSet4NER-v2': 2050, 'HuggingFaceM4/ActivitiyNet_Captions': 2051, 'sileod/discourse_marker_qa': 2052, 'yoshitomo-matsubara/srsd-feynman_medium': 2053, 'BeIR/nfcorpus-generated-queries': 2054, 'BeIR/trec-news-generated-queries': 2055, 'BeIR/robust04-generated-queries': 2056, 'BeIR/quora-generated-queries': 2057, 'valurank/Adult-content-dataset': 2058, 'launch/open_question_type': 2059, 'knkarthick/topicsum': 2060, 'yuningm/citesum': 2061, 'elihoole/asrs-aviation-reports': 2062, 'DeveloperOats/DBPedia_Classes': 2063, 'hoskinson-center/proof-pile': 2064, 'RUCAIBox/Summarization': 2065, 'RUCAIBox/Question-Answering': 2066, 'RUCAIBox/Story-Generation': 2067, 'RUCAIBox/Paraphrase': 2068, 'jakartaresearch/semeval-absa': 2069, 'tner/ttc_dummy': 2071, 'copenlu/citeworth': 2072, 'allenai/multinews_sparse_oracle': 2073, 'allenai/multixscience_sparse_oracle': 2074, 'allenai/multixscience_sparse_mean': 2075, 'allenai/multixscience_sparse_max': 2076, 'allenai/ms2_sparse_oracle': 2077, 'mschi/blogspot_raw': 2078, 'gaurikapse/civis-consultation-summaries': 2079, 'chenghao/cuad_qa': 2080, 'esc-bench/esc-datasets': 2081, 'olm/olm-wikipedia-20221001': 2082, 'allenai/wcep_dense_oracle': 2083, 'dennlinger/wiki-paragraphs': 2084, 'AndyChiang/dgen': 2085, 'esb/diagnostic-dataset': 2086, 'havens2/naacl2022': 2087, 'fkdosilovic/docee-event-classification': 2088, 'DTU54DL/demo-common-whisper': 2089, 'dferndz/cSQuAD1': 2090, 'jpcorb20/multidogo': 2091, 'julien-c/reactiongif': 2092, 'lara-martin/Scifi_TV_Shows': 2093, 'lukesjordan/worldbank-project-documents': 2094, 'mnemlaghi/widdd': 2095, 'mvarma/medwiki': 2096, 'nateraw/beans': 2098, 'nateraw/cats_vs_dogs': 2099, 'nateraw/food101': 2100, 'nateraw/sync_food101': 2101, 'ncats/EpiSet4BinaryClassification': 2102, 'ncats/EpiSet4NER-v1': 2103, 'peixian/equity_evaluation_corpus': 2104, 'rajeshradhakrishnan/malayalam_wiki': 2105, 'softcatala/open-source-english-catalan-corpus': 2106, 'toloka/CrowdSpeech': 2107, 'valurank/12-factor': 2108, 'valurank/PoliticalBias_AllSides_Txt': 2109, 'valurank/PoliticalBias_Sources': 2110, 'valurank/hate-multi': 2111, 'valurank/news-12factor': 2112, 'valurank/offensive-multi': 2113, 'webimmunization/COVID-19-vaccine-attitude-tweets': 2114, 'wpicard/nostradamus-propheties': 2115, 'yuanchuan/annotated_reference_strings': 2116, 'ruanchaves/stan_large': 2117, 'ruanchaves/stan_small': 2118, 'ruanchaves/boun': 2119, 'ruanchaves/dev_stanford': 2120, 'ruanchaves/test_stanford': 2121, 'ruanchaves/snap': 2122, 'z-uo/qasper-squad': 2123, 'SocialGrep/the-antiwork-subreddit-dataset': 2124, 'CLUTRR/v1': 2126, 'malteos/test2': 2132, 'TomTBT/pmc_open_access_xml': 2133, 'SocialGrep/the-reddit-dataset-dataset': 2137, 'SocialGrep/the-reddit-place-dataset': 2139, 'projecte-aina/gencata': 2141, 'mwong/climate-evidence-related': 2142, 'mwong/climate-claim-related': 2143, 'surrey-nlp/PLOD-unfiltered': 2144, 'SocialGrep/the-reddit-irl-dataset': 2145, 'Lexi/spanextract': 2147, 'mwong/climatetext-claim-related-evaluation': 2148, 'mwong/climatetext-evidence-related-evaluation': 2149, 'ylacombe/xsum_factuality': 2150, 'mwong/climatetext-climate_evidence-claim-related-evaluation': 2151, 'mwong/climatetext-claim-climate_evidence-related-evaluation': 2152, 'mwong/climatetext-evidence-claim-pair-related-evaluation': 2153, 'mwong/climatetext-claim-evidence-pair-related-evaluation': 2154, 'patrickvonplaten/librispeech_asr_self_contained': 2155, 'BritishLibraryLabs/web_archive_classification': 2158, 'albertxu/CrosswordQA': 2159, 'SocialGrep/the-reddit-nft-dataset': 2160, 'janck/bigscience-lama': 2162, 'strombergnlp/twitter_pos_vcb': 2163, 'Filippo/osdg_cd': 2164, 'Ukhushn/home-depot': 2165, 'pile-of-law/eoir_privacy': 2166, 'drAbreu/sd-nlp-2': 2168, 'Leyo/TGIF': 2173, 'strombergnlp/named_timexes': 2174, 'domenicrosati/TruthfulQA': 2175, 'Roh/ryanspeech': 2176, 'Leyo/ActivityNet_Captions': 2177, 'IsaacBot/SQuAD-single-sentence-QA': 2178, 'morteza/cogtext': 2179, 'wdc/products-2017': 2180, 'rajeshvarma/QA_on_SLA': 2196, 'statworx/haiku': 2197, 'rajistics/million-headlines': 2198, 'feyzaakyurek/BBNLI': 2199, 'launch/gov_report_qs': 2200, 'DFKI-SLT/wikitext_linked': 2202, 'dianalogan/Marketing-Budget-and-Actual-Sales-Dataset': 2204, 'mehnaazasad/arxiv-co-ga': 2205, 'JeremyAlain/123_test': 2206, 'BeIR/arguana-generated-queries': 2209, 'BeIR/climate-fever-generated-queries': 2210, 'BeIR/dbpedia-entity-generated-queries': 2211, 'wise-east/spolin': 2212, 'yoshitomo-matsubara/srsd-feynman_hard': 2213, 'florentgbelidji/edmunds-car-ratings': 2214, 'olivierdehaene/xkcd': 2215, 'rajistics/auditor_review': 2216, 'BeIR/scifact-generated-queries': 2217, 'BeIR/trec-covid-generated-queries': 2218, 'BeIR/webis-touche2020-generated-queries': 2219, 'BeIR/nq-generated-queries': 2220, 'BeIR/hotpotqa-generated-queries': 2221, 'BeIR/bioasq-generated-queries': 2222, 'icelab/ntrs_meta': 2223, 'iejMac/CLIP-Kinetics700': 2224, 'fever/feverous': 2225, 'Livingwithmachines/hmd-erwt-training': 2226, 'wkrl/cord': 2227, 'launch/reddit_qg': 2228, 'arize-ai/xtreme_en': 2229} dataset_rank['Anthropic/model-written-evals']=13 dataset_rank['Anthropic/hh-rlhf']=14
24,310
28.290361
69
py
sm-vit
sm-vit-main/train.py
# coding=utf-8 from __future__ import absolute_import, division, print_function wnb = False if wnb: import wandb wandb.init(project="sm-vit", entity="xxx") import logging import argparse import os import random import numpy as np from datetime import timedelta import time import torch import torch.distributed as dist from tqdm import tqdm from torch.utils.tensorboard import SummaryWriter from apex import amp from apex.parallel import DistributedDataParallel as DDP from models.modeling import VisionTransformer, CONFIGS from utils.scheduler import WarmupLinearSchedule, WarmupCosineSchedule from utils.data_utils import get_loader from utils.dist_util import get_world_size logger = logging.getLogger(__name__) class AverageMeter(object): """Computes and stores the average and current value""" def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): self.val = val self.sum += val * n self.count += n self.avg = self.sum / self.count def simple_accuracy(preds, labels): return (preds == labels).mean() def save_model(args, model): model_to_save = model.module if hasattr(model, 'module') else model model_checkpoint = os.path.join(args.output_dir, "%s_checkpoint.bin" % args.name) torch.save(model_to_save.state_dict(), model_checkpoint) logger.info("Saved model checkpoint to [DIR: %s]", args.output_dir) def reduce_mean(tensor, nprocs): rt = tensor.clone() dist.all_reduce(rt, op=dist.ReduceOp.SUM) rt /= nprocs return rt def setup(args): # Prepare model config = CONFIGS[args.model_type] if args.dataset == "dogs": num_classes = 120 elif args.dataset == "CUB": num_classes=200 elif args.dataset == "nabirds": num_classes = 555 else: raise Exception(f'Unknown dataset "{args.dataset}"') model = VisionTransformer(config, args.img_size, zero_head=True, num_classes=num_classes, vis=True, smoothing_value=args.smoothing_value, dataset=args.dataset, \ coeff_max=args.coeff_max, contr_loss=args.contr_loss, focal_loss=args.focal_loss) model.load_from(np.load(args.pretrained_dir)) model.to(args.device) num_params = count_parameters(model) logger.info("{}".format(config)) logger.info("Training parameters %s", args) logger.info("Total Parameter: \t%2.1fM" % num_params) return args, model def count_parameters(model): params = sum(p.numel() for p in model.parameters() if p.requires_grad) return params/1000000 def set_seed(args): random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.n_gpu > 0: torch.cuda.manual_seed_all(args.seed) def valid(args, model, writer, test_loader, global_step): # Validation! eval_losses = AverageMeter() logger.info("***** Running Validation *****") logger.info(" Num steps = %d", len(test_loader)) logger.info(" Batch size = %d", args.eval_batch_size) model.eval() all_preds, all_label = [], [] epoch_iterator = tqdm(test_loader, desc="Validating... (loss=X.X)", bar_format="{l_bar}{r_bar}", dynamic_ncols=True, disable=args.local_rank not in [-1, 0]) loss_fct = torch.nn.CrossEntropyLoss() for step, batch in enumerate(epoch_iterator): if wnb: wandb.log({"step": step}) batch = tuple(t.to(args.device) for t in batch) if args.sm_vit: x, y, mask = batch else: x, y = batch with torch.no_grad(): if args.sm_vit: logits = model(x, None, mask)[0] else: logits = model(x)[0] eval_loss = loss_fct(logits, y) if args.contr_loss: eval_loss = eval_loss.mean() eval_losses.update(eval_loss.item()) preds = torch.argmax(logits, dim=-1) if len(all_preds) == 0: all_preds.append(preds.detach().cpu().numpy()) all_label.append(y.detach().cpu().numpy()) else: all_preds[0] = np.append( all_preds[0], preds.detach().cpu().numpy(), axis=0 ) all_label[0] = np.append( all_label[0], y.detach().cpu().numpy(), axis=0 ) epoch_iterator.set_description("Validating... (loss=%2.5f)" % eval_losses.val) all_preds, all_label = all_preds[0], all_label[0] accuracy = simple_accuracy(all_preds, all_label) logger.info("\n") logger.info("Validation Results") logger.info("Global Steps: %d" % global_step) logger.info("Valid Loss: %2.5f" % eval_losses.avg) logger.info("Valid Accuracy: %2.5f" % accuracy) writer.add_scalar("test/accuracy", scalar_value=accuracy, global_step=global_step) if wnb: wandb.log({"acc_test": accuracy}) return accuracy def train(args, model): """ Train the model """ if args.local_rank in [-1, 0]: os.makedirs(args.output_dir, exist_ok=True) writer = SummaryWriter(log_dir=os.path.join("logs", args.name)) best_step=0 args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps # Prepare dataset train_loader, test_loader = get_loader(args) # Prepare optimizer and scheduler optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate, momentum=0.9, weight_decay=args.weight_decay) t_total = args.num_steps if args.decay_type == "cosine": scheduler = WarmupCosineSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total) else: scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total) if args.fp16: model, optimizer = amp.initialize(models=model, optimizers=optimizer, opt_level=args.fp16_opt_level) amp._amp_state.loss_scalers[0]._loss_scale = 2**20 # Distributed training if args.local_rank != -1: model = DDP(model, message_size=250000000, gradient_predivide_factor=get_world_size()) # Train! start_time = time.time() logger.info("***** Running training *****") logger.info(" Total optimization steps = %d", args.num_steps) logger.info(" Instantaneous batch size per GPU = %d", args.train_batch_size) logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d", args.train_batch_size * args.gradient_accumulation_steps * ( torch.distributed.get_world_size() if args.local_rank != -1 else 1)) logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps) model.zero_grad() set_seed(args) # Added here for reproducibility (even between python 2 and 3) losses = AverageMeter() global_step, best_acc = 0, 0 while True: model.train() epoch_iterator = tqdm(train_loader, desc="Training (X / X Steps) (loss=X.X)", bar_format="{l_bar}{r_bar}", dynamic_ncols=True, disable=args.local_rank not in [-1, 0]) all_preds, all_label = [], [] for step, batch in enumerate(epoch_iterator): batch = tuple(t.to(args.device) for t in batch) if args.sm_vit: x, y, mask = batch loss, logits = model(x, y, mask) else: x, y = batch loss, logits = model(x, y) if args.contr_loss: loss = loss.mean() preds = torch.argmax(logits, dim=-1) if len(all_preds) == 0: all_preds.append(preds.detach().cpu().numpy()) all_label.append(y.detach().cpu().numpy()) else: all_preds[0] = np.append( all_preds[0], preds.detach().cpu().numpy(), axis=0 ) all_label[0] = np.append( all_label[0], y.detach().cpu().numpy(), axis=0 ) if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() if (step + 1) % args.gradient_accumulation_steps == 0: losses.update(loss.item()*args.gradient_accumulation_steps) if args.fp16: torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) optimizer.step() scheduler.step() optimizer.zero_grad() global_step += 1 epoch_iterator.set_description( "Training (%d / %d Steps) (loss=%2.5f)" % (global_step, t_total, losses.val) ) if args.local_rank in [-1, 0]: writer.add_scalar("train/loss", scalar_value=losses.val, global_step=global_step) writer.add_scalar("train/lr", scalar_value=scheduler.get_lr()[0], global_step=global_step) if global_step % args.eval_every == 0 and args.local_rank in [-1, 0]: accuracy = valid(args, model, writer, test_loader, global_step) if best_acc < accuracy: save_model(args, model) best_acc = accuracy best_step = global_step logger.info("best accuracy so far: %f" % best_acc) logger.info("best accuracy in step: %f" % best_step) model.train() if global_step % t_total == 0: break all_preds, all_label = all_preds[0], all_label[0] accuracy = simple_accuracy(all_preds, all_label) accuracy = torch.tensor(accuracy).to(args.device) dist.barrier() train_accuracy = reduce_mean(accuracy, args.nprocs) train_accuracy = train_accuracy.detach().cpu().numpy() writer.add_scalar("train/accuracy", scalar_value=train_accuracy, global_step=global_step) if wnb: wandb.log({"acc_train": train_accuracy}) logger.info("train accuracy so far: %f" % train_accuracy) logger.info("best valid accuracy in step: %f" % best_step) losses.reset() if global_step % t_total == 0: break if args.local_rank in [-1, 0]: writer.close() end_time = time.time() logger.info("Best Accuracy: \t%f" % best_acc) logger.info("Total Training Time: \t%f" % ((end_time - start_time) / 3600)) logger.info("End Training!") def main(): parser = argparse.ArgumentParser() # Required parameters parser.add_argument("--name", required=True, default="output", help="Name of this run. Used for monitoring.") parser.add_argument("--dataset", choices=["CUB", "dogs", "nabirds"], default="CUB", help="Which downstream task.") parser.add_argument("--model_type", choices=["ViT-B_16", "ViT-B_32", "ViT-L_16", "ViT-L_32", "ViT-H_14", "R50-ViT-B_16"], default="ViT-B_16", help="Which ViT variant to use.") parser.add_argument("--pretrained_dir", type=str, default="models/pre_trained/ViT-B_16.npz", help="Where to search for pretrained ViT models.") parser.add_argument("--output_dir", default="output", type=str, help="The output directory where checkpoints will be saved.") parser.add_argument("--img_size", default=400, type=int, help="After-crop image resolution") parser.add_argument("--resize_size", default=448, type=int, help="Pre-crop image resolution") parser.add_argument("--train_batch_size", default=16, type=int, help="Total batch size for training.") parser.add_argument("--eval_batch_size", default=16, type=int, help="Total batch size for eval.") parser.add_argument("--eval_every", default=200, type=int, help="Run prediction on validation set every so many steps." "Will always run one evaluation at the end of training.") parser.add_argument("--num_workers", default=4, type=int, help="Number of workers for dataset preparation.") parser.add_argument("--learning_rate", default=3e-2, type=float, help="The initial learning rate for SGD.") parser.add_argument("--weight_decay", default=0, type=float, help="Weight deay if we apply some.") parser.add_argument("--num_steps", default=10000, type=int, help="Total number of training steps to perform.") parser.add_argument("--decay_type", choices=["cosine", "linear"], default="cosine", help="How to decay the learning rate.") parser.add_argument("--warmup_steps", default=500, type=int, help="Step of training to perform learning rate warmup for.") parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument('--gradient_accumulation_steps', type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.") parser.add_argument('--fp16', action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") parser.add_argument('--fp16_opt_level', type=str, default='O2', help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." "See details at https://nvidia.github.io/apex/amp.html") parser.add_argument('--loss_scale', type=float, default=0, help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n" "0 (default value): dynamic loss scaling.\n" "Positive power of 2: static loss scaling value.\n") parser.add_argument('--smoothing_value', type=float, default=0.0, help="Label smoothing value\n") parser.add_argument('--sm_vit', action='store_true', help="Whether to use SM-ViT") parser.add_argument('--coeff_max', type=float, default=0.25, help="Coefficient for attention guiding (see Eq. 3 in the SM-ViT paper). Best for CUB adn NABirds: '0.25', best for St Dogs: '0.3'.\n") parser.add_argument('--low_memory', action='store_true', help="Allows to use less memory (RAM) during input image feeding. False: Slower - Do image pre-processing for the whole dataset at the beginning and store the results in memory. True: Faster - Do pre-processing on-the-go.") parser.add_argument('--contr_loss', action='store_true', help="Whether to use contrastive loss") parser.add_argument('--focal_loss', action='store_true', help="Whether to use focal loss") parser.add_argument('--data_root', type=str, default='./data', # Originall help="Path to the dataset\n") # '/l/users/20020067/Datasets/CUB_200_2011/CUB_200_2011/CUB_200_2011') # CUB # '/l/users/20020067/Datasets/Stanford Dogs/Stanford_Dogs') # dogs # '/l/users/20020067/Datasets/NABirds/NABirds') # NABirds args = parser.parse_args() #args.data_root = '{}/{}'.format(args.data_root, args.dataset) # for future development # Setup CUDA, GPU & distributed training if args.local_rank == -1: device = torch.device("cuda" if torch.cuda.is_available() else "cpu") args.n_gpu = torch.cuda.device_count() else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) torch.distributed.init_process_group(backend='nccl', timeout=timedelta(minutes=60)) args.n_gpu = 1 args.device = device args.nprocs = torch.cuda.device_count() # Setup logging logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S', level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN) logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s" % (args.local_rank, args.device, args.n_gpu, bool(args.local_rank != -1), args.fp16)) # Set seed set_seed(args) # Model & Tokenizer Setup args, model = setup(args) if wnb: wandb.watch(model) # Training train(args, model) if __name__ == "__main__": main()
17,894
39.763098
247
py
sm-vit
sm-vit-main/models/modeling.py
# coding=utf-8 from __future__ import absolute_import from __future__ import division from __future__ import print_function import copy import logging import math from os.path import join as pjoin from re import X from matplotlib.cbook import flatten import torch import torch.nn as nn import numpy as np from torch.nn import CrossEntropyLoss, Dropout, Softmax, Linear, Conv2d, LayerNorm from torch.nn.modules.utils import _pair from scipy import ndimage import torch.nn.functional as F import models.configs as configs debug_mode = False # For debug if debug_mode: import random ATTENTION_Q = "MultiHeadDotProductAttention_1/query" ATTENTION_K = "MultiHeadDotProductAttention_1/key" logger = logging.getLogger(__name__) ATTENTION_V = "MultiHeadDotProductAttention_1/value" ATTENTION_OUT = "MultiHeadDotProductAttention_1/out" FC_0 = "MlpBlock_3/Dense_0" FC_1 = "MlpBlock_3/Dense_1" ATTENTION_NORM = "LayerNorm_0" MLP_NORM = "LayerNorm_2" def np2th(weights, conv=False): """Possibly convert HWIO to OIHW.""" if conv: weights = weights.transpose([3, 2, 0, 1]) return torch.from_numpy(weights) def swish(x): return x * torch.sigmoid(x) class LabelSmoothing(nn.Module): """ NLL loss with label smoothing. """ def __init__(self, smoothing=0.0): """ Constructor for the LabelSmoothing module. :param smoothing: label smoothing factor """ super(LabelSmoothing, self).__init__() self.confidence = 1.0 - smoothing self.smoothing = smoothing def forward(self, x, target): logprobs = torch.nn.functional.log_softmax(x, dim=-1) nll_loss = -logprobs.gather(dim=-1, index=target.unsqueeze(1)) nll_loss = nll_loss.squeeze(1) smooth_loss = -logprobs.mean(dim=-1) loss = self.confidence * nll_loss + self.smoothing * smooth_loss return loss.mean() class FocalLoss(torch.nn.Module): def __init__(self, alpha=1, gamma=2, reduce=True): super(FocalLoss, self).__init__() self.alpha = alpha self.gamma = gamma self.reduce = reduce def forward(self, inputs, targets): BCE_loss = torch.nn.CrossEntropyLoss()(inputs, targets) pt = torch.exp(-BCE_loss) F_loss = self.alpha * (1-pt)**self.gamma * BCE_loss if self.reduce: return torch.mean(F_loss) else: return F_loss ACT2FN = {"gelu": torch.nn.functional.gelu, "relu": torch.nn.functional.relu, "swish": swish} class Attention(nn.Module): def __init__(self, config, vis, coeff_max=0.25): super(Attention, self).__init__() self.coeff_max = coeff_max self.vis = vis self.num_attention_heads = config.transformer["num_heads"] self.attention_head_size = int(config.hidden_size / self.num_attention_heads) self.all_head_size = self.num_attention_heads * self.attention_head_size self.query = Linear(config.hidden_size, self.all_head_size) self.key = Linear(config.hidden_size, self.all_head_size) self.value = Linear(config.hidden_size, self.all_head_size) self.out = Linear(config.hidden_size, config.hidden_size) self.attn_dropout = Dropout(config.transformer["attention_dropout_rate"]) self.proj_dropout = Dropout(config.transformer["attention_dropout_rate"]) self.softmax = Softmax(dim=-1) self.softmax2 = Softmax(dim=-2) def transpose_for_scores(self, x): new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size) x = x.view(*new_x_shape) return x.permute(0, 2, 1, 3) def forward(self, hidden_states, mask=None): mixed_query_layer = self.query(hidden_states) mixed_key_layer = self.key(hidden_states) mixed_value_layer = self.value(hidden_states) query_layer = self.transpose_for_scores(mixed_query_layer) key_layer = self.transpose_for_scores(mixed_key_layer) value_layer = self.transpose_for_scores(mixed_value_layer) attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2)) attention_scores = attention_scores / math.sqrt(self.attention_head_size) if mask is not None: if debug_mode: print_info = True if (random.random() < 0.000001) else False x = random.random() if (x > 0.00005) and (x < 0.00007): print_info = True else: print_info = False else: print_info = False max_as = torch.max(attention_scores[:, :, 0, :], dim=2, keepdim=False)[0] max_as = max_as.to(device='cuda') if print_info: print("mask before:", mask) print("attn scores before:", attention_scores[:, :, 0, :]) print("attn scores max_min before:") print(max_as, torch.min(attention_scores[:, :, 0, :], dim=2, keepdim=False)[0]) print(torch.topk(attention_scores[:, :, 0, :], 5, largest=True), torch.topk(attention_scores[:, :, 0, :], 5, largest=False)) mask_626 = torch.zeros(mask.size(0), (mask.size(1) + 1)) #, dtype=torch.float64) # dtype=torch.double) mask_626 = mask_626.to(device='cuda') mask_626[:, 1:] = mask[:, :] mask_626[:, 0] = 0 if print_info: print("mask626:", mask_626) # positive only, obj + (max * coeff): attention_scores[:, :, 0, :] = \ torch.where( mask_626[:, None, :] < 0.5, \ torch.add( attention_scores[:, :, 0, :], \ torch.mul( max_as[:, :, None] , torch.tensor(self.coeff_max).cuda()) ), \ attention_scores[:, :, 0, :] #.float() ) if print_info: print("attn scores after:", attention_scores[:, :, 0, :]) print("attn scores max_min after:") print(torch.max(attention_scores[:, :, 0, :]), torch.min(attention_scores[:, :, 0, :])) print(torch.topk(attention_scores[:, :, 0, :], 5, largest=True), torch.topk(attention_scores[:, :, 0, :], 5, largest=False)) attention_probs = self.softmax(attention_scores) weights = attention_probs if self.vis else None attention_probs = self.attn_dropout(attention_probs) context_layer = torch.matmul(attention_probs, value_layer) context_layer = context_layer.permute(0, 2, 1, 3).contiguous() new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,) context_layer = context_layer.view(*new_context_layer_shape) attention_output = self.out(context_layer) attention_output = self.proj_dropout(attention_output) return attention_output, weights, self.softmax2(attention_scores)[:,:,:,0] class Mlp(nn.Module): def __init__(self, config): super(Mlp, self).__init__() self.fc1 = Linear(config.hidden_size, config.transformer["mlp_dim"]) self.fc2 = Linear(config.transformer["mlp_dim"], config.hidden_size) self.act_fn = ACT2FN["gelu"] self.dropout = Dropout(config.transformer["dropout_rate"]) self._init_weights() def _init_weights(self): nn.init.xavier_uniform_(self.fc1.weight) nn.init.xavier_uniform_(self.fc2.weight) nn.init.normal_(self.fc1.bias, std=1e-6) nn.init.normal_(self.fc2.bias, std=1e-6) def forward(self, x): x = self.fc1(x) x = self.act_fn(x) x = self.dropout(x) x = self.fc2(x) x = self.dropout(x) return x class Embeddings(nn.Module): """Construct the embeddings from patch, position embeddings. """ def __init__(self, config, img_size, in_channels=3): super(Embeddings, self).__init__() self.hybrid = None img_size = _pair(img_size) # EXPERIMENTAL. Overlapping patches: overlap = False if overlap: slide = 12 # 14 if config.patches.get("grid") is not None: grid_size = config.patches["grid"] patch_size = (img_size[0] // 16 // grid_size[0], img_size[1] // 16 // grid_size[1]) n_patches = (img_size[0] // 16) * (img_size[1] // 16) self.hybrid = True else: patch_size = _pair(config.patches["size"]) if overlap: n_patches = ((img_size[0] - patch_size[0]) // slide + 1) * ((img_size[1] - patch_size[1]) // slide + 1) else: n_patches = (img_size[0] // patch_size[0]) * (img_size[1] // patch_size[1]) self.hybrid = False if overlap: self.patch_embeddings = Conv2d(in_channels=in_channels, out_channels=config.hidden_size, kernel_size=patch_size, stride=(slide, slide) ) else: self.patch_embeddings = Conv2d(in_channels=in_channels, out_channels=config.hidden_size, kernel_size=patch_size, stride=patch_size ) self.position_embeddings = nn.Parameter(torch.zeros(1, n_patches+1, config.hidden_size)) self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size)) self.dropout = Dropout(config.transformer["dropout_rate"]) def forward(self, x): B = x.shape[0] cls_tokens = self.cls_token.expand(B, -1, -1) x = self.patch_embeddings(x) x = x.flatten(2) x = x.transpose(-1, -2) x = torch.cat((cls_tokens, x), dim=1) embeddings = x + self.position_embeddings embeddings = self.dropout(embeddings) return embeddings class Block(nn.Module): def __init__(self, config, vis, coeff_max): super(Block, self).__init__() self.hidden_size = config.hidden_size self.attention_norm = LayerNorm(config.hidden_size, eps=1e-6) self.ffn_norm = LayerNorm(config.hidden_size, eps=1e-6) self.ffn = Mlp(config) self.attn = Attention(config, vis, coeff_max) def forward(self, x, mask=None): h = x x = self.attention_norm(x) x, weights, contribution = self.attn(x, mask) x = x + h h = x x = self.ffn_norm(x) x = self.ffn(x) x = x + h return x, weights, contribution def load_from(self, weights, n_block): ROOT = f"Transformer/encoderblock_{n_block}" with torch.no_grad(): query_weight = np2th(weights[pjoin(ROOT, ATTENTION_Q, "kernel")]).view(self.hidden_size, self.hidden_size).t() key_weight = np2th(weights[pjoin(ROOT, ATTENTION_K, "kernel")]).view(self.hidden_size, self.hidden_size).t() value_weight = np2th(weights[pjoin(ROOT, ATTENTION_V, "kernel")]).view(self.hidden_size, self.hidden_size).t() out_weight = np2th(weights[pjoin(ROOT, ATTENTION_OUT, "kernel")]).view(self.hidden_size, self.hidden_size).t() query_bias = np2th(weights[pjoin(ROOT, ATTENTION_Q, "bias")]).view(-1) key_bias = np2th(weights[pjoin(ROOT, ATTENTION_K, "bias")]).view(-1) value_bias = np2th(weights[pjoin(ROOT, ATTENTION_V, "bias")]).view(-1) out_bias = np2th(weights[pjoin(ROOT, ATTENTION_OUT, "bias")]).view(-1) self.attn.query.weight.copy_(query_weight) self.attn.key.weight.copy_(key_weight) self.attn.value.weight.copy_(value_weight) self.attn.out.weight.copy_(out_weight) self.attn.query.bias.copy_(query_bias) self.attn.key.bias.copy_(key_bias) self.attn.value.bias.copy_(value_bias) self.attn.out.bias.copy_(out_bias) mlp_weight_0 = np2th(weights[pjoin(ROOT, FC_0, "kernel")]).t() mlp_weight_1 = np2th(weights[pjoin(ROOT, FC_1, "kernel")]).t() mlp_bias_0 = np2th(weights[pjoin(ROOT, FC_0, "bias")]).t() mlp_bias_1 = np2th(weights[pjoin(ROOT, FC_1, "bias")]).t() self.ffn.fc1.weight.copy_(mlp_weight_0) self.ffn.fc2.weight.copy_(mlp_weight_1) self.ffn.fc1.bias.copy_(mlp_bias_0) self.ffn.fc2.bias.copy_(mlp_bias_1) self.attention_norm.weight.copy_(np2th(weights[pjoin(ROOT, ATTENTION_NORM, "scale")])) self.attention_norm.bias.copy_(np2th(weights[pjoin(ROOT, ATTENTION_NORM, "bias")])) self.ffn_norm.weight.copy_(np2th(weights[pjoin(ROOT, MLP_NORM, "scale")])) self.ffn_norm.bias.copy_(np2th(weights[pjoin(ROOT, MLP_NORM, "bias")])) class Encoder(nn.Module): def __init__(self, config, vis, coeff_max): super(Encoder, self).__init__() self.vis = vis self.layer = nn.ModuleList() num_layers = config.transformer["num_layers"] self.encoder_norm = LayerNorm(config.hidden_size, eps=1e-6) for _ in range(num_layers): layer = Block(config, vis, coeff_max) self.layer.append(copy.deepcopy(layer)) def forward(self, hidden_states, mask=None): attn_weights = [] contributions = [] tokens = [[] for i in range(hidden_states.shape[0])] for layer_block in self.layer: hidden_states, weights, contribution = layer_block(hidden_states, mask) if self.vis: attn_weights.append(weights) contributions.append(contribution) encoded = self.encoder_norm(hidden_states) return encoded, attn_weights class Transformer(nn.Module): def __init__(self, config, img_size, vis, coeff_max): super(Transformer, self).__init__() self.embeddings = Embeddings(config, img_size=img_size) self.encoder = Encoder(config, vis, coeff_max) def forward(self, input_ids, mask=None): embedding_output = self.embeddings(input_ids) encoded, attn_weights = self.encoder(embedding_output, mask) return encoded, attn_weights class VisionTransformer(nn.Module): def __init__(self, config, img_size=400, num_classes=200, smoothing_value=0, zero_head=False, vis=False, dataset='CUB', coeff_max=0.25, contr_loss=False, focal_loss=False): super(VisionTransformer, self).__init__() self.num_classes = num_classes self.zero_head = zero_head self.smoothing_value = smoothing_value self.classifier = config.classifier self.dataset=dataset self.contr_loss = contr_loss self.focal_loss = focal_loss self.transformer = Transformer(config, img_size, vis, coeff_max) self.head = Linear(config.hidden_size, num_classes) def forward(self, x, labels=None, mask=None): x, attn_weights = self.transformer(x, mask) logits = self.head(x[:, 0]) if labels is not None: if self.smoothing_value == 0: loss_fct = CrossEntropyLoss() else: loss_fct = LabelSmoothing(self.smoothing_value) if self.focal_loss: # enforce another type of loss loss_fct = FocalLoss() ce_loss = loss_fct(logits.view(-1, self.num_classes), labels.view(-1)) if self.contr_loss: contrast_loss = con_loss(x[:, 0], labels.view(-1)) loss = ce_loss + contrast_loss else: loss = ce_loss # FFVT return loss, logits else: return logits, attn_weights def load_from(self, weights): with torch.no_grad(): if self.zero_head: nn.init.zeros_(self.head.weight) nn.init.zeros_(self.head.bias) else: self.head.weight.copy_(np2th(weights["head/kernel"]).t()) self.head.bias.copy_(np2th(weights["head/bias"]).t()) self.transformer.embeddings.patch_embeddings.weight.copy_(np2th(weights["embedding/kernel"], conv=True)) self.transformer.embeddings.patch_embeddings.bias.copy_(np2th(weights["embedding/bias"])) self.transformer.embeddings.cls_token.copy_(np2th(weights["cls"])) self.transformer.encoder.encoder_norm.weight.copy_(np2th(weights["Transformer/encoder_norm/scale"])) self.transformer.encoder.encoder_norm.bias.copy_(np2th(weights["Transformer/encoder_norm/bias"])) posemb = np2th(weights["Transformer/posembed_input/pos_embedding"]) posemb_new = self.transformer.embeddings.position_embeddings if posemb.size() == posemb_new.size(): self.transformer.embeddings.position_embeddings.copy_(posemb) else: logger.info("load_pretrained: resized variant: %s to %s" % (posemb.size(), posemb_new.size())) ntok_new = posemb_new.size(1) if self.classifier == "token": posemb_tok, posemb_grid = posemb[:, :1], posemb[0, 1:] ntok_new -= 1 else: posemb_tok, posemb_grid = posemb[:, :0], posemb[0] gs_old = int(np.sqrt(len(posemb_grid))) gs_new = int(np.sqrt(ntok_new)) print('load_pretrained: grid-size from %s to %s' % (gs_old, gs_new)) posemb_grid = posemb_grid.reshape(gs_old, gs_old, -1) zoom = (gs_new / gs_old, gs_new / gs_old, 1) posemb_grid = ndimage.zoom(posemb_grid, zoom, order=1) posemb_grid = posemb_grid.reshape(1, gs_new * gs_new, -1) posemb = np.concatenate([posemb_tok, posemb_grid], axis=1) self.transformer.embeddings.position_embeddings.copy_(np2th(posemb)) for bname, block in self.transformer.encoder.named_children(): if bname.startswith('ff') == False: for uname, unit in block.named_children(): unit.load_from(weights, n_block=uname) if self.transformer.embeddings.hybrid: self.transformer.embeddings.hybrid_model.root.conv.weight.copy_(np2th(weights["conv_root/kernel"], conv=True)) gn_weight = np2th(weights["gn_root/scale"]).view(-1) gn_bias = np2th(weights["gn_root/bias"]).view(-1) self.transformer.embeddings.hybrid_model.root.gn.weight.copy_(gn_weight) self.transformer.embeddings.hybrid_model.root.gn.bias.copy_(gn_bias) for bname, block in self.transformer.embeddings.hybrid_model.body.named_children(): for uname, unit in block.named_children(): unit.load_from(weights, n_block=bname, n_unit=uname) def con_loss(features, labels): B, _ = features.shape features = F.normalize(features) cos_matrix = features.mm(features.t()) pos_label_matrix = torch.stack([labels == labels[i] for i in range(B)]).float() neg_label_matrix = 1 - pos_label_matrix pos_cos_matrix = 1 - cos_matrix neg_cos_matrix = cos_matrix - 0.4 neg_cos_matrix[neg_cos_matrix < 0] = 0 loss = (pos_cos_matrix * pos_label_matrix).sum() + (neg_cos_matrix * neg_label_matrix).sum() loss /= (B * B) return loss CONFIGS = { 'ViT-B_16': configs.get_b16_config(), 'ViT-B_32': configs.get_b32_config(), 'ViT-L_16': configs.get_l16_config(), 'ViT-L_32': configs.get_l32_config(), 'ViT-H_14': configs.get_h14_config(), 'R50-ViT-B_16': configs.get_r50_b16_config(), 'testing': configs.get_testing(), }
19,835
38.12426
176
py
sm-vit
sm-vit-main/utils/data_utils.py
import logging import torch from torchvision import transforms, datasets from .dataset import * from torch.utils.data import DataLoader, RandomSampler, DistributedSampler, SequentialSampler from PIL import Image from .autoaugment import AutoAugImageNetPolicy import os logger = logging.getLogger(__name__) def get_loader(args): if args.local_rank not in [-1, 0]: torch.distributed.barrier() transform_train = transforms.Compose([ transforms.RandomResizedCrop((args.img_size, args.img_size), scale=(0.05, 1.0)), transforms.ToTensor(), transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]), ]) transform_test = transforms.Compose([ transforms.Resize((args.img_size, args.img_size)), transforms.ToTensor(), transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]), ]) if args.dataset == 'dogs': if args.sm_vit: train_transform=transforms.Compose([ transforms.Resize((args.img_size, args.img_size), Image.BILINEAR), transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4), # transforms.RandomHorizontalFlip(), !!! FLIPPING in dataset.py !!! transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) test_transform=transforms.Compose([ transforms.Resize((args.img_size, args.img_size), Image.BILINEAR), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) else: train_transform=transforms.Compose([ transforms.Resize((args.resize_size, args.resize_size), Image.BILINEAR), transforms.RandomCrop((args.img_size, args.img_size)), transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) test_transform=transforms.Compose([ transforms.Resize((args.resize_size, args.resize_size), Image.BILINEAR), transforms.CenterCrop((args.img_size, args.img_size)), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) trainset = dogs(args.dataset, root=args.data_root, is_train=True, cropped=False, transform=train_transform, download=False, sm_vit=args.sm_vit, low_memory=args.low_memory, img_size=args.img_size ) testset = dogs(args.dataset, root=args.data_root, is_train=False, cropped=False, transform=test_transform, download=False, sm_vit=args.sm_vit, low_memory=args.low_memory, img_size=args.img_size ) elif args.dataset== "CUB": if args.sm_vit: train_transform=transforms.Compose([ transforms.Resize((args.img_size, args.img_size),Image.BILINEAR), transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4), #transforms.RandomHorizontalFlip(), # !!! FLIPPING in dataset.py !!! transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), ]) test_transform=transforms.Compose([ transforms.Resize((args.img_size, args.img_size),Image.BILINEAR), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) else: train_transform=transforms.Compose([ transforms.Resize((args.resize_size, args.resize_size),Image.BILINEAR), transforms.RandomCrop((args.img_size, args.img_size)), transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), ]) test_transform=transforms.Compose([ transforms.Resize((args.resize_size, args.resize_size), Image.BILINEAR), transforms.CenterCrop((args.img_size, args.img_size)), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) trainset = eval(args.dataset)(args.dataset, root=args.data_root, is_train=True, \ transform=train_transform, sm_vit=args.sm_vit, low_memory=args.low_memory, img_size=args.img_size) testset = eval(args.dataset)(args.dataset, root=args.data_root, is_train=False, \ transform = test_transform, sm_vit=args.sm_vit, low_memory=args.low_memory, img_size=args.img_size) elif args.dataset == 'nabirds': if args.sm_vit: train_transform=transforms.Compose([ transforms.Resize((args.img_size, args.img_size), Image.BILINEAR), transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4), # my add (from FFVT) mb try? #transforms.RandomHorizontalFlip(), # !!! FLIPPING in dataset.py !!! transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) test_transform=transforms.Compose([ transforms.Resize((args.img_size, args.img_size), Image.BILINEAR), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) else: train_transform=transforms.Compose([ transforms.Resize((args.resize_size, args.resize_size), Image.BILINEAR), transforms.RandomCrop((args.img_size, args.img_size)), transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) test_transform=transforms.Compose([ transforms.Resize((args.resize_size, args.resize_size), Image.BILINEAR), transforms.CenterCrop((args.img_size, args.img_size)), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) trainset = NABirds(args.dataset, root=args.data_root, is_train=True, \ transform=train_transform, sm_vit=args.sm_vit, low_memory=args.low_memory, img_size=args.img_size) testset = NABirds(args.dataset, root=args.data_root, is_train=False, \ transform=test_transform, sm_vit=args.sm_vit, low_memory=args.low_memory, img_size=args.img_size) ### Not optimised datasets: if args.dataset == 'INat2017': train_transform=transforms.Compose([transforms.Resize((400, 400), Image.BILINEAR), transforms.RandomCrop((304, 304)), transforms.RandomHorizontalFlip(), AutoAugImageNetPolicy(), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) test_transform=transforms.Compose([transforms.Resize((400, 400), Image.BILINEAR), transforms.CenterCrop((304, 304)), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) trainset = INat2017(args.data_root, 'train', train_transform) testset = INat2017(args.data_root, 'val', test_transform) elif args.dataset == 'car': trainset = CarsDataset(os.path.join(args.data_root,'devkit/cars_train_annos.mat'), os.path.join(args.data_root,'cars_train'), os.path.join(args.data_root,'devkit/cars_meta.mat'), # cleaned=os.path.join(data_dir,'cleaned.dat'), transform=transforms.Compose([ transforms.Resize((600, 600), Image.BILINEAR), transforms.RandomCrop((448, 448)), transforms.RandomHorizontalFlip(), AutoAugImageNetPolicy(), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) ) testset = CarsDataset(os.path.join(args.data_root,'cars_test_annos_withlabels.mat'), os.path.join(args.data_root,'cars_test'), os.path.join(args.data_root,'devkit/cars_meta.mat'), # cleaned=os.path.join(data_dir,'cleaned_test.dat'), transform=transforms.Compose([ transforms.Resize((600, 600), Image.BILINEAR), transforms.CenterCrop((448, 448)), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) ) elif args.dataset== "air": train_transform=transforms.Compose([transforms.Resize((args.resize_size, args.resize_size),Image.BILINEAR), transforms.RandomCrop((args.img_size, args.img_size)), transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4), # my add transforms.RandomHorizontalFlip(), #transforms.RandomVerticalFlip(), transforms.ToTensor(), #transforms.Normalize([0.8416, 0.867, 0.8233], [0.2852, 0.246, 0.3262])]) transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) test_transform=transforms.Compose([ transforms.Resize((args.resize_size, args.resize_size), Image.BILINEAR), transforms.CenterCrop((args.img_size, args.img_size)), #transforms.Resize((args.img_size, args.img_size)), transforms.ToTensor(), #transforms.Normalize([0.8416, 0.867, 0.8233], [0.2852, 0.246, 0.3262])]) transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) trainset = FGVC_aircraft(root=args.data_root, is_train=True, transform=train_transform) testset = FGVC_aircraft(root=args.data_root, is_train=False, transform = test_transform) if args.local_rank == 0: torch.distributed.barrier() train_sampler = RandomSampler(trainset) if args.local_rank == -1 else DistributedSampler(trainset) test_sampler = SequentialSampler(testset) train_loader = DataLoader(trainset, sampler=train_sampler, batch_size=args.train_batch_size, num_workers=args.num_workers, pin_memory=True) test_loader = DataLoader(testset, sampler=test_sampler, batch_size=args.eval_batch_size, num_workers=args.num_workers, pin_memory=True) if testset is not None else None return train_loader, test_loader
12,747
46.924812
124
py
sm-vit
sm-vit-main/utils/dataset.py
import os import json from os.path import join import numpy as np import scipy from scipy import io import scipy.misc from PIL import Image import pandas as pd import matplotlib.pyplot as plt import torch from torch.utils.data import Dataset from torchvision.datasets import VisionDataset from torchvision.datasets.folder import default_loader from torchvision.datasets.utils import download_url, list_dir, check_integrity, extract_archive, verify_str_arg # My: from torchvision import transforms from torchvision.utils import save_image import random from torchvision.transforms import functional as F import U2Net from U2Net.u2net_test import mask_hw from skimage import transform as transform_sk import gc # class Generic_smvit_DS(): def generic_preprocess(self, file_list, file_list_full, shape_hw_list, data_len, train_test_list=None): img = [] mask = [] ## For other experiments: if self.ds_name != "CUB": self.gt_bbox = False self.gt_parts = False if self.gt_bbox: bounding_boxes_file = open(os.path.join(self.root, 'bounding_boxes.txt')) bb_list = [] for line in bounding_boxes_file: bb_list_x = line[:-1].split(' ')[-4] bb_list_y = line[:-1].split(' ')[-3] bb_list_w = line[:-1].split(' ')[-2] bb_list_h = line[:-1].split(' ')[-1] bb_list.append( [ int(bb_list_x.split('.')[0]), int(bb_list_y.split('.')[0]), int(bb_list_w.split('.')[0]), int(bb_list_h.split('.')[0]) ] ) bb_list = [x for i, x in zip(train_test_list, bb_list) if i] if self.gt_parts: parts_file = open(os.path.join(self.root, 'parts/part_locs.txt')) PARTS_NUM = 15 parts_list = [] part_t = [] part_count = 0 for line in parts_file: part_t_raw_x = line[:-1].split(' ')[-3] part_t_raw_y = line[:-1].split(' ')[-2] part_t_pres = line[:-1].split(' ')[-1] part_t.append ( [ int(part_t_pres), int(part_t_raw_x.split('.')[0]), int(part_t_raw_y.split('.')[0]) ] ) part_count = part_count + 1 if (part_count >= PARTS_NUM): parts_list.append( part_t ) part_t = [] part_count = 0 parts_list = [x for i, x in zip(train_test_list, parts_list) if i] ## print(f'[INFO] Pre-processing {self.mode} files...') if self.sm_vit: if self.full_ds: mask_u2n_list, x_u2n_list, y_u2n_list, h_u2n_list, w_u2n_list = \ mask_hw(full_ds=self.full_ds, img_path=file_list_full, shape_hw=shape_hw_list) else: # for debug img_path = os.path.join(self.root, self.base_folder, file_list) img_temp = scipy.misc.imread(img_path) h_max_temp = img_temp.shape[0] # y w_max_temp = img_temp.shape[1] # x mask_u2n, x_u2n, y_u2n, h_u2n, w_u2n = \ mask_hw(full_ds=self.full_ds, img_path=img_path, shape_hw=(h_max_temp, w_max_temp)) mask_temp, x, y, h, w = mask_u2n, x_u2n, y_u2n, h_u2n, w_u2n for ind, file in enumerate(file_list[:data_len]): if self.debug: print(f"{self.mode} file:", file) img_temp = scipy.misc.imread(os.path.join(self.root, self.base_folder, file)) ## Downscale large images for memory efficiency if self.ds_name != "CUB": img_temp = (img_temp).astype(np.uint8) if (img_temp.shape[0] > self.max_res) or (img_temp.shape[1] > self.max_res): if self.debug and ind < 10: print("Before:", img_temp.shape[0], img_temp.shape[1]) img_name = ("test/img_before_tr" + str(ind) + ".png") Image.fromarray(img_temp, mode='RGB').save(img_name) if img_temp.shape[0] > img_temp.shape[1]: downscale_coef = img_temp.shape[0] / self.max_res else: downscale_coef = img_temp.shape[1] / self.max_res img_temp = transform_sk.resize( img_temp, ( int((img_temp.shape[0] // downscale_coef)), int((img_temp.shape[1] // downscale_coef)) ), \ mode='constant', anti_aliasing=True, anti_aliasing_sigma=None, preserve_range=True ) if self.debug and ind < 10: print("After:", img_temp.shape[0], img_temp.shape[1]) img_temp = (img_temp).astype(np.uint8) img_name = ("test/img_after_tr" + str(ind) + ".png") Image.fromarray(img_temp, mode='RGB').save(img_name) else: if self.debug and ind < 10: print("Normal:", img_temp.shape[0], img_temp.shape[1]) img_name = ("test/img_normal_tr" + str(ind) + ".png") Image.fromarray(img_temp, mode='RGB').save(img_name) h_max = img_temp.shape[0] # y w_max = img_temp.shape[1] # x #ch_max = img_temp.shape[2] # ch if self.gt_bbox: x, y, w, h = bb_list[ind] # x - distance from top up left (width), y - distance from top up left (height) if self.gt_parts: parts = parts_list[ind] # list of 15 parts with [x, y] center corrdinates #mask_temp = np.zeros((int(h_max), int(w_max))) # Black mask mask_temp = np.ones((int(h_max), int(w_max))) p_part = 16*3 # padding around center point for part_n in range(len(parts)): part = parts[part_n] if part[0] != 0: x_min_p = part[1] - p_part if x_min_p < 0: x_min_p = 0 x_max_p = part[1] + p_part if x_max_p > w_max: x_max_p = w_max y_min_p = part[2] - p_part if y_min_p < 0: y_min_p = 0 y_max_p = part[2] + p_part if y_max_p > h_max: y_max_p = h_max #mask_temp[int(y_min_p):int(y_max_p), int(x_min_p):int(x_max_p)] = 1 # Black mask mask_temp[int(y_min_p):int(y_max_p), int(x_min_p):int(x_max_p)] = 0 if self.sm_vit and self.full_ds: mask_temp = mask_u2n_list[ind] x = x_u2n_list[ind] y = y_u2n_list[ind] h = h_u2n_list[ind] w = w_u2n_list[ind] ## Image and Mask Padding: if self.sm_vit or self.gt_bbox: if self.padding: p = 15 # extra space around bbox else: p = 0 x_min = x - p if x_min < 0: x_min = 0 x_max = x + w + p if x_max > w_max: x_max = w_max y_min = y - p if y_min < 0: y_min = 0 y_max = y + h + p if y_max > h_max: y_max = h_max if h_max <=1: print("[WARNING] bad_h", h_max) if w_max <=1: print("[WARNING] bad_w", w_max) if y_min >= y_max: print("[WARNING] bad_y", "min:", y_min, "max:", y_max) print("[WARNING] y:", y, "h:", h) if x_min >= x_max: print("[WARNING] bad_x", "min:", x_min, "max:", x_max) print("[WARNING] x:", x, "w:", w) ## ## Crop with bbox: if self.rand_crop: #prob_rcrop = 0.25 # 0.07 # 0.3 # 0.5 #rand_crop_mask_temp = bool(random.random() < prob_rcrop) #if rand_crop_mask_temp: h_max_img = img_temp.shape[0] w_max_img = img_temp.shape[1] #h_crop_mid = 368 # 384 (92%), 368 (84%), 352 (77%), 336 (70%), 320 (64%), 304 (57%) h_crop_mid_img = int(h_max_img * 0.88) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) #w_crop_mid = 368 # 384 (92%), 368 (84%), 352 (77%), 336 (70%), 320 (64%), 304 (57%) w_crop_mid_img = int(w_max_img * 0.88) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) h_crop_min_img = random.randint(0, (h_max_img - h_crop_mid_img)) # 40) #, 400-360) #, h - th) w_crop_min_img = random.randint(0, (w_max_img - w_crop_mid_img)) # 40) #, 400-360) #, w - tw) h_crop_max_img = h_crop_mid_img + h_crop_min_img w_crop_max_img = w_crop_mid_img + w_crop_min_img # Crop image with bbox: if len(img_temp.shape) == 3: img_temp = img_temp[int(h_crop_min_img):int(h_crop_max_img), int(w_crop_min_img):int(w_crop_max_img), :] # h, w, ch else: img_temp = img_temp[int(h_crop_min_img):int(h_crop_max_img), int(w_crop_min_img):int(w_crop_max_img)] # h, w # Crop mask with bbox: mask_temp = mask_temp[int(h_crop_min_img):int(h_crop_max_img), int(w_crop_min_img):int(w_crop_max_img)] else: # Crop image with bbox: if len(img_temp.shape) == 3: if self.gt_parts: for j in range(3): img_temp[:, :, j] = img_temp[:, :, j] * mask_temp # Black mask else: #test_img_temp = test_img_temp[int(y):int(y + h), int(x):int(x + w), :] # h, w, ch img_temp = img_temp[int(y_min):int(y_max), int(x_min):int(x_max), :] # h, w, ch else: if self.gt_parts: img_temp[:, :] = img_temp[:, :] * mask_temp # Black mask: else: img_temp = img_temp[int(y_min):int(y_max), int(x_min):int(x_max)] # h, w # Crop mask with bbox: if self.sm_vit or self.gt_bbox: mask_temp = mask_temp[int(y_min):int(y_max), int(x_min):int(x_max)] ## if ( (img_temp.shape[0] != mask_temp.shape[0]) or (img_temp.shape[1] != mask_temp.shape[1]) ): print("[WARNING] Image shape does not match mask shape for sample:", ind, ". \t" , "Found shapes:", img_temp.shape, mask_temp.shape) img.append(img_temp) mask.append(mask_temp) return img, mask def generic_preprocess_lowMem(self, file_list, file_list_full, shape_hw_list): print(f'[INFO] Pre-processing {self.mode} files in the low memory mode...') if self.sm_vit: if self.full_ds: mask_u2n_list, x_u2n_list, y_u2n_list, h_u2n_list, w_u2n_list = \ mask_hw(full_ds=self.full_ds, img_path=file_list_full, shape_hw=shape_hw_list) else: # for debug img_path = os.path.join(self.root, self.base_folder, file_list) img_temp = scipy.misc.imread(img_path) h_max_temp = img_temp.shape[0] # y w_max_temp = img_temp.shape[1] # x mask_u2n, x_u2n, y_u2n, h_u2n, w_u2n = \ mask_hw(full_ds=self.full_ds, img_path=img_path, shape_hw=(h_max_temp, w_max_temp)) mask_temp, x, y, h, w = mask_u2n, x_u2n, y_u2n, h_u2n, w_u2n # mask_u2n_list, x_u2n_list, y_u2n_list, h_u2n_list, w_u2n_list = mask_temp, x, y, h, w return mask_u2n_list, x_u2n_list, y_u2n_list, h_u2n_list, w_u2n_list def generic_getitem(self, index, img, mask): if self.is_train: if self.rand_crop_im_mask: h_max_img = img.shape[0] w_max_img = img.shape[1] h_crop_mid_img = int(h_max_img * 0.88) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) w_crop_mid_img = int(w_max_img * 0.88) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) h_crop_min_img = random.randint(0, (h_max_img - h_crop_mid_img)) # 40) #, 400-360) #, h - th) w_crop_min_img = random.randint(0, (w_max_img - w_crop_mid_img)) # 40) #, 400-360) #, w - tw) h_crop_max_img = h_crop_mid_img + h_crop_min_img w_crop_max_img = w_crop_mid_img + w_crop_min_img # Crop image: if len(img.shape) == 3: img = img[int(h_crop_min_img):int(h_crop_max_img), int(w_crop_min_img):int(w_crop_max_img), :] # h, w, ch else: img = img[int(h_crop_min_img):int(h_crop_max_img), int(w_crop_min_img):int(w_crop_max_img)] # h, w # Crop mask: mask = mask[int(h_crop_min_img):int(h_crop_max_img), int(w_crop_min_img):int(w_crop_max_img)] if len(img.shape) == 2: img = np.stack([img] * 3, 2) if self.ds_name != "CUB": img = (img).astype(np.uint8) img = Image.fromarray(img, mode='RGB') if self.debug and index < 10: img_tem = transforms.ToTensor()(img) img_name = ("test/img_bef" + str(index) + ".png") save_image( img_tem, img_name) ## Image: if self.transform is not None: if self.is_train: if not self.flip_mask_as_image: # normal img = self.transform(img) else: if random.random() < 0.5: flipped = False img = self.transform(img) else: flipped = True transform_img_flip = transforms.Compose([ #transforms.Resize((args.resize_size, args.resize_size),Image.BILINEAR), #transforms.RandomCrop((args.img_size, args.img_size)), transforms.Resize((self.img_size, self.img_size),Image.BILINEAR), # my for bbox transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4), # my add (FFVT) transforms.RandomHorizontalFlip(p=1.0), # !!! FLIPPING in dataset.py !!! transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), ]) img = transform_img_flip(img) else: img = self.transform(img) if self.debug and index < 10: img_name = ("test/img_aft" + str(index) + ".png") save_image( img, img_name) ## Mask: if self.crop_mask: h_max_im = mask.shape[0] w_max_im = mask.shape[1] h_crop_mid = int(h_max_im * 0.84) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) w_crop_mid = int(w_max_im * 0.84) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) cropped = np.ones_like(mask) if self.mid_val: cropped = cropped * 0.125 # (for 0.2) h_crop_min = random.randint(0, (h_max_im - h_crop_mid)) # 40) #, 400-360) #, h - th) w_crop_min = random.randint(0, (w_max_im - w_crop_mid)) # 40) #, 400-360) #, w - tw) h_crop_max = h_crop_mid + h_crop_min w_crop_max = w_crop_mid + w_crop_min cropped[int(h_crop_min):int(h_crop_max), int(w_crop_min):int(w_crop_max)] = 0 mask = mask + cropped if self.mid_val: mask[mask > 1.1] = 1 else: mask[mask > 1] = 1 mask = (mask * 255).astype(np.uint8) mask = Image.fromarray(mask, mode='L') if self.debug and index < 10: mask_tem = transforms.ToTensor()(mask) img_name = ("test/mask_bef" + str(index) + ".png") save_image( mask_tem, img_name) mask_size = int(self.img_size // 16) if self.is_train: if not self.flip_mask_as_image: # normal transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(), #(mode='1'), # non-overlapped: transforms.Resize((mask_size, mask_size), interpolation=Image.NEAREST), #Image.BILINEAR), # interpolation=T.InterpolationMode.NEAREST transforms.ToTensor() ]) else: if flipped: transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(), #(mode='1'), transforms.RandomHorizontalFlip(p=1.0), # non-overlapped: transforms.Resize((mask_size, mask_size), interpolation=Image.NEAREST), #Image.BILINEAR), # interpolation=T.InterpolationMode.NEAREST transforms.ToTensor() ]) else: transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(), #(mode='1'), # non-overlapped: transforms.Resize((mask_size, mask_size), interpolation=Image.NEAREST), #Image.BILINEAR), # interpolation=T.InterpolationMode.NEAREST transforms.ToTensor() ]) else: transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(), #(mode='1'), # non-overlapped: transforms.Resize((mask_size, mask_size), interpolation=Image.NEAREST), #Image.BILINEAR), # interpolation=T.InterpolationMode.NEAREST transforms.ToTensor()]) mask = transform_mask(mask) if self.debug and index < 10: img_name = ("test/mask_aft" + str(index) + ".png") save_image(mask, img_name) mask = torch.flatten(mask) return img, mask def generic_getitem_lowMem(self, index): file_temp = self.file_list[index] img_temp = scipy.misc.imread(os.path.join(self.root, self.base_folder, file_temp)) ## Downscale large images for memory efficiency if self.ds_name != "CUB": self.gt_bbox = False self.gt_parts = False img_temp = (img_temp).astype(np.uint8) if (img_temp.shape[0] > self.max_res) or (img_temp.shape[1] > self.max_res): if self.debug and index < 10: print("Before:", img_temp.shape[0], img_temp.shape[1]) img_name = ("test/img_before_tr" + str(index) + ".png") Image.fromarray(img_temp, mode='RGB').save(img_name) if img_temp.shape[0] > img_temp.shape[1]: downscale_coef = img_temp.shape[0] / self.max_res else: downscale_coef = img_temp.shape[1] / self.max_res img_temp = transform_sk.resize( img_temp, ( int((img_temp.shape[0] // downscale_coef)), int((img_temp.shape[1] // downscale_coef)) ), \ mode='constant', anti_aliasing=True, anti_aliasing_sigma=None, preserve_range=True ) if self.debug and index < 10: print("After:", img_temp.shape[0], img_temp.shape[1]) img_temp = (img_temp).astype(np.uint8) img_name = ("test/img_after_tr" + str(index) + ".png") Image.fromarray(img_temp, mode='RGB').save(img_name) else: if self.debug and index < 10: print("Normal:", img_temp.shape[0], img_temp.shape[1]) img_name = ("test/img_normal_tr" + str(index) + ".png") Image.fromarray(img_temp, mode='RGB').save(img_name) ## h_max = img_temp.shape[0] # y w_max = img_temp.shape[1] # x #ch_max = img_temp.shape[2] # ch mask_temp = self.mask_u2n_list[index] x, y, h, w = self.x_u2n_list[index], self.y_u2n_list[index], self.h_u2n_list[index], self.w_u2n_list[index] ## Image and Mask Padding: if self.sm_vit or self.gt_bbox: if self.padding: p = 15 # extra space around bbox else: p = 0 x_min = x - p if x_min < 0: x_min = 0 x_max = x + w + p if x_max > w_max: x_max = w_max y_min = y - p if y_min < 0: y_min = 0 y_max = y + h + p if y_max > h_max: y_max = h_max if h_max <=1: print("[WARNING] bad_h", h_max) if w_max <=1: print("[WARNING] bad_w", w_max) if y_min >= y_max: print("[WARNING] bad_y", "min:", y_min, "max:", y_max) print("[WARNING] y:", y, "h:", h) if x_min >= x_max: print("[WARNING] bad_x", "min:", x_min, "max:", x_max) print("[WARNING] x:", x, "w:", w) ## ## Crop with bbox: if self.rand_crop: #prob_rcrop = 0.25 # 0.07 # 0.3 # 0.5 #rand_crop_mask_temp = bool(random.random() < prob_rcrop) #if rand_crop_mask_temp: h_max_img = img_temp.shape[0] w_max_img = img_temp.shape[1] #h_crop_mid = 368 # 384 (92%), 368 (84%), 352 (77%), 336 (70%), 320 (64%), 304 (57%) h_crop_mid_img = int(h_max_img * 0.88) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) #w_crop_mid = 368 # 384 (92%), 368 (84%), 352 (77%), 336 (70%), 320 (64%), 304 (57%) w_crop_mid_img = int(w_max_img * 0.88) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) h_crop_min_img = random.randint(0, (h_max_img - h_crop_mid_img)) # 40) #, 400-360) #, h - th) w_crop_min_img = random.randint(0, (w_max_img - w_crop_mid_img)) # 40) #, 400-360) #, w - tw) h_crop_max_img = h_crop_mid_img + h_crop_min_img w_crop_max_img = w_crop_mid_img + w_crop_min_img # Crop image with bbox: if len(img_temp.shape) == 3: img_temp = img_temp[int(h_crop_min_img):int(h_crop_max_img), int(w_crop_min_img):int(w_crop_max_img), :] # h, w, ch else: img_temp = img_temp[int(h_crop_min_img):int(h_crop_max_img), int(w_crop_min_img):int(w_crop_max_img)] # h, w # Crop mask with bbox: mask_temp = mask_temp[int(h_crop_min_img):int(h_crop_max_img), int(w_crop_min_img):int(w_crop_max_img)] else: # Crop image with bbox: if len(img_temp.shape) == 3: if self.gt_parts: for j in range(3): img_temp[:, :, j] = img_temp[:, :, j] * mask_temp # Black mask else: #test_img_temp = test_img_temp[int(y):int(y + h), int(x):int(x + w), :] # h, w, ch img_temp = img_temp[int(y_min):int(y_max), int(x_min):int(x_max), :] # h, w, ch else: if self.gt_parts: img_temp[:, :] = img_temp[:, :] * mask_temp # Black mask: else: img_temp = img_temp[int(y_min):int(y_max), int(x_min):int(x_max)] # h, w # Crop mask with bbox: if self.sm_vit or self.gt_bbox: mask_temp = mask_temp[int(y_min):int(y_max), int(x_min):int(x_max)] ## if ( (img_temp.shape[0] != mask_temp.shape[0]) or (img_temp.shape[1] != mask_temp.shape[1]) ): print("[WARNING] Image shape does not match mask shape for sample:", index, ". \t" , \ "Found shapes:", img_temp.shape, mask_temp.shape) img = img_temp if len(img.shape) == 2: img = np.stack([img] * 3, 2) if self.ds_name != "CUB": img = (img).astype(np.uint8) img = Image.fromarray(img, mode='RGB') if self.debug and index < 10: img_tem = transforms.ToTensor()(img) img_name = ("test/img_bef" + str(index) + ".png") save_image( img_tem, img_name) ## Image: if self.transform is not None: if self.is_train: if not self.flip_mask_as_image: # normal img = self.transform(img) else: if random.random() < 0.5: flipped = False img = self.transform(img) else: flipped = True transform_img_flip = transforms.Compose([ #transforms.Resize((args.resize_size, args.resize_size),Image.BILINEAR), #transforms.RandomCrop((args.img_size, args.img_size)), transforms.Resize((self.img_size, self.img_size),Image.BILINEAR), # my for bbox transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4), # my add (FFVT) transforms.RandomHorizontalFlip(p=1.0), # !!! FLIPPING in dataset.py !!! transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), ]) img = transform_img_flip(img) else: img = self.transform(img) if self.debug and index < 10: img_name = ("test/img_aft" + str(index) + ".png") save_image( img, img_name) ## Mask: mask = mask_temp if self.crop_mask: h_max_im = mask.shape[0] w_max_im = mask.shape[1] h_crop_mid = int(h_max_im * 0.84) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) w_crop_mid = int(w_max_im * 0.84) # 384 (92% - 0.96), 368 (84% - 0.92), 352 (77% - 0.88), 336 (70% - 0.84), 320 (64% - 0.80), 304 (57% - 0.75) cropped = np.ones_like(mask) if self.mid_val: cropped = cropped * 0.125 # (for 0.2) h_crop_min = random.randint(0, (h_max_im - h_crop_mid)) # 40) #, 400-360) #, h - th) w_crop_min = random.randint(0, (w_max_im - w_crop_mid)) # 40) #, 400-360) #, w - tw) h_crop_max = h_crop_mid + h_crop_min w_crop_max = w_crop_mid + w_crop_min cropped[int(h_crop_min):int(h_crop_max), int(w_crop_min):int(w_crop_max)] = 0 mask = mask + cropped if self.mid_val: mask[mask > 1.1] = 1 else: mask[mask > 1] = 1 mask = (mask * 255).astype(np.uint8) mask = Image.fromarray(mask, mode='L') if self.debug and index < 10: mask_tem = transforms.ToTensor()(mask) img_name = ("test/mask_bef" + str(index) + ".png") save_image( mask_tem, img_name) mask_size = int(self.img_size // 16) if self.is_train: if not self.flip_mask_as_image: # normal transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(), #(mode='1'), # non-overlapped: transforms.Resize((mask_size, mask_size), interpolation=Image.NEAREST), #Image.BILINEAR), # interpolation=T.InterpolationMode.NEAREST transforms.ToTensor() ]) else: if flipped: transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(), #(mode='1'), transforms.RandomHorizontalFlip(p=1.0), # non-overlapped: transforms.Resize((mask_size, mask_size), interpolation=Image.NEAREST), #Image.BILINEAR), # interpolation=T.InterpolationMode.NEAREST transforms.ToTensor() ]) else: transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(), #(mode='1'), # non-overlapped: transforms.Resize((mask_size, mask_size), interpolation=Image.NEAREST), #Image.BILINEAR), # interpolation=T.InterpolationMode.NEAREST transforms.ToTensor() ]) else: transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(), #(mode='1'), # non-overlapped: transforms.Resize((mask_size, mask_size), interpolation=Image.NEAREST), #Image.BILINEAR), # interpolation=T.InterpolationMode.NEAREST transforms.ToTensor()]) mask = transform_mask(mask) if self.debug and index < 10: img_name = ("test/mask_aft" + str(index) + ".png") save_image(mask, img_name) mask = torch.flatten(mask) return img, mask class CUB(Generic_smvit_DS): def __init__(self, ds_name, root, is_train=True, data_len=None, transform=None, sm_vit=True, low_memory=True, img_size=400): self.ds_name = ds_name self.img_size = img_size self.max_res = int(self.img_size * 1.5) self.full_ds = True # pre-processing full dataset self.padding = True # image and mask padding self.rand_crop = False # if no other cropping self.flip_mask_as_image = True # if False - turn on RandomHorizontalFlip in data_utils !!! self.rand_crop_im_mask = False # randomly crop both image and mask self.crop_mask = False # randomly crop mask only self.mid_val = False # 3-state mask self.debug = False # for debug info if self.debug: os.makedirs("./test") self.gt_bbox = False # for other experiments self.gt_parts = False # for other experiments self.sm_vit = sm_vit self.low_memory = low_memory if (self.sm_vit + self.gt_bbox + self.gt_parts) > 1 : raise Exception("Only one cropping mode (SM-ViT, bbox, parts) can be chosen") self.root = root self.base_folder = "images" self.transform = transform self.is_train = is_train if self.is_train: self.mode = "Train" else: self.mode = "Test" img_txt_file = open(os.path.join(self.root, 'images.txt')) label_txt_file = open(os.path.join(self.root, 'image_class_labels.txt')) train_val_file = open(os.path.join(self.root, 'train_test_split.txt')) img_name_list = [] for line in img_txt_file: img_name_list.append(line[:-1].split(' ')[-1]) label_list = [] for line in label_txt_file: label_list.append(int(line[:-1].split(' ')[-1]) - 1) train_test_list = [] for line in train_val_file: train_test_list.append(int(line[:-1].split(' ')[-1])) if self.is_train: self.file_list = [x for i, x in zip(train_test_list, img_name_list) if i] file_list_full = [ os.path.join(self.root, self.base_folder, x) for i, x in zip(train_test_list, img_name_list) if i] else: self.file_list = [x for i, x in zip(train_test_list, img_name_list) if not i] file_list_full = [ os.path.join(self.root, self.base_folder, x) for i, x in zip(train_test_list, img_name_list) if not i] if self.sm_vit: print(f"[INFO] Preparing {self.mode} shape_hw list...") shape_hw_list = [] for img_name in self.file_list: img_temp = scipy.misc.imread(os.path.join(self.root, self.base_folder, img_name)) shape_hw_temp = [img_temp.shape[0], img_temp.shape[1]] # h_max (y), w_max (x) shape_hw_list.append(shape_hw_temp) if self.low_memory: self.mask_u2n_list, self.x_u2n_list, self.y_u2n_list, self.h_u2n_list, self.w_u2n_list = \ super(CUB, self).generic_preprocess_lowMem(self.file_list, file_list_full, shape_hw_list ) del shape_hw_list del file_list_full gc.collect() else: self.img, self.mask = \ super(CUB, self).generic_preprocess(self.file_list, file_list_full, shape_hw_list, data_len, train_test_list ) else: self.img = \ [scipy.misc.imread(os.path.join(self.root, self.base_folder, file)) \ for file in self.file_list[:data_len]] if self.is_train: self.label = [x for i, x in zip(train_test_list, label_list) if i][:data_len] else: self.label = [x for i, x in zip(train_test_list, label_list) if not i][:data_len] self.imgname = [x for x in self.file_list[:data_len]] def __getitem__(self, index): if self.sm_vit: if self.low_memory: target, imgname = self.label[index], self.imgname[index] img, mask = super(CUB, self).generic_getitem_lowMem(index) else: img, target, imgname, mask = self.img[index], self.label[index], self.imgname[index], self.mask[index] img, mask = super(CUB, self).generic_getitem(index, img, mask) return img, target, mask else: img, target, imgname = self.img[index], self.label[index], self.imgname[index] if len(img.shape) == 2: img = np.stack([img] * 3, 2) img = Image.fromarray(img, mode='RGB') if self.transform is not None: img = self.transform(img) return img, target def __len__(self): return len(self.label) class dogs(Generic_smvit_DS): #(Dataset): """`Stanford Dogs <http://vision.stanford.edu/aditya86/ImageNetDogs/>`_ Dataset. Args: root (string): Root directory of dataset where directory ``omniglot-py`` exists. cropped (bool, optional): If true, the images will be cropped into the bounding box specified in the annotations transform (callable, optional): A function/transform that takes in an PIL image and returns a transformed version. E.g, ``transforms.RandomCrop`` target_transform (callable, optional): A function/transform that takes in the target and transforms it. download (bool, optional): If true, downloads the dataset tar files from the internet and puts it in root directory. If the tar files are already downloaded, they are not downloaded again. """ folder = 'dog' download_url_prefix = 'http://vision.stanford.edu/aditya86/ImageNetDogs' def __init__(self, ds_name, root, is_train=True, cropped=False, transform=None, target_transform=None, download=False, sm_vit=True, low_memory=True, img_size=400): self.ds_name = ds_name self.img_size = img_size self.max_res = int(self.img_size * 1.5) self.full_ds = True # pre-processing full dataset self.padding = True # image and mask padding self.rand_crop = False # if no other cropping self.flip_mask_as_image = True # if False - turn on RandomHorizontalFlip in data_utils !!! self.rand_crop_im_mask = False # randomly crop both image and mask self.crop_mask = False # randomly crop mask only self.mid_val = False # 3-state mask self.debug = False # for debug info if self.debug: os.makedirs("./test") self.sm_vit = sm_vit self.low_memory = low_memory # self.root = join(os.path.expanduser(root), self.folder) self.root = root self.base_folder = "Images" self.is_train = is_train if self.is_train: self.mode = "Train" else: self.mode = "Test" self.cropped = cropped self.transform = transform self.target_transform = target_transform if download: self.download() split = self.load_split() self.images_folder = join(self.root, 'Images') self.annotations_folder = join(self.root, 'Annotation') self._breeds = list_dir(self.images_folder) if self.cropped: self._breed_annotations = [[(annotation, box, idx) for box in self.get_boxes(join(self.annotations_folder, annotation))] for annotation, idx in split] self._flat_breed_annotations = sum(self._breed_annotations, []) self._flat_breed_images = [(annotation+'.jpg', idx) for annotation, box, idx in self._flat_breed_annotations] else: self._breed_images = [(annotation+'.jpg', idx) for annotation, idx in split] self._flat_breed_images = self._breed_images self.classes = ["Chihuaha", "Japanese Spaniel", "Maltese Dog", "Pekinese", "Shih-Tzu", "Blenheim Spaniel", "Papillon", "Toy Terrier", "Rhodesian Ridgeback", "Afghan Hound", "Basset Hound", "Beagle", "Bloodhound", "Bluetick", "Black-and-tan Coonhound", "Walker Hound", "English Foxhound", "Redbone", "Borzoi", "Irish Wolfhound", "Italian Greyhound", "Whippet", "Ibizian Hound", "Norwegian Elkhound", "Otterhound", "Saluki", "Scottish Deerhound", "Weimaraner", "Staffordshire Bullterrier", "American Staffordshire Terrier", "Bedlington Terrier", "Border Terrier", "Kerry Blue Terrier", "Irish Terrier", "Norfolk Terrier", "Norwich Terrier", "Yorkshire Terrier", "Wirehaired Fox Terrier", "Lakeland Terrier", "Sealyham Terrier", "Airedale", "Cairn", "Australian Terrier", "Dandi Dinmont", "Boston Bull", "Miniature Schnauzer", "Giant Schnauzer", "Standard Schnauzer", "Scotch Terrier", "Tibetan Terrier", "Silky Terrier", "Soft-coated Wheaten Terrier", "West Highland White Terrier", "Lhasa", "Flat-coated Retriever", "Curly-coater Retriever", "Golden Retriever", "Labrador Retriever", "Chesapeake Bay Retriever", "German Short-haired Pointer", "Vizsla", "English Setter", "Irish Setter", "Gordon Setter", "Brittany", "Clumber", "English Springer Spaniel", "Welsh Springer Spaniel", "Cocker Spaniel", "Sussex Spaniel", "Irish Water Spaniel", "Kuvasz", "Schipperke", "Groenendael", "Malinois", "Briard", "Kelpie", "Komondor", "Old English Sheepdog", "Shetland Sheepdog", "Collie", "Border Collie", "Bouvier des Flandres", "Rottweiler", "German Shepard", "Doberman", "Miniature Pinscher", "Greater Swiss Mountain Dog", "Bernese Mountain Dog", "Appenzeller", "EntleBucher", "Boxer", "Bull Mastiff", "Tibetan Mastiff", "French Bulldog", "Great Dane", "Saint Bernard", "Eskimo Dog", "Malamute", "Siberian Husky", "Affenpinscher", "Basenji", "Pug", "Leonberg", "Newfoundland", "Great Pyrenees", "Samoyed", "Pomeranian", "Chow", "Keeshond", "Brabancon Griffon", "Pembroke", "Cardigan", "Toy Poodle", "Miniature Poodle", "Standard Poodle", "Mexican Hairless", "Dingo", "Dhole", "African Hunting Dog"] data_len = None if self.sm_vit: print(f"[INFO] Preparing {self.mode} shape_hw list...") shape_hw_list = [] self.file_list = [] file_list_full = [] for image_name, target_class in self._flat_breed_images: img_name = join(self.images_folder, image_name) img_temp = scipy.misc.imread(os.path.join(img_name)) shape_hw_temp = [img_temp.shape[0], img_temp.shape[1]] # h_max (y), w_max (x) if (shape_hw_temp[0] > self.max_res) or (shape_hw_temp[1] > self.max_res): if shape_hw_temp[0] > shape_hw_temp[1]: downscale_coef = shape_hw_temp[0] / self.max_res else: downscale_coef = shape_hw_temp[1] / self.max_res shape_hw_temp[0] = int(shape_hw_temp[0] // downscale_coef) shape_hw_temp[1] = int(shape_hw_temp[1] // downscale_coef) shape_hw_list.append(shape_hw_temp) self.file_list.append(image_name) file_list_full.append(img_name) if self.low_memory: self.mask_u2n_list, self.x_u2n_list, self.y_u2n_list, self.h_u2n_list, self.w_u2n_list = \ super(dogs, self).generic_preprocess_lowMem(self.file_list, file_list_full, shape_hw_list ) del shape_hw_list del file_list_full gc.collect() else: self.img, self.mask = \ super(dogs, self).generic_preprocess(self.file_list, file_list_full, shape_hw_list, data_len ) if self.is_train: self.label = [x for i, x in self._flat_breed_images][:data_len] else: self.label = [x for i, x in self._flat_breed_images][:data_len] self.imgname = [x for x in self.file_list[:data_len]] def __getitem__(self, index): """ Args: index (int): Index Returns: tuple: (image, target) where target is index of the target character class. """ if self.sm_vit: if self.low_memory: target, imgname = self.label[index], self.imgname[index] img, mask = super(dogs, self).generic_getitem_lowMem(index) else: img, target, imgname, mask = self.img[index], self.label[index], self.imgname[index], self.mask[index] img, mask = super(dogs, self).generic_getitem(index, img, mask) return img, target, mask else: image_name, target = self._flat_breed_images[index] image_path = join(self.images_folder, image_name) img = Image.open(image_path).convert('RGB') if self.cropped: img = img.crop(self._flat_breed_annotations[index][1]) if self.transform: img = self.transform(img) if self.target_transform: target = self.target_transform(target) return img, target def __len__(self): return len(self._flat_breed_images) def download(self): import tarfile if os.path.exists(join(self.root, 'Images')) and os.path.exists(join(self.root, 'Annotation')): if len(os.listdir(join(self.root, 'Images'))) == len(os.listdir(join(self.root, 'Annotation'))) == 120: print('Files already downloaded and verified') return for filename in ['images', 'annotation', 'lists']: tar_filename = filename + '.tar' url = self.download_url_prefix + '/' + tar_filename download_url(url, self.root, tar_filename, None) print('Extracting downloaded file: ' + join(self.root, tar_filename)) with tarfile.open(join(self.root, tar_filename), 'r') as tar_file: tar_file.extractall(self.root) os.remove(join(self.root, tar_filename)) @staticmethod def get_boxes(path): import xml.etree.ElementTree e = xml.etree.ElementTree.parse(path).getroot() boxes = [] for objs in e.iter('object'): boxes.append([int(objs.find('bndbox').find('xmin').text), int(objs.find('bndbox').find('ymin').text), int(objs.find('bndbox').find('xmax').text), int(objs.find('bndbox').find('ymax').text)]) return boxes def load_split(self): if self.is_train: # split = scipy.io.loadmat(join(self.root, 'train_list.mat'))['annotation_list'] # labels = scipy.io.loadmat(join(self.root, 'train_list.mat'))['labels'] split = scipy.io.loadmat(join(self.root, 'splits/train_list.mat'))['annotation_list'] labels = scipy.io.loadmat(join(self.root, 'splits/train_list.mat'))['labels'] else: # split = scipy.io.loadmat(join(self.root, 'test_list.mat'))['annotation_list'] # labels = scipy.io.loadmat(join(self.root, 'test_list.mat'))['labels'] split = scipy.io.loadmat(join(self.root, 'splits/test_list.mat'))['annotation_list'] labels = scipy.io.loadmat(join(self.root, 'splits/test_list.mat'))['labels'] split = [item[0][0] for item in split] labels = [item[0]-1 for item in labels] return list(zip(split, labels)) def stats(self): counts = {} for index in range(len(self._flat_breed_images)): image_name, target_class = self._flat_breed_images[index] if target_class not in counts.keys(): counts[target_class] = 1 else: counts[target_class] += 1 print("%d samples spanning %d classes (avg %f per class)"%(len(self._flat_breed_images), len(counts.keys()), float(len(self._flat_breed_images))/float(len(counts.keys())))) return counts class NABirds(Generic_smvit_DS): #(Dataset): """`NABirds <https://dl.allaboutbirds.org/nabirds>`_ Dataset. Args: root (string): Root directory of the dataset. train (bool, optional): If True, creates dataset from training set, otherwise creates from test set. transform (callable, optional): A function/transform that takes in an PIL image and returns a transformed version. E.g, ``transforms.RandomCrop`` target_transform (callable, optional): A function/transform that takes in the target and transforms it. download (bool, optional): If true, downloads the dataset from the internet and puts it in root directory. If dataset is already downloaded, it is not downloaded again. """ #base_folder = 'nabirds/images' def __init__(self, ds_name, root, is_train=True, data_len=None, transform=None, sm_vit=True, low_memory=True, img_size=448): self.ds_name = ds_name self.img_size = img_size self.max_res = int(self.img_size * 1.25) # 1.5 self.full_ds = True # pre-processing full dataset self.padding = True # image and mask padding self.rand_crop = False # if no other cropping self.flip_mask_as_image = True # if False - turn on RandomHorizontalFlip in data_utils !!! self.rand_crop_im_mask = False # randomly crop both image and mask self.crop_mask = False # randomly crop mask only self.mid_val = False # 3-state mask self.debug = False # for debug info if self.debug: os.makedirs("./test") self.sm_vit = sm_vit self.low_memory = low_memory dataset_path = os.path.join(root) self.root = root self.base_folder = "images" self.loader = default_loader self.transform = transform self.is_train = is_train if self.is_train: self.mode = "Train" else: self.mode = "Test" image_paths = pd.read_csv(os.path.join(dataset_path, 'images.txt'), sep=' ', names=['img_id', 'filepath']) image_class_labels = pd.read_csv(os.path.join(dataset_path, 'image_class_labels.txt'), sep=' ', names=['img_id', 'target']) # Since the raw labels are non-continuous, map them to new ones self.label_map = get_continuous_class_map(image_class_labels['target']) train_test_split = pd.read_csv(os.path.join(dataset_path, 'train_test_split.txt'), sep=' ', names=['img_id', 'is_training_img']) data = image_paths.merge(image_class_labels, on='img_id') self.data = data.merge(train_test_split, on='img_id') # Load in the train / test split if self.is_train: self.data = self.data[self.data.is_training_img == 1] else: self.data = self.data[self.data.is_training_img == 0] # Load in the class data self.class_names = load_class_names(dataset_path) self.class_hierarchy = load_hierarchy(dataset_path) self.data_len = None if self.sm_vit: print(f"[INFO] Preparing {self.mode} shape_hw list...") shape_hw_list = [] self.file_list = [] file_list_full = [] for sample in self.data.iloc: image_name = sample.filepath img_name_full = join(self.root, self.base_folder, image_name) img_temp = scipy.misc.imread(os.path.join(img_name_full)) shape_hw_temp = [img_temp.shape[0], img_temp.shape[1]] # h_max (y), w_max (x) if (shape_hw_temp[0] > self.max_res) or (shape_hw_temp[1] > self.max_res): if shape_hw_temp[0] > shape_hw_temp[1]: downscale_coef = shape_hw_temp[0] / self.max_res else: downscale_coef = shape_hw_temp[1] / self.max_res shape_hw_temp[0] = int(shape_hw_temp[0] // downscale_coef) shape_hw_temp[1] = int(shape_hw_temp[1] // downscale_coef) shape_hw_list.append(shape_hw_temp) self.file_list.append(image_name) file_list_full.append(img_name_full) if self.low_memory: self.mask_u2n_list, self.x_u2n_list, self.y_u2n_list, self.h_u2n_list, self.w_u2n_list = \ super(NABirds, self).generic_preprocess_lowMem(self.file_list, file_list_full, shape_hw_list ) del shape_hw_list del file_list_full gc.collect() else: self.img, self.mask = \ super(NABirds, self).generic_preprocess(self.file_list, file_list_full, shape_hw_list, self.data_len ) if self.is_train: self.label = [ (self.label_map[x.target]) for x in self.data.iloc ][:self.data_len] else: self.label = [ (self.label_map[x.target]) for x in self.data.iloc ][:self.data_len] self.imgname = [x for x in self.file_list[:self.data_len]] def __getitem__(self, index): if self.sm_vit: if self.low_memory: target, imgname = self.label[index], self.imgname[index] img, mask = super(NABirds, self).generic_getitem_lowMem(index) else: img, target, imgname, mask = self.img[index], self.label[index], self.imgname[index], self.mask[index] img, mask = super(NABirds, self).generic_getitem(index, img, mask) return img, target, mask else: sample = self.data.iloc[index] path = os.path.join(self.root, self.base_folder, sample.filepath) target = self.label_map[sample.target] img = self.loader(path) if self.transform is not None: img = self.transform(img) return img, target def __len__(self): return len(self.data) def get_continuous_class_map(class_labels): label_set = set(class_labels) return {k: i for i, k in enumerate(label_set)} def load_class_names(dataset_path=''): names = {} with open(os.path.join(dataset_path, 'classes.txt')) as f: for line in f: pieces = line.strip().split() class_id = pieces[0] names[class_id] = ' '.join(pieces[1:]) return names def load_hierarchy(dataset_path=''): parents = {} with open(os.path.join(dataset_path, 'hierarchy.txt')) as f: for line in f: pieces = line.strip().split() child_id, parent_id = pieces parents[child_id] = parent_id return parents ### Not optimised datasets: class INat2017(VisionDataset): """`iNaturalist 2017 <https://github.com/visipedia/inat_comp/blob/master/2017/README.md>`_ Dataset. Args: root (string): Root directory of the dataset. split (string, optional): The dataset split, supports ``train``, or ``val``. transform (callable, optional): A function/transform that takes in an PIL image and returns a transformed version. E.g, ``transforms.RandomCrop`` target_transform (callable, optional): A function/transform that takes in the target and transforms it. download (bool, optional): If true, downloads the dataset from the internet and puts it in root directory. If dataset is already downloaded, it is not downloaded again. """ base_folder = 'train_val_images/' file_list = { 'imgs': ('https://storage.googleapis.com/asia_inat_data/train_val/train_val_images.tar.gz', 'train_val_images.tar.gz', '7c784ea5e424efaec655bd392f87301f'), 'annos': ('https://storage.googleapis.com/asia_inat_data/train_val/train_val2017.zip', 'train_val2017.zip', '444c835f6459867ad69fcb36478786e7') } def __init__(self, root, split='train', transform=None, target_transform=None, download=False): super(INat2017, self).__init__(root, transform=transform, target_transform=target_transform) self.loader = default_loader self.split = verify_str_arg(split, "split", ("train", "val",)) if self._check_exists(): print('Files already downloaded and verified.') elif download: if not (os.path.exists(os.path.join(self.root, self.file_list['imgs'][1])) and os.path.exists(os.path.join(self.root, self.file_list['annos'][1]))): print('Downloading...') self._download() print('Extracting...') extract_archive(os.path.join(self.root, self.file_list['imgs'][1])) extract_archive(os.path.join(self.root, self.file_list['annos'][1])) else: raise RuntimeError( 'Dataset not found. You can use download=True to download it.') anno_filename = split + '2017.json' with open(os.path.join(self.root, anno_filename), 'r') as fp: all_annos = json.load(fp) self.annos = all_annos['annotations'] self.images = all_annos['images'] def __getitem__(self, index): path = os.path.join(self.root, self.images[index]['file_name']) target = self.annos[index]['category_id'] image = self.loader(path) if self.transform is not None: image = self.transform(image) if self.target_transform is not None: target = self.target_transform(target) return image, target def __len__(self): return len(self.images) def _check_exists(self): return os.path.exists(os.path.join(self.root, self.base_folder)) def _download(self): for url, filename, md5 in self.file_list.values(): download_url(url, root=self.root, filename=filename) if not check_integrity(os.path.join(self.root, filename), md5): raise RuntimeError("File not found or corrupted.") class CarsDataset(Dataset): def __init__(self, mat_anno, data_dir, car_names, cleaned=None, transform=None): """ Args: mat_anno (string): Path to the MATLAB annotation file. data_dir (string): Directory with all the images. transform (callable, optional): Optional transform to be applied on a sample. """ self.full_data_set = io.loadmat(mat_anno) self.car_annotations = self.full_data_set['annotations'] self.car_annotations = self.car_annotations[0] if cleaned is not None: cleaned_annos = [] print("Cleaning up data set (only take pics with rgb chans)...") clean_files = np.loadtxt(cleaned, dtype=str) for c in self.car_annotations: if c[-1][0] in clean_files: cleaned_annos.append(c) self.car_annotations = cleaned_annos self.car_names = scipy.io.loadmat(car_names)['class_names'] self.car_names = np.array(self.car_names[0]) self.data_dir = data_dir self.transform = transform def __len__(self): return len(self.car_annotations) def __getitem__(self, idx): img_name = os.path.join(self.data_dir, self.car_annotations[idx][-1][0]) image = Image.open(img_name).convert('RGB') car_class = self.car_annotations[idx][-2][0][0] car_class = torch.from_numpy(np.array(car_class.astype(np.float32))).long() - 1 assert car_class < 196 if self.transform: image = self.transform(image) # return image, car_class, img_name return image, car_class def map_class(self, id): id = np.ravel(id) ret = self.car_names[id - 1][0][0] return ret def show_batch(self, img_batch, class_batch): for i in range(img_batch.shape[0]): ax = plt.subplot(1, img_batch.shape[0], i + 1) title_str = self.map_class(int(class_batch[i])) img = np.transpose(img_batch[i, ...], (1, 2, 0)) ax.imshow(img) ax.set_title(title_str.__str__(), {'fontsize': 5}) plt.tight_layout() def make_dataset(dir, image_ids, targets): assert(len(image_ids) == len(targets)) images = [] dir = os.path.expanduser(dir) for i in range(len(image_ids)): item = (os.path.join(dir, 'data', 'images', '%s.jpg' % image_ids[i]), targets[i]) images.append(item) return images def find_classes(classes_file): # read classes file, separating out image IDs and class names image_ids = [] targets = [] f = open(classes_file, 'r') for line in f: split_line = line.split(' ') image_ids.append(split_line[0]) targets.append(' '.join(split_line[1:])) f.close() # index class names classes = np.unique(targets) class_to_idx = {classes[i]: i for i in range(len(classes))} targets = [class_to_idx[c] for c in targets] return (image_ids, targets, classes, class_to_idx) class FGVC_aircraft(): def __init__(self, root, is_train=True, data_len=None, transform=None): self.root = root self.is_train = is_train self.transform = transform train_img_path = os.path.join(self.root, 'data', 'images') test_img_path = os.path.join(self.root, 'data', 'images') train_label_file = open(os.path.join(self.root, 'data', 'train.txt')) test_label_file = open(os.path.join(self.root, 'data', 'test.txt')) train_img_label = [] test_img_label = [] for line in train_label_file: train_img_label.append([os.path.join(train_img_path,line[:-1].split(' ')[0]), int(line[:-1].split(' ')[1])-1]) for line in test_label_file: test_img_label.append([os.path.join(test_img_path,line[:-1].split(' ')[0]), int(line[:-1].split(' ')[1])-1]) self.train_img_label = train_img_label[:data_len] self.test_img_label = test_img_label[:data_len] def __getitem__(self, index): if self.is_train: img, target = scipy.misc.imread(self.train_img_label[index][0]), self.train_img_label[index][1] if len(img.shape) == 2: img = np.stack([img] * 3, 2) img = Image.fromarray(img, mode='RGB') if self.transform is not None: img = self.transform(img) else: img, target = scipy.misc.imread(self.test_img_label[index][0]), self.test_img_label[index][1] if len(img.shape) == 2: img = np.stack([img] * 3, 2) img = Image.fromarray(img, mode='RGB') if self.transform is not None: img = self.transform(img) return img, target def __len__(self): if self.is_train: return len(self.train_img_label) else: return len(self.test_img_label)
67,209
39.659407
180
py
sm-vit
sm-vit-main/utils/scheduler.py
import logging import math from torch.optim.lr_scheduler import LambdaLR logger = logging.getLogger(__name__) class ConstantLRSchedule(LambdaLR): """ Constant learning rate schedule. """ def __init__(self, optimizer, last_epoch=-1): super(ConstantLRSchedule, self).__init__(optimizer, lambda _: 1.0, last_epoch=last_epoch) class WarmupConstantSchedule(LambdaLR): """ Linear warmup and then constant. Linearly increases learning rate schedule from 0 to 1 over `warmup_steps` training steps. Keeps learning rate schedule equal to 1. after warmup_steps. """ def __init__(self, optimizer, warmup_steps, last_epoch=-1): self.warmup_steps = warmup_steps super(WarmupConstantSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch) def lr_lambda(self, step): if step < self.warmup_steps: return float(step) / float(max(1.0, self.warmup_steps)) return 1. class WarmupLinearSchedule(LambdaLR): """ Linear warmup and then linear decay. Linearly increases learning rate from 0 to 1 over `warmup_steps` training steps. Linearly decreases learning rate from 1. to 0. over remaining `t_total - warmup_steps` steps. """ def __init__(self, optimizer, warmup_steps, t_total, last_epoch=-1): self.warmup_steps = warmup_steps self.t_total = t_total super(WarmupLinearSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch) def lr_lambda(self, step): if step < self.warmup_steps: return float(step) / float(max(1, self.warmup_steps)) return max(0.0, float(self.t_total - step) / float(max(1.0, self.t_total - self.warmup_steps))) class WarmupCosineSchedule(LambdaLR): """ Linear warmup and then cosine decay. Linearly increases learning rate from 0 to 1 over `warmup_steps` training steps. Decreases learning rate from 1. to 0. over remaining `t_total - warmup_steps` steps following a cosine curve. If `cycles` (default=0.5) is different from default, learning rate follows cosine function after warmup. """ def __init__(self, optimizer, warmup_steps, t_total, cycles=.5, last_epoch=-1): self.warmup_steps = warmup_steps self.t_total = t_total self.cycles = cycles super(WarmupCosineSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch) def lr_lambda(self, step): if step < self.warmup_steps: return float(step) / float(max(1.0, self.warmup_steps)) # progress after warmup progress = float(step - self.warmup_steps) / float(max(1, self.t_total - self.warmup_steps)) return max(0.0, 0.5 * (1. + math.cos(math.pi * float(self.cycles) * 2.0 * progress)))
2,799
42.75
117
py
sm-vit
sm-vit-main/utils/dist_util.py
import torch.distributed as dist def get_rank(): if not dist.is_available(): return 0 if not dist.is_initialized(): return 0 return dist.get_rank() def get_world_size(): if not dist.is_available(): return 1 if not dist.is_initialized(): return 1 return dist.get_world_size() def is_main_process(): return get_rank() == 0 def format_step(step): if isinstance(step, str): return step s = "" if len(step) > 0: s += "Training Epoch: {} ".format(step[0]) if len(step) > 1: s += "Training Iteration: {} ".format(step[1]) if len(step) > 2: s += "Validation Iteration: {} ".format(step[2]) return s
711
21.967742
56
py
sm-vit
sm-vit-main/U2Net/data_loader.py
# data loader from __future__ import print_function, division import glob import torch from skimage import io, transform, color import numpy as np import random import math import matplotlib.pyplot as plt from torch.utils.data import Dataset, DataLoader from torchvision import transforms, utils from PIL import Image #==========================dataset load========================== class RescaleT(object): def __init__(self,output_size): assert isinstance(output_size,(int,tuple)) self.output_size = output_size def __call__(self,sample): imidx, image, label = sample['imidx'], sample['image'],sample['label'] h, w = image.shape[:2] if isinstance(self.output_size,int): if h > w: new_h, new_w = self.output_size*h/w,self.output_size else: new_h, new_w = self.output_size,self.output_size*w/h else: new_h, new_w = self.output_size new_h, new_w = int(new_h), int(new_w) # # #resize the image to new_h x new_w and convert image from range [0,255] to [0,1] # # img = transform.resize(image,(new_h,new_w),mode='constant') # # lbl = transform.resize(label,(new_h,new_w),mode='constant', order=0, preserve_range=True) # img = transform.resize(image,(self.output_size,self.output_size),mode='constant') # lbl = transform.resize(label,(self.output_size,self.output_size),mode='constant', order=0, preserve_range=True) img = transform.resize(image,(self.output_size,self.output_size),mode='constant', anti_aliasing=True, anti_aliasing_sigma=None) lbl = transform.resize(label,(self.output_size,self.output_size),mode='constant', order=0, preserve_range=True, anti_aliasing=True, anti_aliasing_sigma=None) return {'imidx':imidx, 'image':img,'label':lbl} class Rescale(object): def __init__(self,output_size): assert isinstance(output_size,(int,tuple)) self.output_size = output_size def __call__(self,sample): imidx, image, label = sample['imidx'], sample['image'],sample['label'] if random.random() >= 0.5: image = image[::-1] label = label[::-1] h, w = image.shape[:2] if isinstance(self.output_size,int): if h > w: new_h, new_w = self.output_size*h/w,self.output_size else: new_h, new_w = self.output_size,self.output_size*w/h else: new_h, new_w = self.output_size new_h, new_w = int(new_h), int(new_w) # #resize the image to new_h x new_w and convert image from range [0,255] to [0,1] # img = transform.resize(image,(new_h,new_w),mode='constant') # lbl = transform.resize(label,(new_h,new_w),mode='constant', order=0, preserve_range=True) img = transform.resize(image,(new_h,new_w),mode='constant', anti_aliasing=True, anti_aliasing_sigma=None) lbl = transform.resize(label,(new_h,new_w),mode='constant', order=0, preserve_range=True, anti_aliasing=True, anti_aliasing_sigma=None) return {'imidx':imidx, 'image':img,'label':lbl} class RandomCrop(object): def __init__(self,output_size): assert isinstance(output_size, (int, tuple)) if isinstance(output_size, int): self.output_size = (output_size, output_size) else: assert len(output_size) == 2 self.output_size = output_size def __call__(self,sample): imidx, image, label = sample['imidx'], sample['image'], sample['label'] if random.random() >= 0.5: image = image[::-1] label = label[::-1] h, w = image.shape[:2] new_h, new_w = self.output_size top = np.random.randint(0, h - new_h) left = np.random.randint(0, w - new_w) image = image[top: top + new_h, left: left + new_w] label = label[top: top + new_h, left: left + new_w] return {'imidx':imidx,'image':image, 'label':label} class ToTensor(object): """Convert ndarrays in sample to Tensors.""" def __call__(self, sample): imidx, image, label = sample['imidx'], sample['image'], sample['label'] tmpImg = np.zeros((image.shape[0],image.shape[1],3)) tmpLbl = np.zeros(label.shape) image = image/np.max(image) if(np.max(label)<1e-6): label = label else: label = label/np.max(label) if image.shape[2]==1: tmpImg[:,:,0] = (image[:,:,0]-0.485)/0.229 tmpImg[:,:,1] = (image[:,:,0]-0.485)/0.229 tmpImg[:,:,2] = (image[:,:,0]-0.485)/0.229 else: tmpImg[:,:,0] = (image[:,:,0]-0.485)/0.229 tmpImg[:,:,1] = (image[:,:,1]-0.456)/0.224 tmpImg[:,:,2] = (image[:,:,2]-0.406)/0.225 tmpLbl[:,:,0] = label[:,:,0] tmpImg = tmpImg.transpose((2, 0, 1)) tmpLbl = label.transpose((2, 0, 1)) return {'imidx':torch.from_numpy(imidx), 'image': torch.from_numpy(tmpImg), 'label': torch.from_numpy(tmpLbl)} class ToTensorLab(object): """Convert ndarrays in sample to Tensors.""" def __init__(self,flag=0): self.flag = flag def __call__(self, sample): imidx, image, label =sample['imidx'], sample['image'], sample['label'] tmpLbl = np.zeros(label.shape) if(np.max(label)<1e-6): label = label else: label = label/np.max(label) # change the color space if self.flag == 2: # with rgb and Lab colors tmpImg = np.zeros((image.shape[0],image.shape[1],6)) tmpImgt = np.zeros((image.shape[0],image.shape[1],3)) if image.shape[2]==1: tmpImgt[:,:,0] = image[:,:,0] tmpImgt[:,:,1] = image[:,:,0] tmpImgt[:,:,2] = image[:,:,0] else: tmpImgt = image tmpImgtl = color.rgb2lab(tmpImgt) # nomalize image to range [0,1] tmpImg[:,:,0] = (tmpImgt[:,:,0]-np.min(tmpImgt[:,:,0]))/(np.max(tmpImgt[:,:,0])-np.min(tmpImgt[:,:,0])) tmpImg[:,:,1] = (tmpImgt[:,:,1]-np.min(tmpImgt[:,:,1]))/(np.max(tmpImgt[:,:,1])-np.min(tmpImgt[:,:,1])) tmpImg[:,:,2] = (tmpImgt[:,:,2]-np.min(tmpImgt[:,:,2]))/(np.max(tmpImgt[:,:,2])-np.min(tmpImgt[:,:,2])) tmpImg[:,:,3] = (tmpImgtl[:,:,0]-np.min(tmpImgtl[:,:,0]))/(np.max(tmpImgtl[:,:,0])-np.min(tmpImgtl[:,:,0])) tmpImg[:,:,4] = (tmpImgtl[:,:,1]-np.min(tmpImgtl[:,:,1]))/(np.max(tmpImgtl[:,:,1])-np.min(tmpImgtl[:,:,1])) tmpImg[:,:,5] = (tmpImgtl[:,:,2]-np.min(tmpImgtl[:,:,2]))/(np.max(tmpImgtl[:,:,2])-np.min(tmpImgtl[:,:,2])) # tmpImg = tmpImg/(np.max(tmpImg)-np.min(tmpImg)) tmpImg[:,:,0] = (tmpImg[:,:,0]-np.mean(tmpImg[:,:,0]))/np.std(tmpImg[:,:,0]) tmpImg[:,:,1] = (tmpImg[:,:,1]-np.mean(tmpImg[:,:,1]))/np.std(tmpImg[:,:,1]) tmpImg[:,:,2] = (tmpImg[:,:,2]-np.mean(tmpImg[:,:,2]))/np.std(tmpImg[:,:,2]) tmpImg[:,:,3] = (tmpImg[:,:,3]-np.mean(tmpImg[:,:,3]))/np.std(tmpImg[:,:,3]) tmpImg[:,:,4] = (tmpImg[:,:,4]-np.mean(tmpImg[:,:,4]))/np.std(tmpImg[:,:,4]) tmpImg[:,:,5] = (tmpImg[:,:,5]-np.mean(tmpImg[:,:,5]))/np.std(tmpImg[:,:,5]) elif self.flag == 1: #with Lab color tmpImg = np.zeros((image.shape[0],image.shape[1],3)) if image.shape[2]==1: tmpImg[:,:,0] = image[:,:,0] tmpImg[:,:,1] = image[:,:,0] tmpImg[:,:,2] = image[:,:,0] else: tmpImg = image tmpImg = color.rgb2lab(tmpImg) # tmpImg = tmpImg/(np.max(tmpImg)-np.min(tmpImg)) tmpImg[:,:,0] = (tmpImg[:,:,0]-np.min(tmpImg[:,:,0]))/(np.max(tmpImg[:,:,0])-np.min(tmpImg[:,:,0])) tmpImg[:,:,1] = (tmpImg[:,:,1]-np.min(tmpImg[:,:,1]))/(np.max(tmpImg[:,:,1])-np.min(tmpImg[:,:,1])) tmpImg[:,:,2] = (tmpImg[:,:,2]-np.min(tmpImg[:,:,2]))/(np.max(tmpImg[:,:,2])-np.min(tmpImg[:,:,2])) tmpImg[:,:,0] = (tmpImg[:,:,0]-np.mean(tmpImg[:,:,0]))/np.std(tmpImg[:,:,0]) tmpImg[:,:,1] = (tmpImg[:,:,1]-np.mean(tmpImg[:,:,1]))/np.std(tmpImg[:,:,1]) tmpImg[:,:,2] = (tmpImg[:,:,2]-np.mean(tmpImg[:,:,2]))/np.std(tmpImg[:,:,2]) else: # with rgb color tmpImg = np.zeros((image.shape[0],image.shape[1],3)) image = image/np.max(image) if image.shape[2]==1: tmpImg[:,:,0] = (image[:,:,0]-0.485)/0.229 tmpImg[:,:,1] = (image[:,:,0]-0.485)/0.229 tmpImg[:,:,2] = (image[:,:,0]-0.485)/0.229 else: tmpImg[:,:,0] = (image[:,:,0]-0.485)/0.229 tmpImg[:,:,1] = (image[:,:,1]-0.456)/0.224 tmpImg[:,:,2] = (image[:,:,2]-0.406)/0.225 tmpLbl[:,:,0] = label[:,:,0] tmpImg = tmpImg.transpose((2, 0, 1)) tmpLbl = label.transpose((2, 0, 1)) return {'imidx':torch.from_numpy(imidx), 'image': torch.from_numpy(tmpImg), 'label': torch.from_numpy(tmpLbl)} class SalObjDataset(Dataset): def __init__(self,img_name_list,lbl_name_list,transform=None): # self.root_dir = root_dir # self.image_name_list = glob.glob(image_dir+'*.png') # self.label_name_list = glob.glob(label_dir+'*.png') self.image_name_list = img_name_list self.label_name_list = lbl_name_list self.transform = transform def __len__(self): return len(self.image_name_list) def __getitem__(self,idx): # image = Image.open(self.image_name_list[idx])#io.imread(self.image_name_list[idx]) # label = Image.open(self.label_name_list[idx])#io.imread(self.label_name_list[idx]) image = io.imread(self.image_name_list[idx]) imname = self.image_name_list[idx] imidx = np.array([idx]) if(0==len(self.label_name_list)): label_3 = np.zeros(image.shape) else: label_3 = io.imread(self.label_name_list[idx]) label = np.zeros(label_3.shape[0:2]) if(3==len(label_3.shape)): label = label_3[:,:,0] elif(2==len(label_3.shape)): label = label_3 if(3==len(image.shape) and 2==len(label.shape)): label = label[:,:,np.newaxis] elif(2==len(image.shape) and 2==len(label.shape)): image = image[:,:,np.newaxis] label = label[:,:,np.newaxis] sample = {'imidx':imidx, 'image':image, 'label':label} if self.transform: sample = self.transform(sample) return sample
9,327
32.797101
159
py
sm-vit
sm-vit-main/U2Net/u2net_test.py
import os from re import X from skimage import io, transform import torch import torchvision from torch.autograd import Variable import torch.nn as nn import torch.nn.functional as F from torch.utils.data import Dataset, DataLoader from torchvision import transforms#, utils # import torch.optim as optim import numpy as np from PIL import Image import glob from .data_loader import RescaleT from .data_loader import ToTensor from .data_loader import ToTensorLab from .data_loader import SalObjDataset from .model import U2NET # full size version 173.6 MB from .model import U2NETP # small version u2net 4.7 MB #import cv2 import copy from torchvision.utils import save_image # normalize the predicted SOD probability map def normPRED(d): ma = torch.max(d) mi = torch.min(d) dn = (d-mi)/(ma-mi) return dn def save_output(image_name,pred,d_dir): predict = pred predict = predict.squeeze() predict_np = predict.cpu().data.numpy() im = Image.fromarray(predict_np*255).convert('RGB') img_name = image_name.split(os.sep)[-1] image = io.imread(image_name) imo = im.resize((image.shape[1],image.shape[0]),resample=Image.BILINEAR) pb_np = np.array(imo) aaa = img_name.split(".") bbb = aaa[0:-1] imidx = bbb[0] for i in range(1,len(bbb)): imidx = imidx + "." + bbb[i] imo.save(d_dir+imidx+'.png') def mask_hw(full_ds=True, img_path=None, shape_hw=None): print_info = False # --------- 1. get image path and name --------- model_name='u2net' #u2netp if img_path is None: image_dir = os.path.join(os.getcwd(), 'U2Net/images') img_name_list = glob.glob(image_dir + os.sep + '*') if print_info: print("local image") if print_info: print(img_name_list) else: if full_ds: img_name_list = img_path shape_hw_list = shape_hw else: img_name_list = glob.glob(img_path) if print_info: print(img_path) model_dir = os.path.join(os.getcwd(), 'U2Net/model/pre_trained', model_name + '.pth') # --------- 2. dataloader --------- test_salobj_dataset = SalObjDataset(img_name_list = img_name_list, lbl_name_list = [], transform=transforms.Compose([RescaleT(320), ToTensorLab(flag=0)]) ) test_salobj_dataloader = DataLoader(test_salobj_dataset, batch_size=1, shuffle=False, num_workers=4) # 1 # --------- 3. model define --------- if(model_name=='u2net'): net = U2NET(3,1) elif(model_name=='u2netp'): net = U2NETP(3,1) if torch.cuda.is_available(): net.load_state_dict(torch.load(model_dir)) net.cuda() else: net.load_state_dict(torch.load(model_dir, map_location='cpu')) net.eval() # --------- 4. inference for each image --------- mask_out_np_list = [] start_x_list = [] start_y_list = [] h_list = [] w_list = [] bad_mask_count = 0 refined_mask_count = 0 for i_test, data_test in enumerate(test_salobj_dataloader): if print_info: print("U2N:", i_test, img_name_list[i_test]) inputs_test = data_test['image'] inputs_test = inputs_test.type(torch.FloatTensor) if full_ds: shape_hw_i = shape_hw_list[i_test] if torch.cuda.is_available(): inputs_test = Variable(inputs_test.cuda()) else: inputs_test = Variable(inputs_test) with torch.no_grad(): d1,d2,d3,d4,d5,d6,d7= net(inputs_test) # normalization pred = d1[:,0,:,:] pred = normPRED(pred) THRESHOLD = 0.8 # 0.5 # 0.8 # 0.5 #0.8 # for the original mask (better not smaller than 0.7 cuz artifacts) THRESHOLD_resize = 0.2 # 0.1 # 0.2 # for the resized mask THRESHOLD_deep = 0.1 # for the non-detected mask pred = pred[0, :, :] pred_cpu = pred.cpu() out_img = pred_cpu.detach().numpy() out_img_refine = copy.deepcopy(out_img) # BACKGROUND REMOVAL out_img[out_img > THRESHOLD] = 1 out_img[out_img <= THRESHOLD] = 0 out_img = (out_img * 255).astype(np.uint8) out_img = Image.fromarray(out_img, mode='L') # BOUNDING BOX CREATION transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(mode='L'), #(mode='1'), #transforms.Resize((image.shape[1],image.shape[0]), Image.BILINEAR), transforms.Resize((shape_hw_i[0], shape_hw_i[1]), Image.BILINEAR), # shape_hw (0 - height, 1 - width) transforms.ToTensor(), ]) out_img = transform_mask(out_img) out_img = out_img[0, :, :] mask_out = out_img mask_out = mask_out.cpu() mask_out = torch.where( mask_out > THRESHOLD_resize, torch.tensor(0.), torch.tensor(1.)) mask_out_np = mask_out.detach().numpy() out_layer = out_img out_layer = out_layer.cpu() out_layer = torch.where( out_layer > THRESHOLD_resize, torch.tensor(1.), torch.tensor(0.)) out_layer = out_layer.detach().numpy() x_starts = [np.where(out_layer[i]==1)[0][0] if len(np.where(out_layer[i]==1)[0])!=0 \ else out_layer.shape[0]+1 for i in range(out_layer.shape[0])] x_ends = [np.where(out_layer[i]==1)[0][-1] if len(np.where(out_layer[i]==1)[0])!=0 \ else 0 for i in range(out_layer.shape[0])] y_starts = [np.where(out_layer.T[i]==1)[0][0] if len(np.where(out_layer.T[i]==1)[0])!=0 \ else out_layer.T.shape[0]+1 for i in range(out_layer.T.shape[0])] y_ends = [np.where(out_layer.T[i]==1)[0][-1] if len(np.where(out_layer.T[i]==1)[0])!=0 \ else 0 for i in range(out_layer.T.shape[0])] startx = min(x_starts) endx = max(x_ends) starty = min(y_starts) endy = max(y_ends) start = (startx,starty) end = (endx,endy) ## For cases when U2N couldn't detect mask: # [DONE] 1.1 if (end - start) < 30-50px -> decrease the THRESHOLD # [DONE] 1.2 if (start>end) or end==0 ? -> decrease the THRESHOLD # [DONE] 2.1 if no mask found anyway -> create center crop mask (x, y) +-10 % # [DONE] 2.2 + restore h,w from (0,0) to (x, y) +-10 % w_temp = end[0] - start[0] h_temp = end[1] - start[1] mask_px = np.count_nonzero(out_layer > 0.9) # (expected to be == 1.0) if print_info: print("Mask px old:", mask_px) if (end[0] <= start[0]) or (end[1] <= start[1]) or (mask_px < 5000) or (w_temp < 50) or (h_temp < 50) : if print_info: print("[WARNING] Mask was not detected by U2N for image", img_name_list[i_test]) if print_info: print("Trying to refine image and then detect mask again.") if print_info: print("Old x (start, end):", startx, endx) if print_info: print("Old y (start, end):", starty, endy) # img_dir = ("test/" + str(i_test)) # if not os.path.exists(img_dir): # os.makedirs(img_dir, exist_ok=True) # img_name = ("test/" + str(i_test) + "/1mask_init" + str(i_test) + ".png") # img_temp = transforms.ToTensor()(out_img_refine) # save_image(img_temp, img_name) # img_name = ("test/" + str(i_test) + "/2mask_old" + str(i_test) + ".png") # img_temp = transforms.ToTensor()(mask_out_np) # save_image(img_temp, img_name) out_img_refine[out_img_refine > THRESHOLD_deep] = 1 out_img_refine[out_img_refine <= THRESHOLD_deep] = 0 out_img_refine = (out_img_refine * 255).astype(np.uint8) out_img_refine = Image.fromarray(out_img_refine, mode='L') transform_mask = transforms.Compose([ transforms.ToTensor(), transforms.ToPILImage(mode='L'), #(mode='1'), #transforms.Resize((image.shape[1],image.shape[0]), Image.BILINEAR), transforms.Resize((shape_hw_i[0], shape_hw_i[1]), Image.BILINEAR), # shape_hw (0 - height, 1 - width) transforms.ToTensor(), ]) out_img_refine = transform_mask(out_img_refine) out_img_refine = out_img_refine[0, :, :] out_layer_refine = out_img_refine out_layer_refine = out_layer_refine.cpu() out_layer_refine = torch.where( out_img_refine > THRESHOLD_resize, torch.tensor(1.), torch.tensor(0.)) out_layer_refine = out_layer_refine.detach().numpy() x_starts = [np.where(out_layer_refine[i]==1)[0][0] if len(np.where(out_layer_refine[i]==1)[0])!=0 \ else out_layer_refine.shape[0]+1 for i in range(out_layer_refine.shape[0])] x_ends = [np.where(out_layer_refine[i]==1)[0][-1] if len(np.where(out_layer_refine[i]==1)[0])!=0 \ else 0 for i in range(out_layer_refine.shape[0])] y_starts = [np.where(out_layer_refine.T[i]==1)[0][0] if len(np.where(out_layer_refine.T[i]==1)[0])!=0 \ else out_layer_refine.T.shape[0]+1 for i in range(out_layer_refine.T.shape[0])] y_ends = [np.where(out_layer_refine.T[i]==1)[0][-1] if len(np.where(out_layer_refine.T[i]==1)[0])!=0 \ else 0 for i in range(out_layer_refine.T.shape[0])] startx = min(x_starts) endx = max(x_ends) starty = min(y_starts) endy = max(y_ends) start = (startx,starty) end = (endx,endy) if print_info: print("New x (start, end):", startx, endx) if print_info: print("New y (start, end):", starty, endy) w_temp = end[0] - start[0] h_temp = end[1] - start[1] mask_px = np.count_nonzero(out_layer_refine > 0.9) # (expected to be == 1.0) if print_info: print("Mask px new:", mask_px) if (end[0] <= start[0]) or (end[1] <= start[1]) or (mask_px < 5000) or (w_temp < 50) or (h_temp < 50) : if print_info: print("[WARNING] Mask was not deteted by U2N even after refining.") if print_info: print("Changing mask size (0, 0) to img size (", shape_hw_i[1], shape_hw_i[0], ") -10 p/c boundaries: ") if print_info: print("Old x (start, end):", startx, endx) startx = shape_hw_i[1] * 0.1 endx = shape_hw_i[1] * 0.9 # w -> x if print_info: print("New x (start, end):", startx, endx) if print_info: print("Old y (start, end):", starty, endy) starty = shape_hw_i[0] * 0.1 endy = shape_hw_i[0] * 0.9 # h -> y if print_info: print("New y (start, end):", starty, endy) start = (startx,starty) end = (endx,endy) mask_out_np = np.ones((int(shape_hw_i[0]), int(shape_hw_i[1]))) mask_out_np[int(starty):int(endy), int(startx):int(endx)] = 0 # img_name = ("test/" + str(i_test) + "/4mask_new2_" + str(i_test) + ".png") # img_temp = transforms.ToTensor()(mask_out_np) # save_image(img_temp, img_name) bad_mask_count+=1 else: mask_out_refine = out_img_refine mask_out_refine = mask_out_refine.cpu() mask_out_refine = torch.where( mask_out_refine > THRESHOLD_resize, torch.tensor(0.), torch.tensor(1.)) # img_name = ("test/" + str(i_test) + "/3mask_new1_" + str(i_test) + ".png") # #mask_tem = transforms.ToTensor()(mask) # save_image(mask_out_refine, img_name) mask_out_np = mask_out_refine.detach().numpy() refined_mask_count+=1 w = end[0] - start[0] h = end[1] - start[1] # save results to test_results folder # if not os.path.exists(prediction_dir): # os.makedirs(prediction_dir, exist_ok=True) # save_output(img_name_list[i_test], mask_out, prediction_dir) del d1,d2,d3,d4,d5,d6,d7 if print_info: print(start[0], start[1], h, w) mask_out_np_list.append(mask_out_np) start_x_list.append(start[0]) start_y_list.append(start[1]) h_list.append(h) w_list.append(w) if i_test % 1000 == 0: print(i_test) print("Refined masks total:", refined_mask_count) print("Bad masks total:", bad_mask_count) return mask_out_np_list, start_x_list, start_y_list, h_list, w_list if __name__ == "__main__": mask_hw()
13,512
37.719198
139
py
sm-vit
sm-vit-main/U2Net/model/u2net.py
import torch import torch.nn as nn import torch.nn.functional as F class REBNCONV(nn.Module): def __init__(self,in_ch=3,out_ch=3,dirate=1): super(REBNCONV,self).__init__() self.conv_s1 = nn.Conv2d(in_ch,out_ch,3,padding=1*dirate,dilation=1*dirate) self.bn_s1 = nn.BatchNorm2d(out_ch) self.relu_s1 = nn.ReLU(inplace=True) def forward(self,x): hx = x xout = self.relu_s1(self.bn_s1(self.conv_s1(hx))) return xout ## upsample tensor 'src' to have the same spatial size with tensor 'tar' def _upsample_like(src,tar): src = F.upsample(src,size=tar.shape[2:],mode='bilinear') return src ### RSU-7 ### class RSU7(nn.Module):#UNet07DRES(nn.Module): def __init__(self, in_ch=3, mid_ch=12, out_ch=3): super(RSU7,self).__init__() self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool3 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool4 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv5 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool5 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv6 = REBNCONV(mid_ch,mid_ch,dirate=1) self.rebnconv7 = REBNCONV(mid_ch,mid_ch,dirate=2) self.rebnconv6d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv5d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv4d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) def forward(self,x): hx = x hxin = self.rebnconvin(hx) hx1 = self.rebnconv1(hxin) hx = self.pool1(hx1) hx2 = self.rebnconv2(hx) hx = self.pool2(hx2) hx3 = self.rebnconv3(hx) hx = self.pool3(hx3) hx4 = self.rebnconv4(hx) hx = self.pool4(hx4) hx5 = self.rebnconv5(hx) hx = self.pool5(hx5) hx6 = self.rebnconv6(hx) hx7 = self.rebnconv7(hx6) hx6d = self.rebnconv6d(torch.cat((hx7,hx6),1)) hx6dup = _upsample_like(hx6d,hx5) hx5d = self.rebnconv5d(torch.cat((hx6dup,hx5),1)) hx5dup = _upsample_like(hx5d,hx4) hx4d = self.rebnconv4d(torch.cat((hx5dup,hx4),1)) hx4dup = _upsample_like(hx4d,hx3) hx3d = self.rebnconv3d(torch.cat((hx4dup,hx3),1)) hx3dup = _upsample_like(hx3d,hx2) hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1)) hx2dup = _upsample_like(hx2d,hx1) hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1)) return hx1d + hxin ### RSU-6 ### class RSU6(nn.Module):#UNet06DRES(nn.Module): def __init__(self, in_ch=3, mid_ch=12, out_ch=3): super(RSU6,self).__init__() self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool3 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool4 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv5 = REBNCONV(mid_ch,mid_ch,dirate=1) self.rebnconv6 = REBNCONV(mid_ch,mid_ch,dirate=2) self.rebnconv5d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv4d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) def forward(self,x): hx = x hxin = self.rebnconvin(hx) hx1 = self.rebnconv1(hxin) hx = self.pool1(hx1) hx2 = self.rebnconv2(hx) hx = self.pool2(hx2) hx3 = self.rebnconv3(hx) hx = self.pool3(hx3) hx4 = self.rebnconv4(hx) hx = self.pool4(hx4) hx5 = self.rebnconv5(hx) hx6 = self.rebnconv6(hx5) hx5d = self.rebnconv5d(torch.cat((hx6,hx5),1)) hx5dup = _upsample_like(hx5d,hx4) hx4d = self.rebnconv4d(torch.cat((hx5dup,hx4),1)) hx4dup = _upsample_like(hx4d,hx3) hx3d = self.rebnconv3d(torch.cat((hx4dup,hx3),1)) hx3dup = _upsample_like(hx3d,hx2) hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1)) hx2dup = _upsample_like(hx2d,hx1) hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1)) return hx1d + hxin ### RSU-5 ### class RSU5(nn.Module):#UNet05DRES(nn.Module): def __init__(self, in_ch=3, mid_ch=12, out_ch=3): super(RSU5,self).__init__() self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool3 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=1) self.rebnconv5 = REBNCONV(mid_ch,mid_ch,dirate=2) self.rebnconv4d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) def forward(self,x): hx = x hxin = self.rebnconvin(hx) hx1 = self.rebnconv1(hxin) hx = self.pool1(hx1) hx2 = self.rebnconv2(hx) hx = self.pool2(hx2) hx3 = self.rebnconv3(hx) hx = self.pool3(hx3) hx4 = self.rebnconv4(hx) hx5 = self.rebnconv5(hx4) hx4d = self.rebnconv4d(torch.cat((hx5,hx4),1)) hx4dup = _upsample_like(hx4d,hx3) hx3d = self.rebnconv3d(torch.cat((hx4dup,hx3),1)) hx3dup = _upsample_like(hx3d,hx2) hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1)) hx2dup = _upsample_like(hx2d,hx1) hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1)) return hx1d + hxin ### RSU-4 ### class RSU4(nn.Module):#UNet04DRES(nn.Module): def __init__(self, in_ch=3, mid_ch=12, out_ch=3): super(RSU4,self).__init__() self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1) self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1) self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=2) self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1) self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) def forward(self,x): hx = x hxin = self.rebnconvin(hx) hx1 = self.rebnconv1(hxin) hx = self.pool1(hx1) hx2 = self.rebnconv2(hx) hx = self.pool2(hx2) hx3 = self.rebnconv3(hx) hx4 = self.rebnconv4(hx3) hx3d = self.rebnconv3d(torch.cat((hx4,hx3),1)) hx3dup = _upsample_like(hx3d,hx2) hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1)) hx2dup = _upsample_like(hx2d,hx1) hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1)) return hx1d + hxin ### RSU-4F ### class RSU4F(nn.Module):#UNet04FRES(nn.Module): def __init__(self, in_ch=3, mid_ch=12, out_ch=3): super(RSU4F,self).__init__() self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1) self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=2) self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=4) self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=8) self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=4) self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=2) self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1) def forward(self,x): hx = x hxin = self.rebnconvin(hx) hx1 = self.rebnconv1(hxin) hx2 = self.rebnconv2(hx1) hx3 = self.rebnconv3(hx2) hx4 = self.rebnconv4(hx3) hx3d = self.rebnconv3d(torch.cat((hx4,hx3),1)) hx2d = self.rebnconv2d(torch.cat((hx3d,hx2),1)) hx1d = self.rebnconv1d(torch.cat((hx2d,hx1),1)) return hx1d + hxin ##### U^2-Net #### class U2NET(nn.Module): def __init__(self,in_ch=3,out_ch=1): super(U2NET,self).__init__() self.stage1 = RSU7(in_ch,32,64) self.pool12 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage2 = RSU6(64,32,128) self.pool23 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage3 = RSU5(128,64,256) self.pool34 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage4 = RSU4(256,128,512) self.pool45 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage5 = RSU4F(512,256,512) self.pool56 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage6 = RSU4F(512,256,512) # decoder self.stage5d = RSU4F(1024,256,512) self.stage4d = RSU4(1024,128,256) self.stage3d = RSU5(512,64,128) self.stage2d = RSU6(256,32,64) self.stage1d = RSU7(128,16,64) self.side1 = nn.Conv2d(64,out_ch,3,padding=1) self.side2 = nn.Conv2d(64,out_ch,3,padding=1) self.side3 = nn.Conv2d(128,out_ch,3,padding=1) self.side4 = nn.Conv2d(256,out_ch,3,padding=1) self.side5 = nn.Conv2d(512,out_ch,3,padding=1) self.side6 = nn.Conv2d(512,out_ch,3,padding=1) self.outconv = nn.Conv2d(6*out_ch,out_ch,1) def forward(self,x): hx = x #stage 1 hx1 = self.stage1(hx) hx = self.pool12(hx1) #stage 2 hx2 = self.stage2(hx) hx = self.pool23(hx2) #stage 3 hx3 = self.stage3(hx) hx = self.pool34(hx3) #stage 4 hx4 = self.stage4(hx) hx = self.pool45(hx4) #stage 5 hx5 = self.stage5(hx) hx = self.pool56(hx5) #stage 6 hx6 = self.stage6(hx) hx6up = _upsample_like(hx6,hx5) #-------------------- decoder -------------------- hx5d = self.stage5d(torch.cat((hx6up,hx5),1)) hx5dup = _upsample_like(hx5d,hx4) hx4d = self.stage4d(torch.cat((hx5dup,hx4),1)) hx4dup = _upsample_like(hx4d,hx3) hx3d = self.stage3d(torch.cat((hx4dup,hx3),1)) hx3dup = _upsample_like(hx3d,hx2) hx2d = self.stage2d(torch.cat((hx3dup,hx2),1)) hx2dup = _upsample_like(hx2d,hx1) hx1d = self.stage1d(torch.cat((hx2dup,hx1),1)) #side output d1 = self.side1(hx1d) d2 = self.side2(hx2d) d2 = _upsample_like(d2,d1) d3 = self.side3(hx3d) d3 = _upsample_like(d3,d1) d4 = self.side4(hx4d) d4 = _upsample_like(d4,d1) d5 = self.side5(hx5d) d5 = _upsample_like(d5,d1) d6 = self.side6(hx6) d6 = _upsample_like(d6,d1) d0 = self.outconv(torch.cat((d1,d2,d3,d4,d5,d6),1)) return F.sigmoid(d0), F.sigmoid(d1), F.sigmoid(d2), F.sigmoid(d3), F.sigmoid(d4), F.sigmoid(d5), F.sigmoid(d6) ### U^2-Net small ### class U2NETP(nn.Module): def __init__(self,in_ch=3,out_ch=1): super(U2NETP,self).__init__() self.stage1 = RSU7(in_ch,16,64) self.pool12 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage2 = RSU6(64,16,64) self.pool23 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage3 = RSU5(64,16,64) self.pool34 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage4 = RSU4(64,16,64) self.pool45 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage5 = RSU4F(64,16,64) self.pool56 = nn.MaxPool2d(2,stride=2,ceil_mode=True) self.stage6 = RSU4F(64,16,64) # decoder self.stage5d = RSU4F(128,16,64) self.stage4d = RSU4(128,16,64) self.stage3d = RSU5(128,16,64) self.stage2d = RSU6(128,16,64) self.stage1d = RSU7(128,16,64) self.side1 = nn.Conv2d(64,out_ch,3,padding=1) self.side2 = nn.Conv2d(64,out_ch,3,padding=1) self.side3 = nn.Conv2d(64,out_ch,3,padding=1) self.side4 = nn.Conv2d(64,out_ch,3,padding=1) self.side5 = nn.Conv2d(64,out_ch,3,padding=1) self.side6 = nn.Conv2d(64,out_ch,3,padding=1) self.outconv = nn.Conv2d(6*out_ch,out_ch,1) def forward(self,x): hx = x #stage 1 hx1 = self.stage1(hx) hx = self.pool12(hx1) #stage 2 hx2 = self.stage2(hx) hx = self.pool23(hx2) #stage 3 hx3 = self.stage3(hx) hx = self.pool34(hx3) #stage 4 hx4 = self.stage4(hx) hx = self.pool45(hx4) #stage 5 hx5 = self.stage5(hx) hx = self.pool56(hx5) #stage 6 hx6 = self.stage6(hx) hx6up = _upsample_like(hx6,hx5) #decoder hx5d = self.stage5d(torch.cat((hx6up,hx5),1)) hx5dup = _upsample_like(hx5d,hx4) hx4d = self.stage4d(torch.cat((hx5dup,hx4),1)) hx4dup = _upsample_like(hx4d,hx3) hx3d = self.stage3d(torch.cat((hx4dup,hx3),1)) hx3dup = _upsample_like(hx3d,hx2) hx2d = self.stage2d(torch.cat((hx3dup,hx2),1)) hx2dup = _upsample_like(hx2d,hx1) hx1d = self.stage1d(torch.cat((hx2dup,hx1),1)) #side output d1 = self.side1(hx1d) d2 = self.side2(hx2d) d2 = _upsample_like(d2,d1) d3 = self.side3(hx3d) d3 = _upsample_like(d3,d1) d4 = self.side4(hx4d) d4 = _upsample_like(d4,d1) d5 = self.side5(hx5d) d5 = _upsample_like(d5,d1) d6 = self.side6(hx6) d6 = _upsample_like(d6,d1) d0 = self.outconv(torch.cat((d1,d2,d3,d4,d5,d6),1)) return F.sigmoid(d0), F.sigmoid(d1), F.sigmoid(d2), F.sigmoid(d3), F.sigmoid(d4), F.sigmoid(d5), F.sigmoid(d6)
14,719
26.984791
118
py
sm-vit
sm-vit-main/U2Net/model/u2net_refactor.py
import torch import torch.nn as nn import math __all__ = ['U2NET_full', 'U2NET_lite'] def _upsample_like(x, size): return nn.Upsample(size=size, mode='bilinear', align_corners=False)(x) def _size_map(x, height): # {height: size} for Upsample size = list(x.shape[-2:]) sizes = {} for h in range(1, height): sizes[h] = size size = [math.ceil(w / 2) for w in size] return sizes class REBNCONV(nn.Module): def __init__(self, in_ch=3, out_ch=3, dilate=1): super(REBNCONV, self).__init__() self.conv_s1 = nn.Conv2d(in_ch, out_ch, 3, padding=1 * dilate, dilation=1 * dilate) self.bn_s1 = nn.BatchNorm2d(out_ch) self.relu_s1 = nn.ReLU(inplace=True) def forward(self, x): return self.relu_s1(self.bn_s1(self.conv_s1(x))) class RSU(nn.Module): def __init__(self, name, height, in_ch, mid_ch, out_ch, dilated=False): super(RSU, self).__init__() self.name = name self.height = height self.dilated = dilated self._make_layers(height, in_ch, mid_ch, out_ch, dilated) def forward(self, x): sizes = _size_map(x, self.height) x = self.rebnconvin(x) # U-Net like symmetric encoder-decoder structure def unet(x, height=1): if height < self.height: x1 = getattr(self, f'rebnconv{height}')(x) if not self.dilated and height < self.height - 1: x2 = unet(getattr(self, 'downsample')(x1), height + 1) else: x2 = unet(x1, height + 1) x = getattr(self, f'rebnconv{height}d')(torch.cat((x2, x1), 1)) return _upsample_like(x, sizes[height - 1]) if not self.dilated and height > 1 else x else: return getattr(self, f'rebnconv{height}')(x) return x + unet(x) def _make_layers(self, height, in_ch, mid_ch, out_ch, dilated=False): self.add_module('rebnconvin', REBNCONV(in_ch, out_ch)) self.add_module('downsample', nn.MaxPool2d(2, stride=2, ceil_mode=True)) self.add_module(f'rebnconv1', REBNCONV(out_ch, mid_ch)) self.add_module(f'rebnconv1d', REBNCONV(mid_ch * 2, out_ch)) for i in range(2, height): dilate = 1 if not dilated else 2 ** (i - 1) self.add_module(f'rebnconv{i}', REBNCONV(mid_ch, mid_ch, dilate=dilate)) self.add_module(f'rebnconv{i}d', REBNCONV(mid_ch * 2, mid_ch, dilate=dilate)) dilate = 2 if not dilated else 2 ** (height - 1) self.add_module(f'rebnconv{height}', REBNCONV(mid_ch, mid_ch, dilate=dilate)) class U2NET(nn.Module): def __init__(self, cfgs, out_ch): super(U2NET, self).__init__() self.out_ch = out_ch self._make_layers(cfgs) def forward(self, x): sizes = _size_map(x, self.height) maps = [] # storage for maps # side saliency map def unet(x, height=1): if height < 6: x1 = getattr(self, f'stage{height}')(x) x2 = unet(getattr(self, 'downsample')(x1), height + 1) x = getattr(self, f'stage{height}d')(torch.cat((x2, x1), 1)) side(x, height) return _upsample_like(x, sizes[height - 1]) if height > 1 else x else: x = getattr(self, f'stage{height}')(x) side(x, height) return _upsample_like(x, sizes[height - 1]) def side(x, h): # side output saliency map (before sigmoid) x = getattr(self, f'side{h}')(x) x = _upsample_like(x, sizes[1]) maps.append(x) def fuse(): # fuse saliency probability maps maps.reverse() x = torch.cat(maps, 1) x = getattr(self, 'outconv')(x) maps.insert(0, x) return [torch.sigmoid(x) for x in maps] unet(x) maps = fuse() return maps def _make_layers(self, cfgs): self.height = int((len(cfgs) + 1) / 2) self.add_module('downsample', nn.MaxPool2d(2, stride=2, ceil_mode=True)) for k, v in cfgs.items(): # build rsu block self.add_module(k, RSU(v[0], *v[1])) if v[2] > 0: # build side layer self.add_module(f'side{v[0][-1]}', nn.Conv2d(v[2], self.out_ch, 3, padding=1)) # build fuse layer self.add_module('outconv', nn.Conv2d(int(self.height * self.out_ch), self.out_ch, 1)) def U2NET_full(): full = { # cfgs for building RSUs and sides # {stage : [name, (height(L), in_ch, mid_ch, out_ch, dilated), side]} 'stage1': ['En_1', (7, 3, 32, 64), -1], 'stage2': ['En_2', (6, 64, 32, 128), -1], 'stage3': ['En_3', (5, 128, 64, 256), -1], 'stage4': ['En_4', (4, 256, 128, 512), -1], 'stage5': ['En_5', (4, 512, 256, 512, True), -1], 'stage6': ['En_6', (4, 512, 256, 512, True), 512], 'stage5d': ['De_5', (4, 1024, 256, 512, True), 512], 'stage4d': ['De_4', (4, 1024, 128, 256), 256], 'stage3d': ['De_3', (5, 512, 64, 128), 128], 'stage2d': ['De_2', (6, 256, 32, 64), 64], 'stage1d': ['De_1', (7, 128, 16, 64), 64], } return U2NET(cfgs=full, out_ch=1) def U2NET_lite(): lite = { # cfgs for building RSUs and sides # {stage : [name, (height(L), in_ch, mid_ch, out_ch, dilated), side]} 'stage1': ['En_1', (7, 3, 16, 64), -1], 'stage2': ['En_2', (6, 64, 16, 64), -1], 'stage3': ['En_3', (5, 64, 16, 64), -1], 'stage4': ['En_4', (4, 64, 16, 64), -1], 'stage5': ['En_5', (4, 64, 16, 64, True), -1], 'stage6': ['En_6', (4, 64, 16, 64, True), 64], 'stage5d': ['De_5', (4, 128, 16, 64, True), 64], 'stage4d': ['De_4', (4, 128, 16, 64), 64], 'stage3d': ['De_3', (5, 128, 16, 64), 64], 'stage2d': ['De_2', (6, 128, 16, 64), 64], 'stage1d': ['De_1', (7, 128, 16, 64), 64], } return U2NET(cfgs=lite, out_ch=1)
6,097
35.08284
101
py
HighOrderAtten
HighOrderAtten-master/image_model/download_model.py
""" Download the VGG and deep residual model to extract image features. Version: 1.0 Contributor: Jiasen Lu """ import os import argparse import json def download_VGG(): print('Downloading VGG model from http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_19_layers.caffemodel') os.system('wget http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_19_layers.caffemodel') os.system('wget https://gist.githubusercontent.com/ksimonyan/3785162f95cd2d5fee77/raw/bb2b4fe0a9bb0669211cf3d0bc949dfdda173e9e/VGG_ILSVRC_19_layers_deploy.prototxt') def download_deep_residual(): print('Downloading deep residual model from https://d2j0dndfm35trm.cloudfront.net/resnet-200.t7') os.system('wget https://d2j0dndfm35trm.cloudfront.net/resnet-200.t7') os.system('wget https://raw.githubusercontent.com/facebook/fb.resnet.torch/master/datasets/transforms.lua') def main(params): if params['download'] == 'VGG': download_VGG() else: download_deep_residual() if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument('--download', default='VGG', help='VGG or Residual') # input json args = parser.parse_args() params = vars(args) print 'parsed input parameters:' print json.dumps(params, indent = 2) main(params)
1,337
35.162162
169
py
HighOrderAtten
HighOrderAtten-master/data/prepro_vqa.py
''' Preoricess a raw json dataset into hdf5/json files. Caption: Use NLTK or split function to get tokens. ''' from random import shuffle, seed import sys import os.path import argparse import numpy as np import scipy.io import pdb import h5py from nltk.tokenize import word_tokenize import json import re import math def tokenize(sentence): return [i for i in re.split(r"([-.\"',:? !\$#@~()*&\^%;\[\]/\\\+<>\n=])", sentence) if i!='' and i!=' ' and i!='\n']; def prepro_question(dataset, params): # preprocess all the question print 'example processed tokens:' for i,ques in enumerate(dataset): s = ques['question'] if params['token_method'] == 'nltk': txt = word_tokenize(str(s).lower()) else: txt = tokenize(s) ques['processed_tokens'] = txt if i < 10: print txt if i % 1000 == 0: sys.stdout.write("processing %d/%d (%.2f%% done) \r" % (i, len(dataset), i*100.0/len(dataset)) ) sys.stdout.flush() return dataset def build_vocab_question(dataset, params): # build vocabulary for question and answers. count_thr = params['word_count_threshold'] # count up the number of words counts = {} for ques in dataset: for w in ques['processed_tokens']: counts[w] = counts.get(w, 0) + 1 cw = sorted([(count,w) for w,count in counts.iteritems()], reverse=True) print 'top words and their counts:' print '\n'.join(map(str,cw[:20])) # print some stats total_words = sum(counts.itervalues()) print 'total words:', total_words bad_words = [w for w,n in counts.iteritems() if n <= count_thr] vocab = [w for w,n in counts.iteritems() if n > count_thr] bad_count = sum(counts[w] for w in bad_words) print 'number of bad words: %d/%d = %.2f%%' % (len(bad_words), len(counts), len(bad_words)*100.0/len(counts)) print 'number of words in vocab would be %d' % (len(vocab), ) print 'number of UNKs: %d/%d = %.2f%%' % (bad_count, total_words, bad_count*100.0/total_words) # lets now produce the final annotation # additional special UNK token we will use below to map infrequent words to print 'inserting the special UNK token' vocab.append('UNK') for ques in dataset: txt = ques['processed_tokens'] question = [w if counts.get(w,0) > count_thr else 'UNK' for w in txt] ques['final_question'] = question return dataset, vocab def apply_vocab_question(dataset, wtoi): # apply the vocab on test. for ques in dataset: txt = ques['processed_tokens'] question = [w if w in wtoi else 'UNK' for w in txt] ques['final_question'] = question return dataset def get_top_answers(dataset, params): counts = {} for ques in dataset: ans = ques['ans'] counts[ans] = counts.get(ans, 0) + 1 cw = sorted([(count,w) for w,count in counts.iteritems()], reverse=True) print 'top answer and their counts:' print '\n'.join(map(str,cw[:20])) vocab = [] for i in range(params['num_ans']): vocab.append(cw[i][1]) return vocab[:params['num_ans']] def encode_question(dataset, params, wtoi): max_length = params['max_length'] N = len(dataset) label_arrays = np.zeros((N, max_length), dtype='uint32') label_length = np.zeros(N, dtype='uint32') question_id = np.zeros(N, dtype='uint32') question_counter = 0 for i,ques in enumerate(dataset): question_id[question_counter] = ques['ques_id'] label_length[question_counter] = min(max_length, len(ques['final_question'])) # record the length of this sequence question_counter += 1 for k,w in enumerate(ques['final_question']): if k < max_length: label_arrays[i,k] = wtoi[w] return label_arrays, label_length, question_id def encode_answer(dataset, atoi, num_ans): N = len(dataset) ans_arrays = np.zeros(N, dtype='uint32') for i, ques in enumerate(dataset): ans_arrays[i] = atoi.get(ques['ans'], num_ans+1) # -1 means wrong answer. return ans_arrays def encode_mc_answer(dataset, atoi, num_ans): N = len(dataset) mc_ans_arrays = np.zeros((N, 18), dtype='uint32') for i, ques in enumerate(dataset): for j, ans in enumerate(ques['MC_ans']): mc_ans_arrays[i,j] = atoi.get(ans, num_ans+1) return mc_ans_arrays def filter_question(dataset, atoi): new_dataset = [] for i, ques in enumerate(dataset): if ques['ans'] in atoi: new_dataset.append(ques) print 'question number reduce from %d to %d '%(len(dataset), len(new_dataset)) return new_dataset def get_unqiue_img(dataset): count_img = {} N = len(dataset) img_pos = np.zeros(N, dtype='uint32') ques_pos_tmp = {} for ques in dataset: count_img[ques['img_path']] = count_img.get(ques['img_path'], 0) + 1 unique_img = [w for w,n in count_img.iteritems()] imgtoi = {w:i+1 for i,w in enumerate(unique_img)} # add one for torch, since torch start from 1. for i, ques in enumerate(dataset): idx = imgtoi.get(ques['img_path']) img_pos[i] = idx if idx-1 not in ques_pos_tmp: ques_pos_tmp[idx-1] = [] ques_pos_tmp[idx-1].append(i+1) img_N = len(ques_pos_tmp) ques_pos = np.zeros((img_N,3), dtype='uint32') ques_pos_len = np.zeros(img_N, dtype='uint32') for idx, ques_list in ques_pos_tmp.iteritems(): ques_pos_len[idx] = len(ques_list) for j in range(len(ques_list)): ques_pos[idx][j] = ques_list[j] return unique_img, img_pos, ques_pos, ques_pos_len def main(params): # create output h5 file for training set. f = h5py.File(params['output_h5'], "w") if params['input_json']=='': dataset_train = json.load(open(params['input_train_json'], 'r')) #dataset_train = dataset_train[:5000] #dataset_test = dataset_test[:5000] # get top answers top_ans = get_top_answers(dataset_train, params) atoi = {w:i+1 for i,w in enumerate(top_ans)} atoi['error'] = params['num_ans']+1 itoa = {i+1:w for i,w in enumerate(top_ans)} itoa[params['num_ans']+1] = 'error' # filter question, which isn't in the top answers. dataset_train = filter_question(dataset_train, atoi) # tokenization and preprocessing training question dataset_train = prepro_question(dataset_train, params) # create the vocab for question dataset_train, vocab = build_vocab_question(dataset_train, params) itow = {i+1:w for i,w in enumerate(vocab)} # a 1-indexed vocab translation table wtoi = {w:i+1 for i,w in enumerate(vocab)} # inverse table ques_train, ques_length_train, question_id_train = encode_question(dataset_train, params, wtoi) # get the unique image for train unique_img_train, img_pos_train, ques_pos_train, ques_pos_len_train = get_unqiue_img(dataset_train) # get the answer encoding. ans_train = encode_answer(dataset_train, atoi, params['num_ans']) MC_ans_train = encode_mc_answer(dataset_train, atoi, params['num_ans']) N_train = len(dataset_train) split_train = np.zeros(N_train) f.create_dataset("ques_train", dtype='uint32', data=ques_train) f.create_dataset("answers", dtype='uint32', data=ans_train) f.create_dataset("ques_id_train", dtype='uint32', data=question_id_train) f.create_dataset("img_pos_train", dtype='uint32', data=img_pos_train) f.create_dataset("ques_pos_train", dtype='uint32', data=ques_pos_train) f.create_dataset("ques_pos_len_train", dtype='uint32', data=ques_pos_len_train) f.create_dataset("split_train", dtype='uint32', data=split_train) f.create_dataset("ques_len_train", dtype='uint32', data=ques_length_train) f.create_dataset("MC_ans_train", dtype='uint32', data=MC_ans_train) else: loaded_train_data = json.load(open(params['input_json'], 'r')) itow = loaded_train_data['ix_to_word'] wtoi = {v: k for k, v in itow.iteritems()} itoa = loaded_train_data['ix_to_ans'] atoi = {v: k for k, v in itoa.iteritems()} unique_img_train = loaded_train_data['unique_img_train'] dataset_test = json.load(open(params['input_test_json'], 'r')) # tokenization and preprocessing testing question dataset_test = prepro_question(dataset_test, params) dataset_test = apply_vocab_question(dataset_test, wtoi) ques_test, ques_length_test, question_id_test = encode_question(dataset_test, params, wtoi) # get the unique image for test unique_img_test, img_pos_test, ques_pos_test, ques_pos_len_test = get_unqiue_img(dataset_test) if not params['test']: ans_test = encode_answer(dataset_test, atoi, params['num_ans']) #also comment line 238 MC_ans_test = encode_mc_answer(dataset_test, atoi, params['num_ans']) # get the split N_test = len(dataset_test) # since the train image is already suffled, we just use the last val_num image as validation # train = 0, val = 1, test = 2 #split_train[N_train - params['val_num']: N_train] = 1 split_test = np.zeros(N_test) split_test[:] = 2 f.create_dataset("ques_test", dtype='uint32', data=ques_test) if not params['test']: f.create_dataset("ans_test", dtype='uint32', data=ans_test) f.create_dataset("ques_id_test", dtype='uint32', data=question_id_test) f.create_dataset("img_pos_test", dtype='uint32', data=img_pos_test) f.create_dataset("ques_pos_test", dtype='uint32', data=ques_pos_test) f.create_dataset("ques_pos_len_test", dtype='uint32', data=ques_pos_len_test) f.create_dataset("split_test", dtype='uint32', data=split_test) f.create_dataset("ques_len_test", dtype='uint32', data=ques_length_test) f.create_dataset("MC_ans_test", dtype='uint32', data=MC_ans_test) f.close() print 'wrote ', params['output_h5'] # create output json file out = {} out['ix_to_word'] = itow # encode the (1-indexed) vocab out['ix_to_ans'] = itoa out['unique_img_train'] = unique_img_train out['uniuqe_img_test'] = unique_img_test json.dump(out, open(params['output_json'], 'w')) print 'wrote ', params['output_json'] if __name__ == "__main__": parser = argparse.ArgumentParser() # input json parser.add_argument('--input_train_json', default='vqa_raw_train.json', help='input json file to process into hdf5') parser.add_argument('--input_test_json', default='vqa_raw_test.json', help='input json file to process into hdf5') parser.add_argument('--num_ans', default=3000, type=int, help='number of top answers for the final classifications.') parser.add_argument('--input_json', default='' ,help='input existing train perprocess, usefull to process a new test file') parser.add_argument('--output_json', default='vqa_data_prepro.json', help='output json file') parser.add_argument('--output_h5', default='vqa_data_prepro.h5', help='output h5 file') # options parser.add_argument('--max_length', default=15, type=int, help='max length of a caption, in number of words. captions longer than this get clipped.') parser.add_argument('--word_count_threshold', default=0, type=int, help='only words that occur more than this number of times will be put in vocab') parser.add_argument('--token_method', default='nltk', help='token method, nltk is much more slower.') parser.add_argument('--test', default=0 ,type=int, help='token method, nltk is much more slower.') args = parser.parse_args() params = vars(args) # convert to ordinary dict print 'parsed input parameters:' print json.dumps(params, indent = 2) main(params)
11,897
37.882353
153
py
MCEdit-Unified
MCEdit-Unified-master/renderer.py
"""Copyright (c) 2010-2012 David Rio Vierra Permission to use, copy, modify, and/or distribute this software for any purpose with or without fee is hereby granted, provided that the above copyright notice and this permission notice appear in all copies. THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.""" """ renderer.py What is going on in this file? Here is an attempt to show the relationships between classes and their responsibilities MCRenderer: has "position", "origin", optionally "viewFrustum" Loads chunks near position+origin, draws chunks offset by origin Calls visible on viewFrustum to exclude chunks (+) ChunkRenderer Has "chunkPosition", "invalidLayers", "lists" One per chunk and detail level. Creates display lists from BlockRenderers (*) BlockRenderer Has "vertexArrays" One per block type, plus one for low detail and one for Entity BlockRender documentation Each Block renderer renders a particular block types or entities. The block renderer that is chosen to draw that block type(by ID) is the block renderer class that is lowest in the list within the makeRenderstates method. Each blockRenderer is assigned a materialIndex and the blockMaterial parameter indicates what material index each block in the chunk is therefore what block renderer is used to render it. Vertex arrays are arrays of vertices(groups of six elements) and every group of 4 vertices is a quad that will be drawn. Before the vertex arrays will be drawn `.ravel()` will be called (flattened to one dimension arrays). The vertex arrays will draw quads and each vertex elements with the foramt: 0:3 index - xyz values 3:5 index - st(texture coordinates) values 5 index - rgba(colour) value Note: each element of rgba value is a uint8 type(the 4 colour elements makes up 32 bits) to view/change the values use `.view('uint8')` to change the view of the array into uint8 type. To implement a block renderer either makeVertices or makeFaceVertices needs to be implemented. The base class BlockRenderer implements makeVertices in terms of makeFaceVertices, by iterating over the different face directions. The makeVertices function is called on the block renderer to gat a list of vertexArrays that will draw the blocks for a 16x16x16 chunk. parameters: all parameters are in xzy order facingBlockIndices: list of 6, (16, 16, 16) numpy boolean arrays each array corresponds to the blocks within the chunk that has it face exposed in that direction. The direction is the index into the list defined by the constants in pymclevel/faces.py This is used to only draw exposed faces blocks: (16, 16, 16) numpy array of the id of blocks in the chunk blockMaterials: (16, 16, 16) numpy array of the material index of each block in the chunk each material refers to a different block renderer to get the material index for this block renderer `self.materialIndex` blockData: (16, 16, 16) numpy array of the metadata value of each block in the chunk areaBlockLights: (18, 18, 18) numpy array of light value(max of block light and skylight) of the chunk and 1 block 'border' aroun it. texMap: function that takes id, data value and directions and returns texture coordinates returns a list of vertex arrays in the form of float32 numpy arrays. For this chunk. The makeFaceVertices gets an vertexArray for a particular face. parameters: all parameters are in xzy order direction: the face defined by constants in pymclevel/faces.py materialIndices: list of (x, z, y) indices of blocks in this chunks that is of this material(in blocktypes). exposedFaceIndices: list of (x, z, y) indices of blocks that has an exposed face in the direction `direction`. blocks: (16, 16, 16) numpy array of the id of blocks in the chunk blockData: (16, 16, 16) numpy array of the metadata value of each block in the chunk blockLights: (16, 16, 16) numpy array of light values(max of block light and skylight) of the blocks in the chunk chunk. facingBlockLight: (16, 16, 16) numpy array of light values(max of block light and skylight) of the blocks just in front of the face. i.e. if direction = pymclevel.faces.FaceXDecreasing facingBlockLight[1, 0, 0] refers to the light level at position (0, 0, 0) within the chunk. texMap: function that takes id, data value and directions and returns texture coordinates returns a list of vertex arrays in the form of float32 numpy arrays. Fields blocktypes / getBlocktypes(mats) list of block ids the block renderer handles detailLevels what detail level the renderer render at layer what layer is this block renderer in renderstate the render state this block renderer uses Models: There are also several functions that make it easy to translate json models to block renderer. makeVertexTemplatesFromJsonModel: creates a template from information that is in json models rotateTemplate: rotate templates. This is equivalent to the rotation in block states files. makeVerticesFromModel: creates function based on templates to be used for makeVertices function in block renderer. Helper functions: self.MaterialIndices(blockMaterial): Given blockMaterial(parameter in makeVertices) it return a list of (x, z, y) indices of blocks in the chunk that are of this block renderer material(blocktypes). self.makeTemplate(direction, blockIndices): get a vertex array filled with default values for face `direction` and for the block relating to `blockIndices` makeVertexTemplates(xmin=0, ymin=0, zmin=0, xmax=1, ymax=1, zmax=1): returns a numpy array with dimensions (6, 4, 6) filled with values to create a vertex array for a cube. For Entities: renderer's for entities are similar to blocks but: - they extend EntityRendererGeneric class - they are added to the list in calcFacesForChunkRenderer method - makeChunkVertices(chunk) where chunk is a chunk object is called rather than makeVertices there is also a helper method _computeVertices(positions, colors, offset, chunkPosition): parameters: positions locations of entity colors colors of entity boxes offset whether to offset the box chunkPosition chunk position of the chunk creates a vertex array that draws entity boxes """ from collections import defaultdict, deque from datetime import datetime, timedelta from depths import DepthOffset from glutils import gl, Texture from albow.resource import _2478aq_heot import logging import numpy from OpenGL import GL import pymclevel from pymclevel.materials import alphaMaterials, pocketMaterials import sys from config import config # import time def get_materials(): alphaMaterials = pymclevel.materials.alphaMaterials pocketMaterials = pymclevel.materials.pocketMaterials def chunkMarkers(chunkSet): """ Returns a mapping { size: [position, ...] } for different powers of 2 as size. """ sizedChunks = defaultdict(list) size = 1 def all4(cx, cz): cx &= ~size cz &= ~size return [(cx, cz), (cx + size, cz), (cx + size, cz + size), (cx, cz + size)] # lastsize = 6 size = 1 while True: nextsize = size << 1 chunkSet = set(chunkSet) while len(chunkSet): cx, cz = chunkSet.pop() chunkSet.add((cx, cz)) o = all4(cx, cz) others = set(o).intersection(chunkSet) if len(others) == 4: sizedChunks[nextsize].append(o[0]) # Possibly cache append? for c in others: chunkSet.discard(c) else: for c in others: sizedChunks[size].append(c) # Possibly cache append? chunkSet.discard(c) if len(sizedChunks[nextsize]): chunkSet = set(sizedChunks[nextsize]) sizedChunks[nextsize] = [] size <<= 1 else: break return sizedChunks class ChunkRenderer(object): maxlod = 2 minlod = 0 def __init__(self, renderer, chunkPosition): self.renderer = renderer self.blockRenderers = [] self.detailLevel = 0 self.invalidLayers = set(Layer.AllLayers) self.chunkPosition = chunkPosition self.bufferSize = 0 self.renderstateLists = None @property def visibleLayers(self): return self.renderer.visibleLayers def forgetDisplayLists(self, states=None): if self.renderstateLists is not None: # print "Discarded {0}, gained {1} bytes".format(self.chunkPosition,self.bufferSize) for k in states or self.renderstateLists.iterkeys(): a = self.renderstateLists.get(k, []) # print a for i in a: gl.glDeleteLists(i, 1) if states: del self.renderstateLists[states] else: self.renderstateLists = None self.needsRedisplay = True self.renderer.discardMasterList() def debugDraw(self): for blockRenderer in self.blockRenderers: blockRenderer.drawArrays(self.chunkPosition, False) def makeDisplayLists(self): if not self.needsRedisplay: return self.forgetDisplayLists() if not self.blockRenderers: return lists = defaultdict(list) showRedraw = self.renderer.showRedraw if not (showRedraw and self.needsBlockRedraw): GL.glEnableClientState(GL.GL_COLOR_ARRAY) renderers = self.blockRenderers for blockRenderer in renderers: if self.detailLevel not in blockRenderer.detailLevels: continue if blockRenderer.layer not in self.visibleLayers: continue l = blockRenderer.makeArrayList(self.chunkPosition, self.needsBlockRedraw and showRedraw) lists[blockRenderer.renderstate].append(l) if not (showRedraw and self.needsBlockRedraw): GL.glDisableClientState(GL.GL_COLOR_ARRAY) self.needsRedisplay = False self.renderstateLists = lists @property def needsBlockRedraw(self): return Layer.Blocks in self.invalidLayers def invalidate(self, layers=None): if layers is None: layers = Layer.AllLayers if layers: layers = set(layers) self.invalidLayers.update(layers) blockRenderers = [br for br in self.blockRenderers if br.layer is Layer.Blocks or br.layer not in layers] if len(blockRenderers) < len(self.blockRenderers): self.forgetDisplayLists() self.blockRenderers = blockRenderers if self.renderer.showRedraw and Layer.Blocks in layers: self.needsRedisplay = True def calcFaces(self): minlod = self.renderer.detailLevelForChunk(self.chunkPosition) minlod = min(minlod, self.maxlod) if self.detailLevel != minlod: self.forgetDisplayLists() self.detailLevel = minlod self.invalidLayers.add(Layer.Blocks) # discard the standard detail renderers if minlod > 0: blockRenderers = [] append = blockRenderers.append for br in self.blockRenderers: if br.detailLevels != (0,): append(br) self.blockRenderers = blockRenderers if self.renderer.chunkCalculator: for _ in self.renderer.chunkCalculator.calcFacesForChunkRenderer(self): yield else: raise StopIteration def vertexArraysDone(self): bufferSize = 0 for br in self.blockRenderers: bufferSize += br.bufferSize() if self.renderer.alpha != 0xff: br.setAlpha(self.renderer.alpha) self.bufferSize = bufferSize self.invalidLayers = set() self.needsRedisplay = True self.renderer.invalidateMasterList() needsRedisplay = False @property def done(self): return not self.invalidLayers _XYZ = numpy.s_[..., 0:3] _ST = numpy.s_[..., 3:5] _XYZST = numpy.s_[..., :5] _RGBA = numpy.s_[..., 20:24] _RGB = numpy.s_[..., 20:23] _A = numpy.s_[..., 23] def makeVertexTemplatesFromJsonModel(fromVertices, toVertices, uv): """ This is similar to makeVertexTemplates but is a more convenient when reading off of the json model files. :param fromVertices: from :param toVertices: to :param uv: keywords uv map :return: template for a cube """ xmin = fromVertices[0] / 16. xmax = toVertices[0] / 16. ymin = fromVertices[1] / 16. ymax = toVertices[1] / 16. zmin = fromVertices[2] / 16. zmax = toVertices[2] / 16. return numpy.array([ # FaceXIncreasing: [[xmax, ymin, zmax, uv["east"][0], uv["east"][3], 0x0b], [xmax, ymin, zmin, uv["east"][2], uv["east"][3], 0x0b], [xmax, ymax, zmin, uv["east"][2], uv["east"][1], 0x0b], [xmax, ymax, zmax, uv["east"][0], uv["east"][1], 0x0b], ], # FaceXDecreasing: [[xmin, ymin, zmin, uv["west"][0], uv["west"][3], 0x0b], [xmin, ymin, zmax, uv["west"][2], uv["west"][3], 0x0b], [xmin, ymax, zmax, uv["west"][2], uv["west"][1], 0x0b], [xmin, ymax, zmin, uv["west"][0], uv["west"][1], 0x0b]], # FaceYIncreasing: [[xmin, ymax, zmin, uv["up"][0], uv["up"][1], 0x11], # ne [xmin, ymax, zmax, uv["up"][0], uv["up"][3], 0x11], # nw [xmax, ymax, zmax, uv["up"][2], uv["up"][3], 0x11], # sw [xmax, ymax, zmin, uv["up"][2], uv["up"][1], 0x11]], # se # FaceYDecreasing: [[xmin, ymin, zmin, uv["down"][0], uv["down"][3], 0x08], [xmax, ymin, zmin, uv["down"][2], uv["down"][3], 0x08], [xmax, ymin, zmax, uv["down"][2], uv["down"][1], 0x08], [xmin, ymin, zmax, uv["down"][0], uv["down"][1], 0x08]], # FaceZIncreasing: [[xmin, ymin, zmax, uv["south"][0], uv["south"][3], 0x0d], [xmax, ymin, zmax, uv["south"][2], uv["south"][3], 0x0d], [xmax, ymax, zmax, uv["south"][2], uv["south"][1], 0x0d], [xmin, ymax, zmax, uv["south"][0], uv["south"][1], 0x0d]], # FaceZDecreasing: [[xmax, ymin, zmin, uv["north"][0], uv["north"][3], 0x0d], [xmin, ymin, zmin, uv["north"][2], uv["north"][3], 0x0d], [xmin, ymax, zmin, uv["north"][2], uv["north"][1], 0x0d], [xmax, ymax, zmin, uv["north"][0], uv["north"][1], 0x0d], ], ]) def rotateTemplate(template, x=0, y=0): """ Rotate template around x-axis and then around y-axis. Both angles must to multiples of 90. TODO: Add ability for multiples of 45 """ template = template.copy() for _ in xrange(0, x, 90): # y -> -z and z -> y template[..., (1, 2)] = template[..., (2, 1)] template[..., 2] -= 0.5 template[..., 2] *= -1 template[..., 2] += 0.5 for _ in xrange(0, y, 90): # z -> -x and x -> z template[..., (0, 2)] = template[..., (2, 0)] template[..., 0] -= 0.5 template[..., 0] *= -1 template[..., 0] += 0.5 return template def makeVerticesFromModel(templates, dataMask=0): """ Returns a function that creates vertex arrays. This produces vertex arrays based on the passed templates. This doesn't cull any faces based on if they are exposed. :param templates: list of templates to draw :param dataMask: mask to mask the data """ if isinstance(templates, list): templates = numpy.array(templates) if templates.shape == (6, 4, 6): templates = numpy.array([templates]) if len(templates.shape) == 4: templates = templates[numpy.newaxis, ...] elements = templates.shape[0] def makeVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): mask = self.getMaterialIndices(blockMaterials) blockIndices = mask.nonzero() yield data = blockData[mask] data &= dataMask self.vertexArrays = [] append = self.vertexArrays.append for i in xrange(elements): vertexArray = numpy.zeros((len(blockIndices[0]), 6, 4, 6), dtype='float32') for indicies in xrange(3): dimension = (0, 2, 1)[indicies] vertexArray[..., indicies] = blockIndices[dimension][:, numpy.newaxis, numpy.newaxis] # xxx swap z with y using ^ vertexArray[..., 0:5] += templates[i, data][..., 0:5] vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices] & 15)[..., numpy.newaxis, :] vertexArray.view('uint8')[_RGB] = templates[i, data][..., 5][..., numpy.newaxis] vertexArray.view('uint8')[_A] = 0xFF vertexArray.view('uint8')[_RGB] *= areaBlockLights[1:-1, 1:-1, 1:-1][blockIndices][ ..., numpy.newaxis, numpy.newaxis, numpy.newaxis] vertexArray.shape = (vertexArray.shape[0] * 6, 4, 6) yield append(vertexArray) return makeVertices def makeVertexTemplates(xmin=0, ymin=0, zmin=0, xmax=1, ymax=1, zmax=1): return numpy.array([ # FaceXIncreasing: [[xmax, ymin, zmax, (zmin * 16), 16 - (ymin * 16), 0x0b], [xmax, ymin, zmin, (zmax * 16), 16 - (ymin * 16), 0x0b], [xmax, ymax, zmin, (zmax * 16), 16 - (ymax * 16), 0x0b], [xmax, ymax, zmax, (zmin * 16), 16 - (ymax * 16), 0x0b], ], # FaceXDecreasing: [[xmin, ymin, zmin, (zmin * 16), 16 - (ymin * 16), 0x0b], [xmin, ymin, zmax, (zmax * 16), 16 - (ymin * 16), 0x0b], [xmin, ymax, zmax, (zmax * 16), 16 - (ymax * 16), 0x0b], [xmin, ymax, zmin, (zmin * 16), 16 - (ymax * 16), 0x0b]], # FaceYIncreasing: [[xmin, ymax, zmin, xmin * 16, 16 - (zmax * 16), 0x11], # ne [xmin, ymax, zmax, xmin * 16, 16 - (zmin * 16), 0x11], # nw [xmax, ymax, zmax, xmax * 16, 16 - (zmin * 16), 0x11], # sw [xmax, ymax, zmin, xmax * 16, 16 - (zmax * 16), 0x11]], # se # FaceYDecreasing: [[xmin, ymin, zmin, xmin * 16, 16 - (zmax * 16), 0x08], [xmax, ymin, zmin, xmax * 16, 16 - (zmax * 16), 0x08], [xmax, ymin, zmax, xmax * 16, 16 - (zmin * 16), 0x08], [xmin, ymin, zmax, xmin * 16, 16 - (zmin * 16), 0x08]], # FaceZIncreasing: [[xmin, ymin, zmax, xmin * 16, 16 - (ymin * 16), 0x0d], [xmax, ymin, zmax, xmax * 16, 16 - (ymin * 16), 0x0d], [xmax, ymax, zmax, xmax * 16, 16 - (ymax * 16), 0x0d], [xmin, ymax, zmax, xmin * 16, 16 - (ymax * 16), 0x0d]], # FaceZDecreasing: [[xmax, ymin, zmin, xmin * 16, 16 - (ymin * 16), 0x0d], [xmin, ymin, zmin, xmax * 16, 16 - (ymin * 16), 0x0d], [xmin, ymax, zmin, xmax * 16, 16 - (ymax * 16), 0x0d], [xmax, ymax, zmin, xmin * 16, 16 - (ymax * 16), 0x0d], ], ]) elementByteLength = 24 def createPrecomputedVertices(): height = 16 precomputedVertices = [numpy.zeros(shape=(16, 16, height, 4, 6), # x,y,z,s,t,rg, ba dtype='float32') for d in faceVertexTemplates] xArray = numpy.arange(16)[:, numpy.newaxis, numpy.newaxis, numpy.newaxis] zArray = numpy.arange(16)[numpy.newaxis, :, numpy.newaxis, numpy.newaxis] yArray = numpy.arange(height)[numpy.newaxis, numpy.newaxis, :, numpy.newaxis] for dir in xrange(len(faceVertexTemplates)): precomputedVertices[dir][_XYZ][..., 0] = xArray precomputedVertices[dir][_XYZ][..., 1] = yArray precomputedVertices[dir][_XYZ][..., 2] = zArray precomputedVertices[dir][_XYZ] += faceVertexTemplates[dir][..., 0:3] # xyz precomputedVertices[dir][_ST] = faceVertexTemplates[dir][..., 3:5] # s precomputedVertices[dir].view('uint8')[_RGB] = faceVertexTemplates[dir][..., 5, numpy.newaxis] precomputedVertices[dir].view('uint8')[_A] = 0xff return precomputedVertices faceVertexTemplates = makeVertexTemplates() class ChunkCalculator(object): cachedTemplate = None cachedTemplateHeight = 0 whiteLight = numpy.array([[[15] * 16] * 16] * 16, numpy.uint8) precomputedVertices = createPrecomputedVertices() def __init__(self, level): if not hasattr(alphaMaterials, 'Stone'): get_materials() self.stoneid = stoneid = alphaMaterials.Stone.ID self.hiddenOreMaterials[alphaMaterials.Dirt.ID] = stoneid self.hiddenOreMaterials[alphaMaterials.Grass.ID] = stoneid self.hiddenOreMaterials[alphaMaterials.Sand.ID] = stoneid self.hiddenOreMaterials[alphaMaterials.Gravel.ID] = stoneid self.hiddenOreMaterials[alphaMaterials.Netherrack.ID] = stoneid self.level = level self.makeRenderstates(level.materials) # del xArray, zArray, yArray self.nullVertices = numpy.zeros((0,) * len(self.precomputedVertices[0].shape), dtype=self.precomputedVertices[0].dtype) config.settings.fastLeaves.addObserver(self) config.settings.roughGraphics.addObserver(self) class renderstatePlain(object): @classmethod def bind(cls): pass @classmethod def release(cls): pass class renderstateVines(object): @classmethod def bind(cls): GL.glDisable(GL.GL_CULL_FACE) GL.glEnable(GL.GL_ALPHA_TEST) @classmethod def release(cls): GL.glEnable(GL.GL_CULL_FACE) GL.glDisable(GL.GL_ALPHA_TEST) class renderstateLowDetail(object): @classmethod def bind(cls): GL.glDisable(GL.GL_CULL_FACE) GL.glDisable(GL.GL_TEXTURE_2D) @classmethod def release(cls): GL.glEnable(GL.GL_CULL_FACE) GL.glEnable(GL.GL_TEXTURE_2D) class renderstateAlphaTest(object): @classmethod def bind(cls): GL.glEnable(GL.GL_ALPHA_TEST) @classmethod def release(cls): GL.glDisable(GL.GL_ALPHA_TEST) class _renderstateAlphaBlend(object): @classmethod def bind(cls): GL.glEnable(GL.GL_BLEND) @classmethod def release(cls): GL.glDisable(GL.GL_BLEND) class renderstateWater(_renderstateAlphaBlend): pass class renderstateIce(_renderstateAlphaBlend): pass class renderstateEntity(object): @classmethod def bind(cls): GL.glDisable(GL.GL_DEPTH_TEST) # GL.glDisable(GL.GL_CULL_FACE) GL.glDisable(GL.GL_TEXTURE_2D) GL.glEnable(GL.GL_BLEND) @classmethod def release(cls): GL.glEnable(GL.GL_DEPTH_TEST) # GL.glEnable(GL.GL_CULL_FACE) GL.glEnable(GL.GL_TEXTURE_2D) GL.glDisable(GL.GL_BLEND) renderstates = ( renderstatePlain, renderstateVines, renderstateLowDetail, renderstateAlphaTest, renderstateIce, renderstateWater, renderstateEntity, ) def makeRenderstates(self, materials): self.blockRendererClasses = [ GenericBlockRenderer, LeafBlockRenderer, PlantBlockRenderer, TorchBlockRenderer, WaterBlockRenderer, SlabBlockRenderer, ] if materials.name in ("Alpha", "Pocket"): self.blockRendererClasses += [ RailBlockRenderer, LadderBlockRenderer, SnowBlockRenderer, CarpetBlockRenderer, CactusBlockRenderer, PaneBlockRenderer, CakeBlockRenderer, DaylightBlockRenderer, StandingSignRenderer, WallSignBlockRenderer, LeverBlockRenderer, BedBlockRenderer, EnchantingBlockRenderer, RedstoneBlockRenderer, IceBlockRenderer, DoorRenderer, ButtonRenderer, TrapDoorRenderer, FenceBlockRenderer, FenceGateBlockRenderer, StairBlockRenderer, RepeaterBlockRenderer, VineBlockRenderer, PlateBlockRenderer, EndRodRenderer, # button, floor plate, door -> 1-cube features # lever, sign, wall sign, stairs -> 2-cube features # fence # portal ] self.materialMap = materialMap = numpy.zeros((pymclevel.materials.id_limit,), 'uint8') materialMap[1:] = 1 # generic blocks materialCount = 2 for br in self.blockRendererClasses[1:]: # skip generic blocks # materialMap[br.getBlocktypes(materials)] = materialCount materialMap[br(self).getBlocktypes(materials)] = materialCount br.materialIndex = materialCount materialCount += 1 self.exposedMaterialMap = numpy.array(materialMap) self.addTransparentMaterials(self.exposedMaterialMap, materialCount) def addTransparentMaterials(self, mats, materialCount): logging.debug("renderer::ChunkCalculator: Dynamically adding transparent materials.") for b in self.level.materials: yaml = getattr(b, 'yaml', None) if yaml is not None and yaml.get('opacity', 1) < 1: logging.debug("Adding '%s'" % b) mats[b.ID] = materialCount materialCount += 1 logging.debug("renderer::ChunkCalculator: Transparent materials added.") # don't show boundaries between dirt,grass,sand,gravel,or stone. # This hiddenOreMaterial definition shall be delayed after the level is loaded, in order to get the exact ones from the game versionned data. hiddenOreMaterials = numpy.arange(pymclevel.materials.id_limit, dtype='uint16') roughMaterials = numpy.ones((pymclevel.materials.id_limit,), dtype='uint8') roughMaterials[0] = 0 def calcFacesForChunkRenderer(self, cr): if not cr.invalidLayers: return lod = cr.detailLevel cx, cz = cr.chunkPosition level = cr.renderer.level try: chunk = level.getChunk(cx, cz) except Exception as e: if "Session lock lost" in e.message: yield return logging.warn(u"Error reading chunk: %s", e) traceback.print_exc() yield return yield brs = [] append = brs.append classes = ( TileEntityRenderer, MonsterRenderer, ItemRenderer, TileTicksRenderer, TerrainPopulatedRenderer, ChunkBorderRenderer, LowDetailBlockRenderer, OverheadBlockRenderer, ) existingBlockRenderers = dict(((type(b), b) for b in cr.blockRenderers)) for blockRendererClass in classes: if cr.detailLevel not in blockRendererClass.detailLevels: continue if blockRendererClass.layer not in cr.visibleLayers: continue if blockRendererClass.layer not in cr.invalidLayers: if blockRendererClass in existingBlockRenderers: append(existingBlockRenderers[blockRendererClass]) continue br = blockRendererClass(self) br.detailLevel = cr.detailLevel for _ in br.makeChunkVertices(chunk): yield append(br) blockRenderers = [] # Recalculate high detail blocks if needed, otherwise retain the high detail renderers if lod == 0 and Layer.Blocks in cr.invalidLayers: for _ in self.calcHighDetailFaces(cr, blockRenderers): yield else: blockRenderers.extend(br for br in cr.blockRenderers if not isinstance(br, classes)) # Add the layer renderers blockRenderers.extend(brs) cr.blockRenderers = blockRenderers cr.vertexArraysDone() raise StopIteration @staticmethod def getNeighboringChunks(chunk): cx, cz = chunk.chunkPosition level = chunk.world neighboringChunks = {} for dir, dx, dz in ((pymclevel.faces.FaceXDecreasing, -1, 0), (pymclevel.faces.FaceXIncreasing, 1, 0), (pymclevel.faces.FaceZDecreasing, 0, -1), (pymclevel.faces.FaceZIncreasing, 0, 1)): if not level.containsChunk(cx + dx, cz + dz): neighboringChunks[dir] = pymclevel.infiniteworld.ZeroChunk(level.Height) else: try: neighboringChunks[dir] = level.getChunk(cx + dx, cz + dz) except (EnvironmentError, pymclevel.mclevelbase.ChunkNotPresent, pymclevel.mclevelbase.ChunkMalformed): neighboringChunks[dir] = pymclevel.infiniteworld.ZeroChunk(level.Height) return neighboringChunks @staticmethod def getAreaBlocks(chunk, neighboringChunks): chunkWidth, chunkLength, chunkHeight = chunk.Blocks.shape areaBlocks = numpy.zeros((chunkWidth + 2, chunkLength + 2, chunkHeight + 2), numpy.uint16) areaBlocks[1:-1, 1:-1, 1:-1] = chunk.Blocks zeros = numpy.zeros((16, 16, 128), dtype=areaBlocks.dtype) nb_fxd = neighboringChunks[pymclevel.faces.FaceXDecreasing].Blocks if nb_fxd.shape[2] == chunkHeight / 2: nb_fxd = numpy.concatenate((nb_fxd, zeros), axis=2) areaBlocks[:1, 1:-1, 1:-1] = nb_fxd[-1:, :chunkLength, :chunkHeight] nb_fxi = neighboringChunks[pymclevel.faces.FaceXIncreasing].Blocks if nb_fxi.shape[2] == chunkHeight / 2: nb_fxi = numpy.concatenate((nb_fxi, zeros), axis=2) areaBlocks[-1:, 1:-1, 1:-1] = nb_fxi[:1, :chunkLength, :chunkHeight] nb_fzd = neighboringChunks[pymclevel.faces.FaceZDecreasing].Blocks if nb_fzd.shape[2] == chunkHeight / 2: nb_fzd = numpy.concatenate((nb_fzd, zeros), axis=2) areaBlocks[1:-1, :1, 1:-1] = nb_fzd[:chunkWidth, -1:, :chunkHeight] nb_fzi = neighboringChunks[pymclevel.faces.FaceZIncreasing].Blocks if nb_fzi.shape[2] == chunkHeight / 2: nb_fzi = numpy.concatenate((nb_fzi, zeros), axis=2) areaBlocks[1:-1, -1:, 1:-1] = nb_fzi[:chunkWidth, :1, :chunkHeight] return areaBlocks @staticmethod def getFacingBlockIndices(areaBlocks, areaBlockMats): facingBlockIndices = [None] * 6 exposedFacesX = (areaBlockMats[:-1, 1:-1, 1:-1] != areaBlockMats[1:, 1:-1, 1:-1]) facingBlockIndices[pymclevel.faces.FaceXDecreasing] = exposedFacesX[:-1] facingBlockIndices[pymclevel.faces.FaceXIncreasing] = exposedFacesX[1:] exposedFacesZ = (areaBlockMats[1:-1, :-1, 1:-1] != areaBlockMats[1:-1, 1:, 1:-1]) facingBlockIndices[pymclevel.faces.FaceZDecreasing] = exposedFacesZ[:, :-1] facingBlockIndices[pymclevel.faces.FaceZIncreasing] = exposedFacesZ[:, 1:] exposedFacesY = (areaBlockMats[1:-1, 1:-1, :-1] != areaBlockMats[1:-1, 1:-1, 1:]) facingBlockIndices[pymclevel.faces.FaceYDecreasing] = exposedFacesY[:, :, :-1] facingBlockIndices[pymclevel.faces.FaceYIncreasing] = exposedFacesY[:, :, 1:] return facingBlockIndices def getAreaBlockLights(self, chunk, neighboringChunks): chunkWidth, chunkLength, chunkHeight = chunk.Blocks.shape lights = chunk.BlockLight skyLight = chunk.SkyLight finalLight = self.whiteLight if lights is not None: finalLight = lights if skyLight is not None: finalLight = numpy.maximum(skyLight, lights) areaBlockLights = numpy.ones((chunkWidth + 2, chunkLength + 2, chunkHeight + 2), numpy.uint8) areaBlockLights[:] = 15 areaBlockLights[1:-1, 1:-1, 1:-1] = finalLight zeros = numpy.zeros((16, 16, 128), dtype=areaBlockLights.dtype) skyLight, blockLight = neighboringChunks[pymclevel.faces.FaceXDecreasing].SkyLight, neighboringChunks[pymclevel.faces.FaceXDecreasing].BlockLight if skyLight.shape[2] == chunkHeight / 2: skyLight = numpy.concatenate((skyLight, zeros), axis=2) blockLight = numpy.concatenate((blockLight, zeros), axis=2) numpy.maximum(skyLight[-1:, :chunkLength, :chunkHeight], blockLight[-1:, :chunkLength, :chunkHeight], areaBlockLights[0:1, 1:-1, 1:-1]) skyLight, blockLight = neighboringChunks[pymclevel.faces.FaceXIncreasing].SkyLight, neighboringChunks[pymclevel.faces.FaceXIncreasing].BlockLight if skyLight.shape[2] == chunkHeight / 2: skyLight = numpy.concatenate((skyLight, zeros), axis=2) blockLight = numpy.concatenate((blockLight, zeros), axis=2) numpy.maximum(skyLight[:1, :chunkLength, :chunkHeight], blockLight[:1, :chunkLength, :chunkHeight], areaBlockLights[-1:, 1:-1, 1:-1]) skyLight, blockLight = neighboringChunks[pymclevel.faces.FaceZDecreasing].SkyLight, neighboringChunks[pymclevel.faces.FaceZDecreasing].BlockLight if skyLight.shape[2] == chunkHeight / 2: skyLight = numpy.concatenate((skyLight, zeros), axis=2) blockLight = numpy.concatenate((blockLight, zeros), axis=2) numpy.maximum(skyLight[:chunkWidth, -1:, :chunkHeight], blockLight[:chunkWidth, -1:, :chunkHeight], areaBlockLights[1:-1, 0:1, 1:-1]) skyLight, blockLight = neighboringChunks[pymclevel.faces.FaceZIncreasing].SkyLight, neighboringChunks[pymclevel.faces.FaceZIncreasing].BlockLight if skyLight.shape[2] == chunkHeight / 2: skyLight = numpy.concatenate((skyLight, zeros), axis=2) blockLight = numpy.concatenate((blockLight, zeros), axis=2) numpy.maximum(skyLight[:chunkWidth, :1, :chunkHeight], blockLight[:chunkWidth, :1, :chunkHeight], areaBlockLights[1:-1, -1:, 1:-1]) fxd = neighboringChunks[pymclevel.faces.FaceXDecreasing] fxi = neighboringChunks[pymclevel.faces.FaceXIncreasing] fzd = neighboringChunks[pymclevel.faces.FaceZDecreasing] fzi = neighboringChunks[pymclevel.faces.FaceZIncreasing] fxd_skyLight = fxd.SkyLight fxi_skyLight = fxi.SkyLight fzd_skyLight = fzd.SkyLight fzi_skyLight = fzi.SkyLight fxd_blockLight = fxd.BlockLight fxi_blockLight = fxi.BlockLight fzd_blockLight = fzd.BlockLight fzi_blockLight = fzi.BlockLight if fxd_skyLight.shape[2] == chunkHeight / 2: fxd_skyLight = numpy.concatenate((fxd_skyLight, zeros), axis=2) fxd_blockLight = numpy.concatenate((fxd_blockLight, zeros), axis=2) if fxi_skyLight.shape[2] == chunkHeight / 2: fxi_skyLight = numpy.concatenate((fxi_skyLight, zeros), axis=2) fxi_blockLight = numpy.concatenate((fxi_blockLight, zeros), axis=2) if fzd_skyLight.shape[2] == chunkHeight / 2: fzd_skyLight = numpy.concatenate((fzd_skyLight, zeros), axis=2) fzd_blockLight = numpy.concatenate((fzd_blockLight, zeros), axis=2) if fzi_skyLight.shape[2] == chunkHeight / 2: fzi_skyLight = numpy.concatenate((fzi_skyLight, zeros), axis=2) fzi_blockLight = numpy.concatenate((fzi_blockLight, zeros), axis=2) numpy.maximum(fxd_skyLight[-1:, :chunkLength, :chunkHeight], fxd_blockLight[-1:, :chunkLength, :chunkHeight], areaBlockLights[0:1, 1:-1, 1:-1]) numpy.maximum(fxi_skyLight[:1, :chunkLength, :chunkHeight], fxi_blockLight[:1, :chunkLength, :chunkHeight], areaBlockLights[-1:, 1:-1, 1:-1]) numpy.maximum(fzd_skyLight[:chunkWidth, -1:, :chunkHeight], fzd_blockLight[:chunkWidth, -1:, :chunkHeight], areaBlockLights[1:-1, 0:1, 1:-1]) numpy.maximum(fzi_skyLight[:chunkWidth, :1, :chunkHeight], fzi_blockLight[:chunkWidth, :1, :chunkHeight], areaBlockLights[1:-1, -1:, 1:-1]) minimumLight = 4 numpy.clip(areaBlockLights, minimumLight, 16, areaBlockLights) return areaBlockLights def calcHighDetailFaces(self, cr, blockRenderers): """ calculate the geometry for a chunk renderer from its blockMats, data, and lighting array. fills in the cr's blockRenderers with verts for each block facing and material""" # chunkBlocks and chunkLights shall be indexed [x,z,y] to follow infdev's convention cx, cz = cr.chunkPosition level = cr.renderer.level chunk = level.getChunk(cx, cz) # if isinstance(chunk, pymclevel.level.FakeChunk): # return neighboringChunks = self.getNeighboringChunks(chunk) areaBlocks = self.getAreaBlocks(chunk, neighboringChunks) yield areaBlockLights = self.getAreaBlockLights(chunk, neighboringChunks) yield allSlabs = set([b.ID for b in alphaMaterials.allBlocks if "Slab" in b.name]) for slab in allSlabs: slabs = areaBlocks == slab if slabs.any(): areaBlockLights[slabs] = areaBlockLights[:, :, 1:][slabs[:, :, :-1]] yield showHiddenOres = cr.renderer.showHiddenOres if showHiddenOres: facingMats = self.hiddenOreMaterials[areaBlocks] else: facingMats = self.exposedMaterialMap[areaBlocks] yield if self.roughGraphics: areaBlockMats = self.roughMaterials[areaBlocks] else: areaBlockMats = self.materialMap[areaBlocks] facingBlockIndices = self.getFacingBlockIndices(areaBlocks, facingMats) yield for _ in self.computeGeometry(chunk, areaBlockMats, facingBlockIndices, areaBlockLights, cr, blockRenderers): yield def computeGeometry(self, chunk, areaBlockMats, facingBlockIndices, areaBlockLights, chunkRenderer, blockRenderers): blocks, blockData = chunk.Blocks, chunk.Data blockData &= 0xf blockMaterials = areaBlockMats[1:-1, 1:-1, 1:-1] if self.roughGraphics: blockMaterials.clip(0, 1, blockMaterials) else: # Special case for doors # # Each part of a door itself does not have all of the information required # to render, as direction/whether its open is on the lower part and the hinge # side is on the upper part. So here we combine the metadata of the bottom part # with the top to form 0-32 metadata(which would be used in door renderer). # copied = False for door in DoorRenderer.blocktypes: doors = blocks == door if doors.any(): if not copied: # copy if required but only once blockData = blockData.copy() copied = True # only accept lower part one block below upper part valid = doors[:, :, :-1] & doors[:, :, 1:] & (blockData[:, :, :-1] < 8) & (blockData[:, :, 1:] >= 8) mask = valid.nonzero() upper_mask = (mask[0], mask[1], mask[2]+1) blockData[mask] += (blockData[upper_mask] - 8) * 16 blockData[upper_mask] = blockData[mask] + 8 sx = sz = slice(0, 16) asx = asz = slice(0, 18) for y in xrange(0, chunk.world.Height, 16): sy = slice(y, y + 16) asy = slice(y, y + 18) for _ in self.computeCubeGeometry( y, blockRenderers, blocks[sx, sz, sy], blockData[sx, sz, sy], chunk.materials, blockMaterials[sx, sz, sy], [f[sx, sz, sy] for f in facingBlockIndices], areaBlockLights[asx, asz, asy], chunkRenderer): yield def computeCubeGeometry(self, y, blockRenderers, blocks, blockData, materials, blockMaterials, facingBlockIndices, areaBlockLights, chunkRenderer): materialCounts = numpy.bincount(blockMaterials.ravel()) append = blockRenderers.append def texMap(blocks, blockData=0, direction=slice(None)): return materials.blockTextures[blocks, blockData, direction] # xxx slow for blockRendererClass in self.blockRendererClasses: mi = blockRendererClass.materialIndex if mi >= len(materialCounts) or materialCounts[mi] == 0: continue blockRenderer = blockRendererClass(self) blockRenderer.y = y blockRenderer.materials = materials for _ in blockRenderer.makeVertices(facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): yield append(blockRenderer) yield def makeTemplate(self, direction, blockIndices): return self.precomputedVertices[direction][numpy.where(blockIndices)] class Layer: Blocks = "Blocks" Entities = "Entities" Monsters = "Monsters" Items = "Items" TileEntities = "TileEntities" TileTicks = "TileTicks" TerrainPopulated = "TerrainPopulated" ChunkBorder = "ChunkBorder" AllLayers = (Blocks, Entities, Monsters, Items, TileEntities, TileTicks, TerrainPopulated, ChunkBorder) class BlockRenderer(object): detailLevels = (0,) layer = Layer.Blocks directionOffsets = { pymclevel.faces.FaceXDecreasing: numpy.s_[:-2, 1:-1, 1:-1], pymclevel.faces.FaceXIncreasing: numpy.s_[2:, 1:-1, 1:-1], pymclevel.faces.FaceYDecreasing: numpy.s_[1:-1, 1:-1, :-2], pymclevel.faces.FaceYIncreasing: numpy.s_[1:-1, 1:-1, 2:], pymclevel.faces.FaceZDecreasing: numpy.s_[1:-1, :-2, 1:-1], pymclevel.faces.FaceZIncreasing: numpy.s_[1:-1, 2:, 1:-1], } renderstate = ChunkCalculator.renderstateAlphaTest used = False def __init__(self, cc): self.makeTemplate = cc.makeTemplate self.chunkCalculator = cc self.vertexArrays = [] self.materials = cc.level.materials pass def getBlocktypes(self, mats): return self.blocktypes def setAlpha(self, alpha): "alpha is an unsigned byte value" for a in self.vertexArrays: a.view('uint8')[_RGBA][..., 3] = alpha def bufferSize(self): return sum(a.size for a in self.vertexArrays) * 4 def getMaterialIndices(self, blockMaterials): return blockMaterials == self.materialIndex def makeVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): arrays = [] append = arrays.append materialIndices = self.getMaterialIndices(blockMaterials) yield blockLight = areaBlockLights[1:-1, 1:-1, 1:-1] for (direction, exposedFaceIndices) in enumerate(facingBlockIndices): facingBlockLight = areaBlockLights[self.directionOffsets[direction]] vertexArray = self.makeFaceVertices(direction, materialIndices, exposedFaceIndices, blocks, blockData, blockLight, facingBlockLight, texMap) yield if len(vertexArray): append(vertexArray) self.vertexArrays = arrays def makeArrayList(self, chunkPosition, showRedraw): l = gl.glGenLists(1) GL.glNewList(l, GL.GL_COMPILE) self.drawArrays(chunkPosition, showRedraw) GL.glEndList() return l def drawArrays(self, chunkPosition, showRedraw): cx, cz = chunkPosition y = getattr(self, "y", 0) with gl.glPushMatrix(GL.GL_MODELVIEW): GL.glTranslate(cx << 4, y, cz << 4) if showRedraw: GL.glColor(1.0, 0.25, 0.25, 1.0) self.drawVertices() def drawVertices(self): if self.vertexArrays: for buf in self.vertexArrays: self.drawFaceVertices(buf) def drawFaceVertices(self, buf): if not len(buf): return stride = elementByteLength GL.glVertexPointer(3, GL.GL_FLOAT, stride, (buf.ravel())) GL.glTexCoordPointer(2, GL.GL_FLOAT, stride, (buf.ravel()[3:])) GL.glColorPointer(4, GL.GL_UNSIGNED_BYTE, stride, (buf.view(dtype=numpy.uint8).ravel()[20:])) GL.glDrawArrays(GL.GL_QUADS, 0, len(buf) * 4) class EntityRendererGeneric(BlockRenderer): renderstate = ChunkCalculator.renderstateEntity detailLevels = (0, 1, 2) def drawFaceVertices(self, buf): if not len(buf): return stride = elementByteLength GL.glVertexPointer(3, GL.GL_FLOAT, stride, (buf.ravel())) GL.glTexCoordPointer(2, GL.GL_FLOAT, stride, (buf.ravel()[3:])) GL.glColorPointer(4, GL.GL_UNSIGNED_BYTE, stride, (buf.view(dtype=numpy.uint8).ravel()[20:])) GL.glDepthMask(False) GL.glPolygonMode(GL.GL_FRONT_AND_BACK, GL.GL_LINE) GL.glLineWidth(2.0) GL.glDrawArrays(GL.GL_QUADS, 0, len(buf) * 4) GL.glPolygonMode(GL.GL_FRONT_AND_BACK, GL.GL_FILL) GL.glPolygonOffset(DepthOffset.TerrainWire, DepthOffset.TerrainWire) with gl.glEnable(GL.GL_POLYGON_OFFSET_FILL, GL.GL_DEPTH_TEST): GL.glDrawArrays(GL.GL_QUADS, 0, len(buf) * 4) GL.glDepthMask(True) @staticmethod def _computeVertices(positions, colors, offset=False, chunkPosition=(0, 0)): cx, cz = chunkPosition x = cx << 4 z = cz << 4 vertexArray = numpy.zeros(shape=(len(positions), 6, 4, 6), dtype='float32') if positions: positions = numpy.array(positions) positions[:, (0, 2)] -= (x, z) if offset: positions -= 0.5 vertexArray.view('uint8')[_RGBA] = colors vertexArray[_XYZ] = positions[:, numpy.newaxis, numpy.newaxis, :] vertexArray[_XYZ] += faceVertexTemplates[_XYZ] vertexArray.shape = (len(positions) * 6, 4, 6) return vertexArray class TileEntityRenderer(EntityRendererGeneric): layer = Layer.TileEntities def makeChunkVertices(self, chunk): tilePositions = [] append = tilePositions.append for i, ent in enumerate(chunk.TileEntities): if i % 10 == 0: yield if 'x' not in ent: continue append(pymclevel.TileEntity.pos(ent)) tiles = self._computeVertices(tilePositions, (0xff, 0xff, 0x33, 0x44), chunkPosition=chunk.chunkPosition) yield self.vertexArrays = [tiles] class BaseEntityRenderer(EntityRendererGeneric): pass class MonsterRenderer(BaseEntityRenderer): layer = Layer.Entities # xxx Monsters notMonsters = {"Item", "XPOrb", "Painting", "ItemFrame", "ArmorStand"} def makeChunkVertices(self, chunk): monsterPositions = [] append = monsterPositions.append notMonsters = self.chunkCalculator.level.defsIds.mcedit_defs.get('notMonsters', self.notMonsters) for i, ent in enumerate(chunk.Entities): if i % 10 == 0: yield id = ent["id"].value if id in notMonsters: continue pos = pymclevel.Entity.pos(ent) pos[1] += 0.5 append(pos) monsters = self._computeVertices(monsterPositions, (0xff, 0x22, 0x22, 0x44), offset=True, chunkPosition=chunk.chunkPosition) yield self.vertexArrays = [monsters] class EntityRenderer(BaseEntityRenderer): @staticmethod def makeChunkVertices(chunk): yield class ItemRenderer(BaseEntityRenderer): layer = Layer.Items def makeChunkVertices(self, chunk): entityPositions = [] entityColors = [] colorMap = { "Item": (0x22, 0xff, 0x22, 0x5f), "XPOrb": (0x88, 0xff, 0x88, 0x5f), "Painting": (134, 96, 67, 0x5f), "ItemFrame": (134, 96, 67, 0x5f), "ArmorStand": (0x22, 0xff, 0x22, 0x5f), } pos_append = entityPositions.append color_append = entityColors.append defsIds = self.chunkCalculator.level.defsIds mcedit_defs = defsIds.mcedit_defs mcedit_ids = defsIds.mcedit_ids for i, ent in enumerate(chunk.Entities): if i % 10 == 0: yield # Let get the color from the versionned data, and use the 'old' way as fallback color = mcedit_defs.get(mcedit_ids.get(ent["id"].value), {}).get("mapcolor") if color is None: color = colorMap.get(ent["id"].value) if color is None: continue pos = pymclevel.Entity.pos(ent) noRenderDelta = mcedit_defs.get('noRenderDelta', ("Painting", "ItemFrame")) if ent["id"].value not in noRenderDelta: pos[1] += 0.5 pos_append(pos) color_append(color) entities = self._computeVertices(entityPositions, numpy.array(entityColors, dtype='uint8')[:, numpy.newaxis, numpy.newaxis], offset=True, chunkPosition=chunk.chunkPosition) yield self.vertexArrays = [entities] class TileTicksRenderer(EntityRendererGeneric): layer = Layer.TileTicks def makeChunkVertices(self, chunk): if hasattr(chunk, "TileTicks"): self.vertexArrays.append(self._computeVertices([[tick[j].value for j in "xyz"] for i, tick in enumerate(chunk.TileTicks)], (0xff, 0xff, 0xff, 0x44), chunkPosition=chunk.chunkPosition)) yield class TerrainPopulatedRenderer(EntityRendererGeneric): layer = Layer.TerrainPopulated vertexTemplate = numpy.zeros((6, 4, 6), 'float32') vertexTemplate[_XYZ] = faceVertexTemplates[_XYZ] vertexTemplate[_XYZ] *= (16, 256, 16) color = (255, 200, 155) vertexTemplate.view('uint8')[_RGBA] = color + (72,) def drawFaceVertices(self, buf): if not len(buf): return stride = elementByteLength GL.glVertexPointer(3, GL.GL_FLOAT, stride, (buf.ravel())) GL.glTexCoordPointer(2, GL.GL_FLOAT, stride, (buf.ravel()[3:])) GL.glColorPointer(4, GL.GL_UNSIGNED_BYTE, stride, (buf.view(dtype=numpy.uint8).ravel()[20:])) GL.glDepthMask(False) GL.glDisable(GL.GL_CULL_FACE) with gl.glEnable(GL.GL_DEPTH_TEST): GL.glDrawArrays(GL.GL_QUADS, 0, len(buf) * 4) GL.glEnable(GL.GL_CULL_FACE) GL.glPolygonMode(GL.GL_FRONT_AND_BACK, GL.GL_LINE) GL.glLineWidth(1.0) GL.glDrawArrays(GL.GL_QUADS, 0, len(buf) * 4) GL.glLineWidth(2.0) with gl.glEnable(GL.GL_DEPTH_TEST): GL.glDrawArrays(GL.GL_QUADS, 0, len(buf) * 4) GL.glLineWidth(1.0) GL.glPolygonMode(GL.GL_FRONT_AND_BACK, GL.GL_FILL) GL.glDepthMask(True) def makeChunkVertices(self, chunk): neighbors = self.chunkCalculator.getNeighboringChunks(chunk) def getpop(ch): return getattr(ch, "TerrainPopulated", True) pop = getpop(chunk) yield if pop: return visibleFaces = [ getpop(neighbors[pymclevel.faces.FaceXIncreasing]), getpop(neighbors[pymclevel.faces.FaceXDecreasing]), True, True, getpop(neighbors[pymclevel.faces.FaceZIncreasing]), getpop(neighbors[pymclevel.faces.FaceZDecreasing]), ] visibleFaces = numpy.array(visibleFaces, dtype='bool') verts = self.vertexTemplate[visibleFaces] self.vertexArrays.append(verts) yield class ChunkBorderRenderer(EntityRendererGeneric): layer = Layer.ChunkBorder color = (0, 210, 225) vertexTemplate = numpy.zeros((6, 4, 6), 'float32') vertexTemplate[_XYZ] = faceVertexTemplates[_XYZ] vertexTemplate[_XYZ] *= (16, 256, 16) vertexTemplate.view('uint8')[_RGBA] = color + (150,) def makeChunkVertices(self, chunk): visibleFaces = [ True, True, True, True, True, True, ] yield visibleFaces = numpy.array(visibleFaces, dtype='bool') verts = self.vertexTemplate[visibleFaces] self.vertexArrays.append(verts) yield def drawFaceVertices(self, buf): if not len(buf): return stride = elementByteLength GL.glVertexPointer(3, GL.GL_FLOAT, stride, (buf.ravel())) GL.glTexCoordPointer(2, GL.GL_FLOAT, stride, (buf.ravel()[3:])) GL.glColorPointer(4, GL.GL_UNSIGNED_BYTE, stride, (buf.view(dtype=numpy.uint8).ravel()[20:])) GL.glPolygonMode(GL.GL_FRONT_AND_BACK, GL.GL_LINE) GL.glLineWidth(1) with gl.glEnable(GL.GL_DEPTH_TEST): GL.glDrawArrays(GL.GL_QUADS, 0, len(buf) * 4) GL.glLineWidth(2.0) with gl.glEnable(GL.GL_DEPTH_TEST): GL.glDrawArrays(GL.GL_QUADS, 0, len(buf) * 4) GL.glLineWidth(1.0) GL.glPolygonMode(GL.GL_FRONT_AND_BACK, GL.GL_FILL) class LowDetailBlockRenderer(BlockRenderer): renderstate = ChunkCalculator.renderstateLowDetail detailLevels = (1,) def drawFaceVertices(self, buf): if not len(buf): return stride = 16 GL.glVertexPointer(3, GL.GL_FLOAT, stride, numpy.ravel(buf.ravel())) GL.glColorPointer(4, GL.GL_UNSIGNED_BYTE, stride, (buf.view(dtype='uint8').ravel()[12:])) GL.glDisableClientState(GL.GL_TEXTURE_COORD_ARRAY) GL.glDrawArrays(GL.GL_QUADS, 0, len(buf) * 4) GL.glEnableClientState(GL.GL_TEXTURE_COORD_ARRAY) def setAlpha(self, alpha): for va in self.vertexArrays: va.view('uint8')[..., -1] = alpha def makeChunkVertices(self, ch): step = 1 level = ch.world vertexArrays = [] blocks = ch.Blocks heightMap = ch.HeightMap heightMap = heightMap[::step, ::step] blocks = blocks[::step, ::step] if 0 in blocks.shape: return chunkWidth, chunkLength, chunkHeight = blocks.shape blockIndices = numpy.zeros((chunkWidth, chunkLength, chunkHeight), bool) gridaxes = list(numpy.indices((chunkWidth, chunkLength))) h = numpy.swapaxes(heightMap - 1, 0, 1)[:chunkWidth, :chunkLength] numpy.clip(h, 0, chunkHeight - 1, out=h) gridaxes = (gridaxes[0], gridaxes[1], h) depths = numpy.zeros((chunkWidth, chunkLength), dtype='uint16') depths[1:-1, 1:-1] = reduce(numpy.minimum, (h[1:-1, :-2], h[1:-1, 2:], h[:-2, 1:-1]), h[2:, 1:-1]) yield try: topBlocks = blocks[gridaxes] nonAirBlocks = (topBlocks != 0) blockIndices[gridaxes] = nonAirBlocks h += 1 numpy.clip(h, 0, chunkHeight - 1, out=h) overblocks = blocks[gridaxes][nonAirBlocks].ravel() except ValueError as e: raise ValueError(str(e.args) + "Chunk shape: {0}".format(blockIndices.shape), sys.exc_info()[-1]) if nonAirBlocks.any(): blockTypes = blocks[blockIndices] flatcolors = level.materials.flatColors[blockTypes, ch.Data[blockIndices] & 0xf][:, numpy.newaxis, :] x, z, y = blockIndices.nonzero() yield vertexArray = numpy.zeros((len(x), 4, 4), dtype='float32') vertexArray[_XYZ][..., 0] = x[:, numpy.newaxis] vertexArray[_XYZ][..., 1] = y[:, numpy.newaxis] vertexArray[_XYZ][..., 2] = z[:, numpy.newaxis] va0 = numpy.array(vertexArray) va0[..., :3] += faceVertexTemplates[pymclevel.faces.FaceYIncreasing, ..., :3] overmask = overblocks > 0 flatcolors[overmask] = level.materials.flatColors[:, 0][overblocks[overmask]][:, numpy.newaxis] if self.detailLevel == 2: heightfactor = (y / float(2.0 * ch.world.Height)) + 0.5 flatcolors[..., :3] = flatcolors[..., :3].astype(float) * heightfactor[:, numpy.newaxis, numpy.newaxis] _RGBA = numpy.s_[..., 12:16] va0.view('uint8')[_RGBA] = flatcolors va0[_XYZ][:, :, 0] *= step va0[_XYZ][:, :, 2] *= step yield if self.detailLevel == 2: self.vertexArrays = [va0] return va1 = numpy.array(vertexArray) va1[..., :3] += faceVertexTemplates[pymclevel.faces.FaceXIncreasing, ..., :3] va1[_XYZ][:, (0, 1), 1] = depths[nonAirBlocks].ravel()[:, numpy.newaxis] # stretch to floor va1[_XYZ][:, (1, 2), 0] -= 1.0 # turn diagonally va1[_XYZ][:, (2, 3), 1] -= 0.5 # drop down to prevent intersection pixels va1[_XYZ][:, :, 0] *= step va1[_XYZ][:, :, 2] *= step flatcolors = flatcolors.astype(float) * 0.8 va1.view('uint8')[_RGBA] = flatcolors grassmask = topBlocks[nonAirBlocks] == 2 # color grass sides with dirt's color va1.view('uint8')[_RGBA][grassmask] = level.materials.flatColors[:, 0][[3]][:, numpy.newaxis] va2 = numpy.array(va1) va2[_XYZ][:, (1, 2), 0] += step va2[_XYZ][:, (0, 3), 0] -= step vertexArrays = [va1, va2, va0] self.vertexArrays = vertexArrays class OverheadBlockRenderer(LowDetailBlockRenderer): detailLevels = (2,) class GenericBlockRenderer(BlockRenderer): renderstate = ChunkCalculator.renderstateAlphaTest materialIndex = 1 def makeGenericVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): vertexArrays = [] append = vertexArrays.append materialIndices = self.getMaterialIndices(blockMaterials) yield for (direction, exposedFaceIndices) in enumerate(facingBlockIndices): facingBlockLight = areaBlockLights[self.directionOffsets[direction]] blockIndices = materialIndices & exposedFaceIndices theseBlocks = blocks[blockIndices] bdata = blockData[blockIndices] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(theseBlocks, bdata, direction)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] if self.materials.name in ("Alpha", "Pocket"): if direction == pymclevel.faces.FaceYIncreasing: grass = theseBlocks == alphaMaterials.Grass.ID vertexArray.view('uint8')[_RGB][grass] = vertexArray.view('uint8')[_RGB][grass].astype(float) * self.grassColor yield append(vertexArray) self.vertexArrays = vertexArrays grassColor = grassColorDefault = [0.39, 0.71, 0.23] # 62C743 makeVertices = makeGenericVertices class LeafBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [block.ID for block in mats.blocksByType["LEAVES"]] @property def renderstate(self): if self.chunkCalculator.fastLeaves: return ChunkCalculator.renderstatePlain else: return ChunkCalculator.renderstateAlphaTest def makeLeafVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): arrays = [] append = arrays.append materialIndices = self.getMaterialIndices(blockMaterials) yield if self.materials.name in ("Alpha", "Pocket"): if not self.chunkCalculator.fastLeaves: blockIndices = materialIndices data = blockData[blockIndices] data &= 0x3 # ignore decay states leaves = (data == alphaMaterials.Leaves.blockData) pines = (data == alphaMaterials.PineLeaves.blockData) birches = (data == alphaMaterials.BirchLeaves.blockData) jungle = (data == alphaMaterials.JungleLeaves.blockData) acacia = (data == alphaMaterials.AcaciaLeaves.blockData) darkoak = (data == alphaMaterials.DarkOakLeaves.blockData) texes = texMap(blocks[blockIndices], [0], 0) else: blockIndices = materialIndices texes = texMap(blocks[blockIndices], [0], 0) for (direction, exposedFaceIndices) in enumerate(facingBlockIndices): if self.materials.name in ("Alpha", "Pocket"): if self.chunkCalculator.fastLeaves: blockIndices = materialIndices & exposedFaceIndices data = blockData[blockIndices] data &= 0x3 # ignore decay states leaves = (data == alphaMaterials.Leaves.blockData) pines = (data == alphaMaterials.PineLeaves.blockData) birches = (data == alphaMaterials.BirchLeaves.blockData) jungle = (data == alphaMaterials.JungleLeaves.blockData) acacia = (data == alphaMaterials.AcaciaLeaves.blockData) darkoak = (data == alphaMaterials.DarkOakLeaves.blockData) texes = texMap(blocks[blockIndices], data, 0) facingBlockLight = areaBlockLights[self.directionOffsets[direction]] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texes[:, numpy.newaxis] if not self.chunkCalculator.fastLeaves: vertexArray[_ST] -= (0x10, 0x0) vertexArray.view('uint8')[_RGB] *= facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] if self.materials.name in ("Alpha", "Pocket"): vertexArray.view('uint8')[_RGB][leaves] = vertexArray.view('uint8')[_RGB][leaves].astype(float) * self.leafColor vertexArray.view('uint8')[_RGB][pines] = vertexArray.view('uint8')[_RGB][pines].astype(float) * self.pineLeafColor vertexArray.view('uint8')[_RGB][birches] = vertexArray.view('uint8')[_RGB][birches].astype(float) * self.birchLeafColor vertexArray.view('uint8')[_RGB][jungle] = vertexArray.view('uint8')[_RGB][jungle].astype(float) * self.jungleLeafColor vertexArray.view('uint8')[_RGB][acacia] = vertexArray.view('uint8')[_RGB][acacia].astype(float) * self.acaciaLeafColor vertexArray.view('uint8')[_RGB][darkoak] = vertexArray.view('uint8')[_RGB][darkoak].astype(float) * self.darkoakLeafColor yield append(vertexArray) self.vertexArrays = arrays leafColor = leafColorDefault = [0x48 / 255., 0xb5 / 255., 0x18 / 255.] # 48b518 pineLeafColor = pineLeafColorDefault = [0x61 / 255., 0x99 / 255., 0x61 / 255.] # 0x619961 birchLeafColor = birchLeafColorDefault = [0x80 / 255., 0xa7 / 255., 0x55 / 255.] # 0x80a755 jungleLeafColor = jungleLeafColorDefault = [0x48 / 255., 0xb5 / 255., 0x18 / 255.] # 48b518 acaciaLeafColor = acaciaLeafColorDefault = [0x48 / 255., 0xb5 / 255., 0x18 / 255.] # 48b518 darkoakLeafColor = darkoakLeafColorDefault = [0x48 / 255., 0xb5 / 255., 0x18 / 255.] # 48b518 makeVertices = makeLeafVertices class PlantBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): # blocktypes = [6, 37, 38, 39, 40, 59, 83] # if mats.name != "Classic": blocktypes += [31, 32] # shrubs, tall grass # if mats.name == "Alpha": blocktypes += [115] # nether wart blocktypes = [b.ID for b in mats if b.type in ("DECORATION_CROSS", "NETHER_WART", "CROPS", "STEM")] return blocktypes renderstate = ChunkCalculator.renderstateAlphaTest def makePlantVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): arrays = [] append = arrays.append blockIndices = self.getMaterialIndices(blockMaterials) yield theseBlocks = blocks[blockIndices] bdata = blockData[blockIndices] texes = texMap(blocks[blockIndices], bdata, 0) blockLight = areaBlockLights[1:-1, 1:-1, 1:-1] lights = blockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] colorize = None if self.materials.name != "Classic": #so hacky, someone more competent fix this colorize = (theseBlocks == self.materials.TallGrass.ID) & (bdata != 0) colorize2 = (theseBlocks == self.materials.TallFlowers.ID) & (bdata != 0) & ( bdata != 1) & (bdata != 4) & (bdata != 5) for direction in ( pymclevel.faces.FaceXIncreasing, pymclevel.faces.FaceXDecreasing, pymclevel.faces.FaceZIncreasing, pymclevel.faces.FaceZDecreasing): vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): return if direction == pymclevel.faces.FaceXIncreasing: vertexArray[_XYZ][..., 1:3, 0] -= 1 if direction == pymclevel.faces.FaceXDecreasing: vertexArray[_XYZ][..., 1:3, 0] += 1 if direction == pymclevel.faces.FaceZIncreasing: vertexArray[_XYZ][..., 1:3, 2] -= 1 if direction == pymclevel.faces.FaceZDecreasing: vertexArray[_XYZ][..., 1:3, 2] += 1 vertexArray[_ST] += texes[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] = 0xf # ignore precomputed directional light vertexArray.view('uint8')[_RGB] *= lights if colorize is not None: vertexArray.view('uint8')[_RGB][colorize] = vertexArray.view('uint8')[_RGB][colorize].astype(float) * LeafBlockRenderer.leafColor vertexArray.view('uint8')[_RGB][colorize2] = vertexArray.view('uint8')[_RGB][colorize2].astype(float) * LeafBlockRenderer.leafColor append(vertexArray) yield self.vertexArrays = arrays makeVertices = makePlantVertices class TorchBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [block.ID for block in mats.blocksByType["TORCH"]] renderstate = ChunkCalculator.renderstateAlphaTest torchOffsetsStraight = [ [ # FaceXIncreasing (-7 / 16., 0, 0), (-7 / 16., 0, 0), (-7 / 16., 0, 0), (-7 / 16., 0, 0), ], [ # FaceXDecreasing (7 / 16., 0, 0), (7 / 16., 0, 0), (7 / 16., 0, 0), (7 / 16., 0, 0), ], [ # FaceYIncreasing (7 / 16., -6 / 16., 7 / 16.), (7 / 16., -6 / 16., -7 / 16.), (-7 / 16., -6 / 16., -7 / 16.), (-7 / 16., -6 / 16., 7 / 16.), ], [ # FaceYDecreasing (7 / 16., 0., 7 / 16.), (-7 / 16., 0., 7 / 16.), (-7 / 16., 0., -7 / 16.), (7 / 16., 0., -7 / 16.), ], [ # FaceZIncreasing (0, 0, -7 / 16.), (0, 0, -7 / 16.), (0, 0, -7 / 16.), (0, 0, -7 / 16.) ], [ # FaceZDecreasing (0, 0, 7 / 16.), (0, 0, 7 / 16.), (0, 0, 7 / 16.), (0, 0, 7 / 16.) ], ] torchOffsetsSouth = [ [ # FaceXIncreasing (-7 / 16., 3 / 16., 0), (-7 / 16., 3 / 16., 0), (-7 / 16., 3 / 16., 0), (-7 / 16., 3 / 16., 0), ], [ # FaceXDecreasing (7 / 16., 3 / 16., 0), (7 / 16., 3 / 16., 0), (7 / 16., 3 / 16., 0), (7 / 16., 3 / 16., 0), ], [ # FaceYIncreasing (7 / 16., -3 / 16., 7 / 16.), (7 / 16., -3 / 16., -7 / 16.), (-7 / 16., -3 / 16., -7 / 16.), (-7 / 16., -3 / 16., 7 / 16.), ], [ # FaceYDecreasing (7 / 16., 3 / 16., 7 / 16.), (-7 / 16., 3 / 16., 7 / 16.), (-7 / 16., 3 / 16., -7 / 16.), (7 / 16., 3 / 16., -7 / 16.), ], [ # FaceZIncreasing (0, 3 / 16., -7 / 16.), (0, 3 / 16., -7 / 16.), (0, 3 / 16., -7 / 16.), (0, 3 / 16., -7 / 16.) ], [ # FaceZDecreasing (0, 3 / 16., 7 / 16.), (0, 3 / 16., 7 / 16.), (0, 3 / 16., 7 / 16.), (0, 3 / 16., 7 / 16.), ], ] torchOffsetsNorth = torchOffsetsWest = torchOffsetsEast = torchOffsetsSouth torchOffsets = [ torchOffsetsStraight, torchOffsetsSouth, torchOffsetsNorth, torchOffsetsWest, torchOffsetsEast, torchOffsetsStraight, ] + [torchOffsetsStraight] * 10 torchOffsets = numpy.array(torchOffsets, dtype='float32') torchOffsets[1][..., 3, :, 0] -= 0.5 torchOffsets[1][..., 0:2, 0:2, 0] -= 0.5 torchOffsets[1][..., 4:6, 0:2, 0] -= 0.5 torchOffsets[1][..., 0:2, 2:4, 0] -= 0.1 torchOffsets[1][..., 4:6, 2:4, 0] -= 0.1 torchOffsets[1][..., 2, :, 0] -= 0.25 torchOffsets[2][..., 3, :, 0] += 0.5 torchOffsets[2][..., 0:2, 0:2, 0] += 0.5 torchOffsets[2][..., 4:6, 0:2, 0] += 0.5 torchOffsets[2][..., 0:2, 2:4, 0] += 0.1 torchOffsets[2][..., 4:6, 2:4, 0] += 0.1 torchOffsets[2][..., 2, :, 0] += 0.25 torchOffsets[3][..., 3, :, 2] -= 0.5 torchOffsets[3][..., 0:2, 0:2, 2] -= 0.5 torchOffsets[3][..., 4:6, 0:2, 2] -= 0.5 torchOffsets[3][..., 0:2, 2:4, 2] -= 0.1 torchOffsets[3][..., 4:6, 2:4, 2] -= 0.1 torchOffsets[3][..., 2, :, 2] -= 0.25 torchOffsets[4][..., 3, :, 2] += 0.5 torchOffsets[4][..., 0:2, 0:2, 2] += 0.5 torchOffsets[4][..., 4:6, 0:2, 2] += 0.5 torchOffsets[4][..., 0:2, 2:4, 2] += 0.1 torchOffsets[4][..., 4:6, 2:4, 2] += 0.1 torchOffsets[4][..., 2, :, 2] += 0.25 upCoords = ((7, 6), (7, 8), (9, 8), (9, 6)) downCoords = ((7, 14), (7, 16), (9, 16), (9, 14)) def makeTorchVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): blockIndices = self.getMaterialIndices(blockMaterials) torchOffsets = self.torchOffsets[blockData[blockIndices]] texes = texMap(blocks[blockIndices], blockData[blockIndices]) yield arrays = [] append = arrays.append for direction in xrange(6): vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): return vertexArray.view('uint8')[_RGBA] = 0xff vertexArray[_XYZ] += torchOffsets[:, direction] if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_ST] = self.upCoords if direction == pymclevel.faces.FaceYDecreasing: vertexArray[_ST] = self.downCoords vertexArray[_ST] += texes[:, numpy.newaxis, direction] append(vertexArray) yield self.vertexArrays = arrays makeVertices = makeTorchVertices class LeverBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:lever"].ID] leverBaseTemplate = makeVertexTemplatesFromJsonModel((5, 0, 4), (11, 3, 12), { "down": (10, 0, 16, 8), "up": (10, 0, 16, 8), "north": (10, 8, 16, 11), "south": (10, 8, 16, 11), "west": (2, 0, 10, 3), "east": (2, 0, 10, 3) }) leverBaseTemplates = numpy.array([ rotateTemplate(leverBaseTemplate, x=180, y=90), rotateTemplate(leverBaseTemplate, x=90, y=90), rotateTemplate(leverBaseTemplate, x=90, y=270), rotateTemplate(leverBaseTemplate, x=90, y=180), rotateTemplate(leverBaseTemplate, x=270, y=180), leverBaseTemplate, rotateTemplate(leverBaseTemplate, y=90), rotateTemplate(leverBaseTemplate, x=180), rotateTemplate(leverBaseTemplate, x=180, y=90), rotateTemplate(leverBaseTemplate, x=90, y=90), rotateTemplate(leverBaseTemplate, x=270, y=90), rotateTemplate(leverBaseTemplate, x=270), rotateTemplate(leverBaseTemplate, x=270, y=180), leverBaseTemplate, rotateTemplate(leverBaseTemplate, y=90), numpy.zeros((6, 4, 6)), numpy.zeros((6, 4, 6)) ]) leverTemplate = makeVertexTemplatesFromJsonModel((7, 1, 7), (9, 11, 9), { "down": (7, 6, 9, 8), "up": (7, 6, 9, 8), "north": (7, 6, 9, 16), "south": (7, 6, 9, 16), "west": (7, 6, 9, 16), "east": (7, 6, 9, 16) }) leverTemplates = numpy.array([ rotateTemplate(leverTemplate, x=180), rotateTemplate(leverTemplate, x=90, y=90), rotateTemplate(leverTemplate, x=90, y=270), rotateTemplate(leverTemplate, x=90, y=180), rotateTemplate(leverTemplate, x=270, y=180), leverTemplate, rotateTemplate(leverTemplate, y=90), rotateTemplate(leverTemplate, x=180), rotateTemplate(leverTemplate, x=180), rotateTemplate(leverTemplate, x=90, y=90), rotateTemplate(leverTemplate, x=270, y=90), rotateTemplate(leverTemplate, x=270), rotateTemplate(leverTemplate, x=270, y=180), leverTemplate, leverTemplate, numpy.zeros((6, 4, 6)), numpy.zeros((6, 4, 6)) ]) makeVertices = makeVerticesFromModel([leverBaseTemplates, leverTemplates], 15) class RailBlockRenderer(BlockRenderer): renderstate = ChunkCalculator.renderstateAlphaTest def __init__(self, *args, **kwargs): BlockRenderer.__init__(self, *args, **kwargs) self.railTextures = numpy.array([ [(0, 128), (0, 144), (16, 144), (16, 128)], # east-west [(0, 128), (16, 128), (16, 144), (0, 144)], # north-south [(0, 128), (16, 128), (16, 144), (0, 144)], # south-ascending [(0, 128), (16, 128), (16, 144), (0, 144)], # north-ascending [(0, 128), (0, 144), (16, 144), (16, 128)], # east-ascending [(0, 128), (0, 144), (16, 144), (16, 128)], # west-ascending [(0, 112), (0, 128), (16, 128), (16, 112)], # northeast corner [(0, 128), (16, 128), (16, 112), (0, 112)], # southeast corner [(16, 128), (16, 112), (0, 112), (0, 128)], # southwest corner [(16, 112), (0, 112), (0, 128), (16, 128)], # northwest corner [(0, 192), (0, 208), (16, 208), (16, 192)], # unknown [(0, 192), (0, 208), (16, 208), (16, 192)], # unknown [(0, 192), (0, 208), (16, 208), (16, 192)], # unknown [(0, 192), (0, 208), (16, 208), (16, 192)], # unknown [(0, 192), (0, 208), (16, 208), (16, 192)], # unknown [(0, 192), (0, 208), (16, 208), (16, 192)], # unknown ], dtype='float32') self.railTextures -= self.materials.blockTextures[self.materials.Rail.ID, 0, 0] @classmethod def getBlocktypes(cls, mats): return [block.ID for block in mats.blocksByType["SIMPLE_RAIL"]] railOffsets = numpy.array([ [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 1, 1], # south-ascending [1, 1, 0, 0], # north-ascending [1, 0, 0, 1], # east-ascending [0, 1, 1, 0], # west-ascending [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], ], dtype='float32') def makeRailVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): direction = pymclevel.faces.FaceYIncreasing blockIndices = self.getMaterialIndices(blockMaterials) yield bdata = blockData[blockIndices] railBlocks = blocks[blockIndices] tex = texMap(railBlocks, bdata, pymclevel.faces.FaceYIncreasing)[:, numpy.newaxis, :] # disable 'powered' or 'pressed' bit for powered and detector rails bdata[railBlocks != self.materials.Rail.ID] = bdata[railBlocks != self.materials.Rail.ID].astype(int) & ~0x8 vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): return vertexArray[_ST] = self.railTextures[bdata] vertexArray[_ST] += tex vertexArray[_XYZ][..., 1] -= 0.9 vertexArray[_XYZ][..., 1] += self.railOffsets[bdata] blockLight = areaBlockLights[1:-1, 1:-1, 1:-1] vertexArray.view('uint8')[_RGB] *= blockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] yield self.vertexArrays = [vertexArray] makeVertices = makeRailVertices class LadderBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:ladder"].ID] ladderOffsets = numpy.array([ [(0, 0, 0), (0, 0, 0), (0, 0, 0), (0, 0, 0)], [(0, 0, 0), (0, 0, 0), (0, 0, 0), (0, 0, 0)], [(0, -1, 0.9), (0, 0, -0.1), (0, 0, -0.1), (0, -1, 0.9)], # facing east [(0, 0, 0.1), (0, -1, -.9), (0, -1, -.9), (0, 0, 0.1)], # facing west [(.9, -1, 0), (.9, -1, 0), (-.1, 0, 0), (-.1, 0, 0)], # north [(0.1, 0, 0), (0.1, 0, 0), (-.9, -1, 0), (-.9, -1, 0)], # south ] + [[(0, 0, 0), (0, 0, 0), (0, 0, 0), (0, 0, 0)]] * 10, dtype='float32') ladderTextures = numpy.array([ [(0, 192), (0, 208), (16, 208), (16, 192)], # unknown [(0, 192), (0, 208), (16, 208), (16, 192)], # unknown [(64, 96), (64, 80), (48, 80), (48, 96), ], # e [(48, 80), (48, 96), (64, 96), (64, 80), ], # w [(48, 96), (64, 96), (64, 80), (48, 80), ], # n [(64, 80), (48, 80), (48, 96), (64, 96), ], # s ] + [[(0, 192), (0, 208), (16, 208), (16, 192)]] * 10, dtype='float32') def ladderVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): blockIndices = self.getMaterialIndices(blockMaterials) blockLight = areaBlockLights[1:-1, 1:-1, 1:-1] yield bdata = blockData[blockIndices] vertexArray = self.makeTemplate(pymclevel.faces.FaceYIncreasing, blockIndices) if not len(vertexArray): return vertexArray[_ST] = self.ladderTextures[bdata] vertexArray[_XYZ] += self.ladderOffsets[bdata] vertexArray.view('uint8')[_RGB] *= blockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] yield self.vertexArrays = [vertexArray] makeVertices = ladderVertices class WallSignBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:wall_sign"].ID] wallSignTemplate = makeVertexTemplatesFromJsonModel((0, 4.5, 0), (16, 13.5, 2), { "down": (0, 11, 18, 13), "up": (0, 6, 16, 8), "north": (0, 4, 16, 13), "south": (0, 4, 16, 13), "west": (0, 4, 2, 13), "east": (10, 4, 12, 13) }) # I don't know how this sytem works and how it should be structured, but this seem to do the job wallSignTemplates = numpy.array([ wallSignTemplate, wallSignTemplate, rotateTemplate(wallSignTemplate, y=180), wallSignTemplate, rotateTemplate(wallSignTemplate, y=90), rotateTemplate(wallSignTemplate, y=270), numpy.zeros((6, 4, 6)), numpy.zeros((6, 4, 6)) ]) makeVertices = makeVerticesFromModel(wallSignTemplates, 7) class StandingSignRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:standing_sign"].ID] signTemplate = makeVertexTemplatesFromJsonModel((0, 7, 7), (16, 16, 9), { "down": (0, 14, 16, 16), "up": (0, 12, 16, 14), "north": (0, 7, 16, 16), "south": (0, 7, 16, 16), "west": (0, 7, 2, 16), "east": (14, 7, 16, 16) }) signTemplates = numpy.array([ signTemplate, numpy.zeros((6, 4, 6)), numpy.zeros((6, 4, 6)) ]) postTemplate = makeVertexTemplatesFromJsonModel((7, 0, 7), (9, 7, 9), { "down": (7, 0, 9, 6), "up": (7, 0, 9, 6), "north": (7, 0, 9, 6), "south": (7, 0, 9, 6), "west": (7, 0, 9, 6), "east": (7, 0, 9, 6), }) postTemplates = numpy.array([ postTemplate, numpy.zeros((6, 4, 6)), numpy.zeros((6, 4, 6)) ]) makeVertices = makeVerticesFromModel([signTemplates, postTemplates]) class SnowBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:snow_layer"].ID] def makeSnowVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) #snowIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): if direction != pymclevel.faces.FaceYIncreasing: blockIndices = materialIndices & exposedFaceIndices else: blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], 0)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_XYZ][..., 1] -= 0.875 if direction != pymclevel.faces.FaceYIncreasing and direction != pymclevel.faces.FaceYDecreasing: vertexArray[_XYZ][..., 2:4, 1] -= 0.875 vertexArray[_ST][..., 2:4, 1] += 14 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makeSnowVertices class CarpetBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:carpet"].ID, mats["minecraft:waterlily"].ID] #Separate before implementing layers def makeCarpetVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) #snowIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): if direction != pymclevel.faces.FaceYIncreasing: blockIndices = materialIndices & exposedFaceIndices else: blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], 0)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_XYZ][..., 1] -= 0.937 if direction != pymclevel.faces.FaceYIncreasing and direction != pymclevel.faces.FaceYDecreasing: vertexArray[_XYZ][..., 2:4, 1] -= 0.937 vertexArray[_ST][..., 2:4, 1] += 15 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makeCarpetVertices class CactusBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:cactus"].ID] def makeCactusVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], direction)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceXIncreasing: vertexArray[_XYZ][..., 0] -= 0.063 if direction == pymclevel.faces.FaceXDecreasing: vertexArray[_XYZ][..., 0] += 0.063 if direction == pymclevel.faces.FaceZIncreasing: vertexArray[_XYZ][..., 2] -= 0.063 if direction == pymclevel.faces.FaceZDecreasing: vertexArray[_XYZ][..., 2] += 0.063 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makeCactusVertices class PaneBlockRenderer(BlockRenderer): #Basic no thickness panes, add more faces to widen. @classmethod def getBlocktypes(cls, mats): return [block.ID for block in mats.blocksByType["SOLID_PANE"]] def makePaneVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], direction)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceXIncreasing: vertexArray[_XYZ][..., 0] -= 0.5 if direction == pymclevel.faces.FaceXDecreasing: vertexArray[_XYZ][..., 0] += 0.5 if direction == pymclevel.faces.FaceZIncreasing: vertexArray[_XYZ][..., 2] -= 0.5 if direction == pymclevel.faces.FaceZDecreasing: vertexArray[_XYZ][..., 2] += 0.5 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makePaneVertices class PlateBlockRenderer(BlockRenderer): #suggestions to make this the proper shape is appreciated. @classmethod def getBlocktypes(cls, mats): return [block.ID for block in mats.blocksByType["PRESSURE_PLATE"]] def makePlateVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], 0)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_XYZ][..., 1] -= 0.937 if direction != pymclevel.faces.FaceYIncreasing and direction != pymclevel.faces.FaceYDecreasing: vertexArray[_XYZ][..., 2:4, 1] -= 0.937 vertexArray[_ST][..., 2:4, 1] += 15 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makePlateVertices class EnchantingBlockRenderer( BlockRenderer): #Note: Enderportal frame side sprite has been lowered 1 pixel to use this renderer, will need separate renderer for eye. @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:enchanting_table"].ID, mats["minecraft:end_portal_frame"].ID] def makeEnchantingVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): if direction != pymclevel.faces.FaceYIncreasing: blockIndices = materialIndices & exposedFaceIndices else: blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], direction)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_XYZ][..., 1] -= 0.25 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makeEnchantingVertices class DaylightBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:daylight_detector"].ID, mats.DaylightSensorOn.ID] def makeDaylightVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): if direction != pymclevel.faces.FaceYIncreasing: blockIndices = materialIndices & exposedFaceIndices else: blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], direction)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_XYZ][..., 1] -= 0.625 if direction != pymclevel.faces.FaceYIncreasing and direction != pymclevel.faces.FaceYDecreasing: vertexArray[_XYZ][..., 2:4, 1] -= 0.625 vertexArray[_ST][..., 2:4, 1] += 10 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makeDaylightVertices class BedBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:bed"].ID] def makeBedVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): if direction != pymclevel.faces.FaceYIncreasing: blockIndices = materialIndices & exposedFaceIndices else: blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], direction)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_XYZ][..., 1] -= 0.438 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makeBedVertices class CakeBlockRenderer(BlockRenderer): #Only shows whole cakes @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:cake"].ID] def makeCakeVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], direction)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_XYZ][..., 1] -= 0.5 if direction == pymclevel.faces.FaceXIncreasing: vertexArray[_XYZ][..., 0] -= 0.063 if direction == pymclevel.faces.FaceXDecreasing: vertexArray[_XYZ][..., 0] += 0.063 if direction == pymclevel.faces.FaceZIncreasing: vertexArray[_XYZ][..., 2] -= 0.063 if direction == pymclevel.faces.FaceZDecreasing: vertexArray[_XYZ][..., 2] += 0.063 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makeCakeVertices class RepeaterBlockRenderer(BlockRenderer): #Sticks would be nice @classmethod def getBlocktypes(cls, mats): return [block.ID for block in mats.blocksByType["THINSLICE"]] def makeRepeaterVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): materialIndices = self.getMaterialIndices(blockMaterials) arrays = [] append = arrays.append yield for direction, exposedFaceIndices in enumerate(facingBlockIndices): blockIndices = materialIndices facingBlockLight = areaBlockLights[self.directionOffsets[direction]] lights = facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): continue vertexArray[_ST] += texMap(blocks[blockIndices], blockData[blockIndices], direction)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_XYZ][..., 1] -= 0.875 if direction != pymclevel.faces.FaceYIncreasing and direction != pymclevel.faces.FaceYDecreasing: vertexArray[_XYZ][..., 2:4, 1] -= 0.875 vertexArray[_ST][..., 2:4, 1] += 14 append(vertexArray) yield self.vertexArrays = arrays makeVertices = makeRepeaterVertices class RedstoneBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [mats["minecraft:redstone_wire"].ID] def redstoneVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): blockIndices = self.getMaterialIndices(blockMaterials) yield vertexArray = self.makeTemplate(pymclevel.faces.FaceYIncreasing, blockIndices) if not len(vertexArray): return vertexArray[_ST] += self.materials.blockTextures[55, 0, 0] vertexArray[_XYZ][..., 1] -= 0.9 bdata = blockData[blockIndices] bdata <<= 3 # bdata &= 0xe0 bdata[bdata > 0] |= 0x80 vertexArray.view('uint8')[_RGBA][..., 0] = bdata[..., numpy.newaxis] vertexArray.view('uint8')[_RGBA][..., 0:3] = vertexArray.view('uint8')[_RGBA][..., 0:3] * [1, 0, 0] yield self.vertexArrays = [vertexArray] makeVertices = redstoneVertices # button, floor plate, door -> 1-cube features class DoorRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): cls.blocktypes = [block.ID for block in mats.blocksByType["DOOR"]] return cls.blocktypes doorTemplate = makeVertexTemplatesFromJsonModel( (0, 0, 0), (3, 16, 16), { "down": (13, 0, 16, 16), # TODO handle faces that should not appear "up": (13, 0, 16, 16), "north": (3, 0, 0, 16), "south": (0, 0, 3, 16), "west": (0, 0, 16, 16), "east": (16, 0, 0, 16) } ) doorRHTemplate = makeVertexTemplatesFromJsonModel( (0, 0, 0), (3, 16, 16), { "down": (13, 0, 16, 16), # TODO handle faces that should not appear "up": (13, 0, 16, 16), "north": (3, 0, 0, 16), "south": (0, 0, 3, 16), "west": (16, 0, 0, 16), "east": (0, 0, 16, 16) } ) doorTemplates = numpy.array([ # lower hinge left doorTemplate, rotateTemplate(doorTemplate, y=90), rotateTemplate(doorTemplate, y=180), rotateTemplate(doorTemplate, y=270), rotateTemplate(doorRHTemplate, y=90), rotateTemplate(doorRHTemplate, y=180), rotateTemplate(doorRHTemplate, y=270), doorRHTemplate, # upper hinge left doorTemplate, rotateTemplate(doorTemplate, y=90), rotateTemplate(doorTemplate, y=180), rotateTemplate(doorTemplate, y=270), rotateTemplate(doorRHTemplate, y=90), rotateTemplate(doorRHTemplate, y=180), rotateTemplate(doorRHTemplate, y=270), doorRHTemplate, # lower hinge right doorRHTemplate, rotateTemplate(doorRHTemplate, y=90), rotateTemplate(doorRHTemplate, y=180), rotateTemplate(doorRHTemplate, y=270), rotateTemplate(doorTemplate, y=270), doorTemplate, rotateTemplate(doorTemplate, y=90), rotateTemplate(doorTemplate, y=180), # upper hinge right doorRHTemplate, rotateTemplate(doorRHTemplate, y=90), rotateTemplate(doorRHTemplate, y=180), rotateTemplate(doorRHTemplate, y=270), rotateTemplate(doorTemplate, y=270), doorTemplate, rotateTemplate(doorTemplate, y=90), rotateTemplate(doorTemplate, y=180), ]) makeVertices = makeVerticesFromModel(doorTemplates, 31) class ButtonRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [a.ID for a in mats.blocksByType["BUTTON"]] buttonTemplate = makeVertexTemplatesFromJsonModel((5, 0, 6), (11, 2, 10), { "down": (5, 6, 11, 10), "up": (5, 10, 11, 6), "north": (5, 14, 11, 16), "south": (5, 14, 11, 16), "west": (6, 14, 10, 16), "east": (6, 14, 10, 16) }) buttonTemplatePressed = makeVertexTemplatesFromJsonModel((5, 0, 6), (11, 1, 10), { "down": (5, 6, 11, 10), "up": (5, 10, 11, 6), "north": (5, 15, 11, 16), "south": (5, 15, 11, 16), "west": (6, 15, 10, 16), "east": (6, 15, 10, 16) }) buttonTemplates = numpy.array([ rotateTemplate(buttonTemplate, 180, 0), rotateTemplate(buttonTemplate, 90, 90), rotateTemplate(buttonTemplate, 90, 270), rotateTemplate(buttonTemplate, 90, 180), rotateTemplate(buttonTemplate, 90, 0), buttonTemplate, numpy.zeros((6, 4, 6)), numpy.zeros((6, 4, 6)), rotateTemplate(buttonTemplatePressed, 180, 0), rotateTemplate(buttonTemplatePressed, 90, 90), rotateTemplate(buttonTemplatePressed, 90, 270), rotateTemplate(buttonTemplatePressed, 90, 180), rotateTemplate(buttonTemplatePressed, 90, 0), buttonTemplatePressed, numpy.zeros((6, 4, 6)), numpy.zeros((6, 4, 6)), ]) makeVertices = makeVerticesFromModel(buttonTemplates, 15) class TrapDoorRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [a.ID for a in mats.blocksByType["TRAPDOOR"]] openTemplate = makeVertexTemplatesFromJsonModel((0, 0, 13), (16, 16, 16), { "down": (0, 13, 16, 16), "up": (0, 16, 16, 13), "north": (0, 0, 16, 16), "south": (0, 0, 16, 16), "west": (16, 0, 13, 16), "east": (13, 0, 16, 16) }) topTemplate = makeVertexTemplatesFromJsonModel((0, 13, 0), (16, 16, 16), { "down": (0, 0, 16, 16), "up": (0, 0, 16, 16), "north": (0, 16, 16, 13), "south": (0, 16, 16, 13), "west": (0, 16, 16, 13), "east": (0, 16, 16, 13) }) bottomTemplate = makeVertexTemplatesFromJsonModel((0, 0, 0), (16, 3, 16), { "down": (0, 0, 16, 16), "up": (0, 0, 16, 16), "north": (0, 16, 16, 13), "south": (0, 16, 16, 13), "west": (0, 16, 16, 13), "east": (0, 16, 16, 13) }) trapDoorTemplates = numpy.array([ bottomTemplate, bottomTemplate, bottomTemplate, bottomTemplate, openTemplate, rotateTemplate(openTemplate, y=180), rotateTemplate(openTemplate, y=270), rotateTemplate(openTemplate, y=90), topTemplate, topTemplate, topTemplate, topTemplate, openTemplate, rotateTemplate(openTemplate, y=180), rotateTemplate(openTemplate, y=270), rotateTemplate(openTemplate, y=90), ]) makeVertices = makeVerticesFromModel(trapDoorTemplates, 15) class FenceBlockRenderer(BlockRenderer): # def __init__(self, *args, **kwargs): # BlockRenderer.__init__(self, *args, **kwargs) # self.blocktypes = [block.ID for block in self.materials.blocksByType["FENCE"]] fenceTemplates = makeVertexTemplates(3 / 8., 0, 3 / 8., 5 / 8., 1, 5 / 8.) makeVertices = makeVerticesFromModel(fenceTemplates) @classmethod def getBlocktypes(cls, mats): # if mats.name == "Pocket": # cls.blocktypes = cls.blocktypes_pocket # else: # cls.blocktypes = cls.blocktypes_alpha # return cls.blocktypes return [block.ID for block in mats.blocksByType["FENCE"]] class FenceGateBlockRenderer(BlockRenderer): closedFenceTemplates = numpy.array([ makeVertexTemplates(0, 0, 3 / 8., 1, .8, 5 / 8.), makeVertexTemplates(3 / 8., 0, 0, 5 / 8., .8, 1)]) openFenceTemplates = numpy.array([ [makeVertexTemplates(0, 0, 3 / 8., 1 / 8., .8, 1), makeVertexTemplates(7 / 8., 0, 3 / 8., 1, .8, 1)], [makeVertexTemplates(0, 0, 0, 5 / 8., .8, 1 / 8.), makeVertexTemplates(0, 0, 7 / 8., 5 / 8., .8, 1)], [makeVertexTemplates(0, 0, 0, 1 / 8., .8, 5 / 8.), makeVertexTemplates(7 / 8., 0, 0, 1, .8, 5 / 8.)], [makeVertexTemplates(3 / 8., 0, 0, 1, .8, 1 / 8.), makeVertexTemplates(3 / 8., 0, 7 / 8., 1, .8, 1)]]) @classmethod def getBlocktypes(cls, mats): return [a.ID for a in mats.AllStairs] def fenceGateVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): fenceMask = self.getMaterialIndices(blockMaterials) closedGateMask = fenceMask.copy() closedGateMask[blockData & 4 == 4] = 0 openGateMask = fenceMask.copy() openGateMask[blockData & 4 == 0] = 0 closedGateIndices = closedGateMask.nonzero() openGateIndices = openGateMask.nonzero() closedGateData = blockData[closedGateMask] closedGateData &= 1 openGateData = blockData[openGateMask] openGateData &= 3 yield # closed gate vertexArray = numpy.zeros((len(closedGateIndices[0]), 6, 4, 6), dtype='float32') for indicies in xrange(3): dimension = (0, 2, 1)[indicies] vertexArray[..., indicies] = closedGateIndices[dimension][:, numpy.newaxis, numpy.newaxis] # xxx swap z with y using ^ vertexArray[..., 0:5] += self.closedFenceTemplates[closedGateData][..., 0:5] vertexArray[_ST] += texMap(blocks[closedGateIndices], 0)[..., numpy.newaxis, :] vertexArray.view('uint8')[_RGB] = self.closedFenceTemplates[closedGateData][..., 5][..., numpy.newaxis] vertexArray.view('uint8')[_A] = 0xFF vertexArray.view('uint8')[_RGB] *= areaBlockLights[1:-1, 1:-1, 1:-1][closedGateIndices][ ..., numpy.newaxis, numpy.newaxis, numpy.newaxis] vertexArray.shape = (vertexArray.shape[0] * 6, 4, 6) yield self.vertexArrays = [vertexArray] append = self.vertexArrays.append # open gate for i in xrange(2): vertexArray = numpy.zeros((len(openGateIndices[0]), 6, 4, 6), dtype='float32') for indicies in xrange(3): dimension = (0, 2, 1)[indicies] vertexArray[..., indicies] = openGateIndices[dimension][:, numpy.newaxis, numpy.newaxis] # xxx swap z with y using ^ vertexArray[..., 0:5] += self.openFenceTemplates[openGateData, i][..., 0:5] vertexArray[_ST] += texMap(blocks[openGateIndices], 0)[..., numpy.newaxis, :] vertexArray.view('uint8')[_RGB] = self.openFenceTemplates[openGateData, i] \ [..., 5][..., numpy.newaxis] vertexArray.view('uint8')[_A] = 0xFF vertexArray.view('uint8')[_RGB] *= areaBlockLights[1:-1, 1:-1, 1:-1][openGateIndices][ ..., numpy.newaxis, numpy.newaxis, numpy.newaxis] vertexArray.shape = (vertexArray.shape[0] * 6, 4, 6) yield append(vertexArray) makeVertices = fenceGateVertices class StairBlockRenderer(BlockRenderer): @classmethod def getBlocktypes(cls, mats): return [a.ID for a in mats.AllStairs] # South - FaceXIncreasing # North - FaceXDecreasing # West - FaceZIncreasing # East - FaceZDecreasing stairTemplates = numpy.array([makeVertexTemplates(**kw) for kw in [ # South - FaceXIncreasing {"xmin": 0.5}, # North - FaceXDecreasing {"xmax": 0.5}, # West - FaceZIncreasing {"zmin": 0.5}, # East - FaceZDecreasing {"zmax": 0.5}, # Slabtype {"ymax": 0.5}, ] ]) def stairVertices(self, facingBlockIndices, blocks, blockMaterials, blockData, areaBlockLights, texMap): arrays = [] append = arrays.append materialIndices = self.getMaterialIndices(blockMaterials) yield stairBlocks = blocks[materialIndices] stairData = blockData[materialIndices] stairTop = (stairData >> 2).astype(bool) stairData &= 3 x, z, y = materialIndices.nonzero() for _ in ("slab", "step"): vertexArray = numpy.zeros((len(x), 6, 4, 6), dtype='float32') for i in xrange(3): vertexArray[_XYZ][..., i] = (x, y, z)[i][:, numpy.newaxis, numpy.newaxis] if _ == "step": vertexArray[_XYZST] += self.stairTemplates[4][..., :5] vertexArray[_XYZ][..., 1][stairTop] += 0.5 else: vertexArray[_XYZST] += self.stairTemplates[stairData][..., :5] vertexArray[_ST] += texMap(stairBlocks, 0)[..., numpy.newaxis, :] vertexArray.view('uint8')[_RGB] = self.stairTemplates[4][numpy.newaxis, ..., 5, numpy.newaxis] vertexArray.view('uint8')[_RGB] *= 0xf vertexArray.view('uint8')[_A] = 0xff vertexArray.shape = (len(x) * 6, 4, 6) yield append(vertexArray) self.vertexArrays = arrays makeVertices = stairVertices class VineBlockRenderer(BlockRenderer): SouthBit = 1 #FaceZIncreasing WestBit = 2 #FaceXDecreasing NorthBit = 4 #FaceZDecreasing EastBit = 8 #FaceXIncreasing renderstate = ChunkCalculator.renderstateVines def __init__(self, *args, **kwargs): BlockRenderer.__init__(self, *args, **kwargs) self.blocktypes = [self.materials["minecraft:vine"].ID] def vineFaceVertices(self, direction, blockIndices, exposedFaceIndices, blocks, blockData, blockLight, facingBlockLight, texMap): bdata = blockData[blockIndices] blockIndices = numpy.array(blockIndices) if direction == pymclevel.faces.FaceZIncreasing: blockIndices[blockIndices] = (bdata & 1).astype(bool) elif direction == pymclevel.faces.FaceXDecreasing: blockIndices[blockIndices] = (bdata & 2).astype(bool) elif direction == pymclevel.faces.FaceZDecreasing: blockIndices[blockIndices] = (bdata & 4).astype(bool) elif direction == pymclevel.faces.FaceXIncreasing: blockIndices[blockIndices] = (bdata & 8).astype(bool) else: return [] vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): return vertexArray vertexArray[_ST] += texMap(self.blocktypes[0], [0], direction)[:, numpy.newaxis, 0:2] lights = blockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] vertexArray.view('uint8')[_RGB] *= lights vertexArray.view('uint8')[_RGB] = vertexArray.view('uint8')[_RGB].astype(float) * LeafBlockRenderer.leafColor if direction == pymclevel.faces.FaceZIncreasing: vertexArray[_XYZ][..., 2] -= 0.0625 if direction == pymclevel.faces.FaceXDecreasing: vertexArray[_XYZ][..., 0] += 0.0625 if direction == pymclevel.faces.FaceZDecreasing: vertexArray[_XYZ][..., 2] += 0.0625 if direction == pymclevel.faces.FaceXIncreasing: vertexArray[_XYZ][..., 0] -= 0.0625 return vertexArray makeFaceVertices = vineFaceVertices class SlabBlockRenderer(BlockRenderer): def __init__(self, *args, **kwargs): BlockRenderer.__init__(self, *args, **kwargs) materials = self.materials # self.blocktypes = [materials["minecraft:wooden_slab"].ID, # materials["minecraft:stone_slab"].ID, # materials["minecraft:stone_slab2"].ID, # materials["minecraft:purpur_slab"].ID] # print "self.blocktypes", self.blocktypes # print "self.materials.AllSlabs", list(set(a.ID for a in self.materials.AllSlabs if "double" not in a.name.lower())) # print list(set(a for a in self.materials.AllSlabs if "double" not in a.name.lower())) self.blocktypes = list(set(a.ID for a in materials.AllSlabs if "double" not in a.name.lower())) def slabFaceVertices(self, direction, blockIndices, facingBlockLight, blocks, blockData, blockLight, areaBlockLights, texMap): lights = areaBlockLights[blockIndices][..., numpy.newaxis, numpy.newaxis] bdata = blockData[blockIndices] top = (bdata >> 3).astype(bool) bdata &= 7 vertexArray = self.makeTemplate(direction, blockIndices) if not len(vertexArray): return vertexArray vertexArray[_ST] += texMap(blocks[blockIndices], bdata, direction)[:, numpy.newaxis, 0:2] vertexArray.view('uint8')[_RGB] *= lights if direction == pymclevel.faces.FaceYIncreasing: vertexArray[_XYZ][..., 1] -= 0.5 if direction != pymclevel.faces.FaceYIncreasing and direction != pymclevel.faces.FaceYDecreasing: vertexArray[_XYZ][..., 2:4, 1] -= 0.5 vertexArray[_ST][..., 2:4, 1] += 8 vertexArray[_XYZ][..., 1][top] += 0.5 return vertexArray makeFaceVertices = slabFaceVertices # 1.9 renderer's class EndRodRenderer(BlockRenderer): def __init__(self, *args, **kwargs): BlockRenderer.__init__(self, *args, **kwargs) self.blocktypes = [self.materials["minecraft:end_rod"].ID] rodTemplate = makeVertexTemplatesFromJsonModel((7, 1, 7), (9, 16, 9), { "down": (4, 2, 2, 0), "up": (2, 0, 4, 2), "north": (0, 0, 2, 15), "south": (0, 0, 2, 15), "west": (0, 0, 2, 15), "east": (0, 0, 2, 15) }) rodTemplates = numpy.array([ rotateTemplate(rodTemplate, x=180), rodTemplate, rotateTemplate(rodTemplate, x=90), rotateTemplate(rodTemplate, y=180, x=90), rotateTemplate(rodTemplate, y=270, x=90), rotateTemplate(rodTemplate, y=90, x=90), numpy.zeros((6, 4, 6)), numpy.zeros((6, 4, 6)) ]) handleTemplate = makeVertexTemplatesFromJsonModel((6, 0, 6), (10, 1, 10), { "down": (6, 6, 2, 2), "up": (2, 2, 6, 6), "north": (2, 6, 6, 7), "south": (2, 6, 6, 7), "west": (2, 6, 6, 7), "east": (2, 6, 6, 7) }) handleTemplates = numpy.array([ rotateTemplate(handleTemplate, x=180), handleTemplate, rotateTemplate(handleTemplate, x=90), rotateTemplate(handleTemplate, y=180, x=90), rotateTemplate(handleTemplate, y=270, x=90), rotateTemplate(handleTemplate, y=90, x=90), numpy.zeros((6, 4, 6)), numpy.zeros((6, 4, 6)) ]) makeVertices = makeVerticesFromModel([rodTemplates, handleTemplates], 7) class WaterBlockRenderer(BlockRenderer): renderstate = ChunkCalculator.renderstateWater def __init__(self, *args, **kwargs): BlockRenderer.__init__(self, *args, **kwargs) materials = self.materials self.waterID = materials["minecraft:water"].ID self.blocktypes = [materials["minecraft:flowing_water"].ID, self.waterID] @classmethod def getBlocktypes(cls, mats): cls.waterID = mats["minecraft:water"].ID return [mats["minecraft:flowing_water"].ID, cls.waterID] def waterFaceVertices(self, direction, blockIndices, exposedFaceIndices, blocks, blockData, blockLight, facingBlockLight, texMap): blockIndices = blockIndices & exposedFaceIndices vertexArray = self.makeTemplate(direction, blockIndices) vertexArray[_ST] += texMap(self.waterID, 0, 0)[numpy.newaxis, numpy.newaxis] vertexArray.view('uint8')[_RGB] *= facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] return vertexArray makeFaceVertices = waterFaceVertices class IceBlockRenderer(BlockRenderer): renderstate = ChunkCalculator.renderstateIce @classmethod def getBlocktypes(cls, mats): cls.iceID = mats["minecraft:ice"].ID return [cls.iceID] def iceFaceVertices(self, direction, blockIndices, exposedFaceIndices, blocks, blockData, blockLight, facingBlockLight, texMap): blockIndices = blockIndices & exposedFaceIndices vertexArray = self.makeTemplate(direction, blockIndices) vertexArray[_ST] += texMap(self.iceID, 0, 0)[numpy.newaxis, numpy.newaxis] vertexArray.view('uint8')[_RGB] *= facingBlockLight[blockIndices][..., numpy.newaxis, numpy.newaxis] return vertexArray makeFaceVertices = iceFaceVertices from glutils import DisplayList class MCRenderer(object): isPreviewer = False def __init__(self, level=None, alpha=1.0): self.render = True self.origin = (0, 0, 0) self.rotation = 0 self.bufferUsage = 0 self.invalidChunkQueue = deque() self._chunkWorker = None self.chunkRenderers = {} self.loadableChunkMarkers = DisplayList() self.visibleLayers = set(Layer.AllLayers) self.masterLists = None alpha *= 255 self.alpha = (int(alpha) & 0xff) self.chunkStartTime = datetime.now() self.oldChunkStartTime = self.chunkStartTime self.oldPosition = None self.chunkSamples = [timedelta(0, 0, 0)] * 15 self.chunkIterator = None config.settings.fastLeaves.addObserver(self) config.settings.roughGraphics.addObserver(self) config.settings.showHiddenOres.addObserver(self) config.settings.vertexBufferLimit.addObserver(self) config.settings.drawEntities.addObserver(self) config.settings.drawTileEntities.addObserver(self) config.settings.drawTileTicks.addObserver(self) config.settings.drawUnpopulatedChunks.addObserver(self, "drawTerrainPopulated") config.settings.drawChunkBorders.addObserver(self, "drawChunkBorder") config.settings.drawMonsters.addObserver(self) config.settings.drawItems.addObserver(self) config.settings.showChunkRedraw.addObserver(self, "showRedraw") config.settings.spaceHeight.addObserver(self) config.settings.targetFPS.addObserver(self, "targetFPS") for ore in config.settings.hiddableOres.get(): config.settings["showOre{}".format(ore)].addObserver(self, callback=lambda x, id=ore: self.showOre(id, x)) self.level = level if self.level.__class__.__name__ in ("FakeLevel", "MCSchematic"): self.toggleLayer(False, 'ChunkBorder') chunkClass = ChunkRenderer calculatorClass = ChunkCalculator minViewDistance = 2 _viewDistance = 8 needsRedraw = True def toggleLayer(self, val, layer): if val: self.visibleLayers.add(layer) else: self.visibleLayers.discard(layer) for cr in self.chunkRenderers.itervalues(): cr.invalidLayers.add(layer) self.loadNearbyChunks() def layerProperty(layer, default=True): # @NoSelf attr = intern("_draw" + layer) def _get(self): return getattr(self, attr, default) def _set(self, val): if val != _get(self): setattr(self, attr, val) self.toggleLayer(val, layer) return property(_get, _set) drawEntities = layerProperty(Layer.Entities) drawTileEntities = layerProperty(Layer.TileEntities) drawTileTicks = layerProperty(Layer.TileTicks) drawMonsters = layerProperty(Layer.Monsters) drawItems = layerProperty(Layer.Items) drawTerrainPopulated = layerProperty(Layer.TerrainPopulated) drawChunkBorder = layerProperty(Layer.ChunkBorder) def inSpace(self): if self.level is None: return True h = self.position[1] if self.level.dimNo == 1: _2478aq_heot(h) return ((h > self.level.Height + self.spaceHeight) or (h <= -self.spaceHeight)) def chunkDistance(self, cpos): camx, camy, camz = self.position # if the renderer is offset into the world somewhere, adjust for that ox, oy, oz = self.origin camx -= ox camz -= oz camcx = int(numpy.floor(camx)) >> 4 camcz = int(numpy.floor(camz)) >> 4 cx, cz = cpos return max(abs(cx - camcx), abs(cz - camcz)) overheadMode = False def detailLevelForChunk(self, cpos): if self.overheadMode: return 2 if self.isPreviewer: w, l, h = self.level.bounds.size if w + l < 256: return 0 distance = self.chunkDistance(cpos) - self.viewDistance if distance > 0 or self.inSpace(): return 1 return 0 def getViewDistance(self): return self._viewDistance def setViewDistance(self, vd): vd = int(vd) & 0xfffe vd = min(max(vd, self.minViewDistance), config.settings.maxViewDistance.get()) if vd != self._viewDistance: self._viewDistance = vd self.viewDistanceChanged() viewDistance = property(getViewDistance, setViewDistance, None, "View Distance") @property def effectiveViewDistance(self): if self.inSpace(): return self.viewDistance * 4 else: return self.viewDistance * 2 def viewDistanceChanged(self): self.oldPosition = None # xxx self.discardMasterList() self.loadNearbyChunks() self.discardChunksOutsideViewDistance() maxWorkFactor = 64 minWorkFactor = 1 workFactor = 2 chunkCalculator = None _level = None @property def level(self): return self._level @level.setter def level(self, level): """ this probably warrants creating a new renderer """ self.stopWork() self._level = level self.oldPosition = None self.position = (0, 0, 0) self.chunkCalculator = None self.invalidChunkQueue = deque() self.discardAllChunks() self.loadableChunkMarkers.invalidate() if level: self.chunkCalculator = self.calculatorClass(self.level) self.oldPosition = None self.loadNearbyChunks() position = (0, 0, 0) def loadChunksStartingFrom(self, wx, wz, distance=None): # world position if None is self.level: return if self.level.saving: return if distance is None: d = self.effectiveViewDistance else: d = distance self.chunkIterator = self.iterateChunks(wx, wz, d * 2) def iterateChunks(self, x, z, d): cx = x >> 4 cz = z >> 4 yield (cx, cz) step = dir = 1 while True: for i in xrange(step): cx += dir yield (cx, cz) for i in xrange(step): cz += dir yield (cx, cz) step += 1 if step > d and not self.overheadMode: raise StopIteration dir = -dir chunkIterator = None @property def chunkWorker(self): if self._chunkWorker is None: self._chunkWorker = self.makeWorkIterator() return self._chunkWorker def stopWork(self): self._chunkWorker = None def discardAllChunks(self): self.bufferUsage = 0 self.forgetAllDisplayLists() self.chunkRenderers = {} self.oldPosition = None # xxx force reload def discardChunksInBox(self, box): self.discardChunks(box.chunkPositions) def discardChunksOutsideViewDistance(self): if self.overheadMode: return # print "discardChunksOutsideViewDistance" d = self.effectiveViewDistance cx = (self.position[0] - self.origin[0]) / 16 cz = (self.position[2] - self.origin[2]) / 16 origin = (cx - d, cz - d) size = d * 2 if not len(self.chunkRenderers): return (ox, oz) = origin chunks = numpy.fromiter(self.chunkRenderers.iterkeys(), dtype='i,i', count=len(self.chunkRenderers)) chunks.dtype = 'int32' chunks.shape = len(self.chunkRenderers), 2 if size: outsideChunks = chunks[:, 0] < ox - 1 outsideChunks |= chunks[:, 0] > ox + size outsideChunks |= chunks[:, 1] < oz - 1 outsideChunks |= chunks[:, 1] > oz + size chunks = chunks[outsideChunks] self.discardChunks(chunks) def discardChunks(self, chunks): for cx, cz in chunks: self.discardChunk(cx, cz) self.oldPosition = None # xxx force reload def discardChunk(self, cx, cz): " discards the chunk renderer for this chunk and compresses the chunk " if (cx, cz) in self.chunkRenderers: self.bufferUsage -= self.chunkRenderers[cx, cz].bufferSize self.chunkRenderers[cx, cz].forgetDisplayLists() del self.chunkRenderers[cx, cz] _fastLeaves = False @property def fastLeaves(self): return self._fastLeaves @fastLeaves.setter def fastLeaves(self, val): if self._fastLeaves != bool(val): self.discardAllChunks() self._fastLeaves = bool(val) _roughGraphics = False @property def roughGraphics(self): return self._roughGraphics @roughGraphics.setter def roughGraphics(self, val): if self._roughGraphics != bool(val): self.discardAllChunks() self._roughGraphics = bool(val) _showHiddenOres = False @property def showHiddenOres(self): return self._showHiddenOres @showHiddenOres.setter def showHiddenOres(self, val): if self._showHiddenOres != bool(val): self.discardAllChunks() self._showHiddenOres = bool(val) def showOre(self, ore, show): ChunkCalculator.hiddenOreMaterials[ore] = ore if show else 1 if self.showHiddenOres: self.discardAllChunks() def invalidateChunk(self, cx, cz, layers=None): " marks the chunk for regenerating vertex data and display lists " if (cx, cz) in self.chunkRenderers: self.chunkRenderers[(cx, cz)].invalidate(layers) self.invalidChunkQueue.append((cx, cz)) # xxx encapsulate def invalidateChunksInBox(self, box, layers=None): # If the box is at the edge of any chunks, expanding by 1 makes sure the neighboring chunk gets redrawn. box = box.expand(1) self.invalidateChunks(box.chunkPositions, layers) def invalidateEntitiesInBox(self, box): self.invalidateChunks(box.chunkPositions, [Layer.Entities]) def invalidateTileTicksInBox(self, box): self.invalidateChunks(box.chunkPositions, [Layer.TileTicks]) def invalidateChunks(self, chunks, layers=None): for (cx, cz) in chunks: self.invalidateChunk(cx, cz, layers) self.stopWork() self.discardMasterList() self.loadNearbyChunks() def invalidateAllChunks(self, layers=None): self.invalidateChunks(self.chunkRenderers.iterkeys(), layers) def forgetAllDisplayLists(self): for cr in self.chunkRenderers.itervalues(): cr.forgetDisplayLists() def invalidateMasterList(self): self.discardMasterList() shouldRecreateMasterList = True def discardMasterList(self): self.shouldRecreateMasterList = True @property def shouldDrawAll(self): box = self.level.bounds return self.isPreviewer and box.width + box.length < 256 distanceToChunkReload = 32.0 def cameraMovedFarEnough(self): if self.shouldDrawAll: return False if self.oldPosition is None: return True cPos = self.position oldPos = self.oldPosition cameraDelta = self.distanceToChunkReload return any([abs(x - y) > cameraDelta for x, y in zip(cPos, oldPos)]) def loadVisibleChunks(self): """ loads nearby chunks if the camera has moved beyond a certain distance """ # print "loadVisibleChunks" if self.cameraMovedFarEnough(): if datetime.now() - self.lastVisibleLoad > timedelta(0, 0.5): self.discardChunksOutsideViewDistance() self.loadNearbyChunks() self.oldPosition = self.position self.lastVisibleLoad = datetime.now() lastVisibleLoad = datetime.now() def loadNearbyChunks(self): if None is self.level: return # print "loadNearbyChunks" cameraPos = self.position if self.shouldDrawAll: self.loadAllChunks() else: # subtract self.origin to load nearby chunks correctly for preview renderers self.loadChunksStartingFrom(int(cameraPos[0]) - self.origin[0], int(cameraPos[2]) - self.origin[2]) def loadAllChunks(self): box = self.level.bounds self.loadChunksStartingFrom(box.origin[0] + box.width / 2, box.origin[2] + box.length / 2, max(box.width, box.length)) _floorTexture = None @property def floorTexture(self): if self._floorTexture is None: self._floorTexture = Texture(self.makeFloorTex) return self._floorTexture @staticmethod def makeFloorTex(): color0 = (0xff, 0xff, 0xff, 0x22) color1 = (0xff, 0xff, 0xff, 0x44) img = numpy.array([color0, color1, color1, color0], dtype='uint8') GL.glTexParameter(GL.GL_TEXTURE_2D, GL.GL_TEXTURE_MIN_FILTER, GL.GL_NEAREST) GL.glTexParameter(GL.GL_TEXTURE_2D, GL.GL_TEXTURE_MAG_FILTER, GL.GL_NEAREST) GL.glTexImage2D(GL.GL_TEXTURE_2D, 0, GL.GL_RGBA, 2, 2, 0, GL.GL_RGBA, GL.GL_UNSIGNED_BYTE, img) def invalidateChunkMarkers(self): self.loadableChunkMarkers.invalidate() def _drawLoadableChunkMarkers(self): if self.level.chunkCount: chunkSet = set(self.level.allChunks) sizedChunks = chunkMarkers(chunkSet) GL.glPushAttrib(GL.GL_FOG_BIT) GL.glDisable(GL.GL_FOG) GL.glEnable(GL.GL_BLEND) GL.glEnable(GL.GL_POLYGON_OFFSET_FILL) GL.glPolygonOffset(DepthOffset.ChunkMarkers, DepthOffset.ChunkMarkers) GL.glEnable(GL.GL_DEPTH_TEST) GL.glEnableClientState(GL.GL_TEXTURE_COORD_ARRAY) GL.glEnable(GL.GL_TEXTURE_2D) GL.glColor(1.0, 1.0, 1.0, 1.0) self.floorTexture.bind() for size, chunks in sizedChunks.iteritems(): if not len(chunks): continue chunks = numpy.array(chunks, dtype='float32') chunkPosition = numpy.zeros(shape=(chunks.shape[0], 4, 3), dtype='float32') chunkPosition[:, :, (0, 2)] = numpy.array(((0, 0), (0, 1), (1, 1), (1, 0)), dtype='float32') chunkPosition[:, :, (0, 2)] *= size chunkPosition[:, :, (0, 2)] += chunks[:, numpy.newaxis, :] chunkPosition *= 16 GL.glVertexPointer(3, GL.GL_FLOAT, 0, chunkPosition.ravel()) GL.glTexCoordPointer(2, GL.GL_FLOAT, 0, (chunkPosition[..., (0, 2)] * 16).ravel()) GL.glDrawArrays(GL.GL_QUADS, 0, len(chunkPosition) * 4) GL.glDisableClientState(GL.GL_TEXTURE_COORD_ARRAY) GL.glDisable(GL.GL_TEXTURE_2D) GL.glDisable(GL.GL_BLEND) GL.glDisable(GL.GL_DEPTH_TEST) GL.glDisable(GL.GL_POLYGON_OFFSET_FILL) GL.glPopAttrib() def drawLoadableChunkMarkers(self): if not self.isPreviewer or isinstance(self.level, pymclevel.MCInfdevOldLevel): self.loadableChunkMarkers.call(self._drawLoadableChunkMarkers) needsImmediateRedraw = False viewingFrustum = None if "-debuglists" in sys.argv: def createMasterLists(self): pass def callMasterLists(self): for cr in self.chunkRenderers.itervalues(): cr.debugDraw() else: def createMasterLists(self): if self.shouldRecreateMasterList: lists = {} chunkLists = defaultdict(list) chunksPerFrame = 80 shouldRecreateAgain = False for ch in self.chunkRenderers.itervalues(): if chunksPerFrame: if ch.needsRedisplay: chunksPerFrame -= 1 ch.makeDisplayLists() else: shouldRecreateAgain = True if ch.renderstateLists: for rs in ch.renderstateLists: chunkLists[rs] += ch.renderstateLists[rs] for rs in chunkLists: if len(chunkLists[rs]): lists[rs] = numpy.array(chunkLists[rs], dtype='uint32').ravel() self.masterLists = lists self.shouldRecreateMasterList = shouldRecreateAgain self.needsImmediateRedraw = shouldRecreateAgain def callMasterLists(self): for renderstate in self.chunkCalculator.renderstates: if renderstate not in self.masterLists: continue if self.alpha != 0xff and renderstate is not ChunkCalculator.renderstateLowDetail: GL.glEnable(GL.GL_BLEND) renderstate.bind() GL.glCallLists(self.masterLists[renderstate]) renderstate.release() if self.alpha != 0xff and renderstate is not ChunkCalculator.renderstateLowDetail: GL.glDisable(GL.GL_BLEND) errorLimit = 10 def draw(self): self.needsRedraw = False if not self.level: return if not self.chunkCalculator: return if not self.render: return if self.level.materials.name in ("Pocket", "Alpha"): GL.glMatrixMode(GL.GL_TEXTURE) GL.glScalef(1 / 2., 1 / 2., 1 / 2.) with gl.glPushMatrix(GL.GL_MODELVIEW): dx, dy, dz = self.origin GL.glTranslate(dx, dy, dz) GL.glEnable(GL.GL_CULL_FACE) GL.glEnable(GL.GL_DEPTH_TEST) self.level.materials.terrainTexture.bind() GL.glEnable(GL.GL_TEXTURE_2D) GL.glEnableClientState(GL.GL_TEXTURE_COORD_ARRAY) offset = DepthOffset.PreviewRenderer if self.isPreviewer else DepthOffset.Renderer GL.glPolygonOffset(offset, offset) GL.glEnable(GL.GL_POLYGON_OFFSET_FILL) self.createMasterLists() try: self.callMasterLists() except GL.GLError as e: if self.errorLimit: self.errorLimit -= 1 traceback.print_exc() print e GL.glDisable(GL.GL_POLYGON_OFFSET_FILL) GL.glDisable(GL.GL_CULL_FACE) GL.glDisable(GL.GL_DEPTH_TEST) GL.glDisable(GL.GL_TEXTURE_2D) GL.glDisableClientState(GL.GL_TEXTURE_COORD_ARRAY) self.drawLoadableChunkMarkers() if self.level.materials.name in ("Pocket", "Alpha"): GL.glMatrixMode(GL.GL_TEXTURE) GL.glScalef(2., 2., 2.) renderErrorHandled = False def addDebugInfo(self, addDebugString): addDebugString("BU: {0} MB, ".format( self.bufferUsage / 1000000, )) addDebugString("WQ: {0}, ".format(len(self.invalidChunkQueue))) if self.chunkIterator: addDebugString("[LR], ") addDebugString("CR: {0}, ".format(len(self.chunkRenderers), )) def next(self): self.chunkWorker.next() def makeWorkIterator(self): ''' does chunk face and vertex calculation work. returns a generator that can be iterated over for smaller work units.''' try: while True: if self.level is None: raise StopIteration if len(self.invalidChunkQueue) > 1024: self.invalidChunkQueue.clear() if len(self.invalidChunkQueue): c = self.invalidChunkQueue[0] for _ in self.workOnChunk(c): yield self.invalidChunkQueue.popleft() elif self.chunkIterator is None: raise StopIteration else: c = self.chunkIterator.next() if self.vertexBufferLimit: while self.bufferUsage > (0.9 * (self.vertexBufferLimit << 20)): deadChunk = None deadDistance = self.chunkDistance(c) for cr in self.chunkRenderers.itervalues(): dist = self.chunkDistance(cr.chunkPosition) if dist > deadDistance: deadChunk = cr deadDistance = dist if deadChunk is not None: self.discardChunk(*deadChunk.chunkPosition) else: break else: for _ in self.workOnChunk(c): yield else: for _ in self.workOnChunk(c): yield yield finally: self._chunkWorker = None if self.chunkIterator: self.chunkIterator = None vertexBufferLimit = 384 def getChunkRenderer(self, c): if c not in self.chunkRenderers: return self.chunkClass(self, c) return self.chunkRenderers[c] def calcFacesForChunkRenderer(self, cr): self.bufferUsage -= cr.bufferSize calc = cr.calcFaces() work = 0 for _ in calc: yield work += 1 self.chunkDone(cr, work) def workOnChunk(self, c): work = 0 if self.level.containsChunk(*c): cr = self.getChunkRenderer(c) if self.viewingFrustum: if not self.viewingFrustum.visible1([c[0] * 16 + 8, self.level.Height / 2, c[1] * 16 + 8, 1.0], self.level.Height / 2): raise StopIteration faceInfoCalculator = self.calcFacesForChunkRenderer(cr) try: for _ in faceInfoCalculator: work += 1 if (work % MCRenderer.workFactor) == 0: yield self.invalidateMasterList() except Exception as e: traceback.print_exc() fn = c logging.info(u"Skipped chunk {f}: {e}".format(e=e, f=fn)) redrawChunks = 0 def chunkDone(self, chunkRenderer, work): self.chunkRenderers[chunkRenderer.chunkPosition] = chunkRenderer self.bufferUsage += chunkRenderer.bufferSize # print "Chunk {0} used {1} work units".format(chunkRenderer.chunkPosition, work) if not self.needsRedraw: if self.redrawChunks: self.redrawChunks -= 1 if not self.redrawChunks: self.needsRedraw = True else: self.redrawChunks = 2 if work > 0: self.oldChunkStartTime = self.chunkStartTime self.chunkStartTime = datetime.now() self.chunkSamples.pop(0) self.chunkSamples.append(self.chunkStartTime - self.oldChunkStartTime) class PreviewRenderer(MCRenderer): isPreviewer = True def rendermain(): renderer = MCRenderer() renderer.level = pymclevel.mclevel.loadWorld("World1") renderer.viewDistance = 6 renderer.detailLevelForChunk = lambda *x: 0 start = datetime.now() renderer.loadVisibleChunks() try: while True: # for i in range(100): renderer.next() except StopIteration: pass except Exception as e: traceback.print_exc() print repr(e) duration = datetime.now() - start perchunk = duration / len(renderer.chunkRenderers) print "Duration: {0} ({1} chunks per second, {2} per chunk, {3} chunks)".format(duration, 1000000.0 / perchunk.microseconds, perchunk, len(renderer.chunkRenderers)) # display.init( (640, 480), OPENGL | DOUBLEBUF ) from utilities.gl_display_context import GLDisplayContext from OpenGL import GLU import pygame # distance = 4000 GL.glMatrixMode(GL.GL_PROJECTION) GL.glLoadIdentity() GLU.gluPerspective(35, 640.0 / 480.0, 0.5, 4000.0) h = 366 pos = (0, h, 0) look = (0.0001, h - 1, 0.0001) up = (0, 1, 0) GL.glMatrixMode(GL.GL_MODELVIEW) GL.glLoadIdentity() GLU.gluLookAt(pos[0], pos[1], pos[2], look[0], look[1], look[2], up[0], up[1], up[2]) GL.glClearColor(0.0, 0.0, 0.0, 1.0) framestart = datetime.now() frames = 200 for i in xrange(frames): GL.glClear(GL.GL_COLOR_BUFFER_BIT | GL.GL_DEPTH_BUFFER_BIT) renderer.draw() pygame.display.flip() delta = datetime.now() - framestart seconds = delta.seconds + delta.microseconds / 1000000.0 print "{0} frames in {1} ({2} per frame, {3} FPS)".format(frames, delta, delta / frames, frames / seconds) while True: evt = pygame.event.poll() if evt.type == pygame.MOUSEBUTTONDOWN: break # time.sleep(3.0) import traceback if __name__ == "__main__": import cProfile cProfile.run("rendermain()", "mcedit.profile")
146,617
35.572213
153
py
MCEdit-Unified
MCEdit-Unified-master/resource_packs.py
# -*- coding: utf-8 -*- #!# If the comman line parameter '--debug-packs' is given, the logging level is set to debug. #!# Otherwise, it is set to critical. from PIL import Image import zipfile import directories import os import shutil from config import config from cStringIO import StringIO import locale import traceback from utilities.misc import Singleton DEF_ENC = locale.getdefaultlocale()[1] if DEF_ENC is None: DEF_ENC = "UTF-8" try: import resource # @UnresolvedImport resource.setrlimit(resource.RLIMIT_NOFILE, (500,-1)) except: pass #!# Debugging .zip resource pack not loaded bug. import logging level = 50 if '--debug-packs' in os.sys.argv: level = 10 log = logging.getLogger(__name__) log.setLevel(level) #!# def step(slot): ''' Utility method for multiplying the slot by 16 :param slot: Texture slot :type slot: int ''' return slot << 4 ''' Empty comment lines like: # are for texture spaces that I don't know what should go there ''' textureSlots = { # Start Top Row "grass_top": (step(0),step(0)), "stone": (step(1),step(0)), "dirt": (step(2),step(0)), "grass_side": (step(3),step(0)), "planks_oak": (step(4),step(0)), "stone_slab_side": (step(5),step(0)), "stone_slab_top": (step(6),step(0)), "brick": (step(7),step(0)), "tnt_side": (step(8),step(0)), "tnt_top": (step(9),step(0)), "tnt_bottom": (step(10),step(0)), "web": (step(11),step(0)), "flower_rose": (step(12),step(0)), "flower_dandelion": (step(13),step(0)), # "sapling_oak": (step(15),step(0)), "flower_blue_orchid": (step(16),step(0)), "flower_allium": (step(17),step(0)), "flower_houstonia": (step(18),step(0)), "flower_tulip_red": (step(19),step(0)), "sapling_roofed_oak": (step(20),step(0)), # End Top Row # Start Second Row "cobblestone": (step(0),step(1)), "bedrock": (step(1),step(1)), "sand": (step(2),step(1)), "gravel": (step(3),step(1)), "log_oak": (step(4),step(1)), "log_oak_top": (step(5),step(1)), "iron_block": (step(6),step(1)), "gold_block": (step(7),step(1)), "diamond_block": (step(8),step(1)), "emerald_block": (step(9),step(1)), # "red_sand": (step(11),step(1)), "mushroom_red": (step(12),step(1)), "mushroom_brown": (step(13),step(1)), "sapling_jungle": (step(14),step(1)), "fire_layer_0": (step(15),step(1)), "flower_tulip_orange": (step(16),step(1)), "flower_tulip_white": (step(17),step(1)), "flower_tulip_pink": (step(18),step(1)), "flower_oxeye_daisy": (step(19),step(1)), "sapling_acacia": (step(20),step(1)), # End Second Row # Start Third Row "gold_ore": (step(0),step(2)), "iron_ore": (step(1),step(2)), "coal_ore": (step(2),step(2)), "bookshelf": (step(3),step(2)), "cobblestone_mossy": (step(4),step(2)), "obsidian": (step(5),step(2)), # "tallgrass": (step(7),step(2)), # "beacon": (step(9),step(2)), "dropper_front_horizontal": (step(10),step(2)), "crafting_table_top": (step(11),step(2)), "furnace_front_off": (step(12),step(2)), "furnace_side": (step(13),step(2)), "dispenser_front_horizontal": (step(14),step(2)), "fire_layer_1": (step(15),step(2)), # # # # "daylight_detector_side": (step(20),step(2)), # End Third Row # Start Fourth Row "sponge": (step(0), step(3)), "glass": (step(1), step(3)), "diamond_ore": (step(2), step(3)), "redstone_ore": (step(3), step(3)), # # "stonebrick": (step(6), step(3)), "deadbush": (step(7), step(3)), "fern": (step(8), step(3)), "dirt_podzol_top": (step(9), step(3)), "dirt_podzol_side": (step(10), step(3)), "crafting_table_side": (step(11), step(3)), "crafting_table_front": (step(12), step(3)), "furnace_front_on": (step(13), step(3)), "furnace_top": (step(14), step(3)), "sapling_spruce": (step(15), step(3)), # # # # # End Fourth Row # Start Fifth Row "wool_colored_white": (step(0), step(4)), "mob_spawner": (step(1), step(4)), "snow": (step(2), step(4)), "ice": (step(3), step(4)), "grass_side_snowed": (step(4), step(4)), "cactus_top": (step(5), step(4)), "cactus_side": (step(6), step(4)), "cactus_bottom": (step(7), step(4)), "clay": (step(8), step(4)), "reeds": (step(9), step(4)), "jukebox_side": (step(10), step(4)), "jukebox_top": (step(11), step(4)), "waterlily": (step(12), step(4)), "mycelium_side": (step(13), step(4)), "mycelium_top": (step(14), step(4)), "sapling_birch": (step(15), step(4)), # # "dropper_front_vertical": (step(18), step(4)), "daylight_detector_inverted_top": (step(19), step(4)), # End Fifth Row # Start Sixth Row "torch_on": (step(0), step(5)), "door_wood_upper": (step(1), step(5)), "door_iron_upper": (step(2), step(5)), "ladder": (step(3), step(5)), "trapdoor": (step(4), step(5)), "iron_bars": (step(5), step(5)), "farmland_wet": (step(6), step(5)), "farmland_dry": (step(7), step(5)), "wheat_stage_0": (step(8), step(5)), "wheat_stage_1": (step(9), step(5)), "wheat_stage_2": (step(10), step(5)), "wheat_stage_3": (step(11), step(5)), "wheat_stage_4": (step(12), step(5)), "wheat_stage_5": (step(13), step(5)), "wheat_stage_6": (step(14), step(5)), "wheat_stage_7": (step(15), step(5)), # # "dispenser_front_vertical": (step(18), step(5)), # # End Sixth Row # Start Seventh Row "lever": (step(0), step(6)), "door_wood_lower": (step(1), step(6)), "door_iron_lower": (step(2), step(6)), "redstone_torch_on": (step(3), step(6)), "stonebrick_mossy": (step(4), step(6)), "stonebrick_cracked": (step(5), step(6)), "pumpkin_top": (step(6), step(6)), "netherrack": (step(7), step(6)), "soul_sand": (step(8), step(6)), "glowstone": (step(9), step(6)), "piston_top_sticky": (step(10), step(6)), "piston_top_normal": (step(11), step(6)), "piston_side": (step(12), step(6)), "piston_bottom": (step(13), step(6)), "piston_inner": (step(14), step(6)), "pumpkin_stem_disconnected": (step(15), step(6)), # # # # End Seventh Row # Start Eigth Row "rail_normal_turned": (step(0),step(7)), "wool_colored_black": (step(1),step(7)), "wool_colored_gray": (step(2),step(7)), "redstone_torch_off": (step(3),step(7)), "log_spruce": (step(4),step(7)), "log_birch": (step(5),step(7)), "pumpkin_side": (step(6),step(7)), "pumpkin_face_off": (step(7),step(7)), "pumpkin_face_on": (step(8),step(7)), "cake_top": (step(9),step(7)), "cake_side": (step(10),step(7)), "cake_inner": (step(11),step(7)), "cake_bottom": (step(12),step(7)), "mushroom_block_skin_red": (step(13),step(7)), "mushroom_block_skin_brown": (step(14),step(7)), "pumpkin_stem_connected": (step(15),step(7)), # # "repeater_off_west": (step(18),step(7)), # # End Eigth Row # Start Ninth Row "rail_normal": (step(0),step(8)), "wool_colored_red": (step(1),step(8)), "wool_colored_magenta": (step(2),step(8)), "repeater_off_south": (step(3),step(8)), "leaves_spruce": (step(4),step(8)), # "bed_feet_top": (step(6),step(8)), "bed_head_top": (step(7),step(8)), "melon_side": (step(8),step(8)), "melon_top": (step(9),step(8)), # # # "mushroom_block_skin_stem": (step(13),step(8)), "mushroom_block_inside": (step(14),step(8)), "vine": (step(15),step(8)), # # "repeater_off_north": (step(18),step(8)), # # End Ninth Row # Start Tenth Row "lapis_block": (step(0),step(9)), "wool_colored_green": (step(1),step(9)), "wool_colored_lime": (step(2),step(9)), "repeater_on_south": (step(3),step(9)), # "bed_feet_end": (step(5),step(9)), "bed_feet_side": (step(6),step(9)), "bed_head_side": (step(7),step(9)), "bed_head_end": (step(8),step(9)), "log_jungle": (step(9),step(9)), "cauldron_side": (step(10),step(9)), "cauldron_bottom": (step(11),step(9)), "brewing_stand_base": (step(12),step(9)), "brewing_stand": (step(13),step(9)), "endframe_top": (step(14),step(9)), "endframe_side": (step(15),step(9)), "double_plant_sunflower_bottom": (step(16),step(9)), # "repeater_off_east": (step(18),step(9)), "structure_block_data": (step(19),step(9)), "structure_block_corner": (step(20),step(9)), # End Tenth Row # Start Eleventh Row "lapis_ore": (step(0),step(10)), "wool_colored_brown": (step(1),step(10)), "wool_colored_yellow": (step(2),step(10)), "rail_golden": (step(3),step(10)), "redstone_dust_cross": (step(4),step(10)), # "enchanting_table_top": (step(6),step(10)), "dragon_egg": (step(7),step(10)), "cocoa_stage_2": (step(8),step(10)), "cocoa_stage_1": (step(9),step(10)), "cocoa_stage_0": (step(10),step(10)), "emerald_ore": (step(11),step(10)), "trip_wire_source": (step(12),step(10)), "trip_wire": (step(13),step(10)), "endframe_eye": (step(14),step(10)), "end_stone": (step(15),step(10)), "double_plant_syringa_bottom": (step(16),step(10)), "double_plant_syringa_top": (step(17),step(10)), "repeater_on_west": (step(18),step(10)), "structure_block_save": (step(19),step(10)), "structure_block_load": (step(20),step(10)), # End Eleventh Row # Start Twelfth Row "sandstone_top": (step(0),step(11)), "wool_colored_blue": (step(1),step(11)), "wool_colored_light_blue": (step(2),step(11)), "rail_golden_powered": (step(3),step(11)), # "redstone_dust_line": (step(5),step(11)), "enchanting_table_side": (step(6),step(11)), "enchanting_table_bottom": (step(7),step(11)), "command_block": (step(8),step(11)), "itemframe_backround": (step(9),step(11)), "flower_pot": (step(10),step(11)), "comparator_off_south": (step(11),step(11)), "comparator_on_south": (step(12),step(11)), "daylight_detector_top": (step(13),step(11)), "redstone_block": (step(14),step(11)), "quartz_ore": (step(15),step(11)), "double_plant_grass_bottom": (step(16),step(11)), "double_plant_grass_top": (step(17),step(11)), "repeater_on_north": (step(18),step(11)), "command_block_back": (step(19),step(11)), "command_block_conditional": (step(20),step(11)), "command_block_front": (step(21),step(11)), "command_block_side": (step(22),step(11)), "bone_block_top": (step(19),step(11)), # End Twelfth Row # Start Thriteenth Row "sandstone_normal": (step(0),step(12)), "wool_colored_purple": (step(1),step(12)), "wool_colored_pink": (step(2),step(12)), "rail_detector": (step(3),step(12)), "leaves_jungle": (step(4),step(12)), # "planks_spruce": (step(6),step(12)), "planks_jungle": (step(7),step(12)), "carrots_stage_0": (step(8),step(12)), "carrots_stage_1": (step(9),step(12)), "carrots_stage_2": (step(10),step(12)), "carrots_stage_3": (step(11),step(12)), "potatoes_stage_3": (step(12),step(12)), # "piston_right": (step(14),step(12)), "piston_down": (step(15),step(12)), "double_plant_fern_bottom": (step(16),step(12)), "double_plant_fern_top": (step(17),step(12)), "repeater_on_east": (step(18),step(12)), "repeating_command_block_back": (step(19),step(12)), "repeating_command_block_conditional": (step(20),step(12)), "repeating_command_block_front": (step(21),step(12)), "repeating_command_block_side": (step(22),step(12)), "bone_block_side": (step(19),step(12)), # End Thriteenth Row # Start Fourteenth Row "sandstone_bottom": (step(0),step(13)), "wool_colored_cyan": (step(1),step(13)), "wool_colored_orange": (step(2),step(13)), "redstone_lamp_off": (step(3),step(13)), "redstone_lamp_on": (step(4),step(13)), "stonebrick_carved": (step(5),step(13)), "planks_birch": (step(6),step(13)), "anvil_base": (step(7),step(13)), "anvil_top_damaged_1": (step(8),step(13)), "quatrz_block_top": (step(9),step(13)), "rail_activator": (step(10),step(13)), "rail_activator_powered": (step(11),step(13)), "coal_block": (step(12),step(13)), "log_acacia_top": (step(13),step(13)), "piston_left": (step(14),step(13)), "magma": (step(18),step(13)), # "double_plant_rose_bottom": (step(16),step(13)), "double_plant_rose_top": (step(17),step(13)), "chain_command_block_back": (step(19),step(11)), "chain_command_block_conditional": (step(20),step(11)), "chain_command_block_front": (step(21),step(11)), "chain_command_block_side": (step(22),step(11)), # End Fourteenth Row # Start Fifteenth Row "nether_brick": (step(0),step(14)), "wool_colored_silver": (step(1),step(14)), "nether_wart_stage_0": (step(2),step(14)), "nether_wart_stage_1": (step(3),step(14)), "nether_wart_stage_2": (step(4),step(14)), "sandstone_carved": (step(5),step(14)), "sandstone_smooth": (step(6),step(14)), "anvil_top_damaged_0": (step(7),step(14)), "anvil_top_damaged_2": (step(8),step(14)), "log_spruce_top": (step(9),step(14)), "log_birch_top": (step(10),step(14)), "log_jungle_top": (step(11),step(14)), "log_big_oak_top": (step(12),step(14)), "lava_still": (step(13),step(14)), # # "double_plant_paeonia_bottom": (step(16),step(14)), "double_plant_paeonia_top": (step(17),step(14)), "nether_wart_block": (step(18),step(14)), # End Fifteenth Row # Start Sixteenth Row "planks_acacia": (step(0),step(15)), "planks_big_oak": (step(1),step(15)), # "log_acacia": (step(3),step(15)), "log_big_oak": (step(4),step(15)), "hardened_clay": (step(5),step(15)), "portal": (step(6),step(15)), # "quatrz_block_chiseled": (step(8),step(15)), "quartz_block_chiseled_top": (step(9),step(15)), "quartz_block_lines": (step(10),step(15)), "quartz_block_lines_top": (step(11),step(15)), # # # # # "slime": (step(17),step(15)), "red_nether_brick": (step(18),step(15)), # End Sixteenth Row # Start Seventeenth Row "ice_packed": (step(0),step(16)), "hay_block_side": (step(1),step(16)), "hay_block_top": (step(2),step(16)), "iron_trapdoor": (step(3),step(16)), "stone_granite": (step(4),step(16)), "stone_grantie_smooth": (step(5),step(16)), "stone_diorite": (step(6),step(16)), "stone_diorite_smooth": (step(7),step(16)), "stone_andesite": (step(8),step(16)), "stone_andesite_smooth": (step(9),step(16)), # # # # # # # # # "frosted_ice_0": (step(19), step(16)), # End Seventeenth Row # Start Eigteenth Row # # # # # # # # # # # # # # # # "prismarine_bricks": (step(16),step(17)), "prismarine_dark": (step(17),step(17)), "prismarine_rough": (step(18),step(17)), "purpur_pillar": (step(19),step(17)), # End Eigteenth Row # Start Ninteenth Row "hardened_clay_stained_white": (step(0),step(18)), "hardened_clay_stained_orange": (step(1),step(18)), "hardened_clay_stained_magenta": (step(2),step(18)), "hardened_clay_stained_light_blue": (step(3),step(18)), "hardened_clay_stained_yellow": (step(4),step(18)), "hardened_clay_stained_lime": (step(5),step(18)), "hardened_clay_stained_pink": (step(6),step(18)), "hardened_clay_stained_gray": (step(7),step(18)), "hardened_clay_stained_silver": (step(8),step(18)), "hardened_clay_stained_cyan": (step(9),step(18)), "hardened_clay_stained_purple": (step(10),step(18)), "hardened_clay_stained_blue": (step(11),step(18)), "hardened_clay_stained_brown": (step(12),step(18)), "hardened_clay_stained_green": (step(13),step(18)), "hardened_clay_stained_red": (step(14),step(18)), "hardened_clay_stained_black": (step(15),step(18)), "sponge_wet": (step(16),step(18)), "sea_lantern": (step(17),step(18)), "end_bricks": (step(18),step(18)), "purpur_pillar_top": (step(19),step(18)), # End Ninteenth Row # Start Twentieth Row # # # # # # # # # # # # # # # # "hay_block_side_rotated": (step(16),step(19)), "quartz_block_lines_rotated": (step(17),step(19)), "purpur_block": (step(18),step(19)), # End Twentieth Row # Start Twentyfirst Row "red_sandstone_bottom": (step(0),step(20)), "red_sandstone_carved": (step(1),step(20)), "red_sandstone_normal": (step(2),step(20)), "red_sandstone_smooth": (step(3),step(20)), "red_sandstone_top": (step(4),step(20)), "door_spruce_upper": (step(5),step(20)), "door_birch_upper": (step(6),step(20)), "door_jungle_upper": (step(7),step(20)), "door_acacia_upper": (step(8),step(20)), "door_dark_oak_upper": (step(9),step(20)), # # # # # # # # "chorus_plant": (step(13),step(20)), "chorus_flower_dead": (step(14),step(20)), "chorus_flower": (step(15),step(20)), "end_rod": (step(16),step(20)), # End Twentyfirst Row # Start MISC # Start More Bed Textures "bed_head_side_flipped": (step(1),step(21)), "bed_feet_side_flipped": (step(2),step(21)), "bed_head_top_flipped": (step(1),step(22)), "bed_feet_top_flipped": (step(2),step(22)), "bed_feet_top_bottom": (step(3),step(21)), "bed_head_top_bottom": (step(3),step(22)), "bed_feet_top_top": (step(4),step(22)), "bed_head_top_top": (step(4),step(21)), # End More Bed Textures # Start Comparator Block } class MultipartTexture(object): def __init__(self, texture_objects): self.subclasses = [] self.runAnyways = [] self.texture_dict = {} for subcls in self.__class__.__subclasses__(): # This is why I love Python self.subclasses.append(subcls) for cls in self.subclasses: instance = cls(texture_objects) if instance.runAnyway: self.runAnyways.append(instance) else: self.texture_dict[instance.target] = instance class LeverTexture(MultipartTexture): target = "lever" runAnyway = False def __init__(self, texture_objects): self.texture_objects = texture_objects def parse_texture(self): if "lever" not in self.texture_objects or "cobblestone" not in self.texture_objects: return None lever = self.texture_objects["lever"].copy() cobblestone = self.texture_objects["cobblestone"].copy() base_1 = cobblestone.crop((5, 4, 11, 12)) lever.paste(base_1, (10, 0, 16, 8)) base_2 = cobblestone.crop((5, 0, 11, 3)) lever.paste(base_2, (10, 8, 16, 11)) base_3 = cobblestone.crop((4, 0, 12, 3)) lever.paste(base_3, (2, 0, 10, 3)) return lever class StandingSignTexture(MultipartTexture): target = "" runAnyway = True position = (step(20), step(5)) def __init__(self, texture_objects): self.texture_objects = texture_objects def parse_texture(self): if "planks_oak" not in self.texture_objects or "log_oak" not in self.texture_objects: return None planks = self.texture_objects["planks_oak"].copy() log_tex = self.texture_objects["log_oak"].copy() sign = planks.crop((0, 7, 16, 16)) log_tex.paste(sign, (0, 7, 16, 16)) if log_tex.mode != "RGBA": log_tex = log_tex.convert("RGBA") return log_tex class IResourcePack(object): ''' Sets all base variables for a Resource Pack ''' def __init__(self): self.__stop = False texture_path = os.path.join(directories.parentDir, "textures", self._pack_name) self.texture_path = texture_path self._isEmpty = False self._too_big = False self.big_textures_counted = 0 self.big_textures_max = 10 self.block_image = {} self.propogated_textures = [] self.all_texture_slots = [] #self.old_terrain = Image.open(os.path.join(directories.getDataDir(), 'terrain.png')) self.old_terrain = Image.open(directories.getDataFile('terrain.png')) for texx in xrange(0,33): for texy in xrange(0,33): self.all_texture_slots.append((step(texx),step(texy))) self._terrain_name = self._pack_name.replace(" ", "_")+".png" #self._terrain_path = os.path.join("terrain-textures", self._terrain_name.replace(" ", "_")) self._terrain_path = directories.getDataFile(u'terrain-textures', self._terrain_name.replace(u' ', u'_')) @property def pack_name(self): ''' The name of the Resource Pack ''' return self._pack_name @property def terrain_name(self): ''' Name of the parsed texture PNG file ''' return self._terrain_name def terrain_path(self): ''' Path to the parsed PNG file ''' return self._terrain_path @property def isEmpty(self): ''' Returns true if the Resource Pack doesn't replace the minimum amount of textures ''' return self._isEmpty @property def tooBig(self): ''' Returns true if the Resource Pack has a greater resolution than 32x32 ''' return self._too_big def parse_terrain_png(self): ''' Parses each block texture into a usable PNG file like terrain.png ''' multiparts = MultipartTexture(self.block_image) log.debug("Parsing terrain.png") new_terrain = Image.new("RGBA", (512, 512), None) for tex in self.block_image.keys(): if not self.__stop and tex in textureSlots.keys(): try: if tex not in multiparts.texture_dict: image = self.block_image[tex] else: image = multiparts.texture_dict[tex].parse_texture() if image is None: continue log.debug(" Image is %s"%tex) log.debug(" Image mode: %s"%image.mode) if image.mode != "RGBA": try: image = image.convert("RGBA") log.debug(" Image converted to RGBA.") except Exception as ee: print "* * *", tex, ee slot = textureSlots[tex] try: new_terrain.paste(image, slot, image) except Exception as eee: print "* * * new_terrain error:", eee self.propogated_textures.append(slot) log.debug(" Image pasted and propagated.") except Exception as e: try: # Print the resource pack 'raw' name. print "An Exception occurred while trying to parse textures for {}".format(self._pack_name) log.debug("An Exception occurred while trying to parse textures for {}".format(self._pack_name)) except: # I for a reason it fails, print the 'representation' of it. print "An Exception occurred while trying to parse textures for {}".format(repr(self._pack_name)) log.debug("An Exception occurred while trying to parse textures for {}".format(repr(self._pack_name))) traceback.print_stack() print "Exception Message: "+str(e) log.debug("Exception Message: "+str(e)) print "Exception type: "+str(type(e)) log.debug("Exception type: "+str(type(e))) print e self.__stop = True self._isEmpty = True log.debug("Parsing stopped.") log.debug("Resource pack considered as empty.") pass for runAnyway in multiparts.runAnyways: parsed_texture = runAnyway.parse_texture() if parsed_texture is not None: new_terrain.paste(parsed_texture, runAnyway.position, parsed_texture) self.propogated_textures.append(runAnyway.position) copy = self.old_terrain.copy() log.debug("Correcting textures...") for t in self.all_texture_slots: if t not in self.propogated_textures: old_tex = copy.crop((t[0],t[1],t[0]+16,t[1]+16)) new_terrain.paste(old_tex, t, old_tex) log.debug(" Done.") log.debug("Saving %s."%self._terrain_path) new_terrain.save(self._terrain_path) log.debug(" Done.") try: os.remove(self._pack_name.replace(" ", "_")+".png") except: pass if not self.propogated_textures: os.remove(self._terrain_path) self._isEmpty = True log.debug("No propagated textures.\nTexture pack considered as empty.") #print u"{} did not replace any textures".format(self._pack_name) del self.block_image if hasattr(self, 'fps'): log.debug(" Closing file descriptors.") for f in self.fps: f.close() log.debug(" Done") log.debug("Parsing terrain.png ended.") def handle_too_big_packs(self): ''' Removes the parsed PNG file ''' self._too_big = True log.debug("Resource pack is too big.") try: os.remove(self._terrain_path) except: pass del self.block_image class ZipResourcePack(IResourcePack): ''' Represents a single Resource Pack that is in a zip file ''' def __init__(self, zipfileFile, noEncConvert=False): self.zipfile = zipfileFile log.debug("Zip file: %s"%zipfileFile) self._pack_name = os.path.splitext(os.path.split(zipfileFile)[-1])[0] log.debug("Pack name: %s"%self._pack_name) # Define a list of opened textures file objects to be cleaned when operations are finished. self.fps = [] IResourcePack.__init__(self) if not os.path.exists(self._terrain_path): try: self.open_pack() except Exception as e: if 'seek' not in e: print "Error while trying to load one of the resource packs: {}".format(e) def open_pack(self): ''' Opens the zip file and puts texture data into a dictionary, where the key is the texture file name, and the value is a PIL.Image instance ''' zfile = zipfile.ZipFile(self.zipfile) self.fps = [] for name in zfile.infolist(): if name.filename.endswith(".png") and not name.filename.split(os.path.sep)[-1].startswith("._"): filename = "assets/minecraft/textures/blocks" if name.filename.startswith(filename) and name.filename.replace(filename+"/", "").replace(".png","") in textureSlots: log.debug(" Is a possible texture.") block_name = os.path.normpath(name.filename).split(os.path.sep)[-1] block_name = block_name.split(".")[0] log.debug(" Block name: %s"%block_name) log.debug(" Opening %s"%name) fp = zfile.open(name) #!# Sending this 'fp' file descriptor to PIL.Image does not work, because such #!# descriptors are not seekable. #!# But, reading the fd data and writing it to a temporary file seem to work... log.debug(" Done. (%s, seekable: %s, readable: %s)"%(type(fp), fp.seekable(), fp.readable())) log.debug(" Saving fp data to temp file.") fp1 = StringIO() fp1.write(fp.read()) fp.close() fp1.seek(0) log.debug(" Done.") try: possible_texture = Image.open(fp1) log.debug(" File descriptor for %s opened."%block_name) log.debug(" Is %s."%repr(possible_texture.size)) except Exception as e: log.debug(" Can't open descriptor for %s"%block_name) log.debug(" System said:") log.debug(" %s"%repr(e)) if possible_texture.size == (16, 16): self.block_image[block_name] = possible_texture if block_name.startswith("repeater_") or block_name.startswith("comparator_"): self.block_image[block_name+"_west"] = possible_texture.rotate(-90) self.block_image[block_name+"_north"] = possible_texture.rotate(180) self.block_image[block_name+"_east"] = possible_texture.rotate(90) self.block_image[block_name+"_south"] = possible_texture if block_name == "piston_side": self.block_image["piston_up"] = possible_texture self.block_image["piston_left"] = possible_texture.rotate(90) self.block_image["piston_down"] = possible_texture.rotate(180) self.block_image["piston_right"] = possible_texture.rotate(-90) if block_name == "hay_block_side": self.block_image["hay_block_side_rotated"] = possible_texture.rotate(-90) if block_name == "quartz_block_lines": self.block_image["quartz_block_lines_rotated"] = possible_texture.rotate(-90) if block_name.startswith("bed_"): if block_name == "bed_head_side": self.block_image["bed_head_side_flipped"] = possible_texture.transpose(Image.FLIP_LEFT_RIGHT) if block_name == "bed_feet_side": self.block_image["bed_feet_side_flipped"] = possible_texture.transpose(Image.FLIP_LEFT_RIGHT) if block_name == "bed_head_top": self.block_image["bed_head_top_flipped"] = possible_texture.transpose(Image.FLIP_LEFT_RIGHT) self.block_image["bed_head_top_bottom"] = possible_texture.rotate(-90) self.block_image["bed_head_top_top"] = possible_texture.rotate(90) if block_name == "bed_feet_top": self.block_image["bed_feet_top_flipped"] = possible_texture.transpose(Image.FLIP_LEFT_RIGHT) self.block_image["bed_feet_top_bottom"] = possible_texture.rotate(-90) self.block_image["bed_feet_top_top"] = possible_texture.rotate(90) log.debug(" Is loaded.") else: if possible_texture.size == (32, 32): self.block_image[block_name] = possible_texture.resize((16, 16)) log.debug(" Is loaded.") elif possible_texture.size == (64, 64) or possible_texture.size == (128, 128) or possible_texture.size == (256, 256): self.big_textures_counted += 1 log.debug(" Is too big.") else: self.block_image[block_name] = possible_texture.crop((0,0,16,16)) log.debug(" Is loaded.") self.fps.append(fp1) if self.big_textures_counted >= self.big_textures_max: self.handle_too_big_packs() else: try: self.parse_terrain_png() except Exception: traceback.print_exc() log.warning('Encountered an Exception while parsing terrain texture') self._too_big = True pass class FolderResourcePack(IResourcePack): def __init__(self, folder, noEncConvert=False): self._folder = folder self._pack_name = self._folder.replace(" ", "_") IResourcePack.__init__(self) self._full_path = os.path.join(directories.getMinecraftProfileDirectory(directories.getSelectedProfile()), "resourcepacks", self._folder) self.texture_path = os.path.join(directories.parentDir, "textures", self._pack_name) if not os.path.exists(self._terrain_path): self.add_textures() def add_textures(self): ''' Scraps the block textures folder and puts texture data into a dictionary with exactly identical structure as ZipResourcePack ''' base_path = os.path.join(self._full_path, "assets", "minecraft", "textures", "blocks") if os.path.exists(base_path): files = os.listdir(base_path) for tex_file in files: if tex_file.endswith(".png") and not tex_file.startswith("._") and tex_file.replace(".png","") in textureSlots: possible_texture = Image.open(os.path.join(base_path, tex_file)) block_name = tex_file[:-4] if possible_texture.size == (16, 16): self.block_image[block_name] = possible_texture if block_name.startswith("repeater_") or block_name.startswith("comparator_"): self.block_image[block_name+"_west"] = possible_texture.rotate(-90) self.block_image[block_name+"_north"] = possible_texture.rotate(180) self.block_image[block_name+"_east"] = possible_texture.rotate(90) self.block_image[block_name+"_south"] = possible_texture if block_name == "piston_side": self.block_image["piston_up"] = possible_texture self.block_image["piston_left"] = possible_texture.rotate(90) self.block_image["piston_down"] = possible_texture.rotate(180) self.block_image["piston_right"] = possible_texture.rotate(-90) if block_name == "hay_block_side": self.block_image["hay_block_side_rotated"] = possible_texture.rotate(-90) if block_name == "quartz_block_lines": self.block_image["quartz_block_lines_rotated"] = possible_texture.rotate(-90) if block_name.startswith("bed_"): if block_name == "bed_head_side": self.block_image["bed_head_side_flipped"] = possible_texture.transpose(Image.FLIP_LEFT_RIGHT) if block_name == "bed_feet_side": self.block_image["bed_feet_side_flipped"] = possible_texture.transpose(Image.FLIP_LEFT_RIGHT) if block_name == "bed_head_top": self.block_image["bed_head_top_flipped"] = possible_texture.transpose(Image.FLIP_LEFT_RIGHT) self.block_image["bed_head_top_bottom"] = possible_texture.rotate(-90) self.block_image["bed_head_top_top"] = possible_texture.rotate(90) if block_name == "bed_feet_top": self.block_image["bed_feet_top_flipped"] = possible_texture.transpose(Image.FLIP_LEFT_RIGHT) self.block_image["bed_feet_top_bottom"] = possible_texture.rotate(-90) self.block_image["bed_feet_top_top"] = possible_texture.rotate(90) else: if possible_texture.size == (32, 32): self.block_image[block_name] = possible_texture.resize((16, 16)) if possible_texture.size == (64, 64) or possible_texture.size == (128, 128) or possible_texture.size == (256, 256): self.big_textures_counted += 1 else: self.block_image[block_name] = possible_texture.crop((0,0,16,16)) if self.big_textures_counted >= self.big_textures_max: self.handle_too_big_packs() else: try: self.parse_terrain_png() except Exception: traceback.print_exc() log.warning('Encountered an Exception while parsing terrain texture') self._too_big = True pass class DefaultResourcePack(IResourcePack): ''' Represents the default Resource Pack that is always present ''' def __init__(self): self._isEmpty = False self._too_big = False #self._terrain_path = os.path.join(directories.getDataDir(), "terrain.png") self._terrain_path = directories.getDataFile('terrain.png') self._pack_name = "Default" def terrain_path(self): return self._terrain_path @property def isEmpty(self): return self._isEmpty @property def tooBig(self): return self._too_big @Singleton class ResourcePackHandler: ''' A single point to manage which Resource Pack is being used and to provide the paths to each parsed PNG ''' Instance = None def setup_reource_packs(self): ''' Handles parsing of Resource Packs and removing ones that are either have to0 high of a resolution, or don't replace any textures ''' log.debug("Setting up the resource packs.") self._resource_packs = {} try: os.mkdir("terrain-textures") except OSError: pass self._resource_packs["Default Resource Pack"] = DefaultResourcePack() if os.path.exists(os.path.join(directories.getMinecraftProfileDirectory(directories.getSelectedProfile()), "resourcepacks")): log.debug("Gathering zipped packs...") zipResourcePacks = directories.getAllOfAFile(unicode(os.path.join(directories.getMinecraftProfileDirectory(directories.getSelectedProfile()), "resourcepacks")), ".zip") log.debug("Gatering folder packs...") folderResourcePacks = os.listdir(unicode(os.path.join(directories.getMinecraftProfileDirectory(directories.getSelectedProfile()), "resourcepacks"))) log.debug("Processing zipped packs...") for zip_tex_pack in zipResourcePacks: zrp = ZipResourcePack(zip_tex_pack) if not zrp.isEmpty: if not zrp.tooBig: self._resource_packs[zrp.pack_name] = zrp log.debug("Processing folder packs...") for folder_tex_pack in folderResourcePacks: if os.path.isdir(os.path.join(directories.getMinecraftProfileDirectory(directories.getSelectedProfile()), "resourcepacks", folder_tex_pack)): frp = FolderResourcePack(folder_tex_pack) if not frp.isEmpty: if not frp.tooBig: self._resource_packs[frp.pack_name] = frp for tex in self._resource_packs.keys(): pack = self._resource_packs[tex] if not os.path.exists(pack.terrain_path()): del self._resource_packs[tex] try: shutil.rmtree(os.path.join(directories.parentDir, "textures")) except: print "Could not remove \"textures\" directory" pass def __init__(self): try: os.mkdir(os.path.join(directories.parentDir, "textures")) except OSError: pass self.setup_reource_packs() self._selected_resource_pack = config.settings.resourcePack.get() if self._selected_resource_pack not in self._resource_packs.keys(): self.set_selected_resource_pack_name("Default Resource Pack") @property def resource_packs(self): ''' A dictionary of Resource Packs, where the key is the pack's file/folder name, and the value is the path to the parsed PNG ''' return self._resource_packs def get_available_resource_packs(self): ''' Returns the names of all the Resource Packs that can be used ''' return self._resource_packs.keys() def reload_resource_packs(self): ''' Reparses all Resource Packs ''' self.setup_resource_packs() def reparse_resource_pack(self, packName): if packName in self._resource_packs: pack = self._resource_packs[packName] if isinstance(pack, FolderResourcePack): pack.add_textures() elif isinstance(pack, ZipResourcePack): pack.open_pack() def get_selected_resource_pack_name(self): ''' Returns the currently selected Resource Pack's name ''' return self._selected_resource_pack def set_selected_resource_pack_name(self, name): ''' Sets the currently selected Resource Pack :param name: Name of the Resource Pack ''' config.settings.resourcePack.set(name) self._selected_resource_pack = name def get_selected_resource_pack(self): ''' Returns the selected Resource Pack instance. Can be an instance of either DefaultResourcePack, ZipResourcePack or FolderResourcePack ''' return self._resource_packs[self._selected_resource_pack]
42,430
37.963269
180
py
MCEdit-Unified
MCEdit-Unified-master/pymclevel/indev.py
""" Created on Jul 22, 2011 @author: Rio Indev levels: TAG_Compound "MinecraftLevel" { TAG_Compound "Environment" { TAG_Short "SurroundingGroundHeight"// Height of surrounding ground (in blocks) TAG_Byte "SurroundingGroundType" // Block ID of surrounding ground TAG_Short "SurroundingWaterHeight" // Height of surrounding water (in blocks) TAG_Byte "SurroundingWaterType" // Block ID of surrounding water TAG_Short "CloudHeight" // Height of the cloud layer (in blocks) TAG_Int "CloudColor" // Hexadecimal value for the color of the clouds TAG_Int "SkyColor" // Hexadecimal value for the color of the sky TAG_Int "FogColor" // Hexadecimal value for the color of the fog TAG_Byte "SkyBrightness" // The brightness of the sky, from 0 to 100 } TAG_List "Entities" { TAG_Compound { // One of these per entity on the map. // These can change a lot, and are undocumented. // Feel free to play around with them, though. // The most interesting one might be the one with ID "LocalPlayer", which contains the player inventory } } TAG_Compound "Map" { // To access a specific block from either byte array, use the following algorithm: // Index = x + (y * Depth + z) * Width TAG_Short "Width" // Width of the level (along X) TAG_Short "Height" // Height of the level (along Y) TAG_Short "Length" // Length of the level (along Z) TAG_Byte_Array "Blocks" // An array of Length*Height*Width bytes specifying the block types TAG_Byte_Array "Data" // An array of Length*Height*Width bytes with data for each blocks TAG_List "Spawn" // Default spawn position { TAG_Short x // These values are multiplied by 32 before being saved TAG_Short y // That means that the actual values are x/32.0, y/32.0, z/32.0 TAG_Short z } } TAG_Compound "About" { TAG_String "Name" // Level name TAG_String "Author" // Name of the player who made the level TAG_Long "CreatedOn" // Timestamp when the level was first created } } """ from entity import TileEntity from level import MCLevel from logging import getLogger from materials import indevMaterials from numpy import array, swapaxes import nbt import os log = getLogger(__name__) MinecraftLevel = "MinecraftLevel" Environment = "Environment" SurroundingGroundHeight = "SurroundingGroundHeight" SurroundingGroundType = "SurroundingGroundType" SurroundingWaterHeight = "SurroundingWaterHeight" SurroundingWaterType = "SurroundingWaterType" CloudHeight = "CloudHeight" CloudColor = "CloudColor" SkyColor = "SkyColor" FogColor = "FogColor" SkyBrightness = "SkyBrightness" About = "About" Name = "Name" Author = "Author" CreatedOn = "CreatedOn" Spawn = "Spawn" __all__ = ["MCIndevLevel"] from level import EntityLevel class MCIndevLevel(EntityLevel): """ IMPORTANT: self.Blocks and self.Data are indexed with [x,z,y] via axis swapping to be consistent with infinite levels.""" materials = indevMaterials _gamePlatform = 'indev' def setPlayerSpawnPosition(self, pos, player=None): assert len(pos) == 3 self.Spawn = array(pos) def playerSpawnPosition(self, player=None): return self.Spawn def setPlayerPosition(self, pos, player="Ignored"): self.LocalPlayer["Pos"] = nbt.TAG_List([nbt.TAG_Float(p) for p in pos]) def getPlayerPosition(self, player="Ignored"): return array(map(lambda x: x.value, self.LocalPlayer["Pos"])) def setPlayerOrientation(self, yp, player="Ignored"): self.LocalPlayer["Rotation"] = nbt.TAG_List([nbt.TAG_Float(p) for p in yp]) def getPlayerOrientation(self, player="Ignored"): """ returns (yaw, pitch) """ return array(map(lambda x: x.value, self.LocalPlayer["Rotation"])) def setBlockDataAt(self, x, y, z, newdata): if x < 0 or y < 0 or z < 0: return 0 if x >= self.Width or y >= self.Height or z >= self.Length: return 0 self.Data[x, z, y] = (newdata & 0xf) def blockDataAt(self, x, y, z): if x < 0 or y < 0 or z < 0: return 0 if x >= self.Width or y >= self.Height or z >= self.Length: return 0 return self.Data[x, z, y] def blockLightAt(self, x, y, z): if x < 0 or y < 0 or z < 0: return 0 if x >= self.Width or y >= self.Height or z >= self.Length: return 0 return self.BlockLight[x, z, y] def __repr__(self): return u"MCIndevLevel({0}): {1}W {2}L {3}H".format(self.filename, self.Width, self.Length, self.Height) @classmethod def _isTagLevel(cls, root_tag): return "MinecraftLevel" == root_tag.name def __init__(self, root_tag=None, filename=""): self.Width = 0 self.Height = 0 self.Length = 0 self.Blocks = array([], "uint8") self.Data = array([], "uint8") self.Spawn = (0, 0, 0) self.filename = filename if root_tag: self.root_tag = root_tag mapTag = root_tag["Map"] self.Width = mapTag["Width"].value self.Length = mapTag["Length"].value self.Height = mapTag["Height"].value mapTag["Blocks"].value.shape = (self.Height, self.Length, self.Width) self.Blocks = swapaxes(mapTag["Blocks"].value, 0, 2) mapTag["Data"].value.shape = (self.Height, self.Length, self.Width) self.Data = swapaxes(mapTag["Data"].value, 0, 2) self.BlockLight = self.Data & 0xf self.Data >>= 4 self.Spawn = [mapTag[Spawn][i].value for i in xrange(3)] if "Entities" not in root_tag: root_tag["Entities"] = nbt.TAG_List() self.Entities = root_tag["Entities"] # xxx fixup Motion and Pos to match infdev format def numbersToDoubles(ent): for attr in "Motion", "Pos": if attr in ent: ent[attr] = nbt.TAG_List([nbt.TAG_Double(t.value) for t in ent[attr]]) for ent in self.Entities: numbersToDoubles(ent) if "TileEntities" not in root_tag: root_tag["TileEntities"] = nbt.TAG_List() self.TileEntities = root_tag["TileEntities"] # xxx fixup TileEntities positions to match infdev format for te in self.TileEntities: pos = te["Pos"].value (x, y, z) = self.decodePos(pos) TileEntity.setpos(te, (x, y, z)) localPlayerList = [tag for tag in root_tag["Entities"] if tag['id'].value == 'LocalPlayer'] if len(localPlayerList) == 0: # omen doesn't make a player entity playerTag = nbt.TAG_Compound() playerTag['id'] = nbt.TAG_String('LocalPlayer') playerTag['Pos'] = nbt.TAG_List([nbt.TAG_Float(0.), nbt.TAG_Float(64.), nbt.TAG_Float(0.)]) playerTag['Rotation'] = nbt.TAG_List([nbt.TAG_Float(0.), nbt.TAG_Float(45.)]) self.LocalPlayer = playerTag else: self.LocalPlayer = localPlayerList[0] else: log.info(u"Creating new Indev levels is not yet implemented.!") raise ValueError("Can't do that yet") # self.SurroundingGroundHeight = root_tag[Environment][SurroundingGroundHeight].value # self.SurroundingGroundType = root_tag[Environment][SurroundingGroundType].value # self.SurroundingWaterHeight = root_tag[Environment][SurroundingGroundHeight].value # self.SurroundingWaterType = root_tag[Environment][SurroundingWaterType].value # self.CloudHeight = root_tag[Environment][CloudHeight].value # self.CloudColor = root_tag[Environment][CloudColor].value # self.SkyColor = root_tag[Environment][SkyColor].value # self.FogColor = root_tag[Environment][FogColor].value # self.SkyBrightness = root_tag[Environment][SkyBrightness].value # self.TimeOfDay = root_tag[Environment]["TimeOfDay"].value # # # self.Name = self.root_tag[About][Name].value # self.Author = self.root_tag[About][Author].value # self.CreatedOn = self.root_tag[About][CreatedOn].value def rotateLeft(self): MCLevel.rotateLeft(self) self.Data = swapaxes(self.Data, 1, 0)[:, ::-1, :] # x=y; y=-x torchRotation = array([0, 4, 3, 1, 2, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]) torchIndexes = (self.Blocks == self.materials.Torch.ID) log.info(u"Rotating torches: {0}".format(len(torchIndexes.nonzero()[0]))) self.Data[torchIndexes] = torchRotation[self.Data[torchIndexes]] @staticmethod def decodePos(v): b = 10 m = (1 << b) - 1 return v & m, (v >> b) & m, (v >> (2 * b)) @staticmethod def encodePos(x, y, z): b = 10 return x + (y << b) + (z << (2 * b)) def saveToFile(self, filename=None): if filename is None: filename = self.filename if filename is None: log.warn(u"Attempted to save an unnamed file in place") return # you fool! self.Data <<= 4 self.Data |= (self.BlockLight & 0xf) self.Blocks = swapaxes(self.Blocks, 0, 2) self.Data = swapaxes(self.Data, 0, 2) mapTag = nbt.TAG_Compound() mapTag["Width"] = nbt.TAG_Short(self.Width) mapTag["Height"] = nbt.TAG_Short(self.Height) mapTag["Length"] = nbt.TAG_Short(self.Length) mapTag["Blocks"] = nbt.TAG_Byte_Array(self.Blocks) mapTag["Data"] = nbt.TAG_Byte_Array(self.Data) self.Blocks = swapaxes(self.Blocks, 0, 2) self.Data = swapaxes(self.Data, 0, 2) mapTag[Spawn] = nbt.TAG_List([nbt.TAG_Short(i) for i in self.Spawn]) self.root_tag["Map"] = mapTag self.Entities.append(self.LocalPlayer) # fix up Entities imported from Alpha worlds def numbersToFloats(ent): for attr in "Motion", "Pos": if attr in ent: ent[attr] = nbt.TAG_List([nbt.TAG_Double(t.value) for t in ent[attr]]) for ent in self.Entities: numbersToFloats(ent) # fix up TileEntities imported from Alpha worlds. for ent in self.TileEntities: if "Pos" not in ent and all(c in ent for c in 'xyz'): ent["Pos"] = nbt.TAG_Int(self.encodePos(ent['x'].value, ent['y'].value, ent['z'].value)) # output_file = gzip.open(self.filename, "wb", compresslevel=1) try: os.rename(filename, filename + ".old") except Exception: pass try: self.root_tag.save(filename) except: os.rename(filename + ".old", filename) try: os.remove(filename + ".old") except Exception: pass self.Entities.remove(self.LocalPlayer) self.BlockLight = self.Data & 0xf self.Data >>= 4
11,565
34.697531
111
py
MCEdit-Unified
MCEdit-Unified-master/pymclevel/mclevel.py
# -*- coding: utf-8 -*- """ MCLevel interfaces Sample usage: import mclevel # Call mclevel.fromFile to identify and open any of these four file formats: # # Classic levels - gzipped serialized java objects. Returns an instance of MCJavalevel # Indev levels - gzipped NBT data in a single file. Returns an MCIndevLevel # Schematics - gzipped NBT data in a single file. Returns an MCSchematic. # MCSchematics have the special method rotateLeft which will reorient torches, stairs, and other tiles appropriately. # Alpha levels - world folder structure containing level.dat and chunk folders. Single or Multiplayer. # Can accept a path to the world folder or a path to the level.dat. Returns an MCInfdevOldLevel # Load a Classic level. level = mclevel.fromFile("server_level.dat"); # fromFile identified the file type and returned a MCJavaLevel. MCJavaLevel doesn't actually know any java. It guessed the # location of the Blocks array by starting at the end of the file and moving backwards until it only finds valid blocks. # It also doesn't know the dimensions of the level. This is why you have to tell them to MCEdit via the filename. # This works here too: If the file were 512 wide, 512 long, and 128 high, I'd have to name it "server_level_512_512_128.dat" # # This is one area for improvement. # Classic and Indev levels have all of their blocks in one place. blocks = level.Blocks # Sand to glass. blocks[blocks == level.materials.Sand.ID] = level.materials.Glass.ID # Save the file with another name. This only works for non-Alpha levels. level.saveToFile("server_level_glassy.dat"); # Load an Alpha world # Loading an Alpha world immediately scans the folder for chunk files. This takes longer for large worlds. ourworld = mclevel.fromFile("C:\\Minecraft\\OurWorld"); # Convenience method to load a numbered world from the saves folder. world1 = mclevel.loadWorldNumber(1); # Find out which chunks are present. Doing this will scan the chunk folders the # first time it is used. If you already know where you want to be, skip to # world1.getChunk(xPos, zPos) chunkPositions = list(world1.allChunks) # allChunks returns an iterator that yields a (xPos, zPos) tuple for each chunk xPos, zPos = chunkPositions[0]; # retrieve an AnvilChunk object. this object will load and decompress # the chunk as needed, and remember whether it needs to be saved or relighted chunk = world1.getChunk(xPos, zPos) ### Access the data arrays of the chunk like so: # Take note that the array is indexed x, z, y. The last index corresponds to # height or altitude. blockType = chunk.Blocks[0,0,64] chunk.Blocks[0,0,64] = 1 # Access the chunk's Entities and TileEntities as arrays of TAG_Compound as # they appear in the save format. # Entities usually have Pos, Health, and id # TileEntities usually have tileX, tileY, tileZ, and id # For more information, google "Chunk File Format" for entity in chunk.Entities: if entity["id"].value == "Spider": entity["Health"].value = 50 # Accessing one byte at a time from the Blocks array is very slow in Python. # To get around this, we have methods to access multiple bytes at once. # The first technique is slicing. You can use slicing to restrict your access # to certain depth levels, or to extract a column or a larger section from the # array. Standard python slice notation is used. # Set the top half of the array to 0. The : says to use the entire array along # that dimension. The syntax []= indicates we are overwriting part of the array chunk.Blocks[:,:,64:] = 0 # Using [] without = creates a 'view' on part of the array. This is not a # copy, it is a reference to a portion of the original array. midBlocks = chunk.Blocks[:,:,32:64] # Here's a gotcha: You can't just write 'midBlocks = 0' since it will replace # the 'midBlocks' reference itself instead of accessing the array. Instead, do # this to access and overwrite the array using []= syntax. midBlocks[:] = 0 # The second is masking. Using a comparison operator ( <, >, ==, etc ) # against the Blocks array will return a 'mask' that we can use to specify # positions in the array. # Create the mask from the result of the equality test. fireBlocks = ( chunk.Blocks==world.materials.Fire.ID ) # Access Blocks using the mask to set elements. The syntax is the same as # using []= with slices chunk.Blocks[fireBlocks] = world.materials.Leaves.ID # You can also combine mask arrays using logical operations (&, |, ^) and use # the mask to access any other array of the same shape. # Here we turn all trees into birch trees. # Extract a mask from the Blocks array to find the locations of tree trunks. # Or | it with another mask to find the locations of leaves. # Use the combined mask to access the Data array and set those locations to birch # Note that the Data, BlockLight, and SkyLight arrays have been # unpacked from 4-bit arrays to numpy uint8 arrays. This makes them much easier # to work with. treeBlocks = ( chunk.Blocks == world.materials.Wood.ID ) treeBlocks |= ( chunk.Blocks == world.materials.Leaves.ID ) chunk.Data[treeBlocks] = 2 # birch # The chunk doesn't know you've changed any of that data. Call chunkChanged() # to let it know. This will mark the chunk for lighting calculation, # recompression, and writing to disk. It will also immediately recalculate the # chunk's HeightMap and fill the SkyLight only with light falling straight down. # These are relatively fast and were added here to aid MCEdit. chunk.chunkChanged(); # To recalculate all of the dirty lights in the world, call generateLights world.generateLights(); # Move the player and his spawn world.setPlayerPosition( (0, 67, 0) ) # add 3 to make sure his head isn't in the ground. world.setPlayerSpawnPosition( (0, 64, 0) ) # Save the level.dat and any chunks that have been marked for writing to disk # This also compresses any chunks marked for recompression. world.saveInPlace(); # Advanced use: # The getChunkSlices method returns an iterator that returns slices of chunks within the specified range. # the slices are returned as tuples of (chunk, slices, point) # chunk: The AnvilChunk object we're interested in. # slices: A 3-tuple of slice objects that can be used to index chunk's data arrays # point: A 3-tuple of floats representing the relative position of this subslice within the larger slice. # # Take caution: # the point tuple is ordered (x,y,z) in accordance with the tuples used to initialize a bounding box # however, the slices tuple is ordered (x,z,y) for easy indexing into the arrays. # Here is an old version of MCInfdevOldLevel.fillBlocks in its entirety: def fillBlocks(self, box, blockType, blockData = 0): chunkIterator = self.getChunkSlices(box) for (chunk, slices, point) in chunkIterator: chunk.Blocks[slices] = blockType chunk.Data[slices] = blockData chunk.chunkChanged(); Copyright 2010 David Rio Vierra """ from indev import MCIndevLevel from infiniteworld import MCInfdevOldLevel from javalevel import MCJavaLevel from logging import getLogger import logging from directories import minecraftSaveFileDir import nbt from numpy import fromstring import os from pocket import PocketWorld from leveldbpocket import PocketLeveldbWorld from pymclevel import leveldbpocket from schematic import INVEditChest, MCSchematic, ZipSchematic import sys import traceback log = getLogger(__name__) class LoadingError(RuntimeError): pass def fromFile(filename, loadInfinite=True, readonly=False): ''' The preferred method for loading Minecraft levels of any type. pass False to loadInfinite if you'd rather not load infdev levels. ''' log.info(u"Identifying " + filename) if not filename: raise IOError("File not found: " + filename) if not os.path.exists(filename): raise IOError("File not found: " + filename) if ZipSchematic._isLevel(filename): log.info("Zipfile found, attempting zipped infinite level") lev = ZipSchematic(filename) log.info("Detected zipped Infdev level") return lev if PocketWorld._isLevel(filename): return PocketWorld(filename) if MCInfdevOldLevel._isLevel(filename): log.info(u"Detected Infdev level.dat") if loadInfinite: return MCInfdevOldLevel(filename=filename, readonly=readonly) else: raise ValueError("Asked to load {0} which is an infinite level, loadInfinite was False".format( os.path.basename(filename))) if PocketLeveldbWorld._isLevel(filename): if leveldbpocket.leveldb_available: return PocketLeveldbWorld(filename) else: logging.exception("Pocket support has failed") if os.path.isdir(filename): logging.exception("World load failed, trying to open a directory instead of a file") f = file(filename, 'rb') rawdata = f.read() f.close() if len(rawdata) < 4: raise ValueError("{0} is too small! ({1}) ".format(filename, len(rawdata))) data = fromstring(rawdata, dtype='uint8') if not data.any(): raise ValueError("{0} contains only zeroes. This file is damaged beyond repair.") if MCJavaLevel._isDataLevel(data): log.info(u"Detected Java-style level") lev = MCJavaLevel(filename, data) lev.compressed = False return lev # ungzdata = None compressed = True unzippedData = None try: unzippedData = nbt.gunzip(rawdata) except Exception as e: log.info(u"Exception during Gzip operation, assuming {0} uncompressed: {1!r}".format(filename, e)) if unzippedData is None: compressed = False unzippedData = rawdata #data = data = unzippedData if MCJavaLevel._isDataLevel(data): log.info(u"Detected compressed Java-style level") lev = MCJavaLevel(filename, data) lev.compressed = compressed return lev try: root_tag = nbt.load(buf=data) except Exception as e: log.info(u"Error during NBT load: {0!r}".format(e)) log.info(traceback.format_exc()) log.info(u"Fallback: Detected compressed flat block array, yzx ordered ") try: lev = MCJavaLevel(filename, data) lev.compressed = compressed return lev except Exception as e2: raise LoadingError(("Multiple errors encountered", e, e2), sys.exc_info()[2]) else: if MCIndevLevel._isTagLevel(root_tag): log.info(u"Detected Indev .mclevel") return MCIndevLevel(root_tag, filename) if MCSchematic._isTagLevel(root_tag): log.info(u"Detected Schematic.") return MCSchematic(filename=filename) if INVEditChest._isTagLevel(root_tag): log.info(u"Detected INVEdit inventory file") return INVEditChest(root_tag=root_tag, filename=filename) raise IOError("Cannot detect file type.") def loadWorld(name): filename = os.path.join(minecraftSaveFileDir, name) return fromFile(filename) def loadWorldNumber(i): # deprecated filename = u"{0}{1}{2}{3}{1}".format(minecraftSaveFileDir, os.sep, u"World", i) return fromFile(filename)
11,260
35.800654
125
py
MCEdit-Unified
MCEdit-Unified-master/stock-filters/Find.py
# written by texelelf #-# Adding a result pages, and NBT edit stuff from pymclevel import TAG_Byte, TAG_Short, TAG_Int, TAG_Compound, TAG_List, TAG_String, TAG_Double, TAG_Float, TAG_Long, \ TAG_Byte_Array, TAG_Int_Array from pymclevel.box import BoundingBox from albow import alert, ask import ast # Let import the stuff to save files. from mcplatform import askSaveFile from directories import getDocumentsFolder # The RECODR_UNDO is not yet usable... # RECORD_UNDO = False displayName = "Find" tagtypes = {"TAG_Byte": 0, "TAG_Short": 1, "TAG_Int": 2, "TAG_Compound": 3, "TAG_List": 4, "TAG_String": 5, "TAG_Double": 6, "TAG_Float": 7, "TAG_Long": 8, "TAG_Byte_Array": 9, "TAG_Int_Array": 10} tagses = {0: TAG_Byte, 1: TAG_Short, 2: TAG_Int, 3: TAG_Compound, 4: TAG_List, 5: TAG_String, 6: TAG_Double, 7: TAG_Float, 8: TAG_Long, 9: TAG_Byte_Array, 10: TAG_Int_Array, "Any": 11} inputs = [(("Match by:", ("TileEntity", "Entity", "Block")), ("Match block type (for TileEntity searches):", False), ("Match block:", "blocktype"), ("Match block data:", True), ("Match tile entities (for Block searches):", False), ("\"None\" for Name or Value will match any tag's Name or Value respectively.", "label"), ("Match Tag Name:", ("string", "value=None")), ("Match Tag Value:", ("string", "value=None")), ("Case insensitive:", True), ("Match Tag Type:", tuple(("Any",)) + tuple(tagtypes.keys())), ("Operation:", ("Start New Search", "Dump Found Coordinates")), ("Options", "title")), (("Results", "title"), ("", ["NBTTree", {}, 0, False])), # [str name_of_widget_type, dict default_data, int page_to_goback, bool show_load_button] (("Documentation", "title"), ("This filter is designed to search for NBT in either Entities or TileEntities.\n" "It can also be used to search for blocks.\n\"Match by\" determines which type of object " "is prioritized during the search.\nEntites and TileEntities will search relatively quickly, " "while the speed of searching by Block will be directly proportional to the selection size (since every " "single block within the selection will be examined).\n" "All Entity searches will ignore the block settings; TileEntity searches will try to " "\"Match block type\" if checked, and Block searches will try to \"Match tile entity\" tags if " "checked.\nIt is faster to match TileEntity searches with a Block Type than vice versa.\nBlock " "matching can also optionally match block data, e.g. matching all torches, or only torches facing " "a specific direction.\n\"Start New Search\" will re-search through the selected volume, while \"Find Next\" " "will iterate through the search results of the previous search.", "label")) ] tree = None # the tree widget chunks = None # the chunks used to perform the search bbox = None # the bouding box to search in by = None # what is searched: Entities, TileEntities or blocs def set_tree(t): global tree tree = t # Use this method to overwrite the NBT tree default behaviour on mouse clicks def nbttree_mouse_down(e): if e.num_clicks > 1: if tree.selected_item and tree.selected_item[3].startswith('(') and tree.selected_item[3].endswith(')'): s = ast.literal_eval(tree.selected_item[3]) editor.mainViewport.cameraPosition = (s[0] + 0.5, s[1] + 2, s[2] - 1) editor.mainViewport.yaw = 0.0 editor.mainViewport.pitch = 45.0 newBox = BoundingBox(s, (1, 1, 1)) editor.selectionTool.setSelection(newBox) tree.treeRow.__class__.mouse_down(tree.treeRow, e) def get_chunks(): return chunks def get_box(): return bbox def get_by(): return by # Use this method to overwrite the NBT tree 'OK' button default behaviour def nbt_ok_action(): by = get_by() chunks = get_chunks() box = get_box() if by not in (trn._('TileEntity'), trn._('Entity')): return if chunks: for chunk, slices, point in chunks: if by == trn._('TileEntity'): for e in chunk.TileEntities: x = e["x"].value y = e["y"].value z = e["z"].value elif by == trn._('Entity'): for e in chunk.Entities: x = e["Pos"][0].value y = e["Pos"][1].value z = e["Pos"][2].value if (x, y, z) in box: chunk.dirty = True try: search except NameError: search = None def FindTagS(nbtData, name, value, tagtype): if type(nbtData) is TAG_List or type(nbtData) is TAG_Compound: if name in nbtData.name or name == "": if value == "": if type(nbtData) is tagtype or tagtype == 11: print "found in pre-area" return True if type(nbtData) is TAG_List: list = True else: list = False for tag in range(0, len(nbtData)) if list else nbtData.keys(): if type(nbtData[tag]) is TAG_Compound: if FindTagS(nbtData[tag], name, value, tagtype): return True elif type(nbtData[tag]) is TAG_List: if FindTagS(nbtData[tag], name, value, tagtype): return True else: if name in nbtData[tag].name or name == "": if value in unicode(nbtData[tag].value): if type(nbtData[tag]) is tagtype or tagtype == 11: print "found in list/compound" return True else: return False else: if name in nbtData.name or name == "": if value in unicode(nbtData.value): if type(nbtData[tag]) is tagtype or tagtype == 11: print "found outside" return True return False def FindTagI(nbtData, name, value, tagtype): if type(nbtData) is TAG_List or type(nbtData) is TAG_Compound: if name in (u"%s"%nbtData.name).upper() or name == "": if value == "": if type(nbtData) is tagtype or tagtype == 11: return True if type(nbtData) is TAG_List: list = True else: list = False for tag in range(0, len(nbtData)) if list else nbtData.keys(): if type(nbtData[tag]) is TAG_Compound: if FindTagI(nbtData[tag], name, value, tagtype): return True elif type(nbtData[tag]) is TAG_List: if FindTagI(nbtData[tag], name, value, tagtype): return True else: if name in (u"%s"%nbtData[tag].name).upper() or name == "": if value in unicode(nbtData[tag].value).upper(): if type(nbtData[tag]) is tagtype or tagtype == 11: return True else: return False else: if name in (u"%s"%nbtData.name).upper() or name == "": if value in unicode(nbtData.value).upper(): if type(nbtData[tag]) is tagtype or tagtype == 11: return True return False def FindTag(nbtData, name, value, tagtype, caseSensitive): if caseSensitive: return FindTagS(nbtData, name, value, tagtype) else: return FindTagI(nbtData, name, value, tagtype) def perform(level, box, options): global search # Don't forget to 'globalize' these... global chunks global bbox global by bbox = box by = options["Match by:"] matchtype = options["Match block type (for TileEntity searches):"] matchblock = options["Match block:"] matchdata = options["Match block data:"] matchtile = options["Match tile entities (for Block searches):"] matchname = u"" if options["Match Tag Name:"] == "None" else unicode(options["Match Tag Name:"]) matchval = u"" if options["Match Tag Value:"] == "None" else unicode(options["Match Tag Value:"]) caseSensitive = not options["Case insensitive:"] matchtagtype = tagtypes.get(options["Match Tag Type:"], "Any") op = options["Operation:"] datas = [] if not caseSensitive: matchname = matchname.upper() matchval = matchval.upper() if matchtile and matchname == "" and matchval == "": alert("\nInvalid Tag Name and Value; the present values will match every tag of the specified type.") if search is None or op == trn._("Start New Search") or op == trn._("Dump Found Coordinates"): search = [] if not search: if by == trn._("Block"): for x in xrange(box.minx, box.maxx): for z in xrange(box.minz, box.maxz): for y in xrange(box.miny, box.maxy): block = level.blockAt(x, y, z) data = level.blockDataAt(x, y, z) if block == matchblock.ID and (not matchdata or data == matchblock.blockData): pass else: continue if matchtile: tile = level.tileEntityAt(x, y, z) if tile is not None: if not FindTag(tile, matchname, matchval, tagses[matchtagtype], caseSensitive): continue else: continue search.append((x, y, z)) datas.append(data) elif by == trn._("TileEntity"): chunks = [] for (chunk, slices, point) in level.getChunkSlices(box): for e in chunk.TileEntities: x = e["x"].value y = e["y"].value z = e["z"].value if (x, y, z) in box: if matchtype: block = level.blockAt(x, y, z) data = level.blockDataAt(x, y, z) if block == matchblock.ID and (not matchdata or data == matchblock.blockData): pass else: continue if not FindTag(e, matchname, matchval, tagses[matchtagtype], caseSensitive): continue search.append((x, y, z)) datas.append(e) chunks.append([chunk, slices, point]) else: chunks = [] for (chunk, slices, point) in level.getChunkSlices(box): for e in chunk.Entities: x = e["Pos"][0].value y = e["Pos"][1].value z = e["Pos"][2].value if (x, y, z) in box: if FindTag(e, matchname, matchval, tagses[matchtagtype], caseSensitive): search.append((x, y, z)) datas.append(e) chunks.append([chunk, slices, point]) if not search: alert("\nNo matching blocks/tile entities found") else: search.sort() if op == trn._("Dump Found Coordinates"): result = "\n".join("%d, %d, %d" % pos for pos in search) answer = ask(result, height=editor.height, colLabel="Matching Coordinates", responses=["Save", "OK"]) if answer == "Save": fName = askSaveFile(getDocumentsFolder(), "Save to file...", "find.txt", 'TXT\0*.txt\0\0', 'txt') if fName: fData = "# MCEdit find output\n# Search options:\n# Match by: %s\n# Match block type: %s\n# Match block: %s\n# Match block data: %s\n# Match tile entities: %s\n# Match Tag Name:%s\n# Match Tag Value: %s\n# Case insensitive: %s\n# Match Tag Type: %s\n\n%s"%(by, matchtype, matchblock, matchdata, matchtile, matchname, matchval, caseSensitive, matchtagtype, result) open(fName, 'w').write(fData) else: treeData = {} # To set tooltip text to the items the need it, use a dict: {"value": <item to be added to the tree>, "tooltipText": "Some text"} for i in range(len(search)): if by == trn._('Block'): treeData[u"%s"%(search[i],)] = {"value": datas[i], "tooltipText": "Double-click to go to this item."} elif by == trn._('Entity'): treeData[u"%s"%((datas[i]['Pos'][0].value, datas[i]['Pos'][1].value, datas[i]['Pos'][2].value),)] = {"value": datas[i], "tooltipText": "Double-click to go to this item."} else: treeData[u"%s"%((datas[i]['x'].value, datas[i]['y'].value, datas[i]['z'].value),)] = {"value": datas[i], "tooltipText": "Double-click to go to this item."} inputs[1][1][1][1] = {'Data': treeData} options[""](inputs[1])
13,288
43.89527
383
py
MCEdit-Unified
MCEdit-Unified-master/stock-filters/Forester.py
# Version 5 '''This takes a base MineCraft level and adds or edits trees. Place it in the folder where the save files are (usually .../.minecraft/saves) Requires mcInterface.py in the same folder.''' # Here are the variables you can edit. # This is the name of the map to edit. # Make a backup if you are experimenting! LOADNAME = "LevelSave" # How many trees do you want to add? TREECOUNT = 12 # Where do you want the new trees? # X, and Z are the map coordinates X = 66 Z = -315 # How large an area do you want the trees to be in? # for example, RADIUS = 10 will make place trees randomly in # a circular area 20 blocks wide. RADIUS = 80 # NOTE: tree density will be higher in the center than at the edges. # Which shapes would you like the trees to be? # these first three are best suited for small heights, from 5 - 10 # "normal" is the normal minecraft shape, it only gets taller and shorter # "bamboo" a trunk with foliage, it only gets taller and shorter # "palm" a trunk with a fan at the top, only gets taller and shorter # "stickly" selects randomly from "normal", "bamboo" and "palm" # these last five are best suited for very large trees, heights greater than 8 # "round" procedural spherical shaped tree, can scale up to immense size # "cone" procedural, like a pine tree, also can scale up to immense size # "procedural" selects randomly from "round" and "conical" # "rainforest" many slender trees, most at the lower range of the height, # with a few at the upper end. # "mangrove" makes mangrove trees (see PLANTON below). SHAPE = "procedural" # What height should the trees be? # Specifies the average height of the tree # Examples: # 5 is normal minecraft tree # 3 is minecraft tree with foliage flush with the ground # 10 is very tall trees, they will be hard to chop down # NOTE: for round and conical, this affects the foliage size as well. # CENTERHEIGHT is the height of the trees at the center of the area # ie, when radius = 0 CENTERHEIGHT = 55 # EDGEHEIGHT is the height at the trees at the edge of the area. # ie, when radius = RADIUS EDGEHEIGHT = 25 # What should the variation in HEIGHT be? # actual value +- variation # default is 1 # Example: # HEIGHT = 8 and HEIGHTVARIATION = 3 will result in # trunk heights from 5 to 11 # value is clipped to a max of HEIGHT # for a good rainforest, set this value not more than 1/2 of HEIGHT HEIGHTVARIATION = 12 # Do you want branches, trunk, and roots? # True makes all of that # False does not create the trunk and branches, or the roots (even if they are # enabled further down) WOOD = True # Trunk thickness multiplyer # from zero (super thin trunk) to whatever huge number you can think of. # Only works if SHAPE is not a "stickly" subtype # Example: # 1.0 is the default, it makes decently normal sized trunks # 0.3 makes very thin trunks # 4.0 makes a thick trunk (good for HOLLOWTRUNK). # 10.5 will make a huge thick trunk. Not even kidding. Makes spacious # hollow trunks though! TRUNKTHICKNESS = 1.0 # Trunk height, as a fraction of the tree # Only works on "round" shaped trees # Sets the height of the crown, where the trunk ends and splits # Examples: # 0.7 the default value, a bit more than half of the height # 0.3 good for a fan-like tree # 1.0 the trunk will extend to the top of the tree, and there will be no crown # 2.0 the trunk will extend out the top of the foliage, making the tree appear # like a cluster of green grapes impaled on a spike. TRUNKHEIGHT = 0.7 # Do you want the trunk and tree broken off at the top? # removes about half of the top of the trunk, and any foliage # and branches that would attach above it. # Only works if SHAPE is not a "stickly" subtype # This results in trees that are shorter than the height settings # True does that stuff # False makes a normal tree (default) BROKENTRUNK = False # Note, this works well with HOLLOWTRUNK (below) turned on as well. # Do you want the trunk to be hollow (or filled) inside? # Only works with larger sized trunks. # Only works if SHAPE is not a "stickly" subtype # True makes the trunk hollow (or filled with other stuff) # False makes a solid trunk (default) HOLLOWTRUNK = False # Note, this works well with BROKENTRUNK set to true (above) # Further note, you may want to use a large value for TRUNKTHICKNESS # How many branches should there be? # General multiplyer for the number of branches # However, it will not make more branches than foliage clusters # so to garuntee a branch to every foliage cluster, set it very high, like 10000 # this also affects the number of roots, if they are enabled. # Examples: # 1.0 is normal # 0.5 will make half as many branches # 2.0 will make twice as mnay branches # 10000 will make a branch to every foliage cluster (I'm pretty sure) BRANCHDENSITY = 1.0 # do you want roots from the bottom of the tree? # Only works if SHAPE is "round" or "cone" or "procedural" # "yes" roots will penetrate anything, and may enter underground caves. # "tostone" roots will be stopped by stone (default see STOPSROOTS below). # There may be some penetration. # "hanging" will hang downward in air. Good for "floating" type maps # (I really miss "floating" terrain as a default option) # "no" roots will not be generated ROOTS = "tostone" # Do you want root buttresses? # These make the trunk not-round at the base, seen in tropical or old trees. # This option generally makes the trunk larger. # Only works if SHAPE is "round" or "cone" or "procedural" # Options: # True makes root butresses # False leaves them out ROOTBUTTRESSES = True # Do you want leaves on the trees? # True there will be leaves # False there will be no leaves FOLIAGE = True # How thick should the foliage be # General multiplyer for the number of foliage clusters # Examples: # 1.0 is normal # 0.3 will make very sparse spotty trees, half as many foliage clusters # 2.0 will make dense foliage, better for the "rainforests" SHAPE FOLIAGEDENSITY = 1.0 # Limit the tree height to the top of the map? # True the trees will not grow any higher than the top of the map # False the trees may be cut off by the top of the map MAPHEIGHTLIMIT = True # add lights in the middle of foliage clusters # for those huge trees that get so dark underneath # or for enchanted forests that should glow and stuff # Only works if SHAPE is "round" or "cone" or "procedural" # 0 makes just normal trees # 1 adds one light inside the foliage clusters for a bit of light # 2 adds two lights around the base of each cluster, for more light # 4 adds lights all around the base of each cluster for lots of light LIGHTTREE = 0 # Do you want to only place trees near existing trees? # True will only plant new trees near existing trees. # False will not check for existing trees before planting. # NOTE: the taller the tree, the larger the forest needs to be to qualify # OTHER NOTE: this feature has not been extensively tested. # IF YOU HAVE PROBLEMS: SET TO False ONLYINFORESTS = False ##################### # Advanced options! # ##################### # What kind of material should the "wood" be made of? # defaults to 17 WOODMAT = 17 # What data value should the wood blocks have? # Some blocks, like wood, leaves, and cloth change # apperance with different data values # defaults to 0 WOODDATA = 0 # What kind of material should the "leaves" be made of? # defaults to 18 LEAFMAT = 18 # What data value should the leaf blocks have? # Some blocks, like wood, leaves, and cloth change # apperance with different data values # defaults to 0 LEAFDATA = 0 # What kind of material should the "lights" be made of? # defaults to 89 (glowstone) LIGHTMAT = 89 # What data value should the light blocks have? # defaults to 0 LIGHTDATA = 0 # What kind of material would you like the "hollow" trunk filled with? # defaults to 0 (air) TRUNKFILLMAT = 0 # What data value would you like the "hollow" trunk filled with? # defaults to 0 TRUNKFILLDATA = 0 # What kind of blocks should the trees be planted on? # Use the Minecraft index. # Examples # 2 is grass (the default) # 3 is dirt # 1 is stone (an odd choice) # 12 is sand (for beach or desert) # 9 is water (if you want an aquatic forest) # this is a list, and comma seperated. # example: [2, 3] # will plant trees on grass or dirt PLANTON = [2] # What kind of blocks should stop the roots? # a list of block id numbers like PLANTON # Only works if ROOTS = "tostone" # default, [1] (stone) # if you want it to be stopped by other block types, add it to the list STOPSROOTS = [1] # What kind of blocks should stop branches? # same as STOPSROOTS above, but is always turned on # defaults to stone, cobblestone, and glass # set it to [] if you want branches to go through everything STOPSBRANCHES = [1, 4, 20] # How do you want to interpolate from center to edge? # "linear" makes a cone-shaped forest # This is the only option at present INTERPOLATION = "linear" # Do a rough recalculation of the lighting? # Slows it down to do a very rough and incomplete re-light. # If you want to really fix the lighting, use a seperate re-lighting tool. # True do the rough fix # False don't bother LIGHTINGFIX = True # How many times do you want to try to find a location? # it will stop planing after MAXTRIES has been exceeded. # Set to smaller numbers to abort quicker, or larger numbers # if you want to keep trying for a while. # NOTE: the number of trees will not exceed this number # Default: 1000 MAXTRIES = 1000 # Do you want lots of text telling you waht is going on? # True lots of text (default). Good for debugging. # False no text VERBOSE = True ############################################################## # Don't edit below here unless you know what you are doing # ############################################################## # input filtering TREECOUNT = int(TREECOUNT) if TREECOUNT < 0: TREECOUNT = 0 if SHAPE not in ["normal", "bamboo", "palm", "stickly", "round", "cone", "procedural", "rainforest", "mangrove"]: if VERBOSE: print("SHAPE not set correctly, using 'procedural'.") SHAPE = "procedural" if CENTERHEIGHT < 1: CENTERHEIGHT = 1 if EDGEHEIGHT < 1: EDGEHEIGHT = 1 minheight = min(CENTERHEIGHT, EDGEHEIGHT) if HEIGHTVARIATION > minheight: HEIGHTVARIATION = minheight if INTERPOLATION not in ["linear"]: if VERBOSE: print("INTERPOLATION not set correctly, using 'linear'.") INTERPOLATION = "linear" if WOOD not in [True, False]: if VERBOSE: print("WOOD not set correctly, using True") WOOD = True if TRUNKTHICKNESS < 0.0: TRUNKTHICKNESS = 0.0 if TRUNKHEIGHT < 0.0: TRUNKHEIGHT = 0.0 if ROOTS not in ["yes", "tostone", "hanging", "no"]: if VERBOSE: print("ROOTS not set correctly, using 'no' and creating no roots") ROOTS = "no" if ROOTBUTTRESSES not in [True, False]: if VERBOSE: print("ROOTBUTTRESSES not set correctly, using False") ROOTBUTTRESSES = False if FOLIAGE not in [True, False]: if VERBOSE: print("FOLIAGE not set correctly, using True") ROOTBUTTRESSES = True if FOLIAGEDENSITY < 0.0: FOLIAGEDENSITY = 0.0 if BRANCHDENSITY < 0.0: BRANCHDENSITY = 0.0 if MAPHEIGHTLIMIT not in [True, False]: if VERBOSE: print("MAPHEIGHTLIMIT not set correctly, using False") MAPHEIGHTLIMIT = False if LIGHTTREE not in [0, 1, 2, 4]: if VERBOSE: print("LIGHTTREE not set correctly, using 0 for no torches") LIGHTTREE = 0 # assemble the material dictionaries WOODINFO = {'B': WOODMAT, 'D': WOODDATA} LEAFINFO = {'B': LEAFMAT, 'D': LEAFDATA} LIGHTINFO = {'B': LIGHTMAT, 'D': LIGHTDATA} TRUNKFILLINFO = {'B': TRUNKFILLMAT, 'D': TRUNKFILLDATA} # The following is an interface class for .mclevel data for minecraft savefiles. # The following also includes a useful coordinate to index convertor and several # other useful functions. import mcInterface #some handy functions def dist_to_mat(cord, vec, matidxlist, mcmap, invert=False, limit=False): '''travel from cord along vec and return how far it was to a point of matidx the distance is returned in number of iterations. If the edge of the map is reached, then return the number of iterations as well. if invert == True, search for anything other than those in matidxlist ''' assert isinstance(mcmap, mcInterface.SaveFile) block = mcmap.block curcord = [i + .5 for i in cord] iterations = 0 on_map = True while on_map: x = int(curcord[0]) y = int(curcord[1]) z = int(curcord[2]) return_dict = block(x, y, z) if return_dict is None: break else: block_value = return_dict['B'] if (block_value in matidxlist) and (invert is False): break elif (block_value not in matidxlist) and invert: break else: curcord = [curcord[i] + vec[i] for i in range(3)] iterations += 1 if limit and iterations > limit: break return iterations # This is the end of the MCLevel interface. # Now, on to the actual code. from random import random, choice, sample from math import sqrt, sin, cos, pi def calc_column_lighting(x, z, mclevel): '''Recalculate the sky lighting of the column.''' # Begin at the top with sky light level 15. cur_light = 15 # traverse the column until cur_light == 0 # and the existing light values are also zero. y = 255 get_block = mclevel.block set_block = mclevel.set_block get_height = mclevel.retrieve_heightmap set_height = mclevel.set_heightmap #get the current heightmap cur_height = get_height(x, z) # set a flag that the highest point has been updated height_updated = False # if this doesn't exist, the block doesn't exist either, abort. if cur_height is None: return None light_reduction_lookup = {0: 0, 20: 0, 18: 1, 8: 2, 79: 2} while True: #get the block sky light and type block_info = get_block(x, y, z, 'BS') block_light = block_info['S'] block_type = block_info['B'] # update the height map if it hasn't been updated yet, # and the current block reduces light if (not height_updated) and (block_type not in (0, 20)): new_height = y + 1 if new_height == 256: new_height = 255 set_height(x, new_height, z) height_updated = True #compare block with cur_light, escape if both 0 if block_light == 0 and cur_light == 0: break #set the block light if necessary if block_light != cur_light: set_block(x, y, z, {'S': cur_light}) #set the new cur_light if block_type in light_reduction_lookup: # partial light reduction light_reduction = light_reduction_lookup[block_type] else: # full light reduction light_reduction = 16 cur_light += -light_reduction if cur_light < 0: cur_light = 0 #increment and check y y += -1 if y < 0: break class ReLight(object): '''keep track of which squares need to be relit, and then relight them''' def add(self, x, z): coords = (x, z) self.all_columns.add(coords) def calc_lighting(self): mclevel = self.save_file for column_coords in self.all_columns: # recalculate the lighting x = column_coords[0] z = column_coords[1] calc_column_lighting(x, z, mclevel) def __init__(self): self.all_columns = set() self.save_file = None relight_master = ReLight() def assign_value(x, y, z, values, save_file): '''Assign an index value to a location in mcmap. If the index is outside the bounds of the map, return None. If the assignment succeeds, return True. ''' if y > 255: return None result = save_file.set_block(x, y, z, values) if LIGHTINGFIX: relight_master.add(x, z) return result class Tree(object): '''Set up the interface for tree objects. Designed for subclassing. ''' def prepare(self, mcmap): '''initialize the internal values for the Tree object. ''' return None def maketrunk(self, mcmap): '''Generate the trunk and enter it in mcmap. ''' return None def makefoliage(self, mcmap): """Generate the foliage and enter it in mcmap. Note, foliage will disintegrate if there is no log nearby""" return None def copy(self, other): '''Copy the essential values of the other tree object into self. ''' self.pos = other.pos self.height = other.height def __init__(self, pos=[0, 0, 0], height=1): '''Accept values for the position and height of a tree. Store them in self. ''' self.pos = pos self.height = height class StickTree(Tree): '''Set up the trunk for trees with a trunk width of 1 and simple geometry. Designed for sublcassing. Only makes the trunk. ''' def maketrunk(self, mcmap): x = self.pos[0] y = self.pos[1] z = self.pos[2] for i in range(self.height): assign_value(x, y, z, WOODINFO, mcmap) y += 1 class NormalTree(StickTree): '''Set up the foliage for a 'normal' tree. This tree will be a single bulb of foliage above a single width trunk. This shape is very similar to the default Minecraft tree. ''' def makefoliage(self, mcmap): """note, foliage will disintegrate if there is no foliage below, or if there is no "log" block within range 2 (square) at the same level or one level below""" topy = self.pos[1] + self.height - 1 start = topy - 2 end = topy + 2 for y in range(start, end): if y > start + 1: rad = 1 else: rad = 2 for xoff in range(-rad, rad + 1): for zoff in range(-rad, rad + 1): if (random() > 0.618 and abs(xoff) == abs(zoff) and abs(xoff) == rad ): continue x = self.pos[0] + xoff z = self.pos[2] + zoff assign_value(x, y, z, LEAFINFO, mcmap) class BambooTree(StickTree): '''Set up the foliage for a bamboo tree. Make foliage sparse and adjacent to the trunk. ''' def makefoliage(self, mcmap): start = self.pos[1] end = self.pos[1] + self.height + 1 for y in range(start, end): for _ in [0, 1]: xoff = choice([-1, 1]) zoff = choice([-1, 1]) x = self.pos[0] + xoff z = self.pos[2] + zoff assign_value(x, y, z, LEAFINFO, mcmap) class PalmTree(StickTree): '''Set up the foliage for a palm tree. Make foliage stick out in four directions from the top of the trunk. ''' def makefoliage(self, mcmap): y = self.pos[1] + self.height for xoff in range(-2, 3): for zoff in range(-2, 3): if abs(xoff) == abs(zoff): x = self.pos[0] + xoff z = self.pos[2] + zoff assign_value(x, y, z, LEAFINFO, mcmap) class ProceduralTree(Tree): '''Set up the methods for a larger more complicated tree. This tree type has roots, a trunk, and branches all of varying width, and many foliage clusters. MUST BE SUBCLASSED. Specifically, self.foliage_shape must be set. Subclass 'prepare' and 'shapefunc' to make different shaped trees. ''' @staticmethod def crossection(center, radius, diraxis, matidx, mcmap): '''Create a round section of type matidx in mcmap. Passed values: center = [x, y, z] for the coordinates of the center block radius = <number> as the radius of the section. May be a float or int. diraxis: The list index for the axis to make the section perpendicular to. 0 indicates the x axis, 1 the y, 2 the z. The section will extend along the other two axies. matidx = <int> the integer value to make the section out of. mcmap = the array generated by make_mcmap matdata = <int> the integer value to make the block data value. ''' rad = int(radius + .618) if rad <= 0: return None secidx1 = (diraxis - 1) % 3 secidx2 = (1 + diraxis) % 3 coord = [0, 0, 0] for off1 in range(-rad, rad + 1): for off2 in range(-rad, rad + 1): thisdist = sqrt((abs(off1) + .5) ** 2 + (abs(off2) + .5) ** 2) if thisdist > radius: continue pri = center[diraxis] sec1 = center[secidx1] + off1 sec2 = center[secidx2] + off2 coord[diraxis] = pri coord[secidx1] = sec1 coord[secidx2] = sec2 assign_value(coord[0], coord[1], coord[2], matidx, mcmap) def shapefunc(self, y): '''Take y and return a radius for the location of the foliage cluster. If no foliage cluster is to be created, return None Designed for sublcassing. Only makes clusters close to the trunk. ''' if random() < 100. / (self.height ** 2) and y < self.trunkheight: return self.height * .12 return None def foliagecluster(self, center, mcmap): '''generate a round cluster of foliage at the location center. The shape of the cluster is defined by the list self.foliage_shape. This list must be set in a subclass of ProceduralTree. ''' level_radius = self.foliage_shape x = center[0] y = center[1] z = center[2] for i in level_radius: self.crossection([x, y, z], i, 1, LEAFINFO, mcmap) y += 1 def taperedcylinder(self, start, end, startsize, endsize, mcmap, blockdata): '''Create a tapered cylinder in mcmap. start and end are the beginning and ending coordinates of form [x, y, z]. startsize and endsize are the beginning and ending radius. The material of the cylinder is WOODMAT. ''' # delta is the coordinate vector for the difference between # start and end. delta = [int(end[i] - start[i]) for i in range(3)] # primidx is the index (0, 1, or 2 for x, y, z) for the coordinate # which has the largest overall delta. maxdist = max(delta, key=abs) if maxdist == 0: return None primidx = delta.index(maxdist) # secidx1 and secidx2 are the remaining indicies out of [0, 1, 2]. secidx1 = (primidx - 1) % 3 secidx2 = (1 + primidx) % 3 # primsign is the digit 1 or -1 depending on whether the limb is headed # along the positive or negative primidx axis. primsign = int(delta[primidx] / abs(delta[primidx])) # secdelta1 and ...2 are the amount the associated values change # for every step along the prime axis. secdelta1 = delta[secidx1] secfac1 = float(secdelta1) / delta[primidx] secdelta2 = delta[secidx2] secfac2 = float(secdelta2) / delta[primidx] # Initialize coord. These values could be anything, since # they are overwritten. coord = [0, 0, 0] # Loop through each crossection along the primary axis, # from start to end. endoffset = delta[primidx] + primsign for primoffset in range(0, endoffset, primsign): primloc = start[primidx] + primoffset secloc1 = int(start[secidx1] + primoffset * secfac1) secloc2 = int(start[secidx2] + primoffset * secfac2) coord[primidx] = primloc coord[secidx1] = secloc1 coord[secidx2] = secloc2 primdist = abs(delta[primidx]) radius = endsize + (startsize - endsize) * abs(delta[primidx] - primoffset) / primdist self.crossection(coord, radius, primidx, blockdata, mcmap) def makefoliage(self, mcmap): '''Generate the foliage for the tree in mcmap. ''' """note, foliage will disintegrate if there is no foliage below, or if there is no "log" block within range 2 (square) at the same level or one level below""" foliage_coords = self.foliage_cords for coord in foliage_coords: self.foliagecluster(coord, mcmap) for cord in foliage_coords: assign_value(cord[0], cord[1], cord[2], WOODINFO, mcmap) if LIGHTTREE == 1: assign_value(cord[0], cord[1] + 1, cord[2], LIGHTINFO, mcmap) elif LIGHTTREE in [2, 4]: assign_value(cord[0] + 1, cord[1], cord[2], LIGHTINFO, mcmap) assign_value(cord[0] - 1, cord[1], cord[2], LIGHTINFO, mcmap) if LIGHTTREE == 4: assign_value(cord[0], cord[1], cord[2] + 1, LIGHTINFO, mcmap) assign_value(cord[0], cord[1], cord[2] - 1, LIGHTINFO, mcmap) def makebranches(self, mcmap): '''Generate the branches and enter them in mcmap. ''' treeposition = self.pos height = self.height topy = treeposition[1] + int(self.trunkheight + 0.5) # endrad is the base radius of the branches at the trunk endrad = self.trunkradius * (1 - self.trunkheight / height) if endrad < 1.0: endrad = 1.0 for coord in self.foliage_cords: dist = (sqrt(float(coord[0] - treeposition[0]) ** 2 + float(coord[2] - treeposition[2]) ** 2)) ydist = coord[1] - treeposition[1] # value is a magic number that weights the probability # of generating branches properly so that # you get enough on small trees, but not too many # on larger trees. # Very difficult to get right... do not touch! value = (self.branchdensity * 220 * height) / ((ydist + dist) ** 3) if value < random(): continue posy = coord[1] slope = self.branchslope + (0.5 - random()) * .16 if coord[1] - dist * slope > topy: # Another random rejection, for branches between # the top of the trunk and the crown of the tree threshhold = 1 / float(height) if random() < threshhold: continue branchy = topy basesize = endrad else: branchy = posy - dist * slope basesize = (endrad + (self.trunkradius - endrad) * (topy - branchy) / self.trunkheight) startsize = (basesize * (1 + random()) * .618 * (dist / height) ** 0.618) rndr = sqrt(random()) * basesize * 0.618 rndang = random() * 2 * pi rndx = int(rndr * sin(rndang) + 0.5) rndz = int(rndr * cos(rndang) + 0.5) startcoord = [treeposition[0] + rndx, int(branchy), treeposition[2] + rndz] if startsize < 1.0: startsize = 1.0 endsize = 1.0 self.taperedcylinder(startcoord, coord, startsize, endsize, mcmap, WOODINFO) def makeroots(self, rootbases, mcmap): '''generate the roots and enter them in mcmap. rootbases = [[x, z, base_radius], ...] and is the list of locations the roots can originate from, and the size of that location. ''' treeposition = self.pos height = self.height for coord in self.foliage_cords: # First, set the threshhold for randomly selecting this # coordinate for root creation. dist = (sqrt(float(coord[0] - treeposition[0]) ** 2 + float(coord[2] - treeposition[2]) ** 2)) ydist = coord[1] - treeposition[1] value = (self.branchdensity * 220 * height) / ((ydist + dist) ** 3) # Randomly skip roots, based on the above threshold if value < random(): continue # initialize the internal variables from a selection of # starting locations. rootbase = choice(rootbases) rootx = rootbase[0] rootz = rootbase[1] rootbaseradius = rootbase[2] # Offset the root origin location by a random amount # (radialy) from the starting location. rndr = (sqrt(random()) * rootbaseradius * .618) rndang = random() * 2 * pi rndx = int(rndr * sin(rndang) + 0.5) rndz = int(rndr * cos(rndang) + 0.5) rndy = int(random() * rootbaseradius * 0.5) startcoord = [rootx + rndx, treeposition[1] + rndy, rootz + rndz] # offset is the distance from the root base to the root tip. offset = [startcoord[i] - coord[i] for i in range(3)] # If this is a mangrove tree, make the roots longer. if SHAPE == "mangrove": offset = [int(val * 1.618 - 1.5) for val in offset] endcoord = [startcoord[i] + offset[i] for i in range(3)] rootstartsize = (rootbaseradius * 0.618 * abs(offset[1]) / (height * 0.618)) if rootstartsize < 1.0: rootstartsize = 1.0 endsize = 1.0 # If ROOTS is set to "tostone" or "hanging" we need to check # along the distance for collision with existing materials. if ROOTS in ["tostone", "hanging"]: offlength = sqrt(float(offset[0]) ** 2 + float(offset[1]) ** 2 + float(offset[2]) ** 2) if offlength < 1: continue rootmid = endsize # vec is a unit vector along the direction of the root. vec = [offset[i] / offlength for i in range(3)] if ROOTS == "tostone": searchindex = STOPSROOTS elif ROOTS == "hanging": searchindex = [0] # startdist is how many steps to travel before starting to # search for the material. It is used to ensure that large # roots will go some distance before changing directions # or stopping. startdist = int(random() * 6 * sqrt(rootstartsize) + 2.8) # searchstart is the coordinate where the search should begin searchstart = [startcoord[i] + startdist * vec[i] for i in range(3)] # dist stores how far the search went (including searchstart) # before encountering the expected marterial. dist = startdist + dist_to_mat(searchstart, vec, searchindex, mcmap, limit=offlength) # If the distance to the material is less than the length # of the root, change the end point of the root to where # the search found the material. if dist < offlength: # rootmid is the size of the crossection at endcoord. rootmid += (rootstartsize - endsize) * (1 - dist / offlength) # endcoord is the midpoint for hanging roots, # and the endpoint for roots stopped by stone. endcoord = [startcoord[i] + int(vec[i] * dist) for i in range(3)] if ROOTS == "hanging": # remaining_dist is how far the root had left # to go when it was stopped. remaining_dist = offlength - dist # Initialize bottomcord to the stopping point of # the root, and then hang straight down # a distance of remaining_dist. bottomcord = endcoord[:] bottomcord[1] += -int(remaining_dist) # Make the hanging part of the hanging root. self.taperedcylinder(endcoord, bottomcord, rootmid, endsize, mcmap, WOODINFO) # make the beginning part of hanging or "tostone" roots self.taperedcylinder(startcoord, endcoord, rootstartsize, rootmid, mcmap, WOODINFO) # If you aren't searching for stone or air, just make the root. else: self.taperedcylinder(startcoord, endcoord, rootstartsize, endsize, mcmap, WOODINFO) def maketrunk(self, mcmap): '''Generate the trunk, roots, and branches in mcmap. ''' height = self.height trunkheight = self.trunkheight trunkradius = self.trunkradius treeposition = self.pos starty = treeposition[1] midy = treeposition[1] + int(trunkheight * .382) topy = treeposition[1] + int(trunkheight + 0.5) # In this method, x and z are the position of the trunk. x = treeposition[0] z = treeposition[2] end_size_factor = trunkheight / height midrad = trunkradius * (1 - end_size_factor * .5) endrad = trunkradius * (1 - end_size_factor) if endrad < 1.0: endrad = 1.0 if midrad < endrad: midrad = endrad # Make the root buttresses, if indicated if ROOTBUTTRESSES or SHAPE == "mangrove": # The start radius of the trunk should be a little smaller if we # are using root buttresses. startrad = trunkradius * .8 # rootbases is used later in self.makeroots(...) as # starting locations for the roots. rootbases = [[x, z, startrad]] buttress_radius = trunkradius * 0.382 # posradius is how far the root buttresses should be offset # from the trunk. posradius = trunkradius # In mangroves, the root buttresses are much more extended. if SHAPE == "mangrove": posradius *= 2.618 num_of_buttresses = int(sqrt(trunkradius) + 3.5) for i in range(num_of_buttresses): rndang = random() * 2 * pi thisposradius = posradius * (0.9 + random() * .2) # thisx and thisz are the x and z position for the base of # the root buttress. thisx = x + int(thisposradius * sin(rndang)) thisz = z + int(thisposradius * cos(rndang)) # thisbuttressradius is the radius of the buttress. # Currently, root buttresses do not taper. thisbuttressradius = buttress_radius * (0.618 + random()) if thisbuttressradius < 1.0: thisbuttressradius = 1.0 # Make the root buttress. self.taperedcylinder([thisx, starty, thisz], [x, midy, z], thisbuttressradius, thisbuttressradius, mcmap, WOODINFO) # Add this root buttress as a possible location at # which roots can spawn. rootbases += [[thisx, thisz, thisbuttressradius]] else: # If root buttresses are turned off, set the trunk radius # to normal size. startrad = trunkradius rootbases = [[x, z, startrad]] # Make the lower and upper sections of the trunk. self.taperedcylinder([x, starty, z], [x, midy, z], startrad, midrad, mcmap, WOODINFO) self.taperedcylinder([x, midy, z], [x, topy, z], midrad, endrad, mcmap, WOODINFO) #Make the branches self.makebranches(mcmap) #Make the roots, if indicated. if ROOTS in ["yes", "tostone", "hanging"]: self.makeroots(rootbases, mcmap) # Hollow the trunk, if specified # check to make sure that the trunk is large enough to be hollow if trunkradius > 2 and HOLLOWTRUNK: # wall thickness is actually the double the wall thickness # it is a diameter difference, not a radius difference. wall_thickness = (1 + trunkradius * 0.1 * random()) if wall_thickness < 1.3: wall_thickness = 1.3 base_radius = trunkradius - wall_thickness if base_radius < 1: base_radius = 1.0 mid_radius = midrad - wall_thickness top_radius = endrad - wall_thickness # the starting x and y can be offset by up to the wall thickness. base_offset = int(wall_thickness) x_choices = [i for i in range(x - base_offset, x + base_offset + 1)] start_x = choice(x_choices) z_choices = [i for i in range(z - base_offset, z + base_offset + 1)] start_z = choice(z_choices) self.taperedcylinder([start_x, starty, start_z], [x, midy, z], base_radius, mid_radius, mcmap, TRUNKFILLINFO) hollow_top_y = int(topy + trunkradius + 1.5) self.taperedcylinder([x, midy, z], [x, hollow_top_y, z], mid_radius, top_radius, mcmap, TRUNKFILLINFO) def prepare(self, mcmap): '''Initialize the internal values for the Tree object. Primarily, sets up the foliage cluster locations. ''' treeposition = self.pos self.trunkradius = .618 * sqrt(self.height * TRUNKTHICKNESS) if self.trunkradius < 1: self.trunkradius = 1 if BROKENTRUNK: self.trunkheight = self.height * (.3 + random() * .4) yend = int(treeposition[1] + self.trunkheight + .5) else: self.trunkheight = self.height yend = int(treeposition[1] + self.height) self.branchdensity = BRANCHDENSITY / FOLIAGEDENSITY topy = treeposition[1] + int(self.trunkheight + 0.5) foliage_coords = [] ystart = treeposition[1] num_of_clusters_per_y = int(1.5 + (FOLIAGEDENSITY * self.height / 19.) ** 2) if num_of_clusters_per_y < 1: num_of_clusters_per_y = 1 # make sure we don't spend too much time off the top of the map if yend > 255: yend = 255 if ystart > 255: ystart = 255 for y in range(yend, ystart, -1): for i in range(num_of_clusters_per_y): shapefac = self.shapefunc(y - ystart) if shapefac is None: continue r = (sqrt(random()) + .328) * shapefac theta = random() * 2 * pi x = int(r * sin(theta)) + treeposition[0] z = int(r * cos(theta)) + treeposition[2] # if there are values to search in STOPSBRANCHES # then check to see if this cluster is blocked # by stuff, like dirt or rock, or whatever if len(STOPSBRANCHES): dist = (sqrt(float(x - treeposition[0]) ** 2 + float(z - treeposition[2]) ** 2)) slope = self.branchslope if y - dist * slope > topy: # the top of the tree starty = topy else: starty = y - dist * slope # the start position of the search start = [treeposition[0], starty, treeposition[2]] offset = [x - treeposition[0], y - starty, z - treeposition[2]] offlength = sqrt(offset[0] ** 2 + offset[1] ** 2 + offset[2] ** 2) # if the branch is as short as... nothing, don't bother. if offlength < 1: continue # unit vector for the search vec = [offset[i] / offlength for i in range(3)] mat_dist = dist_to_mat(start, vec, STOPSBRANCHES, mcmap, limit=offlength + 3) # after all that, if you find something, don't add # this coordinate to the list if mat_dist < offlength + 2: continue foliage_coords += [[x, y, z]] self.foliage_cords = foliage_coords class RoundTree(ProceduralTree): '''This kind of tree is designed to resemble a deciduous tree. ''' def prepare(self, mcmap): self.branchslope = 0.382 ProceduralTree.prepare(self, mcmap) self.foliage_shape = [2, 3, 3, 2.5, 1.6] self.trunkradius *= 0.8 self.trunkheight *= TRUNKHEIGHT def shapefunc(self, y): twigs = ProceduralTree.shapefunc(self, y) if twigs is not None: return twigs if y < self.height * (.282 + .1 * sqrt(random())): return None radius = self.height / 2. adj = self.height / 2. - y if adj == 0: dist = radius elif abs(adj) >= radius: dist = 0 else: dist = sqrt((radius ** 2) - (adj ** 2)) dist *= .618 return dist class ConeTree(ProceduralTree): '''this kind of tree is designed to resemble a conifer tree. ''' # woodType is the kind of wood the tree has, a data value woodType = 1 def prepare(self, mcmap): self.branchslope = 0.15 ProceduralTree.prepare(self, mcmap) self.foliage_shape = [3, 2.6, 2, 1] self.trunkradius *= 0.5 def shapefunc(self, y): twigs = ProceduralTree.shapefunc(self, y) if twigs is not None: return twigs if y < self.height * (.25 + .05 * sqrt(random())): return None radius = (self.height - y) * 0.382 if radius < 0: radius = 0 return radius class RainforestTree(ProceduralTree): '''This kind of tree is designed to resemble a rainforest tree. ''' def prepare(self, mcmap): self.foliage_shape = [3.4, 2.6] self.branchslope = 1.0 ProceduralTree.prepare(self, mcmap) self.trunkradius *= 0.382 self.trunkheight *= .9 def shapefunc(self, y): if y < self.height * 0.8: if EDGEHEIGHT < self.height: twigs = ProceduralTree.shapefunc(self, y) if (twigs is not None) and random() < 0.07: return twigs return None else: width = self.height * .382 topdist = (self.height - y) / (self.height * 0.2) dist = width * (0.618 + topdist) * (0.618 + random()) * 0.382 return dist class MangroveTree(RoundTree): '''This kind of tree is designed to resemble a mangrove tree. ''' def prepare(self, mcmap): self.branchslope = 1.0 RoundTree.prepare(self, mcmap) self.trunkradius *= 0.618 def shapefunc(self, y): val = RoundTree.shapefunc(self, y) if val is None: return val val *= 1.618 return val def planttrees(mcmap, treelist): '''Take mcmap and add trees to random locations on the surface to treelist. ''' assert isinstance(mcmap, mcInterface.SaveFile) # keep looping until all the trees are placed # calc the radius difference, for interpolation in_out_dif = EDGEHEIGHT - CENTERHEIGHT if VERBOSE: print('Tree Locations: x, y, z, tree height') tries = 0 max_tries = MAXTRIES while len(treelist) < TREECOUNT: if tries > max_tries: if VERBOSE: print("Stopping search for tree locations after {0} tries".format(tries)) print("If you don't have enough trees, check X, Y, RADIUS, and PLANTON") break tries += 1 # choose a location rad_fraction = random() # this is some kind of square interpolation rad_fraction = 1.0 - rad_fraction rad_fraction **= 2 rad_fraction = 1.0 - rad_fraction rad = rad_fraction * RADIUS ang = random() * pi * 2 x = X + int(rad * sin(ang) + .5) z = Z + int(rad * cos(ang) + .5) # check to see if this location is suitable y_top = mcmap.surface_block(x, z) if y_top is None: # this location is off the map! continue if y_top['B'] in PLANTON: # plant the tree on the block above the ground # hence the " + 1" y = y_top['y'] + 1 else: continue # this is linear interpolation also. base_height = CENTERHEIGHT + (in_out_dif * rad_fraction) height_rand = (random() - .5) * 2 * HEIGHTVARIATION height = int(base_height + height_rand) # if the option is set, check the surrounding area for trees if ONLYINFORESTS: '''we are looking for foliage it should show up in the "surface_block" search check every fifth block in a square pattern, offset around the trunk and equal to the trees height if the area is not at least one third foliage, don't build the tree''' # spacing is how far apart each sample should be spacing = 5 # search_size is how many blocks to check # along each axis search_size = 2 + (height // spacing) # check at least 3 x 3 search_size = max([search_size, 3]) # set up the offset values to offset the starting corner offset = ((search_size - 1) * spacing) // 2 # foliage_count is the total number of foliage blocks found foliage_count = 0 # check each sample location for foliage for step_x in range(search_size): # search_x is the x location to search this sample search_x = x - offset + (step_x * spacing) for step_z in range(search_size): # same as for search_x search_z = z - offset + (step_z * spacing) search_block = mcmap.surface_block(search_x, search_z) if search_block is None: continue if search_block['B'] == 18: # this sample contains foliage! # add it to the total foliage_count += 1 #now that we have the total count, find the ratio total_searched = search_size ** 2 foliage_ratio = foliage_count / total_searched # the acceptable amount is about a third acceptable_ratio = .3 if foliage_ratio < acceptable_ratio: # after all that work, there wasn't enough foliage around! # try again! continue # generate the new tree newtree = Tree([x, y, z], height) if VERBOSE: print(x, y, z, height) treelist += [newtree] def processtrees(mcmap, treelist): '''Initalize all of the trees in treelist. Set all of the trees to the right type, and run prepare. If indicated limit the height of the trees to the top of the map. ''' assert isinstance(mcmap, mcInterface.SaveFile) if SHAPE == "stickly": shape_choices = ["normal", "bamboo", "palm"] elif SHAPE == "procedural": shape_choices = ["round", "cone"] else: shape_choices = [SHAPE] # initialize mapheight, just in case mapheight = 255 for i in range(len(treelist)): newshape = choice(shape_choices) if newshape == "normal": newtree = NormalTree() elif newshape == "bamboo": newtree = BambooTree() elif newshape == "palm": newtree = PalmTree() elif newshape == "round": newtree = RoundTree() elif newshape == "cone": newtree = ConeTree() elif newshape == "rainforest": newtree = RainforestTree() elif newshape == "mangrove": newtree = MangroveTree() # Get the height and position of the existing trees in # the list. newtree.copy(treelist[i]) # Now check each tree to ensure that it doesn't stick # out the top of the map. If it does, shorten it until # the top of the foliage just touches the top of the map. if MAPHEIGHTLIMIT: height = newtree.height ybase = newtree.pos[1] if SHAPE == "rainforest": foliageheight = 2 else: foliageheight = 4 if ybase + height + foliageheight > mapheight: newheight = mapheight - ybase - foliageheight newtree.height = newheight # Even if it sticks out the top of the map, every tree # should be at least one unit tall. if newtree.height < 1: newtree.height = 1 newtree.prepare(mcmap) treelist[i] = newtree def main(the_map): '''create the trees ''' treelist = [] if VERBOSE: print("Planting new trees") planttrees(the_map, treelist) if VERBOSE: print("Processing tree changes") processtrees(the_map, treelist) if FOLIAGE: if VERBOSE: print("Generating foliage ") for i in treelist: i.makefoliage(the_map) if VERBOSE: print(' completed') if WOOD: if VERBOSE: print("Generating trunks, roots, and branches ") for i in treelist: i.maketrunk(the_map) if VERBOSE: print(' completed') return None def standalone(): if VERBOSE: print("Importing the map") try: the_map = mcInterface.SaveFile(LOADNAME) except IOError: if VERBOSE: print('File name invalid or save file otherwise corrupted. Aborting') return None main(the_map) if LIGHTINGFIX: if VERBOSE: print("Rough re-lighting the map") relight_master.save_file = the_map relight_master.calc_lighting() if VERBOSE: print("Saving the map, this could be a while") the_map.write() if VERBOSE: print("finished") if __name__ == '__main__': standalone() # to do: # get height limits from map # set "limit height" or somesuch to respect level height limits
51,634
37.163341
89
py
lale
lale-master/setup.py
# Copyright 2019-2023 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import logging import os from datetime import datetime from setuptools import find_packages, setup logger = logging.getLogger(__name__) logger.addHandler(logging.StreamHandler()) try: import builtins # This trick is borrowed from scikit-learn # This is a bit (!) hackish: we are setting a global variable so that the # main lale __init__ can detect if it is being loaded by the setup # routine, to avoid attempting to import components before installation. builtins.__LALE_SETUP__ = True # type: ignore except ImportError: pass with open("README.md", "r", encoding="utf-8") as fh: long_description = fh.read() on_rtd = os.environ.get("READTHEDOCS") == "True" if on_rtd: install_requires = [] else: install_requires = [ "numpy<1.24", "black>=22.1.0", "click==8.0.4", "graphviz", "hyperopt>=0.2,<=0.2.5", "jsonschema", "jsonsubschema>=0.0.6", "scikit-learn>=1.0.0,<=1.2.0", "scipy<1.11.0", "pandas<2.0.0", "packaging", "decorator", "astunparse", "typing-extensions", ] import lale # noqa: E402 # pylint:disable=wrong-import-position if "TRAVIS" in os.environ: now = datetime.now().strftime("%y%m%d%H%M") VERSION = f"{lale.__version__}-{now}" else: VERSION = lale.__version__ extras_require = { "full": [ "mystic", "xgboost<=1.5.1", "lightgbm<4.0.0", "snapml>=1.7.0rc3,<1.12.0", "liac-arff>=2.4.0", "tensorflow>=2.4.0", "smac<=0.10.0", "numba", "aif360>=0.4.0", "protobuf<=3.20.1", "torch>=1.0", "BlackBoxAuditing", "imbalanced-learn", "cvxpy>=1.0", "fairlearn", "h5py", ], "dev": ["pre-commit"], "test": [ "mystic", "joblib", "ipython<8.8.0", "jupyter", "sphinx>=5.0.0", "sphinx_rtd_theme>=0.5.2", "docutils<0.17", "m2r2", "sphinxcontrib.apidoc", "sphinxcontrib-svg2pdfconverter", "pytest", "pyspark", "func_timeout", "category-encoders", "pynisher==0.6.4", ], "fairness": [ "mystic", "liac-arff>=2.4.0", "aif360<0.6.0", "imbalanced-learn", "protobuf<=3.20.1", "BlackBoxAuditing", ], "tutorial": [ "ipython<8.8.0", "jupyter", "xgboost<=1.5.1", "imbalanced-learn", "liac-arff>=2.4.0", "aif360==0.5.0", "protobuf<=3.20.1", "BlackBoxAuditing", "typing-extensions", ], } classifiers = [ "Development Status :: 5 - Production/Stable", "Intended Audience :: Developers", "Intended Audience :: Science/Research", "License :: OSI Approved :: Apache Software License", "Operating System :: MacOS", "Operating System :: Microsoft :: Windows", "Operating System :: POSIX", "Operating System :: Unix", "Programming Language :: Python", "Programming Language :: Python :: 3", "Programming Language :: Python :: 3.7", "Programming Language :: Python :: 3.8", "Programming Language :: Python :: 3.9", "Programming Language :: Python :: 3.10", "Topic :: Software Development", "Topic :: Scientific/Engineering", "Topic :: Scientific/Engineering :: Artificial Intelligence", ] setup( name="lale", version=VERSION, author="Guillaume Baudart, Martin Hirzel, Kiran Kate, Parikshit Ram, Avraham Shinnar", description="Library for Semi-Automated Data Science", long_description=long_description, long_description_content_type="text/markdown", url="https://github.com/IBM/lale", python_requires=">=3.6", package_data={"lale": ["py.typed"]}, packages=find_packages(), license="Apache License 2.0", classifiers=classifiers, install_requires=install_requires, extras_require=extras_require, )
4,556
27.304348
90
py
lale
lale-master/test/test_custom_schemas.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import unittest from test import EnableSchemaValidation from test.mock_module import CustomOrigOperator, UnknownOp import numpy as np from lightgbm import LGBMClassifier as baz from sklearn.decomposition import PCA as sk_PCA from sklearn.linear_model import Lars as sk_lars from xgboost import XGBClassifier as alt_xgb import lale import lale.type_checking from lale import schemas from lale.search.lale_grid_search_cv import get_grid_search_parameter_grids class TestCustomSchema(unittest.TestCase): def setUp(self): import sklearn.decomposition from lale.lib.sklearn import PCA as lale_PCA from lale.operators import make_operator self.sk_pca = make_operator(sklearn.decomposition.PCA, schemas={}) self.ll_pca = lale_PCA self.maxDiff = None def test_override_schemas(self): with EnableSchemaValidation(): init_schemas = self.sk_pca._schemas pca_schemas = self.ll_pca._schemas custom = self.sk_pca.customize_schema(schemas=schemas.JSON(pca_schemas)) self.assertEqual(custom._schemas, pca_schemas) self.assertEqual(self.sk_pca._schemas, init_schemas) self.assertRaises(Exception, self.sk_pca.customize_schema, schemas={}) def test_override_input(self): with EnableSchemaValidation(): init_input_schema = self.sk_pca.get_schema("input_fit") pca_input = self.ll_pca.get_schema("input_fit") custom = self.sk_pca.customize_schema(input_fit=schemas.JSON(pca_input)) self.assertEqual(custom.get_schema("input_fit"), pca_input) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.sk_pca.get_schema("input_fit"), init_input_schema) self.assertRaises(Exception, self.sk_pca.customize_schema, input_fit=42) _ = self.sk_pca.customize_schema(input_foo=pca_input) def test_override_output(self): with EnableSchemaValidation(): init_output_schema = self.sk_pca.get_schema("output_transform") pca_output = self.ll_pca.get_schema("output_transform") custom = self.sk_pca.customize_schema( output_transform=schemas.JSON(pca_output) ) self.assertEqual(custom.get_schema("output_transform"), pca_output) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual( self.sk_pca.get_schema("output_transform"), init_output_schema ) self.assertRaises(Exception, self.sk_pca.customize_schema, output=42) _ = self.sk_pca.customize_schema(output_foo=pca_output) def test_override_output2(self): init_output_schema = self.sk_pca.get_schema("output_transform") pca_output = schemas.AnyOf( [ schemas.Array(schemas.Array(schemas.Float())), schemas.Array(schemas.Float()), ] ) expected = { "anyOf": [ { "type": "array", "items": {"type": "array", "items": {"type": "number"}}, }, {"type": "array", "items": {"type": "number"}}, ] } custom = self.sk_pca.customize_schema(output_transform=pca_output) self.assertEqual(custom.get_schema("output_transform"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.sk_pca.get_schema("output_transform"), init_output_schema) def test_override_bool_param_sk(self): with EnableSchemaValidation(): init = self.sk_pca.hyperparam_schema("whiten") expected = {"default": True, "type": "boolean", "description": "override"} custom = self.sk_pca.customize_schema( whiten=schemas.Bool(default=True, desc="override") ) self.assertEqual(custom.hyperparam_schema("whiten"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.sk_pca.hyperparam_schema("whiten"), init) self.assertRaises(Exception, self.sk_pca.customize_schema, whitenX=42) def test_override_bool_param_ll(self): with EnableSchemaValidation(): init = self.ll_pca.hyperparam_schema("whiten") expected = {"default": True, "type": "boolean"} custom = self.ll_pca.customize_schema(whiten=schemas.Bool(default=True)) self.assertEqual(custom.hyperparam_schema("whiten"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.ll_pca.hyperparam_schema("whiten"), init) self.assertRaises(Exception, self.ll_pca.customize_schema, whitenX=42) def test_override_enum_param(self): with EnableSchemaValidation(): init = self.ll_pca.hyperparam_schema("svd_solver") expected = {"default": "full", "enum": ["auto", "full"]} custom = self.ll_pca.customize_schema( svd_solver=schemas.Enum(default="full", values=["auto", "full"]) ) self.assertEqual(custom.hyperparam_schema("svd_solver"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.ll_pca.hyperparam_schema("svd_solver"), init) def test_override_float_param(self): init = self.ll_pca.hyperparam_schema("tol") expected = { "default": 0.1, "type": "number", "minimum": -10, "maximum": 10, "exclusiveMaximum": True, "exclusiveMinimum": False, } custom = self.ll_pca.customize_schema( tol=schemas.Float( default=0.1, minimum=-10, maximum=10, exclusiveMaximum=True, exclusiveMinimum=False, ) ) self.assertEqual(custom.hyperparam_schema("tol"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.ll_pca.hyperparam_schema("tol"), init) def test_override_int_param(self): init = self.ll_pca.hyperparam_schema("iterated_power") expected = { "default": 1, "type": "integer", "minimum": -10, "maximum": 10, "exclusiveMaximum": True, "exclusiveMinimum": False, } custom = self.ll_pca.customize_schema( iterated_power=schemas.Int( default=1, minimum=-10, maximum=10, exclusiveMaximum=True, exclusiveMinimum=False, ) ) self.assertEqual(custom.hyperparam_schema("iterated_power"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.ll_pca.hyperparam_schema("iterated_power"), init) def test_override_null_param(self): init = self.ll_pca.hyperparam_schema("n_components") expected = {"enum": [None]} custom = self.ll_pca.customize_schema(n_components=schemas.Null()) self.assertEqual(custom.hyperparam_schema("n_components"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.ll_pca.hyperparam_schema("n_components"), init) def test_override_json_param(self): init = self.ll_pca.hyperparam_schema("tol") expected = { "description": "Tol", "type": "number", "minimum": 0.2, "default": 1.0, } custom = self.ll_pca.customize_schema(tol=schemas.JSON(expected)) self.assertEqual(custom.hyperparam_schema("tol"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.ll_pca.hyperparam_schema("tol"), init) def test_override_any_param(self): init = self.ll_pca.hyperparam_schema("iterated_power") expected = { "anyOf": [{"type": "integer"}, {"enum": ["auto", "full"]}], "default": "auto", } custom = self.ll_pca.customize_schema( iterated_power=schemas.AnyOf( [schemas.Int(), schemas.Enum(["auto", "full"])], default="auto" ) ) self.assertEqual(custom.hyperparam_schema("iterated_power"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.ll_pca.hyperparam_schema("iterated_power"), init) def test_override_array_param(self): init = self.sk_pca.hyperparam_schema("copy") expected = { "type": "array", "minItems": 1, "maxItems": 20, "items": {"type": "integer"}, } custom = self.sk_pca.customize_schema( copy=schemas.Array(minItems=1, maxItems=20, items=schemas.Int()) ) self.assertEqual(custom.hyperparam_schema("copy"), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.sk_pca.hyperparam_schema("copy"), init) def test_override_object_param(self): init = self.sk_pca.get_schema("input_fit") expected = { "type": "object", "required": ["X"], "additionalProperties": False, "properties": {"X": {"type": "array", "items": {"type": "number"}}}, } custom = self.sk_pca.customize_schema( input_fit=schemas.Object( required=["X"], additionalProperties=False, X=schemas.Array(schemas.Float()), ) ) self.assertEqual(custom.get_schema("input_fit"), expected) lale.type_checking.validate_is_schema(custom.get_schema("input_fit")) self.assertEqual(self.sk_pca.get_schema("input_fit"), init) def test_add_constraint(self): init_expected = self.sk_pca.hyperparam_schema() expected = { "allOf": [ init_expected["allOf"][0], { "anyOf": [ { "type": "object", "properties": { "n_components": { "not": {"enum": ["mle"]}, } }, }, { "type": "object", "properties": { "svd_solver": {"enum": ["full", "auto"]}, }, }, ] }, ] } custom = self.sk_pca.customize_schema( constraint=schemas.AnyOf( [ schemas.Object(n_components=schemas.Not(schemas.Enum(["mle"]))), schemas.Object(svd_solver=schemas.Enum(["full", "auto"])), ] ) ) self.assertEqual(custom.hyperparam_schema(), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.sk_pca.hyperparam_schema(), init_expected) def test_add_multiple_constraints(self): init_expected = self.sk_pca.hyperparam_schema() expected = { "allOf": [ init_expected["allOf"][0], { "anyOf": [ { "type": "object", "properties": { "n_components": { "not": {"enum": ["mle"]}, } }, }, { "type": "object", "properties": { "svd_solver": {"enum": ["full", "auto"]}, }, }, ] }, { "anyOf": [ { "type": "object", "properties": {"copy": {"enum": [False]}}, }, { "type": "object", "properties": { "whiten": {"enum": [False]}, }, }, ] }, ] } custom = self.sk_pca.customize_schema( constraint=[ schemas.AnyOf( [ schemas.Object(n_components=schemas.Not(schemas.Enum(["mle"]))), schemas.Object(svd_solver=schemas.Enum(["full", "auto"])), ] ), schemas.AnyOf( [ schemas.Object(copy=schemas.Enum([False])), schemas.Object(whiten=schemas.Enum([False])), ] ), ] ) self.assertEqual(custom.hyperparam_schema(), expected) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.sk_pca.hyperparam_schema(), init_expected) def test_override_relevant(self): init = self.ll_pca.hyperparam_schema()["allOf"][0]["relevantToOptimizer"] expected = ["svd_solver"] custom = self.ll_pca.customize_schema(relevantToOptimizer=["svd_solver"]) self.assertEqual( custom.hyperparam_schema()["allOf"][0]["relevantToOptimizer"], expected ) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual( self.ll_pca.hyperparam_schema()["allOf"][0]["relevantToOptimizer"], init ) self.assertRaises( Exception, self.sk_pca.customize_schema, relevantToOptimizer={} ) def test_override_tags(self): with EnableSchemaValidation(): init = self.ll_pca._schemas["tags"] tags = { "pre": ["~categoricals"], "op": ["estimator", "classifier", "interpretable"], "post": ["probabilities"], } custom = self.ll_pca.customize_schema(tags=tags) self.assertEqual(custom._schemas["tags"], tags) lale.type_checking.validate_is_schema(custom._schemas) self.assertEqual(self.ll_pca._schemas["tags"], init) self.assertRaises(Exception, self.sk_pca.customize_schema, tags=42) def test_wrap_imported_operators(self): old_globals = {**globals()} try: from lale.lib.autogen import Lars from lale.lib.lightgbm import LGBMClassifier from lale.lib.xgboost import XGBClassifier lale.wrap_imported_operators( exclude_classes=["sk_PCA"], wrapper_modules=["test.mock_custom_operators"], ) from lale.operators import PlannedIndividualOp op_obj = sk_PCA() self.assertIsInstance(op_obj, sk_PCA) self.assertEqual(alt_xgb._schemas, XGBClassifier._schemas) # type: ignore self.assertEqual(baz._schemas, LGBMClassifier._schemas) # type: ignore self.assertEqual(sk_lars._schemas, Lars._schemas) self.assertIsInstance(CustomOrigOperator, PlannedIndividualOp) finally: for sym, obj in old_globals.items(): globals()[sym] = obj class TestConstraintDropping(unittest.TestCase): def test_constraint_dropping(self): from lale.lib.sklearn import LogisticRegression from lale.operators import make_operator from lale.search.schema2search_space import op_to_search_space orig_schemas = LogisticRegression._schemas mod_schemas = { **orig_schemas, "properties": { **orig_schemas["properties"], "hyperparams": { "allOf": [ s if i == 0 else {**s, "forOptimizer": False} for i, s in enumerate( orig_schemas["properties"]["hyperparams"]["allOf"] ) ] }, }, } orig_space = op_to_search_space(LogisticRegression) mod_op = make_operator(LogisticRegression._impl_class(), mod_schemas) mod_space = op_to_search_space(mod_op) # dropping constraints makes the search space smaller self.assertGreater(len(str(orig_space)), len(str(mod_space))) class TestConstraintMerging(unittest.TestCase): def test_override_float_param1(self): from lale.lib.sklearn import PCA from lale.search.schema2search_space import op_to_search_space from lale.search.search_space import SearchSpaceNumber, SearchSpaceObject pca = PCA.customize_schema( relevantToOptimizer=["tol"], tol=schemas.Float( minimum=0.25, minimumForOptimizer=0, maximum=0.5, maximumForOptimizer=1.0, exclusiveMaximum=True, exclusiveMinimum=False, ), ) search = op_to_search_space(pca) assert isinstance(search, SearchSpaceObject) num_space = list(search.choices)[0][0] assert isinstance(num_space, SearchSpaceNumber) self.assertEqual(num_space.minimum, 0.25) self.assertEqual(num_space.maximum, 0.5) self.assertTrue(num_space.exclusiveMaximum) self.assertFalse(num_space.exclusiveMinimum) def test_override_float_param2(self): from lale.lib.sklearn import PCA from lale.search.schema2search_space import op_to_search_space from lale.search.search_space import SearchSpaceNumber, SearchSpaceObject pca = PCA.customize_schema( relevantToOptimizer=["tol"], tol=schemas.Float( minimum=0, minimumForOptimizer=0.25, maximum=1.5, maximumForOptimizer=1.0, exclusiveMaximumForOptimizer=False, exclusiveMinimumForOptimizer=True, ), ) search = op_to_search_space(pca) assert isinstance(search, SearchSpaceObject) num_space = list(search.choices)[0][0] assert isinstance(num_space, SearchSpaceNumber) self.assertEqual(num_space.minimum, 0.25) self.assertEqual(num_space.maximum, 1.0) self.assertFalse(num_space.exclusiveMaximum) self.assertTrue(num_space.exclusiveMinimum) def test_override_int_param1(self): from lale.lib.sklearn import PCA from lale.search.schema2search_space import op_to_search_space from lale.search.search_space import SearchSpaceNumber, SearchSpaceObject pca = PCA.customize_schema( relevantToOptimizer=["iterated_power"], iterated_power=schemas.Float( minimum=1, minimumForOptimizer=0, maximum=5, maximumForOptimizer=6, exclusiveMaximum=True, exclusiveMinimum=False, ), ) search = op_to_search_space(pca) assert isinstance(search, SearchSpaceObject) num_space = list(search.choices)[0][0] assert isinstance(num_space, SearchSpaceNumber) self.assertEqual(num_space.minimum, 1) self.assertEqual(num_space.maximum, 5) self.assertTrue(num_space.exclusiveMaximum) self.assertFalse(num_space.exclusiveMinimum) def test_override_int_param2(self): from lale.lib.sklearn import PCA from lale.search.schema2search_space import op_to_search_space from lale.search.search_space import SearchSpaceNumber, SearchSpaceObject pca = PCA.customize_schema( relevantToOptimizer=["iterated_power"], iterated_power=schemas.Float( minimum=0, minimumForOptimizer=1, maximum=6, maximumForOptimizer=5, exclusiveMaximumForOptimizer=False, exclusiveMinimumForOptimizer=True, ), ) search = op_to_search_space(pca) assert isinstance(search, SearchSpaceObject) num_space = list(search.choices)[0][0] assert isinstance(num_space, SearchSpaceNumber) self.assertEqual(num_space.minimum, 1) self.assertEqual(num_space.maximum, 5) self.assertFalse(num_space.exclusiveMaximum) self.assertTrue(num_space.exclusiveMinimum) class TestWrapUnknownOps(unittest.TestCase): expected_schema = { "allOf": [ { "type": "object", "relevantToOptimizer": [], "additionalProperties": False, "properties": { "n_neighbors": {"default": 5}, "algorithm": {"default": "auto"}, }, } ] } def test_wrap_from_class(self): from lale.operators import PlannedIndividualOp, make_operator self.assertFalse(isinstance(UnknownOp, PlannedIndividualOp)) Wrapped = make_operator(UnknownOp) self.assertTrue(isinstance(Wrapped, PlannedIndividualOp)) self.assertEqual(Wrapped.hyperparam_schema(), self.expected_schema) instance = Wrapped(n_neighbors=3) self.assertEqual(instance.hyperparams(), {"n_neighbors": 3}) def test_wrapped_from_import(self): old_globals = {**globals()} try: from lale.operators import PlannedIndividualOp self.assertFalse(isinstance(UnknownOp, PlannedIndividualOp)) lale.wrap_imported_operators() self.assertFalse(isinstance(UnknownOp, PlannedIndividualOp)) finally: for sym, obj in old_globals.items(): globals()[sym] = obj def test_wrap_from_instance(self): from sklearn.base import clone from lale.operators import TrainableIndividualOp, make_operator self.assertFalse(isinstance(UnknownOp, TrainableIndividualOp)) instance = UnknownOp(n_neighbors=3) self.assertFalse(isinstance(instance, TrainableIndividualOp)) wrapped = make_operator(instance) self.assertTrue(isinstance(wrapped, TrainableIndividualOp)) assert isinstance( wrapped, TrainableIndividualOp ) # help type checkers that don't know about assertTrue self.assertEqual(wrapped.reduced_hyperparams(), {"n_neighbors": 3}) cloned = clone(wrapped) self.assertTrue(isinstance(cloned, TrainableIndividualOp)) self.assertEqual(cloned.reduced_hyperparams(), {"n_neighbors": 3}) class TestFreeze(unittest.TestCase): def test_individual_op_freeze_trainable(self): from lale.lib.sklearn import LogisticRegression liquid = LogisticRegression(C=0.1, solver="liblinear") self.assertIn("dual", liquid.free_hyperparams()) self.assertFalse(liquid.is_frozen_trainable()) liquid_grid = get_grid_search_parameter_grids(liquid) self.assertTrue(len(liquid_grid) > 1, f"grid size {len(liquid_grid)}") frozen = liquid.freeze_trainable() self.assertEqual(len(frozen.free_hyperparams()), 0) self.assertTrue(frozen.is_frozen_trainable()) frozen_grid = get_grid_search_parameter_grids(frozen) self.assertEqual(len(frozen_grid), 1) def test_pipeline_freeze_trainable(self): from lale.lib.sklearn import PCA, LogisticRegression liquid = PCA() >> LogisticRegression() self.assertFalse(liquid.is_frozen_trainable()) liquid_grid = get_grid_search_parameter_grids(liquid) self.assertTrue(len(liquid_grid) > 1, f"grid size {len(liquid_grid)}") frozen = liquid.freeze_trainable() self.assertTrue(frozen.is_frozen_trainable()) frozen_grid = get_grid_search_parameter_grids(frozen) self.assertEqual(len(frozen_grid), 1) def test_individual_op_freeze_trained(self): from lale.lib.sklearn import KNeighborsClassifier with EnableSchemaValidation(): trainable = KNeighborsClassifier(n_neighbors=1) X = np.array([[0.0], [1.0], [2.0]]) y_old = np.array([0.0, 0.0, 1.0]) y_new = np.array([1.0, 0.0, 0.0]) liquid_old = trainable.fit(X, y_old) self.assertEqual(list(liquid_old.predict(X)), list(y_old)) liquid_new = liquid_old.fit(X, y_new) self.assertEqual(list(liquid_new.predict(X)), list(y_new)) frozen_old = trainable.fit(X, y_old).freeze_trained() self.assertFalse(liquid_old.is_frozen_trained()) self.assertTrue(frozen_old.is_frozen_trained()) self.assertEqual(list(frozen_old.predict(X)), list(y_old)) frozen_new = frozen_old.fit(X, y_new) self.assertEqual(list(frozen_new.predict(X)), list(y_old)) def test_pipeline_freeze_trained(self): from lale.lib.sklearn import LogisticRegression, MinMaxScaler trainable = MinMaxScaler() >> LogisticRegression() X = [[0.0], [1.0], [2.0]] y = [0.0, 0.0, 1.0] liquid = trainable.fit(X, y) frozen = liquid.freeze_trained() self.assertFalse(liquid.is_frozen_trained()) self.assertTrue(frozen.is_frozen_trained()) def test_trained_individual_op_freeze_trainable(self): from lale.lib.sklearn import KNeighborsClassifier from lale.operators import TrainedIndividualOp with EnableSchemaValidation(): trainable = KNeighborsClassifier(n_neighbors=1) X = np.array([[0.0], [1.0], [2.0]]) y_old = np.array([0.0, 0.0, 1.0]) liquid = trainable.fit(X, y_old) self.assertIsInstance(liquid, TrainedIndividualOp) self.assertFalse(liquid.is_frozen_trainable()) self.assertIn("algorithm", liquid.free_hyperparams()) frozen = liquid.freeze_trainable() self.assertIsInstance(frozen, TrainedIndividualOp) self.assertTrue(frozen.is_frozen_trainable()) self.assertFalse(frozen.is_frozen_trained()) self.assertEqual(len(frozen.free_hyperparams()), 0) def test_trained_pipeline_freeze_trainable(self): from lale.lib.sklearn import LogisticRegression, MinMaxScaler from lale.operators import TrainedPipeline trainable = MinMaxScaler() >> LogisticRegression() X = [[0.0], [1.0], [2.0]] y = [0.0, 0.0, 1.0] liquid = trainable.fit(X, y) self.assertIsInstance(liquid, TrainedPipeline) self.assertFalse(liquid.is_frozen_trainable()) frozen = liquid.freeze_trainable() self.assertFalse(liquid.is_frozen_trainable()) self.assertTrue(frozen.is_frozen_trainable()) self.assertIsInstance(frozen, TrainedPipeline)
27,885
40.435364
88
py
lale
lale-master/test/test_lale_lib_versions.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import unittest from test import EnableSchemaValidation import jsonschema import numpy as np import sklearn import sklearn.datasets import sklearn.metrics import sklearn.model_selection import xgboost from lale.lib.lale import Hyperopt from lale.lib.sklearn import ( SVC, DecisionTreeClassifier, DecisionTreeRegressor, ExtraTreesClassifier, ExtraTreesRegressor, FeatureAgglomeration, FunctionTransformer, GradientBoostingClassifier, GradientBoostingRegressor, LinearRegression, LogisticRegression, MLPClassifier, PolynomialFeatures, RandomForestClassifier, RandomForestRegressor, Ridge, VotingClassifier, ) from lale.lib.xgboost import XGBClassifier, XGBRegressor assert sklearn.__version__ == "0.20.3", "This test is for scikit-learn 0.20.3." assert xgboost.__version__ == "0.90", "This test is for XGBoost 0.90." class TestDecisionTreeClassifier(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = DecisionTreeClassifier() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_ccp_alpha(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'ccp_alpha' was unexpected" ): _ = DecisionTreeClassifier(ccp_alpha=0.01) def test_with_hyperopt(self): planned = DecisionTreeClassifier trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestDecisionTreeRegressor(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_diabetes(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = DecisionTreeRegressor() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_ccp_alpha(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'ccp_alpha' was unexpected" ): _ = DecisionTreeRegressor(ccp_alpha=0.01) def test_with_hyperopt(self): planned = DecisionTreeRegressor trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, scoring="r2", cv=3, max_evals=3, ) _ = trained.predict(self.test_X) class TestExtraTreesClassifier(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = ExtraTreesClassifier() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_n_estimators(self): default = ExtraTreesClassifier.get_defaults()["n_estimators"] self.assertEqual(default, 10) def test_ccp_alpha(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'ccp_alpha' was unexpected" ): _ = ExtraTreesClassifier(ccp_alpha=0.01) def test_max_samples(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'max_samples' was unexpected" ): _ = ExtraTreesClassifier(max_samples=0.01) def test_with_hyperopt(self): planned = ExtraTreesClassifier trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestExtraTreesRegressor(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_diabetes(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = ExtraTreesRegressor() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_n_estimators(self): default = ExtraTreesRegressor.get_defaults()["n_estimators"] self.assertEqual(default, 10) def test_ccp_alpha(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'ccp_alpha' was unexpected" ): _ = ExtraTreesRegressor(ccp_alpha=0.01) def test_max_samples(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'max_samples' was unexpected" ): _ = ExtraTreesRegressor(max_samples=0.01) def test_with_hyperopt(self): planned = ExtraTreesRegressor trained = planned.auto_configure( self.train_X, self.train_y, scoring="r2", optimizer=Hyperopt, cv=3, max_evals=3, ) _ = trained.predict(self.test_X) class TestFeatureAgglomeration(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = FeatureAgglomeration() >> LogisticRegression() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_with_hyperopt(self): planned = FeatureAgglomeration >> LogisticRegression trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3, verbose=True, ) _ = trained.predict(self.test_X) class TestFunctionTransformer(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = FunctionTransformer(func=np.log1p) >> LogisticRegression() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_pass_y(self): trainable = ( FunctionTransformer(func=np.log1p, pass_y=False) >> LogisticRegression() ) trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_validate(self): default = FunctionTransformer.get_defaults()["validate"] self.assertEqual(default, True) def test_with_hyperopt(self): planned = FunctionTransformer(func=np.log1p) >> LogisticRegression trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestGradientBoostingClassifier(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = GradientBoostingClassifier() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_ccp_alpha(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'ccp_alpha' was unexpected" ): _ = GradientBoostingClassifier(ccp_alpha=0.01) def test_with_hyperopt(self): planned = GradientBoostingClassifier trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestGradientBoostingRegressor(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_diabetes(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = GradientBoostingRegressor() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_ccp_alpha(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'ccp_alpha' was unexpected" ): _ = GradientBoostingRegressor(ccp_alpha=0.01) def test_with_hyperopt(self): planned = GradientBoostingRegressor trained = planned.auto_configure( self.train_X, self.train_y, scoring="r2", optimizer=Hyperopt, cv=3, max_evals=3, ) _ = trained.predict(self.test_X) class TestLinearRegression(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_diabetes(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = LinearRegression() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_with_hyperopt(self): planned = LinearRegression trained = planned.auto_configure( self.train_X, self.train_y, scoring="r2", optimizer=Hyperopt, cv=3, max_evals=3, ) _ = trained.predict(self.test_X) class TestLogisticRegression(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = LogisticRegression() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_multi_class(self): default = LogisticRegression.get_defaults()["multi_class"] self.assertEqual(default, "ovr") def test_solver(self): default = LogisticRegression.get_defaults()["solver"] self.assertEqual(default, "liblinear") def test_l1_ratio(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'l1_ratio' was unexpected" ): _ = LogisticRegression(l1_ratio=0.2) def test_with_hyperopt(self): planned = LogisticRegression trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestMLPClassifier(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = MLPClassifier() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_max_fun(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'max_fun' was unexpected" ): _ = MLPClassifier(max_fun=1000) def test_with_hyperopt(self): planned = MLPClassifier(max_iter=20) trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestPolynomialFeatures(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = PolynomialFeatures() >> LogisticRegression() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_order(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'order' was unexpected" ): _ = PolynomialFeatures(order="F") >> LogisticRegression() def test_with_hyperopt(self): planned = PolynomialFeatures >> LogisticRegression trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestRandomForestClassifier(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = RandomForestClassifier() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_n_estimators(self): default = RandomForestClassifier.get_defaults()["n_estimators"] self.assertEqual(default, 10) def test_ccp_alpha(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'ccp_alpha' was unexpected" ): _ = RandomForestClassifier(ccp_alpha=0.01) def test_max_samples(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'max_samples' was unexpected" ): _ = RandomForestClassifier(max_samples=0.01) def test_with_hyperopt(self): planned = RandomForestClassifier trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestRandomForestRegressor(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_diabetes(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = RandomForestRegressor() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_n_estimators(self): default = RandomForestRegressor.get_defaults()["n_estimators"] self.assertEqual(default, 10) def test_ccp_alpha(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'ccp_alpha' was unexpected" ): _ = RandomForestRegressor(ccp_alpha=0.01) def test_max_samples(self): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'max_samples' was unexpected" ): _ = RandomForestRegressor(max_samples=0.01) def test_with_hyperopt(self): planned = RandomForestRegressor trained = planned.auto_configure( self.train_X, self.train_y, scoring="r2", optimizer=Hyperopt, cv=3, max_evals=3, ) _ = trained.predict(self.test_X) class TestRidge(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_diabetes(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = Ridge() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_with_hyperopt(self): planned = Ridge trained = planned.auto_configure( self.train_X, self.train_y, scoring="r2", optimizer=Hyperopt, cv=3, max_evals=3, ) _ = trained.predict(self.test_X) class TestSVC(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = SVC() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_gamma(self): default = SVC.get_defaults()["gamma"] self.assertEqual(default, "auto_deprecated") def test_break_ties(self): with EnableSchemaValidation(): with self.assertRaisesRegex( jsonschema.ValidationError, "argument 'break_ties' was unexpected" ): _ = SVC(break_ties=True) def test_with_hyperopt(self): planned = SVC trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestVotingClassifier(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = VotingClassifier( estimators=[("lr", LogisticRegression()), ("dt", DecisionTreeClassifier())] ) trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_estimators(self): trainable = VotingClassifier( estimators=[ ("lr", LogisticRegression()), ("dt", DecisionTreeClassifier()), ("na", None), ] ) trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_with_hyperopt(self): planned = VotingClassifier( estimators=[("lr", LogisticRegression), ("dt", DecisionTreeClassifier)] ) trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestXGBClassifier(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_iris(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = XGBClassifier() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_with_hyperopt(self): planned = XGBClassifier trained = planned.auto_configure( self.train_X, self.train_y, optimizer=Hyperopt, cv=3, max_evals=3 ) _ = trained.predict(self.test_X) class TestXGBRegressor(unittest.TestCase): def setUp(self): X, y = sklearn.datasets.load_diabetes(return_X_y=True) ( self.train_X, self.test_X, self.train_y, self.test_y, ) = sklearn.model_selection.train_test_split(X, y) def test_with_defaults(self): trainable = XGBRegressor() trained = trainable.fit(self.train_X, self.train_y) _ = trained.predict(self.test_X) def test_with_hyperopt(self): planned = XGBRegressor trained = planned.auto_configure( self.train_X, self.train_y, scoring="r2", optimizer=Hyperopt, cv=3, max_evals=3, ) _ = trained.predict(self.test_X)
22,044
30.856936
87
py
lale
lale-master/test/test_pipeline.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import random import unittest import warnings from sklearn.datasets import load_iris from sklearn.metrics import accuracy_score, make_scorer from sklearn.model_selection import train_test_split from sklearn.neural_network import MLPClassifier as SkMLPClassifier from sklearn.pipeline import Pipeline as SkPipeline from sklearn.preprocessing import MinMaxScaler as SkMinMaxScaler from lale.lib.lale import Batching, Hyperopt, NoOp from lale.lib.sklearn import PCA, LogisticRegression, Nystroem from lale.search.lale_grid_search_cv import get_grid_search_parameter_grids class TestBatching(unittest.TestCase): def setUp(self): data = load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_fit(self): import lale.lib.sklearn as lale_sklearn warnings.filterwarnings(action="ignore") pipeline = NoOp() >> Batching( operator=lale_sklearn.MinMaxScaler() >> lale_sklearn.MLPClassifier(random_state=42), batch_size=56, ) trained = pipeline.fit(self.X_train, self.y_train) predictions = trained.predict(self.X_test) lale_accuracy = accuracy_score(self.y_test, predictions) prep = SkMinMaxScaler() trained_prep = prep.partial_fit(self.X_train[0:56, :], self.y_train[0:56]) trained_prep.partial_fit(self.X_train[56:, :], self.y_train[56:]) X_transformed = trained_prep.transform(self.X_train) clf = SkMLPClassifier(random_state=42) import numpy as np trained_clf = clf.partial_fit( X_transformed[0:56, :], self.y_train[0:56], classes=np.unique(self.y_train) ) trained_clf.partial_fit( X_transformed[56:, :], self.y_train[56:], classes=np.unique(self.y_train) ) predictions = trained_clf.predict(trained_prep.transform(self.X_test)) sklearn_accuracy = accuracy_score(self.y_test, predictions) self.assertEqual(lale_accuracy, sklearn_accuracy) def test_fit1(self): warnings.filterwarnings(action="ignore") from lale.lib.sklearn import MinMaxScaler, MLPClassifier pipeline = Batching( operator=MinMaxScaler() >> MLPClassifier(random_state=42), batch_size=56 ) trained = pipeline.fit(self.X_train, self.y_train) predictions = trained.predict(self.X_test) lale_accuracy = accuracy_score(self.y_test, predictions) prep = MinMaxScaler() trained_prep = prep.partial_fit(self.X_train[0:56, :], self.y_train[0:56]) trained_prep.partial_fit(self.X_train[56:, :], self.y_train[56:]) X_transformed = trained_prep.transform(self.X_train) clf = SkMLPClassifier(random_state=42) import numpy as np trained_clf = clf.partial_fit( X_transformed[0:56, :], self.y_train[0:56], classes=np.unique(self.y_train) ) trained_clf.partial_fit( X_transformed[56:, :], self.y_train[56:], classes=np.unique(self.y_train) ) predictions = trained_clf.predict(trained_prep.transform(self.X_test)) sklearn_accuracy = accuracy_score(self.y_test, predictions) self.assertEqual(lale_accuracy, sklearn_accuracy) def test_fit2(self): warnings.filterwarnings(action="ignore") from lale.lib.sklearn import MinMaxScaler pipeline = Batching( operator=MinMaxScaler() >> MinMaxScaler(), batch_size=112, shuffle=False ) trained = pipeline.fit(self.X_train, self.y_train) lale_transforms = trained.transform(self.X_test) prep = SkMinMaxScaler() trained_prep = prep.partial_fit(self.X_train, self.y_train) X_transformed = trained_prep.transform(self.X_train) clf = MinMaxScaler() trained_clf = clf.partial_fit(X_transformed, self.y_train) sklearn_transforms = trained_clf.transform(trained_prep.transform(self.X_test)) for i in range(5): for j in range(2): self.assertAlmostEqual(lale_transforms[i, j], sklearn_transforms[i, j]) def test_fit3(self): from lale.lib.sklearn import MinMaxScaler, MLPClassifier pipeline = PCA() >> Batching( operator=MinMaxScaler() >> MLPClassifier(random_state=42), batch_size=10 ) trained = pipeline.fit(self.X_train, self.y_train) _ = trained.predict(self.X_test) def test_no_partial_fit(self): pipeline = Batching(operator=NoOp() >> LogisticRegression()) _ = pipeline.fit(self.X_train, self.y_train) def test_fit4(self): warnings.filterwarnings(action="ignore") from lale.lib.sklearn import MinMaxScaler, MLPClassifier pipeline = Batching( operator=MinMaxScaler() >> MLPClassifier(random_state=42), batch_size=56, inmemory=True, ) trained = pipeline.fit(self.X_train, self.y_train) predictions = trained.predict(self.X_test) lale_accuracy = accuracy_score(self.y_test, predictions) prep = SkMinMaxScaler() trained_prep = prep.partial_fit(self.X_train[0:56, :], self.y_train[0:56]) trained_prep.partial_fit(self.X_train[56:, :], self.y_train[56:]) X_transformed = trained_prep.transform(self.X_train) clf = SkMLPClassifier(random_state=42) import numpy as np trained_clf = clf.partial_fit( X_transformed[0:56, :], self.y_train[0:56], classes=np.unique(self.y_train) ) trained_clf.partial_fit( X_transformed[56:, :], self.y_train[56:], classes=np.unique(self.y_train) ) predictions = trained_clf.predict(trained_prep.transform(self.X_test)) sklearn_accuracy = accuracy_score(self.y_test, predictions) self.assertEqual(lale_accuracy, sklearn_accuracy) # TODO: Nesting doesn't work yet # def test_nested_pipeline(self): # from lale.lib.sklearn import MinMaxScaler, MLPClassifier # pipeline = Batching(operator = MinMaxScaler() >> Batching(operator = NoOp() >> MLPClassifier(random_state=42)), batch_size = 112) # trained = pipeline.fit(self.X_train, self.y_train) # predictions = trained.predict(self.X_test) # lale_accuracy = accuracy_score(self.y_test, predictions) class TestPipeline(unittest.TestCase): def dont_test_with_gridsearchcv2_auto(self): from sklearn.model_selection import GridSearchCV lr = LogisticRegression(random_state=42) pca = PCA(random_state=42, svd_solver="arpack") trainable = pca >> lr scikit_pipeline = SkPipeline( [ (pca.name(), PCA(random_state=42, svd_solver="arpack")), (lr.name(), LogisticRegression(random_state=42)), ] ) all_parameters = get_grid_search_parameter_grids(trainable, num_samples=1) # otherwise the test takes too long parameters = random.sample(all_parameters, 2) with warnings.catch_warnings(): warnings.simplefilter("ignore") clf = GridSearchCV( scikit_pipeline, parameters, cv=2, scoring=make_scorer(accuracy_score) ) iris = load_iris() clf.fit(iris.data, iris.target) predicted = clf.predict(iris.data) accuracy_with_lale_operators = accuracy_score(iris.target, predicted) from sklearn.decomposition import PCA as SklearnPCA from sklearn.linear_model import LogisticRegression as SklearnLR scikit_pipeline = SkPipeline( [ (pca.name(), SklearnPCA(random_state=42, svd_solver="arpack")), (lr.name(), SklearnLR(random_state=42)), ] ) with warnings.catch_warnings(): warnings.simplefilter("ignore") clf = GridSearchCV( scikit_pipeline, parameters, cv=2, scoring=make_scorer(accuracy_score) ) iris = load_iris() clf.fit(iris.data, iris.target) predicted = clf.predict(iris.data) accuracy_with_scikit_operators = accuracy_score(iris.target, predicted) self.assertEqual(accuracy_with_lale_operators, accuracy_with_scikit_operators) def test_with_gridsearchcv3(self): from sklearn.model_selection import GridSearchCV _ = LogisticRegression() scikit_pipeline = SkPipeline( [("nystroem", Nystroem()), ("lr", LogisticRegression())] ) parameters = {"lr__solver": ("liblinear", "lbfgs"), "lr__penalty": ["l2"]} clf = GridSearchCV( scikit_pipeline, parameters, cv=2, scoring=make_scorer(accuracy_score) ) iris = load_iris() with warnings.catch_warnings(): warnings.simplefilter("ignore") clf.fit(iris.data, iris.target) _ = clf.predict(iris.data) def test_with_gridsearchcv3_auto(self): from sklearn.model_selection import GridSearchCV lr = LogisticRegression() scikit_pipeline = SkPipeline( [(Nystroem().name(), Nystroem()), (lr.name(), LogisticRegression())] ) all_parameters = get_grid_search_parameter_grids( Nystroem() >> lr, num_samples=1 ) # otherwise the test takes too long parameters = random.sample(all_parameters, 2) with warnings.catch_warnings(): warnings.simplefilter("ignore") clf = GridSearchCV( scikit_pipeline, parameters, cv=2, scoring=make_scorer(accuracy_score) ) iris = load_iris() clf.fit(iris.data, iris.target) _ = clf.predict(iris.data) def test_with_gridsearchcv3_auto_wrapped(self): pipeline = Nystroem() >> LogisticRegression() with warnings.catch_warnings(): warnings.simplefilter("ignore") from lale.lib.lale import GridSearchCV clf = GridSearchCV( estimator=pipeline, lale_num_samples=1, lale_num_grids=1, cv=2, scoring=make_scorer(accuracy_score), ) iris = load_iris() clf.fit(iris.data, iris.target) _ = clf.predict(iris.data) class TestBatching2(unittest.TestCase): def setUp(self): data = load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_batching_with_hyperopt(self): from lale.lib.sklearn import MinMaxScaler, SGDClassifier pipeline = Batching(operator=MinMaxScaler() >> SGDClassifier()) trained = pipeline.auto_configure( self.X_train, self.y_train, optimizer=Hyperopt, max_evals=1 ) _ = trained.predict(self.X_test) class TestImportFromSklearnWithCognito(unittest.TestCase): def test_import_from_sklearn(self): pipeline_str = """from lale.lib.autoai_libs import NumpyColumnSelector from lale.lib.autoai_libs import CompressStrings from lale.lib.autoai_libs import NumpyReplaceMissingValues from lale.lib.autoai_libs import NumpyReplaceUnknownValues from lale.lib.autoai_libs import boolean2float from lale.lib.autoai_libs import CatImputer from lale.lib.autoai_libs import CatEncoder import numpy as np from lale.lib.autoai_libs import float32_transform from lale.operators import make_pipeline from lale.lib.autoai_libs import FloatStr2Float from lale.lib.autoai_libs import NumImputer from lale.lib.autoai_libs import OptStandardScaler from lale.operators import make_union from lale.lib.autoai_libs import NumpyPermuteArray from lale.lib.autoai_libs import TA1 import autoai_libs.utils.fc_methods from lale.lib.autoai_libs import FS1 from xgboost import XGBRegressor numpy_column_selector_0 = NumpyColumnSelector(columns=[1]) compress_strings = CompressStrings(compress_type='hash', dtypes_list=['int_num'], missing_values_reference_list=['', '-', '?', float('nan')], misslist_list=[[]]) numpy_replace_missing_values_0 = NumpyReplaceMissingValues(filling_values=float('nan'), missing_values=[]) numpy_replace_unknown_values = NumpyReplaceUnknownValues(filling_values=float('nan'), filling_values_list=[float('nan')], known_values_list=[[36, 45, 56, 67, 68, 75, 78, 89]], missing_values_reference_list=['', '-', '?', float('nan')]) cat_imputer = CatImputer(missing_values=float('nan'), sklearn_version_family='20', strategy='most_frequent') cat_encoder = CatEncoder(dtype=np.float64, handle_unknown='error', sklearn_version_family='20') pipeline_0 = make_pipeline(numpy_column_selector_0, compress_strings, numpy_replace_missing_values_0, numpy_replace_unknown_values, boolean2float(), cat_imputer, cat_encoder, float32_transform()) numpy_column_selector_1 = NumpyColumnSelector(columns=[0]) float_str2_float = FloatStr2Float(dtypes_list=['int_num'], missing_values_reference_list=[]) numpy_replace_missing_values_1 = NumpyReplaceMissingValues(filling_values=float('nan'), missing_values=[]) num_imputer = NumImputer(missing_values=float('nan'), strategy='median') opt_standard_scaler = OptStandardScaler(num_scaler_copy=None, num_scaler_with_mean=None, num_scaler_with_std=None, use_scaler_flag=False) pipeline_1 = make_pipeline(numpy_column_selector_1, float_str2_float, numpy_replace_missing_values_1, num_imputer, opt_standard_scaler, float32_transform()) union = make_union(pipeline_0, pipeline_1) numpy_permute_array = NumpyPermuteArray(axis=0, permutation_indices=[1, 0]) ta1_0 = TA1(fun=np.tan, name='tan', datatypes=['float'], feat_constraints=[autoai_libs.utils.fc_methods.is_not_categorical], col_names=['age', 'weight'], col_dtypes=[np.dtype('float32'), np.dtype('float32')]) fs1_0 = FS1(cols_ids_must_keep=range(0, 2), additional_col_count_to_keep=4, ptype='regression') ta1_1 = TA1(fun=np.square, name='square', datatypes=['numeric'], feat_constraints=[autoai_libs.utils.fc_methods.is_not_categorical], col_names=['age', 'weight', 'tan(age)'], col_dtypes=[np.dtype('float32'), np.dtype('float32'), np.dtype('float32')]) fs1_1 = FS1(cols_ids_must_keep=range(0, 2), additional_col_count_to_keep=4, ptype='regression') ta1_2 = TA1(fun=np.sin, name='sin', datatypes=['float'], feat_constraints=[autoai_libs.utils.fc_methods.is_not_categorical], col_names=['age', 'weight', 'tan(age)', 'square(age)', 'square(tan(age))'], col_dtypes=[np.dtype('float32'), np.dtype('float32'), np.dtype('float32'), np.dtype('float32'), np.dtype('float32')]) fs1_2 = FS1(cols_ids_must_keep=range(0, 2), additional_col_count_to_keep=4, ptype='regression') xgb_regressor = XGBRegressor(missing=float('nan'), n_jobs=4, random_state=33, silent=True, verbosity=0) pipeline = make_pipeline(union, numpy_permute_array, ta1_0, fs1_0, ta1_1, fs1_1, ta1_2, fs1_2, xgb_regressor) """ globals2 = {} # This call to exec should be safe since we are using a fixed (completely specified) string exec(pipeline_str, globals2) # nosec pipeline2 = globals2["pipeline"] sklearn_pipeline = pipeline2.export_to_sklearn_pipeline() from lale import helpers _ = helpers.import_from_sklearn_pipeline(sklearn_pipeline) class TestExportToSklearnForEstimator(unittest.TestCase): def setUp(self): data = load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def create_pipeline(self): from sklearn.decomposition import PCA as SkPCA from sklearn.pipeline import make_pipeline pipeline = make_pipeline(SkPCA(), LogisticRegression()) return pipeline def test_import_export_trained(self): import numpy as np from lale.helpers import import_from_sklearn_pipeline pipeline = self.create_pipeline() self.assertEqual(isinstance(pipeline, SkPipeline), True) pipeline.fit(self.X_train, self.y_train) predictions_before = pipeline.predict(self.X_test) lale_pipeline = import_from_sklearn_pipeline(pipeline) predictions_after = lale_pipeline.predict(self.X_test) sklearn_pipeline = lale_pipeline.export_to_sklearn_pipeline() predictions_after_1 = sklearn_pipeline.predict(self.X_test) self.assertEqual(np.all(predictions_before == predictions_after), True) self.assertEqual(np.all(predictions_before == predictions_after_1), True) def test_import_export_trainable(self): from sklearn.exceptions import NotFittedError from lale.helpers import import_from_sklearn_pipeline pipeline = self.create_pipeline() self.assertEqual(isinstance(pipeline, SkPipeline), True) pipeline.fit(self.X_train, self.y_train) lale_pipeline = import_from_sklearn_pipeline(pipeline, fitted=False) with self.assertRaises(ValueError): lale_pipeline.predict(self.X_test) sklearn_pipeline = lale_pipeline.export_to_sklearn_pipeline() with self.assertRaises(NotFittedError): sklearn_pipeline.predict(self.X_test)
17,752
43.717884
318
py
lale
lale-master/test/test_core_regressors.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import unittest from jsonschema.exceptions import ValidationError import lale.lib.lale import lale.type_checking from lale.lib.lale import NoOp from lale.lib.sklearn import ( ExtraTreesRegressor, GradientBoostingRegressor, RandomForestRegressor, Ridge, SGDRegressor, ) class TestRegression(unittest.TestCase): def setUp(self): from sklearn.datasets import make_regression from sklearn.model_selection import train_test_split X, y = make_regression( n_samples=200, n_features=4, n_informative=2, random_state=0, shuffle=False ) self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def create_function_test_regressor(clf_name): def test_regressor(self): X_train, y_train = self.X_train, self.y_train import importlib module_name = ".".join(clf_name.split(".")[0:-1]) class_name = clf_name.split(".")[-1] module = importlib.import_module(module_name) class_ = getattr(module, class_name) regr = None if class_name in ["StackingRegressor", "VotingRegressor"]: regr = class_(estimators=[("base", SGDRegressor())]) else: regr = class_() # test_schemas_are_schemas lale.type_checking.validate_is_schema(regr.input_schema_fit()) lale.type_checking.validate_is_schema(regr.input_schema_predict()) lale.type_checking.validate_is_schema(regr.output_schema_predict()) lale.type_checking.validate_is_schema(regr.hyperparam_schema()) # test_init_fit_predict trained = regr.fit(self.X_train, self.y_train) _ = trained.predict(self.X_test) # test score _ = trained.score(self.X_test, self.y_test) # test_predict_on_trainable trained = regr.fit(X_train, y_train) regr.predict(X_train) # test_to_json regr.to_json() # test_in_a_pipeline pipeline = NoOp() >> regr trained = pipeline.fit(self.X_train, self.y_train) _ = trained.predict(self.X_test) # test_with_hyperopt if isinstance(regr, Ridge): # type: ignore from lale.lib.lale import Hyperopt hyperopt = Hyperopt(estimator=pipeline, max_evals=1) trained = hyperopt.fit(self.X_train, self.y_train) _ = trained.predict(self.X_test) test_regressor.__name__ = f"test_{clf_name.split('.')[-1]}" return test_regressor regressors = [ "lale.lib.sklearn.BaggingRegressor", "lale.lib.sklearn.DummyRegressor", "lale.lib.sklearn.RandomForestRegressor", "lale.lib.sklearn.DecisionTreeRegressor", "lale.lib.sklearn.ExtraTreesRegressor", "lale.lib.sklearn.GradientBoostingRegressor", "lale.lib.sklearn.LinearRegression", "lale.lib.sklearn.Ridge", "lale.lib.lightgbm.LGBMRegressor", "lale.lib.xgboost.XGBRegressor", "lale.lib.sklearn.AdaBoostRegressor", "lale.lib.sklearn.SGDRegressor", "lale.lib.sklearn.SVR", "lale.lib.sklearn.KNeighborsRegressor", "lale.lib.sklearn.LinearSVR", "lale.lib.sklearn.StackingRegressor", "lale.lib.sklearn.VotingRegressor", ] for clf in regressors: setattr( TestRegression, f"test_{clf.rsplit('.', maxsplit=1)[-1]}", create_function_test_regressor(clf), ) class TestSpuriousSideConstraintsRegression(unittest.TestCase): # This was prompted by a bug, keeping it as it may help with support for other sklearn versions def setUp(self): from sklearn.datasets import make_regression from sklearn.model_selection import train_test_split X, y = make_regression( n_features=4, n_informative=2, random_state=0, shuffle=False ) self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_gradient_boost_regressor(self): reg = GradientBoostingRegressor( alpha=0.9789984970831765, criterion="friedman_mse", init=None, learning_rate=0.1, loss="squared_error", ) reg.fit(self.X_train, self.y_train) def test_sgd_regressor(self): reg = SGDRegressor(loss="squared_error", epsilon=0.2) reg.fit(self.X_train, self.y_train) def test_sgd_regressor_1(self): reg = SGDRegressor(learning_rate="optimal", eta0=0.2) reg.fit(self.X_train, self.y_train) def test_sgd_regressor_2(self): reg = SGDRegressor(early_stopping=False, validation_fraction=0.2) reg.fit(self.X_train, self.y_train) def test_sgd_regressor_3(self): reg = SGDRegressor(l1_ratio=0.2, penalty="l1") reg.fit(self.X_train, self.y_train) class TestFriedmanMSE(unittest.TestCase): # This was prompted buy a bug, keeping it as it may help with support for other sklearn versions def setUp(self): from sklearn.datasets import make_regression from sklearn.model_selection import train_test_split X, y = make_regression( n_features=4, n_informative=2, random_state=0, shuffle=False ) self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_rfr(self): import sklearn if sklearn.__version__ < "1.0": reg = RandomForestRegressor( bootstrap=True, criterion="friedman_mse", max_depth=4, max_features=0.9832410473940374, max_leaf_nodes=None, min_impurity_decrease=0.0, min_impurity_split=None, min_samples_leaf=3, min_samples_split=2, min_weight_fraction_leaf=0.0, n_estimators=29, n_jobs=4, oob_score=False, random_state=33, verbose=0, warm_start=False, ) reg.fit(self.X_train, self.y_train) def test_etr(self): import sklearn if sklearn.__version__ < "1.0": reg = ExtraTreesRegressor( bootstrap=True, criterion="friedman_mse", max_depth=4, max_features=0.9832410473940374, max_leaf_nodes=None, min_impurity_decrease=0.0, min_impurity_split=None, min_samples_leaf=3, min_samples_split=2, min_weight_fraction_leaf=0.0, n_estimators=29, n_jobs=4, oob_score=False, random_state=33, verbose=0, warm_start=False, ) reg.fit(self.X_train, self.y_train) class TestRidge(unittest.TestCase): # This was prompted by a bug, keeping it as it may help with support for other sklearn versions def setUp(self): from sklearn.datasets import make_regression from sklearn.model_selection import train_test_split X, y = make_regression( n_features=4, n_informative=2, random_state=0, shuffle=False ) self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_positive(self): import sklearn from lale.settings import set_disable_data_schema_validation set_disable_data_schema_validation(False) if sklearn.__version__ > "1.0": reg = Ridge(solver="lbfgs", positive=True) reg.fit(self.X_train, self.y_train) with self.assertRaises(ValidationError): reg = Ridge(solver="saga", positive=True) reg = Ridge(solver="auto", positive=True) reg.fit(self.X_train, self.y_train) with self.assertRaises(ValidationError): reg = Ridge(solver="lbfgs", positive=False) reg = Ridge(solver="auto", positive=False) reg.fit(self.X_train, self.y_train)
8,557
32.826087
100
py
lale
lale-master/test/test_core_pipeline.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import pickle import traceback import typing import unittest import numpy as np import sklearn.datasets import sklearn.pipeline from sklearn.feature_selection import SelectKBest as SkSelectKBest from sklearn.metrics import accuracy_score, r2_score from sklearn.model_selection import train_test_split import lale.datasets.openml import lale.helpers from lale.helpers import import_from_sklearn_pipeline from lale.lib.lale import ConcatFeatures, NoOp from lale.lib.sklearn import ( PCA, AdaBoostClassifier, GaussianNB, IsolationForest, KNeighborsClassifier, LinearRegression, LinearSVC, LogisticRegression, Nystroem, OneHotEncoder, PassiveAggressiveClassifier, SelectKBest, SGDClassifier, StandardScaler, ) from lale.lib.xgboost import XGBClassifier from lale.operators import ( TrainableIndividualOp, TrainablePipeline, TrainedIndividualOp, TrainedPipeline, make_choice, make_pipeline, make_union, ) class TestCreation(unittest.TestCase): def setUp(self): data = sklearn.datasets.load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_pipeline_create(self): from lale.operators import Pipeline pipeline = Pipeline(([("pca1", PCA()), ("lr1", LogisticRegression())])) trained = pipeline.fit(self.X_train, self.y_train) predictions = trained.predict(self.X_test) accuracy_score(self.y_test, predictions) def test_pipeline_create_trainable(self): from lale.lib.sklearn import Pipeline as SkPipeline pipeline = SkPipeline(steps=[("pca1", PCA()), ("lr1", LogisticRegression())]) self.assertIsInstance(pipeline, TrainableIndividualOp) trained = pipeline.fit(self.X_train, self.y_train) pca_trained, lr_trained = [op for _, op in trained.hyperparams()["steps"]] self.assertIsInstance(pca_trained, TrainedIndividualOp) self.assertIsInstance(lr_trained, TrainedIndividualOp) predictions = trained.predict(self.X_test) accuracy_score(self.y_test, predictions) def test_pipeline_create_trained(self): from lale.lib.sklearn import Pipeline as SkPipeline orig_trainable = PCA() >> LogisticRegression() orig_trained = orig_trainable.fit(self.X_train, self.y_train) self.assertIsInstance(orig_trained, TrainedPipeline) pca_trained, lr_trained = orig_trained.steps_list() pre_trained = SkPipeline(steps=[("pca1", pca_trained), ("lr1", lr_trained)]) self.assertIsInstance(pre_trained, TrainedIndividualOp) predictions = pre_trained.predict(self.X_test) accuracy_score(self.y_test, predictions) def test_pipeline_clone(self): from sklearn.base import clone from lale.operators import Pipeline pipeline = Pipeline(([("pca1", PCA()), ("lr1", LogisticRegression())])) trained = pipeline.fit(self.X_train, self.y_train) predictions = trained.predict(self.X_test) orig_acc = accuracy_score(self.y_test, predictions) cloned_pipeline = clone(pipeline) trained = cloned_pipeline.fit(self.X_train, self.y_train) predictions = trained.predict(self.X_test) cloned_acc = accuracy_score(self.y_test, predictions) self.assertEqual(orig_acc, cloned_acc) def test_make_pipeline(self): tfm = PCA(n_components=10) clf = LogisticRegression(random_state=42) trainable = make_pipeline(tfm, clf) digits = sklearn.datasets.load_digits() trained = trainable.fit(digits.data, digits.target) _ = trained.predict(digits.data) def test_compose2(self): tfm = PCA(n_components=10) clf = LogisticRegression(random_state=42) trainable = tfm >> clf digits = sklearn.datasets.load_digits() trained = trainable.fit(digits.data, digits.target) _ = trained.predict(digits.data) def test_compose3(self): nys = Nystroem(n_components=15) pca = PCA(n_components=10) lr = LogisticRegression(random_state=42) trainable = nys >> pca >> lr digits = sklearn.datasets.load_digits() trained = trainable.fit(digits.data, digits.target) _ = trained.predict(digits.data) def test_pca_nys_lr(self): nys = Nystroem(n_components=15) pca = PCA(n_components=10) lr = LogisticRegression(random_state=42) trainable = make_union(nys, pca) >> lr digits = sklearn.datasets.load_digits() trained = trainable.fit(digits.data, digits.target) _ = trained.predict(digits.data) def test_compose4(self): digits = sklearn.datasets.load_digits() _ = digits ohe = OneHotEncoder(handle_unknown=OneHotEncoder.enum.handle_unknown.ignore) ohe.get_params() no_op = NoOp() pca = PCA() nys = Nystroem() lr = LogisticRegression() knn = KNeighborsClassifier() step1 = ohe | no_op step2 = pca | nys step3 = lr | knn model_plan = step1 >> step2 >> step3 _ = model_plan # TODO: optimize on this plan and then fit and predict def test_compose5(self): ohe = OneHotEncoder(handle_unknown=OneHotEncoder.enum.handle_unknown.ignore) digits = sklearn.datasets.load_digits() lr = LogisticRegression() lr_trained = lr.fit(digits.data, digits.target) lr_trained.predict(digits.data) pipeline1 = ohe >> lr pipeline1_trained = pipeline1.fit(digits.data, digits.target) pipeline1_trained.predict(digits.data) def test_compare_with_sklearn(self): tfm = PCA() clf = LogisticRegression( LogisticRegression.enum.solver.saga, LogisticRegression.enum.multi_class.auto, ) trainable = make_pipeline(tfm, clf) digits = sklearn.datasets.load_digits() trained = trainable.fit(digits.data, digits.target) predicted = trained.predict(digits.data) from sklearn.decomposition import PCA as SklearnPCA from sklearn.linear_model import LogisticRegression as SklearnLR sklearn_pipeline = sklearn.pipeline.make_pipeline( SklearnPCA(), SklearnLR(solver="saga", multi_class="auto") ) sklearn_pipeline.fit(digits.data, digits.target) predicted_sklearn = sklearn_pipeline.predict(digits.data) lale_score = accuracy_score(digits.target, predicted) scikit_score = accuracy_score(digits.target, predicted_sklearn) self.assertEqual(lale_score, scikit_score) class TestImportExport(unittest.TestCase): def setUp(self): data = sklearn.datasets.load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) @classmethod def get_sklearn_params(cls, op): lale_sklearn_impl = op._impl_instance() wrapped_model = getattr(lale_sklearn_impl, "_wrapped_model", None) if wrapped_model is not None: lale_sklearn_impl = wrapped_model return lale_sklearn_impl.get_params() def assert_equal_predictions(self, pipeline1, pipeline2): trained = pipeline1.fit(self.X_train, self.y_train) predictions1 = trained.predict(self.X_test) trained = pipeline2.fit(self.X_train, self.y_train) predictions2 = trained.predict(self.X_test) for i, p1 in enumerate(predictions1): self.assertEqual(p1, predictions2[i]) def test_import_from_sklearn_pipeline(self): from sklearn.feature_selection import f_regression from sklearn.pipeline import Pipeline from sklearn.svm import SVC as SklearnSVC anova_filter = SkSelectKBest(f_regression, k=3) clf = SklearnSVC(kernel="linear") sklearn_pipeline = Pipeline([("anova", anova_filter), ("svc", clf)]) lale_pipeline = typing.cast( TrainablePipeline, import_from_sklearn_pipeline(sklearn_pipeline), ) for i, pipeline_step in enumerate(sklearn_pipeline.named_steps): sklearn_step_params = sklearn_pipeline.named_steps[ pipeline_step ].get_params() lale_sklearn_params = self.get_sklearn_params(lale_pipeline.steps_list()[i]) self.assertEqual(sklearn_step_params, lale_sklearn_params) self.assert_equal_predictions(sklearn_pipeline, lale_pipeline) def test_import_from_sklearn_pipeline1(self): from sklearn.decomposition import PCA as SklearnPCA from sklearn.neighbors import KNeighborsClassifier as SklearnKNN sklearn_pipeline = sklearn.pipeline.make_pipeline( SklearnPCA(n_components=3), SklearnKNN() ) lale_pipeline = typing.cast( TrainablePipeline, import_from_sklearn_pipeline(sklearn_pipeline), ) for i, pipeline_step in enumerate(sklearn_pipeline.named_steps): sklearn_step_params = sklearn_pipeline.named_steps[ pipeline_step ].get_params() lale_sklearn_params = self.get_sklearn_params(lale_pipeline.steps_list()[i]) self.assertEqual(sklearn_step_params, lale_sklearn_params) self.assert_equal_predictions(sklearn_pipeline, lale_pipeline) def test_import_from_sklearn_pipeline_feature_union(self): from sklearn.decomposition import PCA as SklearnPCA from sklearn.kernel_approximation import Nystroem as SklearnNystroem from sklearn.neighbors import KNeighborsClassifier as SklearnKNN from sklearn.pipeline import FeatureUnion union = FeatureUnion( [ ("pca", SklearnPCA(n_components=1)), ("nys", SklearnNystroem(n_components=2, random_state=42)), ] ) sklearn_pipeline = sklearn.pipeline.make_pipeline(union, SklearnKNN()) lale_pipeline = typing.cast( TrainablePipeline, import_from_sklearn_pipeline(sklearn_pipeline), ) self.assertEqual(len(lale_pipeline.edges()), 3) self.assertIsInstance(lale_pipeline.edges()[0][0], PCA) # type: ignore self.assertIsInstance(lale_pipeline.edges()[0][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[1][0], Nystroem) # type: ignore self.assertIsInstance(lale_pipeline.edges()[1][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[2][0], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[2][1], KNeighborsClassifier) # type: ignore self.assert_equal_predictions(sklearn_pipeline, lale_pipeline) def test_import_from_sklearn_pipeline_nested_pipeline(self): from sklearn.decomposition import PCA as SklearnPCA from sklearn.kernel_approximation import Nystroem as SklearnNystroem from sklearn.neighbors import KNeighborsClassifier as SklearnKNN from sklearn.pipeline import FeatureUnion union = FeatureUnion( [ ( "selectkbest_pca", sklearn.pipeline.make_pipeline( SkSelectKBest(k=3), SklearnPCA(n_components=1) ), ), ("nys", SklearnNystroem(n_components=2, random_state=42)), ] ) sklearn_pipeline = sklearn.pipeline.make_pipeline(union, SklearnKNN()) lale_pipeline = typing.cast( TrainablePipeline, import_from_sklearn_pipeline(sklearn_pipeline), ) self.assertEqual(len(lale_pipeline.edges()), 4) # These assertions assume topological sort self.assertIsInstance(lale_pipeline.edges()[0][0], SelectKBest) # type: ignore self.assertIsInstance(lale_pipeline.edges()[0][1], PCA) # type: ignore self.assertIsInstance(lale_pipeline.edges()[1][0], PCA) # type: ignore self.assertIsInstance(lale_pipeline.edges()[1][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[2][0], Nystroem) # type: ignore self.assertIsInstance(lale_pipeline.edges()[2][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[3][0], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[3][1], KNeighborsClassifier) # type: ignore self.assert_equal_predictions(sklearn_pipeline, lale_pipeline) def test_import_from_sklearn_pipeline_nested_pipeline1(self): from sklearn.decomposition import PCA as SklearnPCA from sklearn.kernel_approximation import Nystroem as SklearnNystroem from sklearn.neighbors import KNeighborsClassifier as SklearnKNN from sklearn.pipeline import FeatureUnion union = FeatureUnion( [ ( "selectkbest_pca", sklearn.pipeline.make_pipeline( SkSelectKBest(k=3), FeatureUnion( [ ("pca", SklearnPCA(n_components=1)), ( "nested_pipeline", sklearn.pipeline.make_pipeline( SkSelectKBest(k=2), SklearnNystroem() ), ), ] ), ), ), ("nys", SklearnNystroem(n_components=2, random_state=42)), ] ) sklearn_pipeline = sklearn.pipeline.make_pipeline(union, SklearnKNN()) lale_pipeline = typing.cast( TrainablePipeline, import_from_sklearn_pipeline(sklearn_pipeline), ) self.assertEqual(len(lale_pipeline.edges()), 8) # These assertions assume topological sort, which may not be unique. So the assertions are brittle. self.assertIsInstance(lale_pipeline.edges()[0][0], SelectKBest) # type: ignore self.assertIsInstance(lale_pipeline.edges()[0][1], PCA) # type: ignore self.assertIsInstance(lale_pipeline.edges()[1][0], SelectKBest) # type: ignore self.assertIsInstance(lale_pipeline.edges()[1][1], SelectKBest) # type: ignore self.assertIsInstance(lale_pipeline.edges()[2][0], SelectKBest) # type: ignore self.assertIsInstance(lale_pipeline.edges()[2][1], Nystroem) # type: ignore self.assertIsInstance(lale_pipeline.edges()[3][0], PCA) # type: ignore self.assertIsInstance(lale_pipeline.edges()[3][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[4][0], Nystroem) # type: ignore self.assertIsInstance(lale_pipeline.edges()[4][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[5][0], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[5][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[6][0], Nystroem) # type: ignore self.assertIsInstance(lale_pipeline.edges()[6][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[7][0], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[7][1], KNeighborsClassifier) # type: ignore self.assert_equal_predictions(sklearn_pipeline, lale_pipeline) def test_import_from_sklearn_pipeline_nested_pipeline2(self): from sklearn.decomposition import PCA as SklearnPCA from sklearn.kernel_approximation import Nystroem as SklearnNystroem from sklearn.neighbors import KNeighborsClassifier as SklearnKNN from sklearn.pipeline import FeatureUnion union = FeatureUnion( [ ( "selectkbest_pca", sklearn.pipeline.make_pipeline( SkSelectKBest(k=3), sklearn.pipeline.make_pipeline( SkSelectKBest(k=2), SklearnPCA() ), ), ), ("nys", SklearnNystroem(n_components=2, random_state=42)), ] ) sklearn_pipeline = sklearn.pipeline.make_pipeline(union, SklearnKNN()) lale_pipeline = typing.cast( TrainablePipeline, import_from_sklearn_pipeline(sklearn_pipeline), ) self.assertEqual(len(lale_pipeline.edges()), 5) self.assertIsInstance(lale_pipeline.edges()[0][0], SelectKBest) # type: ignore self.assertIsInstance(lale_pipeline.edges()[0][1], SelectKBest) # type: ignore self.assertIsInstance(lale_pipeline.edges()[1][0], SelectKBest) # type: ignore self.assertIsInstance(lale_pipeline.edges()[1][1], PCA) # type: ignore self.assertIsInstance(lale_pipeline.edges()[2][0], PCA) # type: ignore self.assertIsInstance(lale_pipeline.edges()[2][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[3][0], Nystroem) # type: ignore self.assertIsInstance(lale_pipeline.edges()[3][1], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[4][0], ConcatFeatures) # type: ignore self.assertIsInstance(lale_pipeline.edges()[4][1], KNeighborsClassifier) # type: ignore self.assert_equal_predictions(sklearn_pipeline, lale_pipeline) def test_import_from_sklearn_pipeline_noop(self): from sklearn.ensemble import GradientBoostingClassifier from sklearn.pipeline import Pipeline pipe = Pipeline([("noop", None), ("gbc", GradientBoostingClassifier())]) _ = import_from_sklearn_pipeline(pipe) def test_import_from_sklearn_pipeline_noop1(self): from sklearn.ensemble import GradientBoostingClassifier from sklearn.pipeline import Pipeline pipe = Pipeline([("noop", NoOp()), ("gbc", GradientBoostingClassifier())]) _ = import_from_sklearn_pipeline(pipe) def test_import_from_sklearn_pipeline_no_wrapper(self): from sklearn.neighbors import LocalOutlierFactor from sklearn.pipeline import make_pipeline as sk_make_pipeline sklearn_pipeline = sk_make_pipeline(PCA(), LocalOutlierFactor()) _ = import_from_sklearn_pipeline(sklearn_pipeline, fitted=False) def test_import_from_sklearn_pipeline_higherorder(self): from sklearn.ensemble import VotingClassifier as VC from sklearn.feature_selection import f_regression from sklearn.pipeline import Pipeline from sklearn.svm import SVC as SklearnSVC anova_filter = SkSelectKBest(f_regression, k=3) clf = SklearnSVC(kernel="linear") sklearn_pipeline = Pipeline( [("anova", anova_filter), ("vc_svc", VC(estimators=[("clf", clf)]))] ) lale_pipeline = typing.cast( TrainablePipeline, import_from_sklearn_pipeline(sklearn_pipeline), ) # for i, pipeline_step in enumerate(sklearn_pipeline.named_steps): # sklearn_step_params = sklearn_pipeline.named_steps[ # pipeline_step # ].get_params() # lale_sklearn_params = self.get_sklearn_params(lale_pipeline.steps_list()[i]) # self.assertEqual(sklearn_step_params, lale_sklearn_params) self.assert_equal_predictions(sklearn_pipeline, lale_pipeline) def test_export_to_sklearn_pipeline(self): lale_pipeline = PCA(n_components=3) >> KNeighborsClassifier() trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) sklearn_pipeline = trained_lale_pipeline.export_to_sklearn_pipeline() for i, pipeline_step in enumerate(sklearn_pipeline.named_steps): sklearn_step_params = sklearn_pipeline.named_steps[ pipeline_step ].get_params() lale_sklearn_params = self.get_sklearn_params( trained_lale_pipeline.steps_list()[i] ) self.assertEqual(sklearn_step_params, lale_sklearn_params) self.assert_equal_predictions(sklearn_pipeline, trained_lale_pipeline) def test_export_to_sklearn_pipeline1(self): lale_pipeline = SkSelectKBest(k=3) >> KNeighborsClassifier() trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) sklearn_pipeline = trained_lale_pipeline.export_to_sklearn_pipeline() for i, pipeline_step in enumerate(sklearn_pipeline.named_steps): sklearn_step_params = type(sklearn_pipeline.named_steps[pipeline_step]) lale_sklearn_params = ( type(trained_lale_pipeline.steps_list()[i]._impl._wrapped_model) if hasattr( trained_lale_pipeline.steps_list()[i]._impl, "_wrapped_model" ) else type(trained_lale_pipeline.steps_list()[i]._impl) ) self.assertEqual(sklearn_step_params, lale_sklearn_params) self.assert_equal_predictions(sklearn_pipeline, trained_lale_pipeline) def test_export_to_sklearn_pipeline2(self): from sklearn.pipeline import FeatureUnion lale_pipeline = ( ( ( (PCA(svd_solver="randomized", random_state=42) & SkSelectKBest(k=3)) >> ConcatFeatures() ) & Nystroem(random_state=42) ) >> ConcatFeatures() >> KNeighborsClassifier() ) trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) sklearn_pipeline = trained_lale_pipeline.export_to_sklearn_pipeline() self.assertIsInstance( sklearn_pipeline.named_steps["featureunion"], FeatureUnion ) from sklearn.neighbors import KNeighborsClassifier as SklearnKNN self.assertIsInstance( sklearn_pipeline.named_steps["kneighborsclassifier"], SklearnKNN ) self.assert_equal_predictions(sklearn_pipeline, trained_lale_pipeline) def test_export_to_sklearn_pipeline3(self): from sklearn.pipeline import FeatureUnion lale_pipeline = ( ( (PCA() >> SkSelectKBest(k=2)) & (Nystroem(random_state=42) >> SkSelectKBest(k=3)) & (SkSelectKBest(k=3)) ) >> ConcatFeatures() >> SkSelectKBest(k=2) >> LogisticRegression() ) trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) sklearn_pipeline = trained_lale_pipeline.export_to_sklearn_pipeline() self.assertIsInstance( sklearn_pipeline.named_steps["featureunion"], FeatureUnion ) self.assertIsInstance( sklearn_pipeline.named_steps["selectkbest"], SkSelectKBest ) from sklearn.linear_model import LogisticRegression as SklearnLR self.assertIsInstance( sklearn_pipeline.named_steps["logisticregression"], SklearnLR ) self.assert_equal_predictions(sklearn_pipeline, trained_lale_pipeline) def test_export_to_sklearn_pipeline4(self): lale_pipeline = make_pipeline(LogisticRegression()) trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) sklearn_pipeline = trained_lale_pipeline.export_to_sklearn_pipeline() from sklearn.linear_model import LogisticRegression as SklearnLR self.assertIsInstance( sklearn_pipeline.named_steps["logisticregression"], SklearnLR ) self.assert_equal_predictions(sklearn_pipeline, trained_lale_pipeline) def test_export_to_sklearn_pipeline5(self): lale_pipeline = PCA() >> (XGBClassifier() | SGDClassifier()) with self.assertRaises(ValueError): _ = lale_pipeline.export_to_sklearn_pipeline() def test_export_to_pickle(self): lale_pipeline = make_pipeline(LogisticRegression()) trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) pickle.dumps(lale_pipeline) pickle.dumps(trained_lale_pipeline) def test_import_from_sklearn_pipeline2(self): from sklearn.feature_selection import f_regression from sklearn.pipeline import Pipeline from sklearn.svm import SVC as SklearnSVC anova_filter = SkSelectKBest(f_regression, k=3) clf = SklearnSVC(kernel="linear") sklearn_pipeline = Pipeline([("anova", anova_filter), ("svc", clf)]) sklearn_pipeline.fit(self.X_train, self.y_train) lale_pipeline = typing.cast( TrainedPipeline, import_from_sklearn_pipeline(sklearn_pipeline), ) lale_pipeline.predict(self.X_test) def test_import_from_sklearn_pipeline3(self): from sklearn.feature_selection import f_regression from sklearn.pipeline import Pipeline from sklearn.svm import SVC as SklearnSVC anova_filter = SkSelectKBest(f_regression, k=3) clf = SklearnSVC(kernel="linear") sklearn_pipeline = Pipeline([("anova", anova_filter), ("svc", clf)]) lale_pipeline = typing.cast( TrainablePipeline, import_from_sklearn_pipeline(sklearn_pipeline, fitted=False), ) with self.assertRaises( ValueError ): # fitted=False returns a Trainable, so calling predict is invalid. lale_pipeline.predict(self.X_test) def test_export_to_sklearn_pipeline_with_noop_1(self): lale_pipeline = NoOp() >> PCA(n_components=3) >> KNeighborsClassifier() trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) sklearn_pipeline = trained_lale_pipeline.export_to_sklearn_pipeline() self.assert_equal_predictions(sklearn_pipeline, trained_lale_pipeline) def test_export_to_sklearn_pipeline_with_noop_2(self): lale_pipeline = PCA(n_components=3) >> NoOp() >> KNeighborsClassifier() trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) sklearn_pipeline = trained_lale_pipeline.export_to_sklearn_pipeline() self.assert_equal_predictions(sklearn_pipeline, trained_lale_pipeline) def test_export_to_sklearn_pipeline_with_noop_3(self): # This test is probably unnecessary, but doesn't harm at this point lale_pipeline = PCA(n_components=3) >> KNeighborsClassifier() >> NoOp() trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) _ = trained_lale_pipeline.export_to_sklearn_pipeline() def test_export_to_sklearn_pipeline_with_noop_4(self): lale_pipeline = NoOp() >> KNeighborsClassifier() trained_lale_pipeline = lale_pipeline.fit(self.X_train, self.y_train) sklearn_pipeline = trained_lale_pipeline.export_to_sklearn_pipeline() self.assert_equal_predictions(sklearn_pipeline, trained_lale_pipeline) class TestComposition(unittest.TestCase): def setUp(self): data = sklearn.datasets.load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_two_estimators_predict(self): pipeline = ( StandardScaler() >> (PCA() & Nystroem() & LogisticRegression()) >> ConcatFeatures() >> NoOp() >> LogisticRegression() ) trained = pipeline.fit(self.X_train, self.y_train) trained.predict(self.X_test) def test_two_estimators_predict1(self): pipeline = ( StandardScaler() >> (PCA() & Nystroem() & PassiveAggressiveClassifier()) >> ConcatFeatures() >> NoOp() >> PassiveAggressiveClassifier() ) trained = pipeline.fit(self.X_train, self.y_train) trained.predict(self.X_test) def test_two_estimators_predict_proba(self): pipeline = ( StandardScaler() >> (PCA() & Nystroem() & LogisticRegression()) >> ConcatFeatures() >> NoOp() >> LogisticRegression() ) trained = pipeline.fit(self.X_train, self.y_train) trained.predict_proba(self.X_test) def test_two_estimators_predict_proba1(self): pipeline = ( StandardScaler() >> (PCA() & Nystroem() & GaussianNB()) >> ConcatFeatures() >> NoOp() >> GaussianNB() ) pipeline.fit(self.X_train, self.y_train) pipeline.predict_proba(self.X_test) def test_multiple_estimators_predict_predict_proba(self): pipeline = ( StandardScaler() >> (LogisticRegression() & PCA()) >> ConcatFeatures() >> (NoOp() & LinearSVC()) >> ConcatFeatures() >> KNeighborsClassifier() ) pipeline.fit(self.X_train, self.y_train) _ = pipeline.predict_proba(self.X_test) _ = pipeline.predict(self.X_test) def test_two_transformers(self): tfm1 = PCA() tfm2 = Nystroem() trainable = tfm1 >> tfm2 digits = sklearn.datasets.load_digits() trained = trainable.fit(digits.data, digits.target) _ = trained.transform(digits.data) def test_duplicate_instances(self): tfm = PCA() clf = LogisticRegression( LogisticRegression.enum.solver.lbfgs, LogisticRegression.enum.multi_class.auto, ) with self.assertRaises(ValueError): _ = make_pipeline(tfm, tfm, clf) def test_increase_num_rows_predict(self): from test.mock_custom_operators import IncreaseRows increase_rows = IncreaseRows() trainable = increase_rows >> LogisticRegression() iris = sklearn.datasets.load_iris() X, y = iris.data, iris.target trained = trainable.fit(X, y) y_pred = trained.predict(X) self.assertEqual(len(y_pred), len(y) + increase_rows.impl.n_rows) def test_increase_num_rows_transform_X_y(self): from test.mock_custom_operators import IncreaseRows increase_rows_4 = IncreaseRows(n_rows=4) increase_rows_2 = IncreaseRows(n_rows=2) trainable = increase_rows_4 >> increase_rows_2 iris = sklearn.datasets.load_iris() X, y = iris.data, iris.target trained = trainable.fit(X, y) output_X, output_y = trained.transform_X_y(X, y) self.assertEqual(output_X.shape[0], X.shape[0] + 4 + 2) self.assertEqual(output_X.shape[1], X.shape[1]) self.assertEqual(output_y.shape[0], y.shape[0] + 4 + 2) def test_remove_last1(self): pipeline = ( StandardScaler() >> (PCA() & Nystroem() & PassiveAggressiveClassifier()) >> ConcatFeatures() >> NoOp() >> PassiveAggressiveClassifier() ) new_pipeline = pipeline.remove_last() self.assertEqual(len(new_pipeline._steps), 6) self.assertEqual(len(pipeline._steps), 7) def test_remove_last2(self): pipeline = ( StandardScaler() >> (PCA() & Nystroem() & PassiveAggressiveClassifier()) >> ConcatFeatures() >> NoOp() >> (PassiveAggressiveClassifier() & LogisticRegression()) ) with self.assertRaises(ValueError): pipeline.remove_last() def test_remove_last3(self): pipeline = ( StandardScaler() >> (PCA() & Nystroem() & PassiveAggressiveClassifier()) >> ConcatFeatures() >> NoOp() >> PassiveAggressiveClassifier() ) pipeline.remove_last().freeze_trainable() def test_remove_last4(self): pipeline = ( StandardScaler() >> (PCA() & Nystroem() & PassiveAggressiveClassifier()) >> ConcatFeatures() >> NoOp() >> PassiveAggressiveClassifier() ) new_pipeline = pipeline.remove_last(inplace=True) self.assertEqual(len(new_pipeline._steps), 6) self.assertEqual(len(pipeline._steps), 6) def test_remove_last5(self): pipeline = ( StandardScaler() >> (PCA() & Nystroem() & PassiveAggressiveClassifier()) >> ConcatFeatures() >> NoOp() >> PassiveAggressiveClassifier() ) pipeline.remove_last(inplace=True).freeze_trainable() class TestAutoPipeline(unittest.TestCase): def _fit_predict(self, prediction_type, all_X, all_y, verbose=True): if verbose: _file_name, _line, fn_name, _text = traceback.extract_stack()[-2] print(f"--- TestAutoPipeline.{fn_name}() ---") from lale.lib.lale import AutoPipeline train_X, test_X, train_y, test_y = train_test_split(all_X, all_y) trainable = AutoPipeline( prediction_type=prediction_type, max_evals=10, verbose=verbose ) trained = trainable.fit(train_X, train_y) predicted = trained.predict(test_X) if prediction_type == "regression": score = f"r2 score {r2_score(test_y, predicted):.2f}" else: score = f"accuracy {accuracy_score(test_y, predicted):.1%}" if verbose: print(score) pipe = trained.get_pipeline() assert pipe is not None print(pipe.pretty_print(show_imports=False)) def test_sklearn_iris(self): # classification, only numbers, no missing values all_X, all_y = sklearn.datasets.load_iris(return_X_y=True) self._fit_predict("classification", all_X, all_y) def test_sklearn_digits(self): # classification, numbers but some appear categorical, no missing values all_X, all_y = sklearn.datasets.load_digits(return_X_y=True) self._fit_predict("classification", all_X, all_y) def test_sklearn_boston(self): # regression, categoricals+numbers, no missing values from lale.datasets.util import load_boston all_X, all_y = load_boston(return_X_y=True) self._fit_predict("regression", all_X, all_y) def test_sklearn_diabetes(self): # regression, categoricals+numbers, no missing values all_X, all_y = sklearn.datasets.load_diabetes(return_X_y=True) self._fit_predict("regression", all_X, all_y) def test_openml_creditg(self): # classification, categoricals+numbers incl. string, no missing values (orig_train_X, orig_train_y), _ = lale.datasets.openml.fetch( "credit-g", "classification", preprocess=False ) subsample_X, _, subsample_y, _ = train_test_split( orig_train_X, orig_train_y, train_size=0.05 ) self._fit_predict("classification", subsample_X, subsample_y) def test_missing_iris(self): # classification, only numbers, synthetically added missing values all_X, all_y = sklearn.datasets.load_iris(return_X_y=True) with_missing_X = lale.helpers.add_missing_values(all_X) with self.assertRaisesRegex(ValueError, "Input.*contains NaN"): lr_trainable = LogisticRegression() _ = lr_trainable.fit(with_missing_X, all_y) self._fit_predict("classification", with_missing_X, all_y) def test_missing_boston(self): # regression, categoricals+numbers, synthetically added missing values from lale.datasets.util import load_boston all_X, all_y = load_boston(return_X_y=True) with_missing_X = lale.helpers.add_missing_values(all_X) with self.assertRaisesRegex(ValueError, "Input.*contains NaN"): lr_trainable = LinearRegression() _ = lr_trainable.fit(with_missing_X, all_y) self._fit_predict("regression", with_missing_X, all_y) def test_missing_creditg(self): # classification, categoricals+numbers incl. string, synth. missing (orig_train_X, orig_train_y), _ = lale.datasets.openml.fetch( "credit-g", "classification", preprocess=False ) subsample_X, _, subsample_y, _ = train_test_split( orig_train_X, orig_train_y, train_size=0.05 ) with_missing_X = lale.helpers.add_missing_values(subsample_X) self._fit_predict("classification", with_missing_X, subsample_y) class TestOperatorChoice(unittest.TestCase): def test_make_choice_with_instance(self): from sklearn.datasets import load_iris iris = load_iris() X, y = iris.data, iris.target tfm = PCA() | Nystroem() | NoOp() with self.assertRaises(AttributeError): # we are trying to trigger a runtime error here, so we ignore the static warning _ = tfm.fit(X, y) # type: ignore _ = (OneHotEncoder | NoOp) >> tfm >> (LogisticRegression | KNeighborsClassifier) _ = ( (OneHotEncoder | NoOp) >> (PCA | Nystroem) >> (LogisticRegression | KNeighborsClassifier) ) _ = ( make_choice(OneHotEncoder, NoOp) >> make_choice(PCA, Nystroem) >> make_choice(LogisticRegression, KNeighborsClassifier) ) class TestScore(unittest.TestCase): def setUp(self): data = sklearn.datasets.load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_trained_pipeline(self): trainable_pipeline = StandardScaler() >> LogisticRegression() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) score = trained_pipeline.score(self.X_test, self.y_test) predictions = trained_pipeline.predict(self.X_test) accuracy = accuracy_score(self.y_test, predictions) self.assertEqual(accuracy, score) def test_trainable_pipeline(self): trainable_pipeline = StandardScaler() >> LogisticRegression() trainable_pipeline.fit(self.X_train, self.y_train) score = trainable_pipeline.score(self.X_test, self.y_test) predictions = trainable_pipeline.predict(self.X_test) accuracy = accuracy_score(self.y_test, predictions) self.assertEqual(accuracy, score) def test_planned_pipeline(self): planned_pipeline = StandardScaler >> LogisticRegression with self.assertRaises(AttributeError): planned_pipeline.score(self.X_test, self.y_test) # type: ignore class TestScoreSamples(unittest.TestCase): def setUp(self): data = sklearn.datasets.load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) import warnings warnings.filterwarnings("ignore") def test_trained_pipeline(self): trainable_pipeline = StandardScaler() >> IsolationForest() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) _ = trained_pipeline.score_samples(self.X_test) def test_trainable_pipeline(self): trainable_pipeline = StandardScaler() >> IsolationForest() trainable_pipeline.fit(self.X_train, self.y_train) with self.assertWarns(DeprecationWarning): _ = trainable_pipeline.score_samples(self.X_test) def test_planned_pipeline(self): planned_pipeline = StandardScaler >> IsolationForest with self.assertRaises(AttributeError): planned_pipeline.score_samples(self.X_test) # type: ignore def test_with_incompatible_estimator(self): trainable_pipeline = StandardScaler() >> LogisticRegression() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) with self.assertRaises(AttributeError): _ = trained_pipeline.score_samples(self.X_test) class TestPredictLogProba(unittest.TestCase): def setUp(self): data = sklearn.datasets.load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) import warnings warnings.filterwarnings("ignore") def test_trained_pipeline(self): trainable_pipeline = StandardScaler() >> AdaBoostClassifier() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) _ = trained_pipeline.predict_log_proba(self.X_test) def test_trainable_pipeline(self): trainable_pipeline = StandardScaler() >> AdaBoostClassifier() trainable_pipeline.fit(self.X_train, self.y_train) with self.assertWarns(DeprecationWarning): _ = trainable_pipeline.predict_log_proba(self.X_test) def test_planned_pipeline(self): planned_pipeline = StandardScaler >> AdaBoostClassifier with self.assertRaises(AttributeError): planned_pipeline.predict_log_proba(self.X_test) # type: ignore def test_with_incompatible_estimator(self): trainable_pipeline = StandardScaler() >> IsolationForest() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) with self.assertRaises(AttributeError): _ = trained_pipeline.predict_log_proba(self.X_test) def test_with_incompatible_estimator_1(self): trainable_pipeline = IsolationForest() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) with self.assertRaises(AttributeError): _ = trained_pipeline.predict_log_proba(self.X_test) class TestPartialFit(unittest.TestCase): def setUp(self): data = sklearn.datasets.load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) import warnings warnings.filterwarnings("ignore") def test_first_call(self): trainable_pipeline = StandardScaler() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline.freeze_trained() >> SGDClassifier() new_trained_pipeline = new_pipeline.partial_fit( self.X_train, self.y_train, classes=[0, 1, 2] ) _ = new_trained_pipeline.predict(self.X_test) def test_multiple_calls_with_classes(self): trainable_pipeline = StandardScaler() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline.freeze_trained() >> SGDClassifier() new_trained_pipeline = new_pipeline.partial_fit( self.X_train, self.y_train, classes=[0, 1, 2] ) new_trained_pipeline = new_trained_pipeline.partial_fit( self.X_test, self.y_test, classes=[0, 1, 2] ) _ = new_trained_pipeline.predict(self.X_test) def _last_impl_has(self, op, attr): last = op.get_last() assert last is not None return hasattr(last._impl, attr) def test_second_call_without_classes(self): trainable_pipeline = StandardScaler() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline.freeze_trained() >> SGDClassifier() new_trained_pipeline = new_pipeline.partial_fit( self.X_train, self.y_train, classes=[0, 1, 2] ) # Once SGDClassifier is trained, it has a classes_ attribute. self.assertTrue(self._last_impl_has(new_trained_pipeline, "classes_")) new_trained_pipeline = new_trained_pipeline.partial_fit( self.X_test, self.y_test ) _ = new_trained_pipeline.predict(self.X_test) def test_second_call_with_different_classes(self): trainable_pipeline = StandardScaler() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline.freeze_trained() >> SGDClassifier() new_trained_pipeline = new_pipeline.partial_fit( self.X_train, self.y_train, classes=[0, 1, 2] ) # Once SGDClassifier is trained, it has a classes_ attribute. self.assertTrue(self._last_impl_has(new_trained_pipeline, "classes_")) subset_labels = self.y_test[np.where(self.y_test != 0)] subset_X = self.X_test[0 : len(subset_labels)] new_trained_pipeline = new_trained_pipeline.partial_fit(subset_X, subset_labels) _ = new_trained_pipeline.predict(self.X_test) def test_second_call_with_different_classes_trainable(self): trainable_pipeline = StandardScaler() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline.freeze_trained() >> SGDClassifier() new_pipeline.partial_fit(self.X_train, self.y_train, classes=[0, 1, 2]) # Once SGDClassifier is trained, it has a classes_ attribute. self.assertTrue(self._last_impl_has(new_pipeline._trained, "classes_")) subset_labels = self.y_test[np.where(self.y_test != 0)] subset_X = self.X_test[0 : len(subset_labels)] new_trained_pipeline = new_pipeline.partial_fit(subset_X, subset_labels) _ = new_trained_pipeline.predict(self.X_test) def test_call_on_trainable(self): trainable_pipeline = StandardScaler() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline.freeze_trained() >> SGDClassifier() new_pipeline.partial_fit(self.X_train, self.y_train, classes=[0, 1, 2]) new_pipeline.pretty_print() new_trained_pipeline = new_pipeline.partial_fit( self.X_test, self.y_test, classes=[0, 1, 2] ) self.assertEqual(new_trained_pipeline, new_pipeline._trained) _ = new_trained_pipeline.predict(self.X_test) new_pipeline.partial_fit(self.X_train, self.y_train, classes=[0, 1, 2]) def test_call_on_trainable_with_freeze_trained_prefix(self): trainable_pipeline = StandardScaler() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline >> SGDClassifier() new_pipeline.partial_fit(self.X_train, self.y_train, classes=[0, 1, 2]) new_pipeline.pretty_print() new_trained_pipeline = new_pipeline.partial_fit( self.X_test, self.y_test, classes=[0, 1, 2] ) self.assertEqual(new_trained_pipeline, new_pipeline._trained) _ = new_trained_pipeline.predict(self.X_test) new_pipeline.partial_fit(self.X_train, self.y_train, classes=[0, 1, 2]) def test_call_on_trainable_with_freeze_trained_prefix_false(self): trainable_pipeline = StandardScaler() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline >> SGDClassifier() with self.assertRaises(ValueError): new_pipeline.partial_fit( self.X_train, self.y_train, freeze_trained_prefix=False, classes=[0, 1, 2], ) def test_call_on_trained_with_freeze_trained_prefix(self): trainable_pipeline = StandardScaler() >> SGDClassifier() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline new_pipeline.partial_fit(self.X_train, self.y_train, classes=[0, 1, 2]) new_pipeline.pretty_print() new_trained_pipeline = new_pipeline.partial_fit( self.X_test, self.y_test, classes=[0, 1, 2] ) _ = new_trained_pipeline.predict(self.X_test) new_pipeline.partial_fit(self.X_train, self.y_train, classes=[0, 1, 2]) def test_call_on_trained_with_freeze_trained_prefix_false(self): trainable_pipeline = StandardScaler() >> SGDClassifier() trained_pipeline = trainable_pipeline.fit(self.X_train, self.y_train) new_pipeline = trained_pipeline with self.assertRaises(ValueError): new_pipeline.partial_fit( self.X_train, self.y_train, freeze_trained_prefix=False, classes=[0, 1, 2], )
48,983
42.348673
107
py
lale
lale-master/test/test_autoai_libs.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import unittest import numpy as np import pandas as pd import sklearn.datasets import sklearn.model_selection import lale.lib.autoai_libs # from lale.datasets.uci import fetch_household_power_consumption from lale.lib.autoai_libs import float32_transform from lale.lib.lale import Hyperopt from lale.lib.sklearn import LogisticRegression as LR from lale.lib.xgboost.xgb_classifier import XGBClassifier class TestAutoaiLibs(unittest.TestCase): @classmethod def setUpClass(cls): iris = sklearn.datasets.load_iris() iris_X, iris_y = iris.data, iris.target ( iris_train_X, iris_test_X, iris_train_y, iris_test_y, ) = sklearn.model_selection.train_test_split(iris_X, iris_y) cls._iris = { "train_X": iris_train_X, "train_y": iris_train_y, "test_X": iris_test_X, "test_y": iris_test_y, } def doTest(self, trainable, train_X, train_y, test_X, test_y): trained = trainable.fit(train_X, train_y) _ = trained.transform(test_X) with self.assertWarns(DeprecationWarning): trainable.transform(train_X) trainable.to_json() trainable_pipeline = trainable >> float32_transform() >> LR() trained_pipeline = trainable_pipeline.fit(train_X, train_y) trained_pipeline.predict(test_X) hyperopt = Hyperopt(estimator=trainable_pipeline, max_evals=1, verbose=True) trained_hyperopt = hyperopt.fit(train_X, train_y) trained_hyperopt.predict(test_X) def test_NumpyColumnSelector(self): trainable = lale.lib.autoai_libs.NumpyColumnSelector() self.doTest(trainable, **self._iris) def test_NumpyColumnSelector_pandas(self): iris_X, iris_y = sklearn.datasets.load_iris(return_X_y=True, as_frame=True) keys = ["train_X", "test_X", "train_y", "test_y"] splits = sklearn.model_selection.train_test_split(iris_X, iris_y) iris = dict(zip(keys, splits)) self.assertIsInstance(iris["train_X"], pd.DataFrame) trainable = lale.lib.autoai_libs.NumpyColumnSelector(columns=[0, 2, 3]) self.doTest(trainable, **iris) def test_CompressStrings(self): n_columns = self._iris["train_X"].shape[1] trainable = lale.lib.autoai_libs.CompressStrings( dtypes_list=["int_num" for i in range(n_columns)], misslist_list=[[] for i in range(n_columns)], ) self.doTest(trainable, **self._iris) def test_NumpyReplaceMissingValues(self): trainable = lale.lib.autoai_libs.NumpyReplaceMissingValues() self.doTest(trainable, **self._iris) def test_NumpyReplaceUnknownValues(self): trainable = lale.lib.autoai_libs.NumpyReplaceUnknownValues(filling_values=42.0) self.doTest(trainable, **self._iris) def test_boolean2float(self): trainable = lale.lib.autoai_libs.boolean2float() self.doTest(trainable, **self._iris) def test_CatImputer(self): trainable = lale.lib.autoai_libs.CatImputer() self.doTest(trainable, **self._iris) def test_CatEncoder(self): trainable = lale.lib.autoai_libs.CatEncoder( encoding="ordinal", categories="auto", dtype="float64", handle_unknown="ignore", ) self.doTest(trainable, **self._iris) def test_float32_transform(self): trainable = lale.lib.autoai_libs.float32_transform() self.doTest(trainable, **self._iris) def test_FloatStr2Float(self): n_columns = self._iris["train_X"].shape[1] trainable = lale.lib.autoai_libs.FloatStr2Float( dtypes_list=["int_num" for i in range(n_columns)] ) self.doTest(trainable, **self._iris) def test_OptStandardScaler(self): trainable = lale.lib.autoai_libs.OptStandardScaler() self.doTest(trainable, **self._iris) def test_NumImputer(self): trainable = lale.lib.autoai_libs.NumImputer() self.doTest(trainable, **self._iris) def test_NumpyPermuteArray(self): trainable = lale.lib.autoai_libs.NumpyPermuteArray( axis=0, permutation_indices=[2, 0, 1, 3] ) self.doTest(trainable, **self._iris) def test_TNoOp(self): from autoai_libs.utils.fc_methods import is_not_categorical trainable = lale.lib.autoai_libs.TNoOp( fun=np.rint, name="do nothing", datatypes=["numeric"], feat_constraints=[is_not_categorical], ) self.doTest(trainable, **self._iris) def test_TA1(self): from autoai_libs.utils.fc_methods import is_not_categorical float32 = np.dtype("float32") trainable = lale.lib.autoai_libs.TA1( fun=np.rint, name="round", datatypes=["numeric"], feat_constraints=[is_not_categorical], col_names=["a", "b", "c", "d"], col_dtypes=[float32, float32, float32, float32], ) self.doTest(trainable, **self._iris) def test_TA2(self): from autoai_libs.utils.fc_methods import is_not_categorical float32 = np.dtype("float32") trainable = lale.lib.autoai_libs.TA2( fun=np.add, name="sum", datatypes1=["numeric"], feat_constraints1=[is_not_categorical], datatypes2=["numeric"], feat_constraints2=[is_not_categorical], col_names=["a", "b", "c", "d"], col_dtypes=[float32, float32, float32, float32], ) self.doTest(trainable, **self._iris) def test_TB1(self): from autoai_libs.utils.fc_methods import is_not_categorical from sklearn.preprocessing import StandardScaler float32 = np.dtype("float32") trainable = lale.lib.autoai_libs.TB1( tans_class=StandardScaler, name="stdscaler", datatypes=["numeric"], feat_constraints=[is_not_categorical], col_names=["a", "b", "c", "d"], col_dtypes=[float32, float32, float32, float32], ) self.doTest(trainable, **self._iris) def test_TB2(self): pass # TODO: not sure how to instantiate, what to pass for tans_class def test_TAM(self): from autoai_libs.cognito.transforms.transform_extras import ( IsolationForestAnomaly, ) float32 = np.dtype("float32") trainable = lale.lib.autoai_libs.TAM( tans_class=IsolationForestAnomaly, name="isoforestanomaly", col_names=["a", "b", "c", "d"], col_dtypes=[float32, float32, float32, float32], ) self.doTest(trainable, **self._iris) def test_TGen(self): from autoai_libs.cognito.transforms.transform_extras import NXOR from autoai_libs.utils.fc_methods import is_not_categorical float32 = np.dtype("float32") trainable = lale.lib.autoai_libs.TGen( fun=NXOR, name="nxor", arg_count=2, datatypes_list=[["numeric"], ["numeric"]], feat_constraints_list=[[is_not_categorical], [is_not_categorical]], col_names=["a", "b", "c", "d"], col_dtypes=[float32, float32, float32, float32], ) self.doTest(trainable, **self._iris) def test_FS1(self): trainable = lale.lib.autoai_libs.FS1( cols_ids_must_keep=[1], additional_col_count_to_keep=3, ptype="classification", ) self.doTest(trainable, **self._iris) def test_FS2(self): from sklearn.ensemble import ExtraTreesClassifier trainable = lale.lib.autoai_libs.FS2( cols_ids_must_keep=[1], additional_col_count_to_keep=3, ptype="classification", eval_algo=ExtraTreesClassifier, ) self.doTest(trainable, **self._iris) def test_ColumnSelector(self): trainable = lale.lib.autoai_libs.ColumnSelector() self.doTest(trainable, **self._iris) def test_ColumnSelector_pandas(self): iris_X, iris_y = sklearn.datasets.load_iris(return_X_y=True, as_frame=True) keys = ["train_X", "test_X", "train_y", "test_y"] splits = sklearn.model_selection.train_test_split(iris_X, iris_y) iris = dict(zip(keys, splits)) self.assertIsInstance(iris["train_X"], pd.DataFrame) trainable = lale.lib.autoai_libs.ColumnSelector(columns_indices_list=[0, 2, 3]) self.doTest(trainable, **iris) class TestAutoaiLibsText(unittest.TestCase): def setUp(self): from sklearn.datasets import fetch_20newsgroups cats = ["alt.atheism", "sci.space"] newsgroups_train = fetch_20newsgroups(subset="train", categories=cats) self.train_X, self.train_y = ( np.array(newsgroups_train.data), newsgroups_train.target, ) self.train_X = np.reshape(self.train_X, (self.train_X.shape[0], 1)) newsgroups_test = fetch_20newsgroups(subset="test", categories=cats) self.test_X, self.test_y = ( np.array(newsgroups_test.data), newsgroups_test.target, ) self.test_X = np.reshape(self.test_X, (self.test_X.shape[0], 1)) def doTest(self, trainable, train_X, train_y, test_X, test_y): trained = trainable.fit(train_X, train_y) _ = trained.transform(test_X) with self.assertWarns(DeprecationWarning): trainable.transform(train_X) trainable.to_json() trainable_pipeline = trainable >> float32_transform() >> XGBClassifier() trained_pipeline = trainable_pipeline.fit(train_X, train_y) trained_pipeline.predict(test_X) hyperopt = Hyperopt(estimator=trainable_pipeline, max_evals=1, verbose=True) trained_hyperopt = hyperopt.fit(train_X, train_y) trained_hyperopt.predict(test_X) @unittest.skip( "skipping for now because this does not work with the latest xgboost." ) def test_TextTransformer(self): trainable = lale.lib.autoai_libs.TextTransformer( drop_columns=True, columns_to_be_deleted=[0, 1], text_processing_options={"word2vec": {"output_dim": 5}}, ) self.doTest(trainable, self.train_X, self.train_y, self.test_X, self.test_y) @unittest.skip( "skipping for now because this does not work with the latest xgboost." ) def test_Word2VecTransformer(self): trainable = lale.lib.autoai_libs.Word2VecTransformer( drop_columns=True, output_dim=5 ) self.doTest(trainable, self.train_X, self.train_y, self.test_X, self.test_y) # class TestDateTransformer(unittest.TestCase): # @classmethod # def setUpClass(cls): # data = fetch_household_power_consumption() # data = data.iloc[:5000, [0, 2, 3, 4, 5]] # cls.X_train = data.iloc[-1000:] # cls.X_test = data.iloc[:-1000] # def test_01_all_mini_options_with_headers(self): # transformer = lale.lib.autoai_libs.DateTransformer( # options=["all"], column_headers_list=self.X_train.columns.values.tolist() # ) # fitted_transformer = transformer.fit(self.X_train.values) # X_test_transformed = fitted_transformer.transform(self.X_test.values) # X_train_transformed = fitted_transformer.transform(self.X_train.values) # header_list = fitted_transformer.impl.new_column_headers_list # print(f"New columns: {header_list}, new shape: {X_train_transformed.shape}") # self.assertEqual( # X_train_transformed.shape[1], # X_test_transformed.shape[1], # f"Number of columns after transform is different.:{X_train_transformed.shape[1]}, {X_test_transformed.shape[1]}", # ) # def test_02_all_options_without_headers(self): # transformer = lale.lib.autoai_libs.DateTransformer(options=["all"]) # fitted_transformer = transformer.fit(self.X_train.values) # X_train = fitted_transformer.transform(self.X_train.values) # X_test = transformer.transform(self.X_test.values) # header_list = fitted_transformer.impl.new_column_headers_list # print(f"New columns: {header_list}") # self.assertEqual( # X_train.shape[1], X_test.shape[1], msg="Shape after transform is different." # ) # def test_03_specific_options_and_delete_source_columns(self): # transformer = lale.lib.autoai_libs.DateTransformer( # options=["FloatTimestamp", "DayOfWeek", "Hour", "Minute"], # delete_source_columns=True, # column_headers_list=self.X_train.columns.values.tolist(), # ) # fitted_transformer = transformer.fit(self.X_train.values) # X_train = fitted_transformer.transform(self.X_train.values) # X_test = transformer.transform(self.X_test.values) # header_list = fitted_transformer.impl.new_column_headers_list # print(f"New columns: {header_list}") # self.assertEqual( # X_train.shape[1], X_test.shape[1], msg="Shape after transform is different." # ) # def test_04_option_Datetime_and_delete_source_columns(self): # transformer = lale.lib.autoai_libs.DateTransformer( # options=["Datetime"], # delete_source_columns=True, # column_headers_list=self.X_train.columns.values.tolist(), # ) # fitted_transformer = transformer.fit(self.X_train.values) # X_train = fitted_transformer.transform(self.X_train.values) # X_test = transformer.transform(self.X_test.values) # header_list = fitted_transformer.impl.new_column_headers_list # print(f"New columns: {header_list}") # self.assertEqual( # X_train.shape[1], X_test.shape[1], msg="Shape after transform is different." # )
14,626
38.005333
127
py
lale
lale-master/test/test_core_classifiers.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import unittest import warnings from test import EnableSchemaValidation import jsonschema from sklearn.datasets import load_iris import lale.lib.lale import lale.type_checking from lale.lib.lale import NoOp from lale.lib.sklearn import ( PCA, SVC, IsolationForest, KMeans, KNeighborsClassifier, LogisticRegression, MLPClassifier, Nystroem, PassiveAggressiveClassifier, RidgeClassifier, SGDClassifier, SimpleImputer, VotingClassifier, ) from lale.search.lale_grid_search_cv import get_grid_search_parameter_grids from lale.settings import set_disable_data_schema_validation set_disable_data_schema_validation(False) class TestClassification(unittest.TestCase): def setUp(self): from sklearn.model_selection import train_test_split data = load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def create_function_test_classifier(clf_name): def test_classifier(self): X_train, y_train = self.X_train, self.y_train import importlib module_name = ".".join(clf_name.split(".")[0:-1]) class_name = clf_name.split(".")[-1] module = importlib.import_module(module_name) class_ = getattr(module, class_name) clf = class_() # test_schemas_are_schemas lale.type_checking.validate_is_schema(clf.input_schema_fit()) lale.type_checking.validate_is_schema(clf.input_schema_predict()) lale.type_checking.validate_is_schema(clf.output_schema_predict()) lale.type_checking.validate_is_schema(clf.hyperparam_schema()) # test_init_fit_predict trained = clf.fit(self.X_train, self.y_train) _ = trained.predict(self.X_test) # test score if not isinstance( clf, IsolationForest # type: ignore ): # IsolationForest does not define score _ = trained.score(self.X_test, self.y_test) from lale.lib.sklearn.gradient_boosting_classifier import ( GradientBoostingClassifier, ) if isinstance(clf, GradientBoostingClassifier): # type: ignore # because exponential loss does not work with iris dataset as it is not binary classification from lale import schemas clf = clf.customize_schema( loss=schemas.Enum(default="deviance", values=["deviance"]) ) # test_with_hyperopt from lale.lib.lale import Hyperopt hyperopt = Hyperopt(estimator=clf, max_evals=1, verbose=True) trained = hyperopt.fit(self.X_train, self.y_train) _ = trained.predict(self.X_test) # test_cross_validation from lale.helpers import cross_val_score cv_results = cross_val_score(clf, X_train, y_train, cv=2) self.assertEqual(len(cv_results), 2) # test_with_gridsearchcv_auto_wrapped from sklearn.metrics import accuracy_score, make_scorer with warnings.catch_warnings(): warnings.simplefilter("ignore") grid_search = lale.lib.lale.GridSearchCV( estimator=clf, lale_num_samples=1, lale_num_grids=1, cv=2, scoring=make_scorer(accuracy_score), ) grid_search.fit(X_train, y_train) # test_predict_on_trainable trained = clf.fit(X_train, y_train) clf.predict(X_train) # test_to_json clf.to_json() # test_in_a_pipeline pipeline = NoOp() >> clf trained = pipeline.fit(self.X_train, self.y_train) _ = trained.predict(self.X_test) test_classifier.__name__ = f"test_{clf_to_test.rsplit('.', maxsplit=1)[-1]}" return test_classifier classifiers = [ "lale.lib.sklearn.DummyClassifier", "lale.lib.sklearn.RandomForestClassifier", "lale.lib.sklearn.DecisionTreeClassifier", "lale.lib.sklearn.ExtraTreesClassifier", "lale.lib.sklearn.GradientBoostingClassifier", "lale.lib.sklearn.GaussianNB", "lale.lib.sklearn.QuadraticDiscriminantAnalysis", "lale.lib.lightgbm.LGBMClassifier", "lale.lib.xgboost.XGBClassifier", "lale.lib.sklearn.KNeighborsClassifier", "lale.lib.sklearn.LinearSVC", "lale.lib.sklearn.LogisticRegression", "lale.lib.sklearn.MLPClassifier", "lale.lib.sklearn.SVC", "lale.lib.sklearn.Perceptron", "lale.lib.sklearn.PassiveAggressiveClassifier", "lale.lib.sklearn.MultinomialNB", "lale.lib.sklearn.AdaBoostClassifier", "lale.lib.sklearn.SGDClassifier", "lale.lib.sklearn.RidgeClassifier", "lale.lib.sklearn.IsolationForest", "lale.lib.sklearn.KMeans", ] for clf_to_test in classifiers: setattr( TestClassification, f"test_{clf_to_test.rsplit('.', maxsplit=1)[-1]}", create_function_test_classifier(clf_to_test), ) class TestMLPClassifier(unittest.TestCase): def test_with_multioutput_targets(self): import numpy as np from sklearn.datasets import make_classification from sklearn.utils import shuffle X, y1 = make_classification( n_samples=10, n_features=100, n_informative=30, n_classes=3, random_state=1 ) y2 = shuffle(y1, random_state=1) y3 = shuffle(y1, random_state=2) Y = np.vstack((y1, y2, y3)).T trainable = KNeighborsClassifier() trained = trainable.fit(X, Y) _ = trained.predict(X) def test_predict_proba(self): trainable = MLPClassifier() iris = load_iris() trained = trainable.fit(iris.data, iris.target) # with self.assertWarns(DeprecationWarning): _ = trainable.predict_proba(iris.data) _ = trained.predict_proba(iris.data) class TestVotingClassifier(unittest.TestCase): def setUp(self): from sklearn.model_selection import train_test_split data = load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) warnings.filterwarnings("ignore") def test_with_lale_classifiers(self): clf = VotingClassifier( estimators=[("knn", KNeighborsClassifier()), ("lr", LogisticRegression())] ) trained = clf.fit(self.X_train, self.y_train) trained.predict(self.X_test) def test_with_lale_pipeline(self): clf = VotingClassifier( estimators=[ ("knn", KNeighborsClassifier()), ("pca_lr", PCA() >> LogisticRegression()), ] ) trained = clf.fit(self.X_train, self.y_train) trained.predict(self.X_test) def test_with_hyperopt(self): from lale.lib.lale import Hyperopt clf = VotingClassifier( estimators=[("knn", KNeighborsClassifier()), ("lr", LogisticRegression())] ) _ = clf.auto_configure(self.X_train, self.y_train, Hyperopt, max_evals=1) def test_with_gridsearch(self): from sklearn.metrics import accuracy_score, make_scorer from lale.lib.lale import GridSearchCV clf = VotingClassifier( estimators=[("knn", KNeighborsClassifier()), ("rc", RidgeClassifier())], voting="hard", ) _ = clf.auto_configure( self.X_train, self.y_train, GridSearchCV, lale_num_samples=1, lale_num_grids=1, cv=2, scoring=make_scorer(accuracy_score), ) @unittest.skip("TODO: get this working with sklearn 0.23") def test_with_observed_gridsearch(self): from sklearn.metrics import accuracy_score, make_scorer from lale.lib.lale import GridSearchCV from lale.lib.lale.observing import LoggingObserver clf = VotingClassifier( estimators=[("knn", KNeighborsClassifier()), ("rc", RidgeClassifier())], voting="hard", ) _ = clf.auto_configure( self.X_train, self.y_train, GridSearchCV, lale_num_samples=1, lale_num_grids=1, cv=2, scoring=make_scorer(accuracy_score), observer=LoggingObserver, ) class TestBaggingClassifier(unittest.TestCase): def setUp(self): from sklearn.model_selection import train_test_split data = load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) warnings.filterwarnings("ignore") def test_with_lale_classifiers(self): from lale.lib.sklearn import BaggingClassifier clf = BaggingClassifier(base_estimator=LogisticRegression()) trained = clf.fit(self.X_train, self.y_train) trained.predict(self.X_test) def test_with_lale_pipeline(self): from lale.lib.sklearn import BaggingClassifier clf = BaggingClassifier(base_estimator=PCA() >> LogisticRegression()) trained = clf.fit(self.X_train, self.y_train) trained.predict(self.X_test) def test_with_hyperopt(self): from lale.lib.lale import Hyperopt from lale.lib.sklearn import BaggingClassifier clf = BaggingClassifier(base_estimator=LogisticRegression()) trained = clf.auto_configure(self.X_train, self.y_train, Hyperopt, max_evals=1) print(trained.to_json()) def test_pipeline_with_hyperopt(self): from lale.lib.lale import Hyperopt from lale.lib.sklearn import BaggingClassifier clf = BaggingClassifier(base_estimator=PCA() >> LogisticRegression()) _ = clf.auto_configure(self.X_train, self.y_train, Hyperopt, max_evals=1) def test_pipeline_choice_with_hyperopt(self): from lale.lib.lale import Hyperopt from lale.lib.sklearn import BaggingClassifier clf = BaggingClassifier( base_estimator=PCA() >> (LogisticRegression() | KNeighborsClassifier()) ) _ = clf.auto_configure(self.X_train, self.y_train, Hyperopt, max_evals=1) def test_predict_log_proba(self): from lale.lib.sklearn import BaggingClassifier clf = BaggingClassifier(base_estimator=PCA() >> LogisticRegression()) trained = clf.fit(self.X_train, self.y_train) trained.predict_log_proba(self.X_test) def test_predict_log_proba_trained_trainable(self): from lale.lib.sklearn import BaggingClassifier clf = BaggingClassifier() clf.fit(self.X_train, self.y_train) with self.assertWarns(DeprecationWarning): clf.predict_log_proba(self.X_test) def test_predict_log_proba_trainable(self): from lale.lib.sklearn import BaggingClassifier clf = BaggingClassifier(base_estimator=PCA() >> LogisticRegression()) with self.assertRaises(ValueError): clf.predict_log_proba(self.X_test) class TestStackingClassifier(unittest.TestCase): def setUp(self): from sklearn.model_selection import train_test_split data = load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_with_lale_classifiers(self): from lale.lib.sklearn import StackingClassifier clf = StackingClassifier(estimators=[("base", LogisticRegression())]) trained = clf.fit(self.X_train, self.y_train) trained.predict(self.X_test) def test_with_lale_pipeline(self): from lale.lib.sklearn import StackingClassifier clf = StackingClassifier(estimators=[("base", PCA() >> LogisticRegression())]) trained = clf.fit(self.X_train, self.y_train) trained.predict(self.X_test) def test_with_hyperopt(self): from lale.lib.lale import Hyperopt from lale.lib.sklearn import StackingClassifier clf = StackingClassifier( estimators=[("base", LogisticRegression())], final_estimator=LogisticRegression(), ) trained = clf.auto_configure(self.X_train, self.y_train, Hyperopt, max_evals=1) print(trained.to_json()) def test_pipeline_with_hyperopt(self): from lale.lib.lale import Hyperopt from lale.lib.sklearn import StackingClassifier clf = StackingClassifier( estimators=[("base", PCA() >> LogisticRegression())], final_estimator=LogisticRegression(), ) _ = clf.auto_configure(self.X_train, self.y_train, Hyperopt, max_evals=1) def test_pipeline_choice_with_hyperopt(self): from lale.lib.lale import Hyperopt from lale.lib.sklearn import StackingClassifier clf = StackingClassifier( estimators=[ ("base", PCA() >> (LogisticRegression() | KNeighborsClassifier())) ] ) _ = clf.auto_configure(self.X_train, self.y_train, Hyperopt, max_evals=1) class TestSpuriousSideConstraintsClassification(unittest.TestCase): # This was prompted buy a bug, keeping it as it may help with support for other sklearn versions def setUp(self): from sklearn.model_selection import train_test_split data = load_iris() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) def test_sgd_classifier(self): reg = SGDClassifier(loss="squared_error", epsilon=0.2) reg.fit(self.X_train, self.y_train) def test_sgd_classifier_1(self): reg = SGDClassifier(learning_rate="optimal", eta0=0.2) reg.fit(self.X_train, self.y_train) def test_sgd_classifier_2(self): reg = SGDClassifier(early_stopping=False, validation_fraction=0.2) reg.fit(self.X_train, self.y_train) def test_sgd_classifier_3(self): reg = SGDClassifier(l1_ratio=0.2, penalty="l1") reg.fit(self.X_train, self.y_train) def test_mlp_classifier(self): reg = MLPClassifier(early_stopping=False, validation_fraction=0.2) reg.fit(self.X_train, self.y_train) def test_mlp_classifier_1(self): reg = MLPClassifier(beta_1=0.8, solver="sgd") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_2b(self): reg = MLPClassifier(beta_2=0.8, solver="sgd") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_2e(self): reg = MLPClassifier(epsilon=0.8, solver="sgd") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_3(self): reg = MLPClassifier(n_iter_no_change=100, solver="lbfgs") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_4(self): reg = MLPClassifier(early_stopping=True, solver="lbfgs") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_5(self): reg = MLPClassifier(nesterovs_momentum=False, solver="lbfgs") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_6(self): reg = MLPClassifier(momentum=0.8, solver="lbfgs") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_7(self): reg = MLPClassifier(shuffle=False, solver="lbfgs") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_8(self): reg = MLPClassifier(learning_rate="invscaling", solver="lbfgs") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_9(self): reg = MLPClassifier(learning_rate_init=0.002, solver="lbfgs") reg.fit(self.X_train, self.y_train) def test_mlp_classifier_10(self): reg = MLPClassifier(learning_rate="invscaling", power_t=0.4, solver="lbfgs") reg.fit(self.X_train, self.y_train) def test_passive_aggressive_classifier(self): reg = PassiveAggressiveClassifier(validation_fraction=0.4, early_stopping=False) reg.fit(self.X_train, self.y_train) def test_svc(self): reg = SVC(kernel="linear", gamma=1) reg.fit(self.X_train, self.y_train) def test_simple_imputer(self): reg = SimpleImputer(strategy="mean", fill_value=10) reg.fit(self.X_train, self.y_train) def test_nystroem(self): reg = Nystroem(kernel="cosine", gamma=0.1) reg.fit(self.X_train, self.y_train) def test_nystroem_1(self): reg = Nystroem(kernel="cosine", coef0=0.1) reg.fit(self.X_train, self.y_train) def test_nystroem_2(self): reg = Nystroem(kernel="cosine", degree=2) reg.fit(self.X_train, self.y_train) def test_ridge_classifier(self): reg = RidgeClassifier(fit_intercept=False) reg.fit(self.X_train, self.y_train) def test_ridge_classifier_1(self): reg = RidgeClassifier(solver="svd", max_iter=10) reg.fit(self.X_train, self.y_train) class TestKNeighborsClassifier(unittest.TestCase): def test_with_multioutput_targets(self): import numpy as np from sklearn.datasets import make_classification from sklearn.utils import shuffle X, y1 = make_classification( n_samples=10, n_features=100, n_informative=30, n_classes=3, random_state=1 ) y2 = shuffle(y1, random_state=1) y3 = shuffle(y1, random_state=2) Y = np.vstack((y1, y2, y3)).T trainable = KNeighborsClassifier() trained = trainable.fit(X, Y) _ = trained.predict(X) def test_predict_proba(self): trainable = KNeighborsClassifier() iris = load_iris() trained = trainable.fit(iris.data, iris.target) # with self.assertWarns(DeprecationWarning): _ = trainable.predict_proba(iris.data) _ = trained.predict_proba(iris.data) class TestLogisticRegression(unittest.TestCase): def test_hyperparam_keyword_enum(self): _ = LogisticRegression( LogisticRegression.enum.penalty.l1, C=0.1, solver=LogisticRegression.enum.solver.saga, ) def test_hyperparam_exclusive_min(self): with EnableSchemaValidation(): with self.assertRaises(jsonschema.ValidationError): _ = LogisticRegression(LogisticRegression.enum.penalty.l1, C=0.0) def test_hyperparam_penalty_solver_dependence(self): with EnableSchemaValidation(): with self.assertRaises(jsonschema.ValidationError): _ = LogisticRegression( LogisticRegression.enum.penalty.l1, LogisticRegression.enum.solver.newton_cg, ) def test_hyperparam_dual_penalty_solver_dependence(self): with EnableSchemaValidation(): with self.assertRaises(jsonschema.ValidationError): _ = LogisticRegression( LogisticRegression.enum.penalty.l2, LogisticRegression.enum.solver.sag, dual=True, ) def test_sample_weight(self): import numpy as np trainable_lr = LogisticRegression(n_jobs=1) iris = load_iris() trained_lr = trainable_lr.fit( iris.data, iris.target, sample_weight=np.arange(len(iris.target)) ) _ = trained_lr.predict(iris.data) def test_predict_proba(self): import numpy as np trainable_lr = LogisticRegression(n_jobs=1) iris = load_iris() trained_lr = trainable_lr.fit( iris.data, iris.target, sample_weight=np.arange(len(iris.target)) ) # with self.assertWarns(DeprecationWarning): _ = trainable_lr.predict_proba(iris.data) _ = trained_lr.predict_proba(iris.data) def test_decision_function(self): import numpy as np trainable_lr = LogisticRegression(n_jobs=1) iris = load_iris() trained_lr = trainable_lr.fit( iris.data, iris.target, sample_weight=np.arange(len(iris.target)) ) _ = trained_lr.decision_function(iris.data) def test_with_sklearn_gridsearchcv(self): from sklearn.metrics import accuracy_score, make_scorer from sklearn.model_selection import GridSearchCV lr = LogisticRegression() parameters = {"solver": ("liblinear", "lbfgs"), "penalty": ["l2"]} with warnings.catch_warnings(): warnings.simplefilter("ignore") clf = GridSearchCV( lr, parameters, cv=5, scoring=make_scorer(accuracy_score) ) iris = load_iris() clf.fit(iris.data, iris.target) def test_with_randomizedsearchcv(self): import numpy as np from scipy.stats.distributions import uniform from sklearn.metrics import accuracy_score, make_scorer from sklearn.model_selection import RandomizedSearchCV lr = LogisticRegression() ranges, _cat_idx = lr.get_param_ranges() # specify parameters and distributions to sample from # the loguniform distribution needs to be taken care of properly param_dist = {"solver": ranges["solver"], "C": uniform(0.03125, np.log(32768))} # run randomized search n_iter_search = 5 with warnings.catch_warnings(): warnings.simplefilter("ignore") random_search = RandomizedSearchCV( lr, param_distributions=param_dist, n_iter=n_iter_search, cv=5, scoring=make_scorer(accuracy_score), ) iris = load_iris() random_search.fit(iris.data, iris.target) def test_grid_search_on_trained(self): from sklearn.metrics import accuracy_score, make_scorer from sklearn.model_selection import GridSearchCV iris = load_iris() X, y = iris.data, iris.target lr = LogisticRegression() trained = lr.fit(X, y) parameters = {"solver": ("liblinear", "lbfgs"), "penalty": ["l2"]} _ = GridSearchCV(trained, parameters, cv=5, scoring=make_scorer(accuracy_score)) def test_grid_search_on_trained_auto(self): from sklearn.metrics import accuracy_score, make_scorer from sklearn.model_selection import GridSearchCV iris = load_iris() X, y = iris.data, iris.target lr = LogisticRegression() trained = lr.fit(X, y) parameters = get_grid_search_parameter_grids(lr, num_samples=2) _ = GridSearchCV(trained, parameters, cv=5, scoring=make_scorer(accuracy_score)) def test_doc(self): from test.mock_custom_operators import MyLR import sklearn.datasets import sklearn.utils iris = load_iris() X_all, y_all = sklearn.utils.shuffle(iris.data, iris.target, random_state=42) X_train, y_train = X_all[10:], y_all[10:] X_test, y_test = X_all[:10], y_all[:10] print(f"expected {y_test}") warnings.filterwarnings("ignore", category=FutureWarning) trainable = MyLR(solver="lbfgs", C=0.1) trained = trainable.fit(X_train, y_train) predictions = trained.predict(X_test) print(f"actual {predictions}") class TestIsolationForest(unittest.TestCase): def setUp(self): from sklearn.model_selection import train_test_split from lale.datasets.util import load_boston data = load_boston() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) warnings.filterwarnings("ignore") def test_with_no_y(self): clf = IsolationForest() trained = clf.fit(self.X_train) trained.predict(self.X_test) def test_with_hyperopt(self): def my_scorer(estimator, X, y=None): return 1 from lale.lib.lale import Hyperopt hyperopt = Hyperopt( estimator=IsolationForest(max_features=1.0, max_samples=1.0), max_evals=5, verbose=True, scoring=my_scorer, ) trained = hyperopt.fit(self.X_train) _ = trained.predict(self.X_test) def test_decision_function_1(self): def my_scorer(estimator, X, y=None): return 1 from lale.lib.lale import Hyperopt hyperopt = Hyperopt( estimator=IsolationForest(max_features=1.0, max_samples=1.0), max_evals=5, verbose=True, scoring=my_scorer, ) trained = hyperopt.fit(self.X_train) pipeline = trained.get_pipeline() assert pipeline is not None _ = pipeline.decision_function(self.X_test) def test_decision_function_2(self): def my_scorer(estimator, X, y=None): return 1 from lale.lib.lale import Hyperopt from lale.lib.sklearn import MinMaxScaler hyperopt = Hyperopt( estimator=MinMaxScaler() >> IsolationForest(max_features=1.0, max_samples=1.0), max_evals=5, verbose=True, scoring=my_scorer, ) trained = hyperopt.fit(self.X_train) pipeline = trained.get_pipeline() assert pipeline is not None _ = pipeline.decision_function(self.X_test) def test_score_samples(self): clf = IsolationForest() trained = clf.fit(self.X_train) trained.score_samples(self.X_test) def test_score_samples_trainable(self): clf = IsolationForest() with self.assertRaises(ValueError): clf.score_samples(self.X_test) def test_score_samples_trained_trainable(self): clf = IsolationForest() clf.fit(self.X_train) with self.assertWarns(DeprecationWarning): clf.score_samples(self.X_test) class TestKMeans(unittest.TestCase): def setUp(self): from sklearn.model_selection import train_test_split from lale.datasets.util import load_boston data = load_boston() X, y = data.data, data.target self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(X, y) warnings.filterwarnings("ignore") def test_with_no_y(self): clf = KMeans() trained = clf.fit(self.X_train) trained.predict(self.X_test) def test_with_hyperopt(self): from lale.lib.lale import Hyperopt def my_scorer(estimator, X, y=None): return 1 hyperopt = Hyperopt( estimator=KMeans(n_clusters=3), max_evals=5, verbose=True, scoring=my_scorer ) trained = hyperopt.fit(self.X_train) _ = trained.predict(self.X_test)
27,160
33.6
105
py
lale
lale-master/test/test_relational_sklearn.py
# Copyright 2021-2023 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import itertools import math import numbers import os.path import re import tempfile import unittest import urllib.request from typing import Any, Dict, List, Tuple, cast import jsonschema import numpy as np import pandas as pd import sklearn import sklearn.datasets from category_encoders import HashingEncoder as SkHashingEncoder from sklearn.feature_selection import SelectKBest as SkSelectKBest from sklearn.impute import SimpleImputer as SkSimpleImputer from sklearn.metrics import accuracy_score as sk_accuracy_score from sklearn.metrics import balanced_accuracy_score as sk_balanced_accuracy_score from sklearn.metrics import f1_score as sk_f1_score from sklearn.metrics import make_scorer from sklearn.metrics import r2_score as sk_r2_score from sklearn.model_selection import KFold from sklearn.model_selection import cross_val_score as sk_cross_val_score from sklearn.model_selection import cross_validate as sk_cross_validate from sklearn.pipeline import make_pipeline as sk_make_pipeline from sklearn.preprocessing import MinMaxScaler as SkMinMaxScaler from sklearn.preprocessing import OneHotEncoder as SkOneHotEncoder from sklearn.preprocessing import OrdinalEncoder as SkOrdinalEncoder from sklearn.preprocessing import StandardScaler as SkStandardScaler from sklearn.preprocessing import scale as sk_scale import lale.datasets import lale.datasets.openml import lale.datasets.openml.openml_datasets import lale.lib.aif360 import lale.type_checking from lale.datasets import pandas2spark from lale.datasets.data_schemas import ( SparkDataFrameWithIndex, add_table_name, forward_metadata, get_index_name, ) from lale.datasets.multitable.fetch_datasets import fetch_go_sales_dataset from lale.expressions import it from lale.helpers import _ensure_pandas, create_data_loader, datatype_param_type from lale.lib.category_encoders import TargetEncoder as SkTargetEncoder from lale.lib.lightgbm import LGBMClassifier, LGBMRegressor from lale.lib.rasl import BatchedBaggingClassifier, ConcatFeatures, Convert from lale.lib.rasl import HashingEncoder as RaslHashingEncoder from lale.lib.rasl import Map from lale.lib.rasl import MinMaxScaler as RaslMinMaxScaler from lale.lib.rasl import OneHotEncoder as RaslOneHotEncoder from lale.lib.rasl import OrdinalEncoder as RaslOrdinalEncoder from lale.lib.rasl import PrioBatch, PrioStep, Project, Scan from lale.lib.rasl import SelectKBest as RaslSelectKBest from lale.lib.rasl import SimpleImputer as RaslSimpleImputer from lale.lib.rasl import StandardScaler as RaslStandardScaler from lale.lib.rasl import TargetEncoder as RaslTargetEncoder from lale.lib.rasl import accuracy_score as rasl_accuracy_score from lale.lib.rasl import balanced_accuracy_score as rasl_balanced_accuracy_score from lale.lib.rasl import categorical from lale.lib.rasl import cross_val_score as rasl_cross_val_score from lale.lib.rasl import cross_validate as rasl_cross_validate from lale.lib.rasl import csv_data_loader from lale.lib.rasl import f1_score as rasl_f1_score from lale.lib.rasl import fit_with_batches from lale.lib.rasl import get_scorer as rasl_get_scorer from lale.lib.rasl import mockup_data_loader, openml_data_loader from lale.lib.rasl import r2_score as rasl_r2_score from lale.lib.rasl.standard_scaler import scale as rasl_scale from lale.lib.sklearn import ( DecisionTreeClassifier, LinearRegression, LogisticRegression, RandomForestClassifier, SGDClassifier, ) from lale.lib.xgboost import XGBClassifier, XGBRegressor from lale.operators import TrainedPipeline assert sklearn.__version__ >= "1.0", sklearn.__version__ class TestDatasets(unittest.TestCase): def test_openml_creditg_arff(self): batches = openml_data_loader("credit-g", 340) n_rows_found = 0 n_batches_found = 0 for bX, by in batches: n_batches_found += 1 n_rows_batch, n_columns_batch = bX.shape n_rows_found += n_rows_batch self.assertEqual(n_rows_batch, len(by)) self.assertEqual(n_columns_batch, 20) self.assertEqual(n_batches_found, 3) self.assertEqual(n_rows_found, 1000) def test_autoai_creditg_csv(self): with tempfile.TemporaryDirectory() as tmpdir_name: url = "https://raw.githubusercontent.com/pmservice/wml-sample-models/master/autoai/credit-risk-prediction/data/german_credit_data_biased_training.csv" file_name = os.path.join(tmpdir_name, "credit-g.csv") # this request is to a hardcoded https url, so does not risk leaking local data urllib.request.urlretrieve(url, file_name) # nosec assert os.path.exists(file_name) n_rows = 5000 n_batches = 3 rows_per_batch = (n_rows + n_batches - 1) // n_batches batches = csv_data_loader(file_name, "Risk", rows_per_batch) n_rows_found = 0 n_batches_found = 0 for bX, by in batches: n_batches_found += 1 n_rows_batch, n_columns_batch = bX.shape n_rows_found += n_rows_batch self.assertEqual(n_rows_batch, len(by)) self.assertEqual(n_columns_batch, 20) self.assertEqual(n_batches_found, n_batches) self.assertEqual(n_rows_found, n_rows) def _check_trained_min_max_scaler(test, op1, op2, msg): test.assertEqual(list(op1.data_min_), list(op2.data_min_), msg) test.assertEqual(list(op1.data_max_), list(op2.data_max_), msg) test.assertEqual(list(op1.data_range_), list(op2.data_range_), msg) test.assertEqual(list(op1.scale_), list(op2.scale_), msg) test.assertEqual(list(op1.min_), list(op2.min_), msg) test.assertEqual(op1.n_features_in_, op2.n_features_in_, msg) test.assertEqual(op1.n_samples_seen_, op2.n_samples_seen_, msg) class TestMinMaxScaler(unittest.TestCase): @classmethod def setUpClass(cls): targets: List[datatype_param_type] = ["pandas", "spark"] cls.tgt2datasets = {tgt: fetch_go_sales_dataset(tgt) for tgt in targets} def test_get_params(self): sk_scaler = SkMinMaxScaler() rasl_scaler = RaslMinMaxScaler() sk_params = sk_scaler.get_params() rasl_params = rasl_scaler.get_params() self.assertDictContainsSubset(sk_params, rasl_params) def test_error(self): _ = RaslMinMaxScaler(clip=True) # should raise no error with self.assertRaisesRegex( jsonschema.ValidationError, re.compile(r"MinMaxScaler\(copy=False\)", re.MULTILINE | re.DOTALL), ): _ = RaslMinMaxScaler(copy=False) def test_fit(self): columns = ["Product number", "Quantity", "Retailer code"] pandas_data = self.tgt2datasets["pandas"][0][columns] sk_scaler = SkMinMaxScaler() sk_trained = sk_scaler.fit(pandas_data) rasl_scaler = RaslMinMaxScaler() for tgt, go_sales in self.tgt2datasets.items(): data = go_sales[0][columns] if tgt == "spark": data = SparkDataFrameWithIndex(data) rasl_trained = rasl_scaler.fit(data) _check_trained_min_max_scaler(self, sk_trained, rasl_trained.impl, "pandas") def test_transform(self): columns = ["Product number", "Quantity", "Retailer code"] pandas_data = self.tgt2datasets["pandas"][0][columns] sk_scaler = SkMinMaxScaler() sk_trained = sk_scaler.fit(pandas_data) sk_transformed = sk_trained.transform(pandas_data) rasl_scaler = RaslMinMaxScaler() for tgt, go_sales in self.tgt2datasets.items(): data = go_sales[0][columns] if tgt == "spark": data = SparkDataFrameWithIndex(data) rasl_trained = rasl_scaler.fit(data) rasl_transformed = rasl_trained.transform(data) if tgt == "spark": self.assertEqual(get_index_name(rasl_transformed), "index") rasl_transformed = _ensure_pandas(rasl_transformed) self.assertAlmostEqual(sk_transformed[0, 0], rasl_transformed.iloc[0, 0]) self.assertAlmostEqual(sk_transformed[0, 1], rasl_transformed.iloc[0, 1]) self.assertAlmostEqual(sk_transformed[0, 2], rasl_transformed.iloc[0, 2]) self.assertAlmostEqual(sk_transformed[10, 0], rasl_transformed.iloc[10, 0]) self.assertAlmostEqual(sk_transformed[10, 1], rasl_transformed.iloc[10, 1]) self.assertAlmostEqual(sk_transformed[10, 2], rasl_transformed.iloc[10, 2]) self.assertAlmostEqual(sk_transformed[20, 0], rasl_transformed.iloc[20, 0]) self.assertAlmostEqual(sk_transformed[20, 1], rasl_transformed.iloc[20, 1]) self.assertAlmostEqual(sk_transformed[20, 2], rasl_transformed.iloc[20, 2]) def test_transform_clipped(self): columns = ["Product number", "Quantity", "Retailer code"] pandas_data = self.tgt2datasets["pandas"][0][columns] sk_scaler = SkMinMaxScaler(clip=True) sk_trained = sk_scaler.fit(pandas_data) sk_transformed = sk_trained.transform(pandas_data) rasl_scaler = RaslMinMaxScaler(clip=True) for tgt, go_sales in self.tgt2datasets.items(): data = go_sales[0][columns] if tgt == "spark": data = SparkDataFrameWithIndex(data) rasl_trained = rasl_scaler.fit(data) rasl_transformed = rasl_trained.transform(data) if tgt == "spark": self.assertEqual(get_index_name(rasl_transformed), "index") rasl_transformed = _ensure_pandas(rasl_transformed) self.assertAlmostEqual(sk_transformed[0, 0], rasl_transformed.iloc[0, 0]) self.assertAlmostEqual(sk_transformed[0, 1], rasl_transformed.iloc[0, 1]) self.assertAlmostEqual(sk_transformed[0, 2], rasl_transformed.iloc[0, 2]) self.assertAlmostEqual(sk_transformed[10, 0], rasl_transformed.iloc[10, 0]) self.assertAlmostEqual(sk_transformed[10, 1], rasl_transformed.iloc[10, 1]) self.assertAlmostEqual(sk_transformed[10, 2], rasl_transformed.iloc[10, 2]) self.assertAlmostEqual(sk_transformed[20, 0], rasl_transformed.iloc[20, 0]) self.assertAlmostEqual(sk_transformed[20, 1], rasl_transformed.iloc[20, 1]) self.assertAlmostEqual(sk_transformed[20, 2], rasl_transformed.iloc[20, 2]) def test_zero_scale(self): pandas_data = pd.DataFrame({"a": [0.5]}) sk_scaler = SkMinMaxScaler() sk_trained = sk_scaler.fit(pandas_data) sk_transformed = sk_trained.transform(pandas_data) rasl_scaler = RaslMinMaxScaler() for tgt, _ in self.tgt2datasets.items(): data = Convert(astype=tgt).transform(pandas_data) rasl_trained = rasl_scaler.fit(data) rasl_transformed = rasl_trained.transform(data) rasl_transformed = _ensure_pandas(rasl_transformed) self.assertAlmostEqual(sk_transformed[0, 0], rasl_transformed.iloc[0, 0]) def test_fit_range(self): columns = ["Product number", "Quantity", "Retailer code"] pandas_data = self.tgt2datasets["pandas"][0][columns] sk_scaler = SkMinMaxScaler(feature_range=(-5, 5)) sk_trained = sk_scaler.fit(pandas_data) for tgt, go_sales in self.tgt2datasets.items(): data = go_sales[0][columns] if tgt == "spark": data = SparkDataFrameWithIndex(data) rasl_scaler = RaslMinMaxScaler(feature_range=(-5, 5)) rasl_trained = rasl_scaler.fit(data) _check_trained_min_max_scaler(self, sk_trained, rasl_trained.impl, "pandas") def test_transform_range(self): columns = ["Product number", "Quantity", "Retailer code"] pandas_data = self.tgt2datasets["pandas"][0][columns] sk_scaler = SkMinMaxScaler(feature_range=(-5, 5)) sk_trained = sk_scaler.fit(pandas_data) sk_transformed = sk_trained.transform(pandas_data) rasl_scaler = RaslMinMaxScaler(feature_range=(-5, 5)) for tgt, go_sales in self.tgt2datasets.items(): data = go_sales[0][columns] if tgt == "spark": data = SparkDataFrameWithIndex(data) rasl_trained = rasl_scaler.fit(data) rasl_transformed = rasl_trained.transform(data) rasl_transformed = _ensure_pandas(rasl_transformed) self.assertAlmostEqual(sk_transformed[0, 0], rasl_transformed.iloc[0, 0]) self.assertAlmostEqual(sk_transformed[0, 1], rasl_transformed.iloc[0, 1]) self.assertAlmostEqual(sk_transformed[0, 2], rasl_transformed.iloc[0, 2]) self.assertAlmostEqual(sk_transformed[10, 0], rasl_transformed.iloc[10, 0]) self.assertAlmostEqual(sk_transformed[10, 1], rasl_transformed.iloc[10, 1]) self.assertAlmostEqual(sk_transformed[10, 2], rasl_transformed.iloc[10, 2]) self.assertAlmostEqual(sk_transformed[20, 0], rasl_transformed.iloc[20, 0]) self.assertAlmostEqual(sk_transformed[20, 1], rasl_transformed.iloc[20, 1]) self.assertAlmostEqual(sk_transformed[20, 2], rasl_transformed.iloc[20, 2]) def test_partial_fit(self): columns = ["Product number", "Quantity", "Retailer code"] data = self.tgt2datasets["pandas"][0][columns] for tgt in self.tgt2datasets.keys(): sk_scaler = SkMinMaxScaler() rasl_scaler = RaslMinMaxScaler() for lower, upper in [[0, 10], [10, 100], [100, data.shape[0]]]: data_so_far = data[0:upper] data_delta = data[lower:upper] if tgt == "pandas": pass elif tgt == "spark": data_delta = pandas2spark(data_delta) else: assert False sk_trained = sk_scaler.fit(data_so_far) rasl_trained = rasl_scaler.partial_fit(data_delta) _check_trained_min_max_scaler(self, sk_trained, rasl_trained.impl, tgt) class TestPipeline(unittest.TestCase): @classmethod def setUpClass(cls): from sklearn.datasets import load_iris from sklearn.model_selection import train_test_split targets = ["pandas", "spark"] cls.tgt2datasets = {tgt: {} for tgt in targets} def add_df(name, df): cls.tgt2datasets["pandas"][name] = df cls.tgt2datasets["spark"][name] = pandas2spark(df) X, y = load_iris(as_frame=True, return_X_y=True) X_train, X_test, y_train, y_test = train_test_split(X, y) add_df("X_train", X_train) add_df("X_test", X_test) add_df("y_train", y_train) add_df("y_test", y_test) def test_pipeline(self): for _tgt, datasets in self.tgt2datasets.items(): X_train, X_test = (datasets["X_train"], datasets["X_test"]) y_train = self.tgt2datasets["pandas"]["y_train"] pipeline = ( RaslMinMaxScaler() >> Convert(astype="pandas") >> LogisticRegression() ) trained = pipeline.fit(X_train, y_train) _ = trained.predict(X_test) _ = trained.predict(X_test) def _check_trained_select_k_best(self, sk_trained, rasl_trained, msg=""): self.assertEqual(len(sk_trained.scores_), len(rasl_trained.impl.scores_)) self.assertEqual(len(sk_trained.scores_), len(sk_trained.pvalues_)) self.assertEqual(len(sk_trained.scores_), len(rasl_trained.impl.pvalues_)) for i, (sk_score, rasl_score, sk_pvalue, rasl_pvalue) in enumerate( zip( sk_trained.scores_, rasl_trained.impl.scores_, sk_trained.pvalues_, rasl_trained.impl.pvalues_, ) ): if not (np.isnan(sk_score) and np.isnan(rasl_score)): self.assertAlmostEqual( sk_score, rasl_score, msg=f"{msg}: {i}", ) if not (np.isnan(sk_pvalue) and np.isnan(rasl_pvalue)): self.assertAlmostEqual( sk_pvalue, rasl_pvalue, msg=f"{msg}: {i}", ) self.assertEqual(sk_trained.n_features_in_, rasl_trained.impl.n_features_in_, msg) class TestSelectKBest(unittest.TestCase): @classmethod def setUpClass(cls): from sklearn.datasets import load_digits targets = ["pandas", "spark"] cls.tgt2datasets = {tgt: {} for tgt in targets} def add_df(name, df): cls.tgt2datasets["pandas"][name] = df cls.tgt2datasets["spark"][name] = pandas2spark(df) X, y = load_digits(return_X_y=True, as_frame=True) X = add_table_name(X, "X") y = add_table_name(y, "y") add_df("X", X) add_df("y", y) def test_fit(self): sk_trainable = SkSelectKBest(k=20) X, y = self.tgt2datasets["pandas"]["X"], self.tgt2datasets["pandas"]["y"] sk_trained = sk_trainable.fit(X, y) rasl_trainable = RaslSelectKBest(k=20) for tgt, datasets in self.tgt2datasets.items(): X, y = datasets["X"], datasets["y"] rasl_trained = rasl_trainable.fit(X, y) _check_trained_select_k_best(self, sk_trained, rasl_trained, tgt) def test_transform(self): sk_trainable = SkSelectKBest(k=20) X, y = self.tgt2datasets["pandas"]["X"], self.tgt2datasets["pandas"]["y"] sk_trained = sk_trainable.fit(X, y) sk_transformed = sk_trained.transform(X) rasl_trainable = RaslSelectKBest(k=20) for tgt, datasets in self.tgt2datasets.items(): X, y = datasets["X"], datasets["y"] rasl_trained = rasl_trainable.fit(X, y) rasl_transformed = rasl_trained.transform(X) _check_trained_select_k_best(self, sk_trained, rasl_trained, tgt) rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertAlmostEqual( sk_transformed[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], msg=(row_idx, col_idx), ) def test_partial_fit(self): rasl_trainable = RaslSelectKBest(k=20) X, y = self.tgt2datasets["pandas"]["X"], self.tgt2datasets["pandas"]["y"] for lower, upper in [[0, 100], [100, 200], [200, X.shape[0]]]: X_so_far, y_so_far = X[0:upper], y[0:upper] sk_trainable = SkSelectKBest(k=20) sk_trained = sk_trainable.fit(X_so_far, y_so_far) X_delta, y_delta = X[lower:upper], y[lower:upper] rasl_trained = rasl_trainable.partial_fit(X_delta, y_delta) _check_trained_select_k_best( self, sk_trained, rasl_trained, f"lower: {lower}, upper: {upper}" ) def _check_trained_ordinal_encoder(test, op1, op2, msg): if hasattr(op1, "feature_names_in_"): test.assertEqual(list(op1.feature_names_in_), list(op2.feature_names_in_), msg) test.assertEqual(len(op1.categories_), len(op2.categories_), msg) for cat1, cat2 in zip(op1.categories_, op2.categories_): test.assertEqual(len(cat1), len(cat2), msg) for num1, num2 in zip(cat1, cat2): if isinstance(num1, numbers.Number) and math.isnan(num2): test.assertTrue(math.isnan(num2), msg) else: test.assertEqual(num1, num2, msg) class TestOrdinalEncoder(unittest.TestCase): @classmethod def setUpClass(cls): targets: List[datatype_param_type] = ["pandas", "spark"] cls.tgt2gosales = {tgt: fetch_go_sales_dataset(tgt) for tgt in targets} cls.tgt2creditg = { tgt: lale.datasets.openml.fetch( "credit-g", "classification", preprocess=False, astype=tgt, ) for tgt in targets } def _check_last_trained(self, op1, op2, msg): _check_trained_ordinal_encoder( self, op1.get_last().impl, op2.get_last().impl, msg ) def test_fit(self): prefix = Scan(table=it.go_daily_sales) >> Map( columns={"retailer": it["Retailer code"], "method": it["Order method code"]} ) encoder_args = {"handle_unknown": "use_encoded_value", "unknown_value": np.nan} rasl_trainable = prefix >> RaslOrdinalEncoder(**encoder_args) sk_trainable = prefix >> SkOrdinalEncoder(**encoder_args) sk_trained = sk_trainable.fit(self.tgt2gosales["pandas"]) for tgt, datasets in self.tgt2gosales.items(): rasl_trained = rasl_trainable.fit(datasets) self._check_last_trained(sk_trained, rasl_trained, tgt) def test_partial_fit(self): prefix = Scan(table=it.go_daily_sales) >> Map( columns={"retailer": it["Retailer code"], "method": it["Order method code"]} ) pandas_data = prefix.transform(self.tgt2gosales["pandas"]) encoder_args = {"handle_unknown": "use_encoded_value", "unknown_value": np.nan} for tgt in self.tgt2gosales.keys(): rasl_op = RaslOrdinalEncoder(**encoder_args) for lower, upper in [[0, 10], [10, 100], [100, pandas_data.shape[0]]]: data_so_far = pandas_data[0:upper] sk_op = SkOrdinalEncoder(**encoder_args).fit(data_so_far) data_delta = pandas_data[lower:upper] if tgt == "pandas": pass elif tgt == "spark": data_delta = pandas2spark(data_delta) else: assert False rasl_op = rasl_op.partial_fit(data_delta) _check_trained_ordinal_encoder( self, sk_op, rasl_op.impl, f"tgt {tgt}, lower {lower}, upper {upper}", ) def test_transform(self): prefix = Scan(table=it.go_daily_sales) >> Map( columns={"retailer": it["Retailer code"], "method": it["Order method code"]} ) encoder_args = {"handle_unknown": "use_encoded_value", "unknown_value": np.nan} rasl_trainable = prefix >> RaslOrdinalEncoder(**encoder_args) sk_trainable = prefix >> SkOrdinalEncoder(**encoder_args) sk_trained = sk_trainable.fit(self.tgt2gosales["pandas"]) sk_transformed = sk_trained.transform(self.tgt2gosales["pandas"]) for tgt, datasets in self.tgt2gosales.items(): rasl_trained = rasl_trainable.fit(datasets) self._check_last_trained(sk_trained, rasl_trained, tgt) rasl_transformed = rasl_trained.transform(datasets) if tgt == "spark": self.assertEqual(get_index_name(rasl_transformed), "index") rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertEqual( sk_transformed[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], (row_idx, col_idx, tgt), ) def test_predict(self): (train_X_pd, train_y_pd), (test_X_pd, _test_y_pd) = self.tgt2creditg["pandas"] cat_columns = categorical()(train_X_pd) prefix = Map(columns={c: it[c] for c in cat_columns}) to_pd = Convert(astype="pandas") lr = LogisticRegression() encoder_args = {"handle_unknown": "use_encoded_value", "unknown_value": -1} sk_trainable = prefix >> SkOrdinalEncoder(**encoder_args) >> lr sk_trained = sk_trainable.fit(train_X_pd, train_y_pd) sk_predicted = sk_trained.predict(test_X_pd) rasl_trainable = prefix >> RaslOrdinalEncoder(**encoder_args) >> to_pd >> lr for tgt, dataset in self.tgt2creditg.items(): (train_X, train_y), (test_X, _test_y) = dataset rasl_trained = rasl_trainable.fit(train_X, train_y) rasl_predicted = rasl_trained.predict(test_X) self.assertEqual(sk_predicted.shape, rasl_predicted.shape, tgt) self.assertEqual(sk_predicted.tolist(), rasl_predicted.tolist(), tgt) def _check_trained_one_hot_encoder(test, op1, op2, msg): if hasattr(op1, "feature_names_in_"): test.assertEqual(list(op1.feature_names_in_), list(op2.feature_names_in_), msg) test.assertEqual(len(op1.categories_), len(op2.categories_), msg) for cat1, cat2 in zip(op1.categories_, op2.categories_): test.assertEqual(list(cat1), list(cat2), msg) class TestOneHotEncoder(unittest.TestCase): @classmethod def setUpClass(cls): targets = ["pandas", "spark"] cls.tgt2creditg = cast( Dict[str, Any], { tgt: lale.datasets.openml.fetch( "credit-g", "classification", preprocess=False, astype=tgt, ) for tgt in targets }, ) def _check_last_trained(self, op1, op2, msg): _check_trained_one_hot_encoder( self, op1.get_last().impl, op2.get_last().impl, msg ) def test_fit(self): (train_X_pd, _), (_, _) = self.tgt2creditg["pandas"] cat_columns = categorical()(train_X_pd) prefix = Map(columns={c: it[c] for c in cat_columns}) rasl_trainable = prefix >> RaslOneHotEncoder() sk_trainable = prefix >> SkOneHotEncoder() sk_trained = sk_trainable.fit(train_X_pd) for tgt, dataset in self.tgt2creditg.items(): (train_X, _train_y), (_test_X, _test_y) = dataset rasl_trained = rasl_trainable.fit(train_X) self._check_last_trained(sk_trained, rasl_trained, tgt) def test_partial_fit(self): (train_X_pd, _), (_, _) = self.tgt2creditg["pandas"] cat_columns = categorical()(train_X_pd) prefix = Map(columns={c: it[c] for c in cat_columns}) for tgt in self.tgt2creditg.keys(): rasl_pipe = prefix >> RaslOneHotEncoder() for lower, upper in [[0, 10], [10, 100], [100, train_X_pd.shape[0]]]: data_so_far = train_X_pd[0:upper] sk_pipe = prefix >> SkOrdinalEncoder() sk_pipe = sk_pipe.fit(data_so_far) data_delta = train_X_pd[lower:upper] if tgt == "pandas": pass elif tgt == "spark": data_delta = pandas2spark(data_delta) else: assert False rasl_pipe = rasl_pipe.partial_fit(data_delta) self._check_last_trained( sk_pipe, rasl_pipe, (tgt, lower, upper), ) def test_transform(self): (train_X_pd, _train_y_pd), (test_X_pd, _test_y_pd) = self.tgt2creditg["pandas"] cat_columns = categorical()(train_X_pd) prefix = Map(columns={c: it[c] for c in cat_columns}) rasl_trainable = prefix >> RaslOneHotEncoder(sparse=False) sk_trainable = prefix >> SkOneHotEncoder(sparse=False) sk_trained = sk_trainable.fit(train_X_pd) sk_transformed = sk_trained.transform(test_X_pd) for tgt, dataset in self.tgt2creditg.items(): (train_X, _train_y), (test_X, _test_y) = dataset rasl_trained = rasl_trainable.fit(train_X) self._check_last_trained(sk_trained, rasl_trained, tgt) rasl_transformed = rasl_trained.transform(test_X) if tgt == "spark": self.assertEqual(get_index_name(rasl_transformed), "index") rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertEqual( sk_transformed[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], (row_idx, col_idx, tgt), ) def test_predict(self): (train_X_pd, train_y_pd), (test_X_pd, _test_y_pd) = self.tgt2creditg["pandas"] cat_columns = categorical()(train_X_pd) prefix = Map(columns={c: it[c] for c in cat_columns}) to_pd = Convert(astype="pandas") lr = LogisticRegression() sk_trainable = prefix >> SkOneHotEncoder(sparse=False) >> lr sk_trained = sk_trainable.fit(train_X_pd, train_y_pd) sk_predicted = sk_trained.predict(test_X_pd) rasl_trainable = prefix >> RaslOneHotEncoder(sparse=False) >> to_pd >> lr for tgt, dataset in self.tgt2creditg.items(): (train_X, train_y), (test_X, _test_y) = dataset rasl_trained = rasl_trainable.fit(train_X, train_y) rasl_predicted = rasl_trained.predict(test_X) self.assertEqual(sk_predicted.shape, rasl_predicted.shape, tgt) self.assertEqual(sk_predicted.tolist(), rasl_predicted.tolist(), tgt) def _get_feature_names_out(op): """later version of category_encoder's HashingEncoder changed the attribute name""" fnames = getattr(op, "feature_names", None) if fnames is not None: return fnames fnames = getattr(op, "get_feature_names_out", None) assert fnames is not None return fnames() def _check_trained_hashing_encoder(test, op1, op2, msg): test.assertEqual( list(_get_feature_names_out(op1)), list(_get_feature_names_out(op2)), msg ) class TestHashingEncoder(unittest.TestCase): @classmethod def setUpClass(cls): targets = ["pandas", "spark"] cls.tgt2creditg = cast( Dict[str, Any], { tgt: lale.datasets.openml.fetch( "credit-g", "classification", preprocess=False, astype=tgt, ) for tgt in targets }, ) def _check_last_trained(self, op1, op2, msg): _check_trained_hashing_encoder( self, op1.get_last().impl, op2.get_last().impl, msg ) def test_fit(self): (train_X_pd, _), (_, _) = self.tgt2creditg["pandas"] cat_columns = categorical()(train_X_pd) prefix = Map(columns={c: it[c] for c in cat_columns}) rasl_trainable = prefix >> RaslHashingEncoder() sk_trainable = prefix >> SkHashingEncoder() sk_trained = sk_trainable.fit(train_X_pd) # TODO: test with multiple batches for tgt, dataset in self.tgt2creditg.items(): (train_X, _train_y), (_test_X, _test_y) = dataset rasl_trained = rasl_trainable.fit(train_X) self._check_last_trained(sk_trained, rasl_trained, tgt) def test_transform(self): (train_X_pd, _train_y_pd), (test_X_pd, _test_y_pd) = self.tgt2creditg["pandas"] cat_columns = categorical()(train_X_pd) prefix = Map(columns={c: it[c] for c in cat_columns}) rasl_trainable = prefix >> RaslHashingEncoder() sk_trainable = prefix >> SkHashingEncoder() sk_trained = sk_trainable.fit(train_X_pd) sk_transformed = sk_trained.transform(test_X_pd) for tgt, dataset in self.tgt2creditg.items(): (train_X, _train_y), (test_X, _test_y) = dataset rasl_trained = rasl_trainable.fit(train_X) self._check_last_trained(sk_trained, rasl_trained, tgt) rasl_transformed = rasl_trained.transform(test_X) if tgt == "spark": self.assertEqual(get_index_name(rasl_transformed), "index") rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertEqual( sk_transformed.iloc[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], (row_idx, col_idx, tgt), ) def test_predict(self): (train_X_pd, train_y_pd), (test_X_pd, _test_y_pd) = self.tgt2creditg["pandas"] cat_columns = categorical()(train_X_pd) prefix = Map(columns={c: it[c] for c in cat_columns}) to_pd = Convert(astype="pandas") lr = LogisticRegression() sk_trainable = prefix >> SkHashingEncoder() >> lr sk_trained = sk_trainable.fit(train_X_pd, train_y_pd) sk_predicted = sk_trained.predict(test_X_pd) rasl_trainable = prefix >> RaslHashingEncoder() >> to_pd >> lr for tgt, dataset in self.tgt2creditg.items(): (train_X, train_y), (test_X, _test_y) = dataset rasl_trained = rasl_trainable.fit(train_X, train_y) rasl_predicted = rasl_trained.predict(test_X) self.assertEqual(sk_predicted.shape, rasl_predicted.shape, tgt) self.assertEqual(sk_predicted.tolist(), rasl_predicted.tolist(), tgt) def _check_trained_target_encoder(test, op1, op2, msg): names1, names2 = _get_feature_names_out(op1), _get_feature_names_out(op2) test.assertListEqual(list(names1), list(names2), msg) test.assertSequenceEqual(op1.mapping.keys(), op2._col2cat2value.keys(), msg) for col in op1.mapping.keys(): op1_cat2val = op1.mapping[col].to_dict() op2_cat2val = op2._col2cat2value[col] expected_keys = list(range(1, 1 + len(op2_cat2val))) + [-1, -2] test.assertListEqual(list(op1_cat2val.keys()), expected_keys, (col, msg)) test.assertAlmostEqual(op1_cat2val[-1], op2._prior, msg=(col, msg)) for i, cat in enumerate(op2_cat2val.keys()): test.assertAlmostEqual( op1_cat2val[i + 1], op2_cat2val[cat], msg=(col, i, cat, msg) ) class TestTargetEncoder(unittest.TestCase): @classmethod def setUpClass(cls): targets = ["pandas", "spark"] dataset_names = [ "tae", # 3-class classification "cloud", # regression "credit-g", # binary classification ] experiments_dict = lale.datasets.openml.openml_datasets.experiments_dict cls.datasets = cast( Dict[Tuple[str, str], Any], { (tgt, dataset_name): lale.datasets.openml.fetch( dataset_name, experiments_dict[dataset_name]["task_type"], preprocess=False, astype=tgt, ) for tgt, dataset_name in itertools.product(targets, dataset_names) }, ) def test_fit(self): for (tgt, dataset_name), dataset in self.datasets.items(): (train_X, train_y), (_, _) = dataset (train_X_pd, train_y_pd), (_, _) = self.datasets["pandas", dataset_name] sk_trainable = SkTargetEncoder() sk_trained = sk_trainable.fit(train_X_pd, train_y_pd) experiments_dict = lale.datasets.openml.openml_datasets.experiments_dict if experiments_dict[dataset_name]["task_type"] == "regression": rasl_trainable = RaslTargetEncoder() else: classes = sorted(list(train_y_pd.unique())) rasl_trainable = RaslTargetEncoder(classes=classes) rasl_trained = rasl_trainable.fit(train_X, train_y) _check_trained_target_encoder( self, sk_trained.impl, rasl_trained.impl, (tgt, dataset_name) ) def test_partial_fit(self): for (tgt, dataset_name), dataset in self.datasets.items(): if tgt != "pandas": continue (train_X, train_y), (_, _) = dataset experiments_dict = lale.datasets.openml.openml_datasets.experiments_dict if experiments_dict[dataset_name]["task_type"] == "regression": classes = None else: classes = sorted(list(_ensure_pandas(train_y).unique())) rasl_trainable = RaslTargetEncoder(classes=classes) rasl_trained = rasl_trainable.fit(train_X, train_y) rasl_batched = RaslTargetEncoder(classes=classes) for batch_X, batch_y in mockup_data_loader(train_X, train_y, 3, tgt): rasl_batched = rasl_batched.partial_fit(batch_X, batch_y) op1, op2 = rasl_trained.impl, rasl_batched.impl self.assertAlmostEqual(op1._prior, op2._prior, msg=(tgt, dataset_name)) self.assertSequenceEqual( op1._col2cat2value.keys(), op2._col2cat2value.keys(), msg=(tgt, dataset_name), ) for col in op1._col2cat2value: self.assertSequenceEqual( op1._col2cat2value[col].keys(), op2._col2cat2value[col].keys(), msg=(tgt, dataset_name, col), ) for cat in op1._col2cat2value[col]: self.assertAlmostEqual( op1._col2cat2value[col][cat], op2._col2cat2value[col][cat], msg=(tgt, dataset_name, col, cat), ) def test_transform(self): for (tgt, dataset_name), dataset in self.datasets.items(): (train_X, train_y), (test_X, _) = dataset (train_X_pd, train_y_pd), (test_X_pd, _) = self.datasets[ "pandas", dataset_name ] sk_trainable = SkTargetEncoder() sk_trained = sk_trainable.fit(train_X_pd, train_y_pd) sk_transformed = sk_trained.transform(test_X_pd) experiments_dict = lale.datasets.openml.openml_datasets.experiments_dict if experiments_dict[dataset_name]["task_type"] == "regression": rasl_trainable = RaslTargetEncoder(classes=None) else: classes = sorted(list(train_y_pd.unique())) rasl_trainable = RaslTargetEncoder(classes=classes) rasl_trained = rasl_trainable.fit(train_X, train_y) _check_trained_target_encoder( self, sk_trained.impl, rasl_trained.impl, (tgt, dataset_name) ) rasl_transformed = rasl_trained.transform(test_X) if tgt == "spark": self.assertEqual( get_index_name(rasl_transformed), "index", (tgt, dataset_name) ) rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual( sk_transformed.shape, rasl_transformed.shape, (tgt, dataset_name) ) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertEqual( sk_transformed.iloc[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], (tgt, dataset_name, row_idx, col_idx), ) def test_predict(self): for (tgt, dataset_name), dataset in self.datasets.items(): (train_X, train_y), (test_X, _) = dataset (train_X_pd, train_y_pd), (test_X_pd, _) = self.datasets[ "pandas", dataset_name ] experiments_dict = lale.datasets.openml.openml_datasets.experiments_dict if experiments_dict[dataset_name]["task_type"] == "regression": classes = None est = LinearRegression() else: classes = sorted(list(train_y_pd.unique())) est = LogisticRegression() sk_trainable = SkTargetEncoder() >> est sk_trained = sk_trainable.fit(train_X_pd, train_y_pd) sk_predicted = sk_trained.predict(test_X_pd) rasl_trainable = ( RaslTargetEncoder(classes=classes) >> Convert(astype="pandas") >> est ) rasl_trained = rasl_trainable.fit(train_X, train_y) rasl_predicted = rasl_trained.predict(test_X) self.assertEqual( sk_predicted.shape, rasl_predicted.shape, (tgt, dataset_name) ) self.assertListEqual( sk_predicted.tolist(), rasl_predicted.tolist(), (tgt, dataset_name) ) def _check_trained_simple_imputer(test, op1, op2, msg): if hasattr(op1, "feature_names_in_"): test.assertEqual(list(op1.feature_names_in_), list(op2.feature_names_in_), msg) if hasattr(op1, "statistics_"): test.assertEqual(len(op1.statistics_), len(op2.statistics_), msg) for stats1, stats2 in zip(op1.statistics_, op2.statistics_): test.assertEqual(stats1, stats2, msg) if hasattr(op1, "n_features_in_"): test.assertEqual(op1.n_features_in_, op2.n_features_in_, msg) if hasattr(op1, "indicator_"): test.assertEqual(op1.indicator_, op2.indicator_, msg) class TestSimpleImputer(unittest.TestCase): def setUp(self): targets = ["pandas", "spark"] self.tgt2adult = { tgt: lale.datasets.openml.fetch( "adult", "classification", preprocess=False, astype=tgt, ) for tgt in targets } def _fill_missing_value(self, col_name, value, missing_value): for tgt, datasets in self.tgt2adult.items(): (train_X, train_y), (test_X, test_y) = datasets if tgt == "pandas": train_X.loc[ train_X[col_name] == value, col_name ] = missing_value # type:ignore test_X.loc[ test_X[col_name] == value, col_name ] = missing_value # type:ignore elif tgt == "spark": from pyspark.sql.functions import col, when train_X_new = train_X.withColumn( col_name, when(col(col_name) == value, missing_value).otherwise( col(col_name) ), ) test_X_new = test_X.withColumn( col_name, when(col(col_name) == value, missing_value).otherwise( col(col_name) ), ) train_X = forward_metadata(train_X, train_X_new) test_X = forward_metadata(test_X, test_X_new) else: assert False self.tgt2adult[tgt] = (train_X, train_y), (test_X, test_y) def test_fit_transform_numeric_nan_missing(self): self._fill_missing_value("age", 36.0, np.nan) num_columns = ["age", "fnlwgt", "education-num"] prefix = Map(columns={c: it[c] for c in num_columns}) hyperparams = [ {"strategy": "mean"}, {"strategy": "median"}, {"strategy": "most_frequent"}, {"strategy": "constant", "fill_value": 99}, ] for hyperparam in hyperparams: rasl_imputer = RaslSimpleImputer(**hyperparam) sk_imputer = SkSimpleImputer(**hyperparam) rasl_trainable = prefix >> rasl_imputer sk_trainable = prefix >> sk_imputer sk_trained = sk_trainable.fit(self.tgt2adult["pandas"][0][0]) sk_transformed = sk_trained.transform(self.tgt2adult["pandas"][1][0]) sk_statistics_ = sk_trained.steps[-1][1].impl.statistics_ for tgt, dataset in self.tgt2adult.items(): (train_X, _), (test_X, _) = dataset rasl_trained = rasl_trainable.fit(train_X) # test the fit succeeded. rasl_statistics_ = rasl_trained.steps[-1][1].impl.statistics_ self.assertEqual( len(sk_statistics_), len(rasl_statistics_), (hyperparam, tgt) ) for i in range(sk_statistics_.shape[0]): self.assertAlmostEqual( sk_statistics_[i], rasl_statistics_[i], msg=(i, hyperparam, tgt) ) rasl_transformed = rasl_trained.transform(test_X) if tgt == "spark": self.assertEqual(get_index_name(rasl_transformed), "index") rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertAlmostEqual( sk_transformed[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], msg=(row_idx, col_idx, tgt), ) _check_trained_simple_imputer(self, sk_imputer, rasl_imputer, tgt) def test_fit_transform_numeric_nonan_missing(self): self._fill_missing_value("age", 36.0, -1) num_columns = ["age", "fnlwgt", "education-num"] prefix = Map(columns={c: it[c] for c in num_columns}) hyperparams = [ {"strategy": "mean"}, {"strategy": "median"}, {"strategy": "most_frequent"}, {"strategy": "constant", "fill_value": 99}, ] for hyperparam in hyperparams: rasl_trainable = prefix >> RaslSimpleImputer( missing_values=-1, **hyperparam ) sk_trainable = prefix >> SkSimpleImputer(missing_values=-1, **hyperparam) sk_trained = sk_trainable.fit(self.tgt2adult["pandas"][0][0]) sk_transformed = sk_trained.transform(self.tgt2adult["pandas"][1][0]) sk_statistics_ = sk_trained.get_last().impl.statistics_ for tgt, dataset in self.tgt2adult.items(): (train_X, _), (test_X, _) = dataset rasl_trained = rasl_trainable.fit(train_X) # test the fit succeeded. rasl_statistics_ = rasl_trained.get_last().impl.statistics_ # type: ignore self.assertEqual(len(sk_statistics_), len(rasl_statistics_), tgt) self.assertEqual(list(sk_statistics_), list(rasl_statistics_), tgt) rasl_transformed = rasl_trained.transform(test_X) if tgt == "spark": self.assertEqual(get_index_name(rasl_transformed), "index") rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertEqual( sk_transformed[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], (row_idx, col_idx, tgt), ) def test_predict(self): self._fill_missing_value("age", 36.0, np.nan) (train_X_pd, train_y_pd), (test_X_pd, _test_y_pd) = self.tgt2adult["pandas"] num_columns = ["age", "fnlwgt", "education-num"] prefix = Map(columns={c: it[c] for c in num_columns}) to_pd = Convert(astype="pandas") lr = LogisticRegression() imputer_args = {"strategy": "mean"} sk_trainable = prefix >> SkSimpleImputer(**imputer_args) >> lr sk_trained = sk_trainable.fit(train_X_pd, train_y_pd) sk_predicted = sk_trained.predict(test_X_pd) rasl_trainable = prefix >> RaslSimpleImputer(**imputer_args) >> to_pd >> lr for tgt, dataset in self.tgt2adult.items(): (train_X, train_y), (test_X, _test_y) = dataset rasl_trained = rasl_trainable.fit(train_X, train_y) rasl_predicted = rasl_trained.predict(test_X) self.assertEqual(sk_predicted.shape, rasl_predicted.shape, tgt) self.assertEqual(sk_predicted.tolist(), rasl_predicted.tolist(), tgt) def test_invalid_datatype_strategy(self): sk_trainable = SkSimpleImputer() with self.assertRaises(ValueError): sk_trainable.fit(self.tgt2adult["pandas"][0][0]) rasl_trainable = RaslSimpleImputer() for tgt, dataset in self.tgt2adult.items(): (train_X, _), (_, _) = dataset if tgt == "spark": # Skip test because of timeout! continue with self.assertRaises(ValueError): _ = rasl_trainable.fit(train_X) def test_default_numeric_fill_value(self): self._fill_missing_value("age", 36.0, np.nan) num_columns = ["age", "fnlwgt", "education-num"] prefix = Map(columns={c: it[c] for c in num_columns}) hyperparams = [{"strategy": "constant"}] for hyperparam in hyperparams: rasl_trainable = prefix >> RaslSimpleImputer(**hyperparam) sk_trainable = prefix >> SkSimpleImputer(**hyperparam) sk_trained = sk_trainable.fit(self.tgt2adult["pandas"][0][0]) sk_transformed = sk_trained.transform(self.tgt2adult["pandas"][1][0]) sk_statistics_ = sk_trained.steps[-1][1].impl.statistics_ for tgt, dataset in self.tgt2adult.items(): (train_X, _), (test_X, _) = dataset rasl_trained = rasl_trainable.fit(train_X) # test the fit succeeded. rasl_statistics_ = rasl_trained.steps[-1][1].impl.statistics_ self.assertEqual(len(sk_statistics_), len(rasl_statistics_), tgt) self.assertEqual(list(sk_statistics_), list(rasl_statistics_), tgt) rasl_transformed = rasl_trained.transform(test_X) rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertEqual( sk_transformed[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], (row_idx, col_idx, tgt), ) def test_default_string_fill_value(self): self._fill_missing_value("education", "Prof-school", np.nan) str_columns = ["workclass", "education", "capital-gain"] prefix = Map(columns={c: it[c] for c in str_columns}) hyperparams = [{"strategy": "constant"}] for hyperparam in hyperparams: rasl_trainable = prefix >> RaslSimpleImputer(**hyperparam) sk_trainable = prefix >> SkSimpleImputer(**hyperparam) sk_trained = sk_trainable.fit(self.tgt2adult["pandas"][0][0]) sk_statistics_ = sk_trained.steps[-1][1].impl.statistics_ for tgt, dataset in self.tgt2adult.items(): (train_X, _), (test_X, _) = dataset rasl_trained = rasl_trainable.fit(train_X) # test the fit succeeded. rasl_statistics_ = rasl_trained.steps[-1][1].impl.statistics_ self.assertEqual(len(sk_statistics_), len(rasl_statistics_), tgt) self.assertEqual(list(sk_statistics_), list(rasl_statistics_), tgt) rasl_transformed = rasl_trained.transform(test_X) rasl_transformed = _ensure_pandas(rasl_transformed) # Note that for this test case, the output of sklearn transform does not # match rasl transform. There is at least one row which has a None # value and pandas replace treats it as nan and replaces it. # Sklearn which uses numpy does not replace a None. # So we just test that `missing_value` is the default assigned. self.assertEqual(rasl_transformed.iloc[1, 1], "missing_value") def test_multiple_modes_numeric(self): # Sklearn SimpleImputer says: for strategy `most_frequent`, # if there is more than one such value, only the smallest is returned. data = [[1, 10], [2, 14], [3, 15], [4, 15], [5, 14], [6, np.nan]] df = pd.DataFrame(data, columns=["Id", "Age"]) hyperparam = {"strategy": "most_frequent"} sk_trainable = SkSimpleImputer(**hyperparam) rasl_trainable = RaslSimpleImputer(**hyperparam) sk_trained = sk_trainable.fit(df) rasl_trained = rasl_trainable.fit(df) self.assertEqual( len(sk_trained.statistics_), len(rasl_trained.impl.statistics_), "pandas" ) self.assertEqual([6, 15], list(rasl_trained.impl.statistics_), "pandas") spark_df = pandas2spark(df) rasl_trained = rasl_trainable.fit(spark_df) self.assertEqual( len(sk_trained.statistics_), len(rasl_trained.impl.statistics_), "spark" ) self.assertIn(rasl_trained.impl.statistics_[1], [14, 15]) def test_multiple_modes_string(self): # Sklearn SimpleImputer says: for strategy `most_frequent`, # if there is more than one such value, only the smallest is returned. data = [ ["a", "t"], ["b", "f"], ["b", "m"], ["c", "f"], ["c", "m"], ["f", "missing"], ] df = pd.DataFrame(data, columns=["Id", "Gender"]) hyperparam = {"strategy": "most_frequent", "missing_values": "missing"} sk_trainable = SkSimpleImputer(**hyperparam) rasl_trainable = RaslSimpleImputer(**hyperparam) sk_trained = sk_trainable.fit(df) rasl_trained = rasl_trainable.fit(df) self.assertEqual( len(sk_trained.statistics_), len(rasl_trained.impl.statistics_), "pandas" ) self.assertEqual( list(["c", "m"]), list(rasl_trained.impl.statistics_), "pandas" ) spark_df = pandas2spark(df) rasl_trained = rasl_trainable.fit(spark_df) self.assertEqual( len(sk_trained.statistics_), len(rasl_trained.impl.statistics_), "spark" ) self.assertIn(rasl_trained.impl.statistics_[1], ["f", "m"]) def test_valid_partial_fit(self): self._fill_missing_value("age", 36.0, -1) num_columns = ["age", "fnlwgt", "education-num"] prefix = Map(columns={c: it[c] for c in num_columns}) hyperparams = [ {"strategy": "mean"}, {"strategy": "constant", "fill_value": 99}, ] for hyperparam in hyperparams: rasl_trainable = prefix >> RaslSimpleImputer( missing_values=-1, **hyperparam ) sk_trainable = prefix >> SkSimpleImputer(missing_values=-1, **hyperparam) sk_trained = sk_trainable.fit(self.tgt2adult["pandas"][0][0]) sk_transformed = sk_trained.transform(self.tgt2adult["pandas"][1][0]) sk_statistics_ = sk_trained.get_last().impl.statistics_ (train_X, _), (test_X, _) = self.tgt2adult["pandas"] data1_pandas = train_X.iloc[:10] data2_pandas = train_X.iloc[10:100] data3_pandas = train_X.iloc[100:] test_X_pandas = test_X for tgt in self.tgt2adult.keys(): if tgt == "pandas": data1 = data1_pandas data2 = data2_pandas data3 = data3_pandas test_X = test_X_pandas elif tgt == "spark": data1 = pandas2spark(data1_pandas) # type:ignore data2 = pandas2spark(data2_pandas) # type:ignore data3 = pandas2spark(data3_pandas) # type:ignore test_X = pandas2spark(test_X_pandas) # type:ignore else: assert False rasl_trainable = prefix >> RaslSimpleImputer( missing_values=-1, **hyperparam ) rasl_trained = rasl_trainable.partial_fit(data1) rasl_trained = rasl_trained.partial_fit(data2) rasl_trained = rasl_trained.partial_fit(data3) # test the fit succeeded. rasl_statistics_ = rasl_trained.get_last().impl.statistics_ # type: ignore self.assertEqual(len(sk_statistics_), len(rasl_statistics_), tgt) for sk_stats, rasl_stats in zip(sk_statistics_, rasl_statistics_): self.assertEqual(sk_stats, rasl_stats) rasl_transformed = rasl_trained.transform(test_X) rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertEqual( sk_transformed[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], (row_idx, col_idx, tgt), ) def test_invalid_partial_fit(self): num_columns = ["age", "fnlwgt", "education-num"] prefix = Map(columns={c: it[c] for c in num_columns}) hyperparams = [ {"strategy": "median"}, {"strategy": "most_frequent"}, ] for hyperparam in hyperparams: rasl_trainable = prefix >> RaslSimpleImputer( missing_values=-1, **hyperparam ) (train_X, _), (_, _) = self.tgt2adult["pandas"] with self.assertRaises(ValueError): _ = rasl_trainable.partial_fit(train_X) def _check_trained_standard_scaler(test, op1, op2, msg): if hasattr(op1, "feature_names_in_"): test.assertEqual(list(op1.feature_names_in_), list(op2.feature_names_in_), msg) test.assertEqual(op1.n_features_in_, op2.n_features_in_, msg) test.assertEqual(op1.n_samples_seen_, op2.n_samples_seen_, msg) if op1.mean_ is None: test.assertIsNone(op2.mean_, msg) else: test.assertIsNotNone(op2.mean_, msg) test.assertEqual(len(op1.mean_), len(op2.mean_), msg) for mean1, mean2 in zip(op1.mean_, op2.mean_): test.assertAlmostEqual(mean1, mean2, msg=msg) if op1.var_ is None: test.assertIsNone(op2.var_, msg) else: test.assertIsNotNone(op2.var_, msg) test.assertEqual(len(op1.var_), len(op2.var_), msg) for var1, var2 in zip(op1.var_, op2.var_): test.assertAlmostEqual(var1, var2, msg=msg) if op1.scale_ is None: test.assertIsNone(op2.scale_, msg) else: test.assertIsNotNone(op2.scale_, msg) test.assertEqual(len(op1.scale_), len(op2.scale_), msg) for scale1, scale2 in zip(op1.scale_, op2.scale_): test.assertAlmostEqual(scale1, scale2, msg=msg) class TestStandardScaler(unittest.TestCase): @classmethod def setUpClass(cls): targets = ["pandas", "spark"] cls.tgt2creditg = cast( Dict[str, Any], { tgt: lale.datasets.openml.fetch( "credit-g", "classification", preprocess=True, astype=tgt, ) for tgt in targets }, ) def test_fit(self): (train_X_pd, _), (_, _) = self.tgt2creditg["pandas"] sk_trainable = SkStandardScaler() sk_trained = sk_trainable.fit(train_X_pd) rasl_trainable = RaslStandardScaler() for tgt, dataset in self.tgt2creditg.items(): (train_X, _), (_, _) = dataset rasl_trained = rasl_trainable.fit(train_X) _check_trained_standard_scaler(self, sk_trained, rasl_trained.impl, tgt) def test_partial_fit(self): (train_X_pd, _), (_, _) = self.tgt2creditg["pandas"] for tgt in self.tgt2creditg.keys(): rasl_op = RaslStandardScaler() for lower, upper in [[0, 10], [10, 100], [100, train_X_pd.shape[0]]]: data_so_far = train_X_pd[0:upper] sk_op = SkStandardScaler() sk_op = sk_op.fit(data_so_far) data_delta = train_X_pd[lower:upper] if tgt == "pandas": pass elif tgt == "spark": data_delta = pandas2spark(data_delta) else: assert False rasl_op = rasl_op.partial_fit(data_delta) _check_trained_standard_scaler( self, sk_op, rasl_op.impl, (tgt, lower, upper) ) def test_transform(self): (train_X_pd, _), (test_X_pd, _) = self.tgt2creditg["pandas"] sk_trainable = SkStandardScaler() sk_trained = sk_trainable.fit(train_X_pd) sk_transformed = sk_trained.transform(test_X_pd) rasl_trainable = RaslStandardScaler() for tgt, dataset in self.tgt2creditg.items(): (train_X, _), (test_X, _) = dataset rasl_trained = rasl_trainable.fit(train_X) _check_trained_standard_scaler(self, sk_trained, rasl_trained.impl, tgt) rasl_transformed = rasl_trained.transform(test_X) if tgt == "spark": self.assertEqual(get_index_name(rasl_transformed), "index") rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertAlmostEqual( sk_transformed[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], msg=(row_idx, col_idx, tgt), ) def test_scale(self): (X_pd, _), _ = self.tgt2creditg["pandas"] sk_transformed = sk_scale(X_pd) for tgt, dataset in self.tgt2creditg.items(): (X, _), _ = dataset rasl_transformed = rasl_scale(X) if tgt == "spark": self.assertEqual(get_index_name(rasl_transformed), "index") rasl_transformed = _ensure_pandas(rasl_transformed) self.assertEqual(sk_transformed.shape, rasl_transformed.shape, tgt) for row_idx in range(sk_transformed.shape[0]): for col_idx in range(sk_transformed.shape[1]): self.assertAlmostEqual( sk_transformed[row_idx, col_idx], rasl_transformed.iloc[row_idx, col_idx], msg=(row_idx, col_idx, tgt), ) def test_predict(self): (train_X_pd, train_y_pd), (test_X_pd, _test_y_pd) = self.tgt2creditg["pandas"] to_pd = Convert(astype="pandas") lr = LogisticRegression() sk_trainable = SkStandardScaler() >> lr sk_trained = sk_trainable.fit(train_X_pd, train_y_pd) sk_predicted = sk_trained.predict(test_X_pd) rasl_trainable = RaslStandardScaler() >> to_pd >> lr for tgt, dataset in self.tgt2creditg.items(): (train_X, train_y), (test_X, _test_y) = dataset rasl_trained = rasl_trainable.fit(train_X, train_y) rasl_predicted = rasl_trained.predict(test_X) self.assertEqual(sk_predicted.shape, rasl_predicted.shape, tgt) self.assertEqual(sk_predicted.tolist(), rasl_predicted.tolist(), tgt) class _BatchTestingKFold: def __init__(self, n_batches, n_splits): self.n_batches = n_batches self.n_splits = n_splits def get_n_splits(self, X=None, y=None, groups=None): return self.n_splits def split(self, X, y=None, groups=None): # re-arrange batched data [[d0,e0,f0], [d1,e1,f1], [d2,e2,f2]] # into non-batched data [d0+d1+d2, e0+e1+e2, f0+f1+f2] result = [([], []) for _ in range(self.n_splits)] cv = KFold(self.n_splits) batches = mockup_data_loader(X, y, self.n_batches, "pandas") for bX, by in batches: for fold, (_, test) in enumerate(cv.split(bX, by)): remapped = bX.index[test] for f in range(self.n_splits): if f != fold: assert set(result[f][0]).isdisjoint(set(remapped)) result[f][0].extend(remapped) assert set(result[fold][1]).isdisjoint(set(remapped)) result[fold][1].extend(remapped) return result class _BatchTestingCallback: def __init__(self): self.n_calls = 0 def __call__( self, score_train, score_valid, n_batches_scanned, end_of_scanned_batches ): self.n_calls += 1 assert self.n_calls == n_batches_scanned, (self.n_calls, n_batches_scanned) assert not end_of_scanned_batches class TestTaskGraphs(unittest.TestCase): @classmethod def setUpClass(cls): X, y, fairness_info = lale.lib.aif360.fetch_creditg_df(preprocess=False) X = Project(columns=categorical()).fit(X).transform(X) fairness_info = { # remove numeric protected attribute age "favorable_labels": fairness_info["favorable_labels"], "protected_attributes": fairness_info["protected_attributes"][:1], } cls.creditg = X, y, fairness_info @classmethod def _make_sk_trainable(cls, final_est): if final_est == "sgd": est = SGDClassifier(random_state=97) elif final_est == "rfc": est = RandomForestClassifier(random_state=97) else: assert False, final_est return sk_make_pipeline( SkOrdinalEncoder(handle_unknown="use_encoded_value", unknown_value=-1), SkMinMaxScaler(), est, ) @classmethod def _make_rasl_trainable(cls, final_est): if final_est == "sgd": est = SGDClassifier(random_state=97) elif final_est == "rfc": est = RandomForestClassifier(random_state=97) else: assert False, final_est return ( RaslOrdinalEncoder(handle_unknown="use_encoded_value", unknown_value=-1) >> RaslMinMaxScaler() >> est ) def test_fit_no_batching(self): train_X, train_y, _ = self.creditg sk_trainable = self._make_sk_trainable("sgd") sk_trained = sk_trainable.fit(train_X, train_y) rasl_trainable = self._make_rasl_trainable("sgd") rasl_trained = rasl_trainable.fit(train_X, train_y) _check_trained_ordinal_encoder( self, sk_trained.steps[0][1], rasl_trained.steps[0][1].impl, "pandas" ) _check_trained_min_max_scaler( self, sk_trained.steps[1][1], rasl_trained.steps[1][1].impl, "pandas" ) def test_fit_batching(self): train_X, train_y, _ = self.creditg train_data_space = train_X.memory_usage().sum() + train_y.memory_usage() sk_trainable = self._make_sk_trainable("sgd") sk_trained = sk_trainable.fit(train_X, train_y) unique_class_labels = list(train_y.unique()) for n_batches in [1, 3]: for prio in [PrioStep(), PrioBatch()]: batches = mockup_data_loader(train_X, train_y, n_batches, "pandas") rasl_trainable = self._make_rasl_trainable("sgd") rasl_trained = fit_with_batches( pipeline=rasl_trainable, batches_train=batches, batches_valid=None, scoring=None, unique_class_labels=unique_class_labels, max_resident=3 * math.ceil(train_data_space / n_batches), prio=prio, partial_transform=False, verbose=0, progress_callback=None, ) _check_trained_ordinal_encoder( self, sk_trained.steps[0][1], rasl_trained.steps[0][1].impl, (n_batches, type(prio)), ) _check_trained_min_max_scaler( self, sk_trained.steps[1][1], rasl_trained.steps[1][1].impl, (n_batches, type(prio)), ) def test_partial_transform(self): train_X, train_y, _ = self.creditg unique_class_labels = list(train_y.unique()) for n_batches in [1, 3]: batches = mockup_data_loader(train_X, train_y, n_batches, "pandas") rasl_trainable = self._make_rasl_trainable("sgd") progress_callback = _BatchTestingCallback() _ = fit_with_batches( pipeline=rasl_trainable, batches_train=batches, batches_valid=None, scoring=rasl_get_scorer("accuracy"), unique_class_labels=unique_class_labels, max_resident=None, prio=PrioBatch(), partial_transform=True, verbose=0, progress_callback=progress_callback, ) self.assertEqual(progress_callback.n_calls, n_batches) def test_frozen_prefix(self): train_X, train_y, _ = self.creditg unique_class_labels = list(train_y.unique()) n_batches = 3 batches = mockup_data_loader(train_X, train_y, n_batches, "pandas") first_batch = next(iter(batches)) trainable1 = self._make_rasl_trainable("sgd") trained1 = fit_with_batches( pipeline=trainable1, batches_train=[first_batch], batches_valid=None, scoring=None, unique_class_labels=unique_class_labels, max_resident=None, prio=PrioBatch(), partial_transform=False, verbose=0, progress_callback=None, ) prefix2 = trained1.remove_last().freeze_trained() suffix2 = trained1.get_last() trainable2 = prefix2 >> suffix2 assert isinstance(trainable2, TrainedPipeline) progress_callback = _BatchTestingCallback() _ = fit_with_batches( pipeline=trainable2, batches_train=batches, batches_valid=None, scoring=rasl_get_scorer("accuracy"), unique_class_labels=unique_class_labels, max_resident=None, prio=PrioBatch(), partial_transform="score", verbose=0, progress_callback=progress_callback, ) self.assertEqual(progress_callback.n_calls, n_batches - 1) def test_cross_val_score_accuracy(self): X, y, _ = self.creditg n_splits = 3 for n_batches in [1, 3]: with self.assertWarnsRegex(DeprecationWarning, "trainable operator"): sk_scores = sk_cross_val_score( estimator=self._make_sk_trainable("rfc"), X=X, y=y, scoring=make_scorer(sk_accuracy_score), cv=_BatchTestingKFold(n_batches, n_splits), ) rasl_scores = rasl_cross_val_score( pipeline=self._make_rasl_trainable("rfc"), batches=mockup_data_loader(X, y, n_batches, "pandas"), scoring=rasl_get_scorer("accuracy"), cv=KFold(n_splits), unique_class_labels=list(y.unique()), max_resident=None, prio=PrioBatch(), same_fold=True, verbose=0, ) for sk_s, rasl_s in zip(sk_scores, rasl_scores): self.assertAlmostEqual(sk_s, rasl_s, msg=n_batches) def test_cross_val_score_disparate_impact(self): X, y, fairness_info = self.creditg disparate_impact_scorer = lale.lib.aif360.disparate_impact(**fairness_info) n_splits = 3 for n_batches in [1, 3]: with self.assertWarnsRegex(DeprecationWarning, "trainable operator"): sk_scores = sk_cross_val_score( estimator=self._make_sk_trainable("rfc"), X=X, y=y, scoring=disparate_impact_scorer, cv=_BatchTestingKFold(n_batches, n_splits), ) rasl_scores = rasl_cross_val_score( pipeline=self._make_rasl_trainable("rfc"), batches=mockup_data_loader(X, y, n_batches, "pandas"), scoring=disparate_impact_scorer, cv=KFold(n_splits), unique_class_labels=list(y.unique()), max_resident=None, prio=PrioBatch(), same_fold=True, verbose=0, ) for sk_s, rasl_s in zip(sk_scores, rasl_scores): self.assertAlmostEqual(sk_s, rasl_s, msg=n_batches) def test_cross_validate(self): X, y, _ = self.creditg n_splits = 3 for n_batches in [1, 3]: with self.assertWarnsRegex(DeprecationWarning, "trainable operator"): sk_scr = sk_cross_validate( estimator=self._make_sk_trainable("rfc"), X=X, y=y, scoring=make_scorer(sk_accuracy_score), cv=_BatchTestingKFold(n_batches, n_splits), return_estimator=True, ) rasl_scr = rasl_cross_validate( pipeline=self._make_rasl_trainable("rfc"), batches=mockup_data_loader(X, y, n_batches, "pandas"), scoring=rasl_get_scorer("accuracy"), cv=KFold(n_splits), unique_class_labels=list(y.unique()), max_resident=None, prio=PrioBatch(), same_fold=True, return_estimator=True, verbose=0, ) for sk_s, rasl_s in zip(sk_scr["test_score"], rasl_scr["test_score"]): self.assertAlmostEqual(cast(float, sk_s), cast(float, rasl_s)) for sk_e, rasl_e in zip(sk_scr["estimator"], rasl_scr["estimator"]): _check_trained_ordinal_encoder( self, sk_e.steps[0][1], cast(TrainedPipeline, rasl_e).steps[0][1].impl, n_batches, ) _check_trained_min_max_scaler( self, sk_e.steps[1][1], cast(TrainedPipeline, rasl_e).steps[1][1].impl, n_batches, ) class TestTaskGraphsWithConcat(unittest.TestCase): @classmethod def setUpClass(cls): (train_X, train_y), (test_X, test_y) = lale.datasets.openml.fetch( "credit-g", "classification", preprocess=False ) cls.cat_columns = [ "checking_status", "credit_history", "purpose", "savings_status", "employment", "personal_status", "other_parties", "property_magnitude", "other_payment_plans", "housing", "job", "own_telephone", "foreign_worker", ] cls.num_columns = [ "duration", "credit_amount", "installment_commitment", "residence_since", "age", "existing_credits", "num_dependents", ] cls.creditg = (train_X, train_y), (test_X, test_y) @classmethod def _make_sk_trainable(cls, final_est): from sklearn.compose import ColumnTransformer as SkColumnTransformer from sklearn.ensemble import RandomForestClassifier as SkRandomForestClassifier from sklearn.linear_model import SGDClassifier as SkSGDClassifier if final_est == "sgd": est = SkSGDClassifier(random_state=97) elif final_est == "rfc": est = SkRandomForestClassifier(random_state=97) else: assert False, final_est return sk_make_pipeline( SkColumnTransformer( [ ( "prep_cat", SkOrdinalEncoder( handle_unknown="use_encoded_value", unknown_value=-1 ), cls.cat_columns, ), ( "prep_num", sk_make_pipeline( SkSimpleImputer(strategy="mean"), SkMinMaxScaler(), "passthrough", ), cls.num_columns, ), ] ), est, ) @classmethod def _make_rasl_trainable(cls, final_est): if final_est == "sgd": est = SGDClassifier(random_state=97) elif final_est == "rfc": est = RandomForestClassifier(random_state=97) else: assert False, final_est return ( ( ( Project(columns=cls.cat_columns) >> RaslOrdinalEncoder( handle_unknown="use_encoded_value", unknown_value=-1 ) ) & ( Project(columns=cls.num_columns) >> RaslSimpleImputer(strategy="mean") >> RaslMinMaxScaler() ) ) >> ConcatFeatures() >> est ) def test_fit_no_batching(self): (train_X, train_y), _ = self.creditg sk_trainable = self._make_sk_trainable("sgd") sk_trained = sk_trainable.fit(train_X, train_y) rasl_trainable = self._make_rasl_trainable("sgd") rasl_trained = rasl_trainable.fit(train_X, train_y) _check_trained_ordinal_encoder( self, sk_trained.steps[0][1].transformers_[0][1], rasl_trained.steps[1][1].impl, "pandas", ) _check_trained_min_max_scaler( self, sk_trained.steps[0][1].transformers_[1][1].steps[1][1], rasl_trained.steps[4][1].impl, "pandas", ) def test_fit_batching(self): (train_X, train_y), _ = self.creditg train_data_space = train_X.memory_usage().sum() + train_y.memory_usage() sk_trainable = self._make_sk_trainable("sgd") sk_trained = sk_trainable.fit(train_X, train_y) unique_class_labels = list(train_y.unique()) for n_batches in [1, 3]: for prio in [PrioStep(), PrioBatch()]: batches = create_data_loader( train_X, train_y, math.ceil(len(train_y) / n_batches) ) self.assertEqual(n_batches, len(batches)) rasl_trainable = self._make_rasl_trainable("sgd") rasl_trained = fit_with_batches( pipeline=rasl_trainable, batches_train=batches, # type: ignore batches_valid=None, scoring=None, unique_class_labels=unique_class_labels, max_resident=3 * math.ceil(train_data_space / n_batches), prio=prio, partial_transform=False, verbose=0, progress_callback=None, ) _check_trained_ordinal_encoder( self, sk_trained.steps[0][1].transformers_[0][1], rasl_trained.steps[1][1].impl, (n_batches, type(prio)), ) _check_trained_min_max_scaler( self, sk_trained.steps[0][1].transformers_[1][1].steps[1][1], rasl_trained.steps[4][1].impl, (n_batches, type(prio)), ) def test_cross_val_score(self): (X, y), _ = self.creditg sk_scores = sk_cross_val_score( estimator=self._make_sk_trainable("rfc"), X=X, y=y, scoring=make_scorer(sk_accuracy_score), cv=KFold(3), ) for n_batches in [1, 3]: rasl_scores = rasl_cross_val_score( pipeline=self._make_rasl_trainable("rfc"), batches=mockup_data_loader(X, y, n_batches, "pandas"), scoring=rasl_get_scorer("accuracy"), cv=KFold(3), unique_class_labels=list(y.unique()), max_resident=None, prio=PrioBatch(), same_fold=True, verbose=0, ) if n_batches == 1: for sk_s, rasl_s in zip(sk_scores, rasl_scores): self.assertAlmostEqual(sk_s, rasl_s) class TestTaskGraphsWithCategoricalConcat(unittest.TestCase): @classmethod def setUpClass(cls): (train_X, train_y), (test_X, test_y) = lale.datasets.openml.fetch( "credit-g", "classification", preprocess=False ) cls.cat_columns = [ "checking_status", "credit_history", "purpose", "savings_status", "employment", "personal_status", "other_parties", "property_magnitude", "other_payment_plans", "housing", "job", "own_telephone", "foreign_worker", ] cls.num_columns = [ "duration", "credit_amount", "installment_commitment", "residence_since", "age", "existing_credits", "num_dependents", ] cls.creditg = (train_X, train_y), (test_X, test_y) @classmethod def _make_sk_trainable(cls, final_est): from sklearn.compose import ColumnTransformer as SkColumnTransformer from sklearn.ensemble import RandomForestClassifier as SkRandomForestClassifier from sklearn.linear_model import SGDClassifier as SkSGDClassifier if final_est == "sgd": est = SkSGDClassifier(random_state=97) elif final_est == "rfc": est = SkRandomForestClassifier(random_state=97) else: assert False, final_est return sk_make_pipeline( SkColumnTransformer( [ ( "prep_cat", SkOrdinalEncoder( handle_unknown="use_encoded_value", unknown_value=-1 ), cls.cat_columns, ), ( "prep_num", sk_make_pipeline( SkSimpleImputer(strategy="mean"), SkMinMaxScaler(), "passthrough", ), cls.num_columns, ), ] ), est, ) @classmethod def _make_rasl_trainable(cls, final_est): if final_est == "sgd": est = SGDClassifier(random_state=97) elif final_est == "rfc": est = RandomForestClassifier(random_state=97) else: assert False, final_est return ( ( ( Project(columns=categorical(11), drop_columns={"type": "number"}) >> RaslOrdinalEncoder( handle_unknown="use_encoded_value", unknown_value=-1 ) ) & ( Project(columns={"type": "number"}) >> RaslSimpleImputer(strategy="mean") >> RaslMinMaxScaler() ) ) >> ConcatFeatures() >> est ) def test_fit_no_batching(self): (train_X, train_y), _ = self.creditg sk_trainable = self._make_sk_trainable("sgd") sk_trained = sk_trainable.fit(train_X, train_y) rasl_trainable = self._make_rasl_trainable("sgd") rasl_trained = rasl_trainable.fit(train_X, train_y) _check_trained_ordinal_encoder( self, sk_trained.steps[0][1].transformers_[0][1], rasl_trained.steps[1][1].impl, "pandas", ) _check_trained_min_max_scaler( self, sk_trained.steps[0][1].transformers_[1][1].steps[1][1], rasl_trained.steps[4][1].impl, "pandas", ) def test_fit_batching(self): (train_X, train_y), _ = self.creditg train_data_space = train_X.memory_usage().sum() + train_y.memory_usage() sk_trainable = self._make_sk_trainable("sgd") sk_trained = sk_trainable.fit(train_X, train_y) unique_class_labels = list(train_y.unique()) for n_batches in [1, 3]: for prio in [PrioStep(), PrioBatch()]: batches = create_data_loader( train_X, train_y, math.ceil(len(train_y) / n_batches) ) self.assertEqual(n_batches, len(batches)) rasl_trainable = self._make_rasl_trainable("sgd") rasl_trained = fit_with_batches( pipeline=rasl_trainable, batches_train=batches, # type: ignore batches_valid=None, scoring=None, unique_class_labels=unique_class_labels, max_resident=3 * math.ceil(train_data_space / n_batches), prio=prio, partial_transform=False, verbose=0, progress_callback=None, ) _check_trained_ordinal_encoder( self, sk_trained.steps[0][1].transformers_[0][1], rasl_trained.steps[1][1].impl, (n_batches, type(prio)), ) _check_trained_min_max_scaler( self, sk_trained.steps[0][1].transformers_[1][1].steps[1][1], rasl_trained.steps[4][1].impl, (n_batches, type(prio)), ) def test_cross_val_score(self): (X, y), _ = self.creditg sk_scores = sk_cross_val_score( estimator=self._make_sk_trainable("rfc"), X=X, y=y, scoring=make_scorer(sk_accuracy_score), cv=KFold(3), ) for n_batches in [1, 3]: rasl_scores = rasl_cross_val_score( pipeline=self._make_rasl_trainable("rfc"), batches=mockup_data_loader(X, y, n_batches, "pandas"), scoring=rasl_get_scorer("accuracy"), cv=KFold(3), unique_class_labels=list(y.unique()), max_resident=None, prio=PrioBatch(), same_fold=True, verbose=0, ) if n_batches == 1: for sk_s, rasl_s in zip(sk_scores, rasl_scores): self.assertAlmostEqual(sk_s, rasl_s) class TestTaskGraphsSpark(unittest.TestCase): @classmethod def setUpClass(cls): X, y, _ = lale.lib.aif360.fetch_creditg_df(preprocess=False) X = Project(columns=categorical()).fit(X).transform(X) cls.creditg = X, y @classmethod def _make_sk_trainable(cls): return sk_make_pipeline( SkOrdinalEncoder(handle_unknown="use_encoded_value", unknown_value=-1), SkMinMaxScaler(), Convert(astype="pandas"), RandomForestClassifier(random_state=97), ) @classmethod def _make_rasl_trainable(cls): return ( RaslOrdinalEncoder(handle_unknown="use_encoded_value", unknown_value=-1) >> RaslMinMaxScaler() >> Convert(astype="pandas") >> RandomForestClassifier(random_state=97) ) def test_fit_with_batches(self): train_X, train_y = self.creditg sk_trainable = self._make_sk_trainable() sk_trained = sk_trainable.fit(train_X, train_y) unique_class_labels = list(train_y.unique()) datatype_list: List[datatype_param_type] = ["pandas", "spark"] for tgt, n_batches in itertools.product(datatype_list, [1, 3]): rasl_trained = fit_with_batches( pipeline=self._make_rasl_trainable(), batches_train=mockup_data_loader(train_X, train_y, n_batches, tgt), batches_valid=None, scoring=None, unique_class_labels=unique_class_labels, max_resident=None, prio=PrioBatch(), partial_transform=False, verbose=0, progress_callback=None, ) _check_trained_ordinal_encoder( self, sk_trained.steps[0][1], rasl_trained.steps[0][1].impl, tgt ) _check_trained_min_max_scaler( self, sk_trained.steps[1][1], rasl_trained.steps[1][1].impl, tgt ) def test_cross_val_score(self): X, y = self.creditg with self.assertWarnsRegex(DeprecationWarning, "trainable operator"): sk_scores = sk_cross_val_score( estimator=self._make_sk_trainable(), X=X, y=y, scoring=make_scorer(sk_accuracy_score), cv=KFold(3), ) unique_class_labels = list(y.unique()) datatype_list: List[datatype_param_type] = ["pandas", "spark"] for tgt, n_batches in itertools.product(datatype_list, [1, 3]): rasl_scores = rasl_cross_val_score( pipeline=self._make_rasl_trainable(), batches=mockup_data_loader(X, y, n_batches, tgt), scoring=rasl_get_scorer("accuracy"), cv=KFold(3), unique_class_labels=unique_class_labels, max_resident=None, prio=PrioBatch(), same_fold=True, verbose=0, ) if n_batches == 1: for sk_s, rasl_s in zip(sk_scores, rasl_scores): self.assertAlmostEqual(sk_s, rasl_s, msg=(tgt, n_batches)) def test_cross_validate(self): X, y = self.creditg with self.assertWarnsRegex(DeprecationWarning, "trainable operator"): sk_scores = sk_cross_validate( estimator=self._make_sk_trainable(), X=X, y=y, scoring=make_scorer(sk_accuracy_score), cv=KFold(3), return_estimator=True, ) unique_class_labels = list(y.unique()) datatype_list: List[datatype_param_type] = ["pandas", "spark"] for tgt, n_batches in itertools.product(datatype_list, [1, 3]): rasl_scores = rasl_cross_validate( pipeline=self._make_rasl_trainable(), batches=mockup_data_loader(X, y, n_batches, tgt), scoring=rasl_get_scorer("accuracy"), cv=KFold(3), unique_class_labels=unique_class_labels, max_resident=None, prio=PrioBatch(), same_fold=True, return_estimator=True, verbose=0, ) msg = tgt, n_batches if n_batches == 1: for sk_s, rasl_s in zip( sk_scores["test_score"], rasl_scores["test_score"] ): self.assertAlmostEqual( cast(float, sk_s), cast(float, rasl_s), msg=msg ) for sk_e, rasl_e in zip( sk_scores["estimator"], rasl_scores["estimator"] ): rasl_steps = cast(TrainedPipeline, rasl_e).steps _check_trained_ordinal_encoder( self, sk_e.steps[0][1], rasl_steps[0][1].impl, msg=msg, ) _check_trained_min_max_scaler( self, sk_e.steps[1][1], rasl_steps[1][1].impl, msg=msg, ) class TestMetrics(unittest.TestCase): @classmethod def setUpClass(cls): cls.creditg = lale.datasets.openml.fetch( "credit-g", "classification", preprocess=True, astype="pandas" ) def test_accuracy(self): (train_X, train_y), (test_X, test_y) = self.creditg est = LogisticRegression().fit(train_X, train_y) y_pred = est.predict(test_X) sk_score = sk_accuracy_score(test_y, y_pred) self.assertAlmostEqual(sk_score, rasl_accuracy_score(test_y, y_pred)) rasl_scorer = rasl_get_scorer("accuracy") self.assertAlmostEqual(sk_score, rasl_scorer(est, test_X, test_y)) batches = mockup_data_loader(test_X, test_y, 3, "pandas") self.assertAlmostEqual( sk_score, rasl_scorer.score_estimator_batched(est, batches) ) def test_balanced_accuracy(self): (train_X, train_y), (test_X, test_y) = self.creditg est = LogisticRegression().fit(train_X, train_y) y_pred = est.predict(test_X) sk_score = sk_balanced_accuracy_score(test_y, y_pred) self.assertAlmostEqual(sk_score, rasl_balanced_accuracy_score(test_y, y_pred)) rasl_scorer = rasl_get_scorer("balanced_accuracy") self.assertAlmostEqual(sk_score, rasl_scorer(est, test_X, test_y)) batches = mockup_data_loader(test_X, test_y, 3, "pandas") self.assertAlmostEqual( sk_score, rasl_scorer.score_estimator_batched(est, batches) ) def test_f1(self): (train_X, train_y), (test_X, test_y) = self.creditg est = LogisticRegression().fit(train_X, train_y) y_pred = est.predict(test_X) sk_score = sk_f1_score(test_y, y_pred, pos_label=1) self.assertAlmostEqual(sk_score, rasl_f1_score(test_y, y_pred, pos_label=1)) rasl_scorer = rasl_get_scorer("f1", pos_label=1) self.assertAlmostEqual(sk_score, rasl_scorer(est, test_X, test_y)) batches = mockup_data_loader(test_X, test_y, 3, "pandas") self.assertAlmostEqual( sk_score, rasl_scorer.score_estimator_batched(est, batches) ) def test_r2_score(self): (train_X, train_y), (test_X, test_y) = self.creditg est = LogisticRegression().fit(train_X, train_y) y_pred = est.predict(test_X) sk_score = sk_r2_score(test_y, y_pred) self.assertAlmostEqual(sk_score, rasl_r2_score(test_y, y_pred)) rasl_scorer = rasl_get_scorer("r2") self.assertAlmostEqual(sk_score, rasl_scorer(est, test_X, test_y)) batches = mockup_data_loader(test_X, test_y, 3, "pandas") self.assertAlmostEqual( sk_score, rasl_scorer.score_estimator_batched(est, batches) ) class TestBatchedBaggingClassifier(unittest.TestCase): @classmethod def setUpClass(cls): cls.creditg = lale.datasets.openml.fetch( "credit-g", "classification", preprocess=False, astype="pandas" ) @classmethod def _make_sk_trainable(cls): from sklearn.tree import DecisionTreeClassifier as SkDecisionTreeClassifier return sk_make_pipeline( SkOrdinalEncoder(handle_unknown="use_encoded_value", unknown_value=-1), SkMinMaxScaler(), SkDecisionTreeClassifier(random_state=97, max_features="auto"), ) @classmethod def _make_rasl_trainable(cls, final_est): if final_est == "bagging_monoid": est = BatchedBaggingClassifier( base_estimator=DecisionTreeClassifier( random_state=97, max_features="auto" ) ) else: assert False, final_est return ( RaslOrdinalEncoder(handle_unknown="use_encoded_value", unknown_value=-1) >> RaslMinMaxScaler() >> est ) def test_classifier(self): (X_train, y_train), (X_test, _y_test) = self.creditg import warnings clf = BatchedBaggingClassifier() # test_schemas_are_schemas lale.type_checking.validate_is_schema(clf.input_schema_fit()) lale.type_checking.validate_is_schema(clf.input_schema_predict()) lale.type_checking.validate_is_schema(clf.output_schema_predict()) lale.type_checking.validate_is_schema(clf.hyperparam_schema()) # test_init_fit_predict pipeline = self._make_rasl_trainable("bagging_monoid") trained = pipeline.fit(X_train, y_train) _ = trained.predict(X_test) # test_with_hyperopt from lale.lib.lale import Hyperopt (X_train, y_train), (X_test, _) = lale.datasets.openml.fetch( "credit-g", "classification", preprocess=True, astype="pandas" ) hyperopt = Hyperopt(estimator=pipeline, max_evals=1, verbose=True) trained = hyperopt.fit(X_train, y_train) _ = trained.predict(X_test) # test_cross_validation from lale.helpers import cross_val_score cv_results = cross_val_score(pipeline, X_train, y_train, cv=2) self.assertEqual(len(cv_results), 2) # test_with_gridsearchcv_auto_wrapped with warnings.catch_warnings(): warnings.simplefilter("ignore") from lale.lib.lale import GridSearchCV grid_search = GridSearchCV( estimator=pipeline, lale_num_samples=1, lale_num_grids=1, cv=2, scoring=make_scorer(sk_accuracy_score), ) grid_search.fit(X_train, y_train) # test_predict_on_trainable trained = clf.fit(X_train, y_train) clf.predict(X_train) # test_to_json clf.to_json() def test_fit_batching(self): (train_X, train_y), (test_X, test_y) = self.creditg train_data_space = train_X.memory_usage().sum() + train_y.memory_usage() unique_class_labels = list(train_y.unique()) for n_batches in [1, 3]: for prio in [PrioStep(), PrioBatch()]: batches = mockup_data_loader(train_X, train_y, n_batches, "pandas") rasl_trainable = self._make_rasl_trainable("bagging_monoid") rasl_trained = fit_with_batches( pipeline=rasl_trainable, batches_train=batches, batches_valid=None, scoring=None, unique_class_labels=unique_class_labels, max_resident=3 * math.ceil(train_data_space / n_batches), prio=prio, partial_transform=False, verbose=0, progress_callback=None, ) predictions = rasl_trained.predict(test_X) rasl_acc = sk_accuracy_score(test_y, predictions) if n_batches == 1: sk_pipeline = self._make_sk_trainable() sk_pipeline.fit(train_X, train_y) predictions = sk_pipeline.predict(test_X) sk_acc = sk_accuracy_score(test_y, predictions) self.assertEqual(rasl_acc, sk_acc) def test_cross_val_score(self): (train_X, train_y), (_, _) = self.creditg for n_batches in [1, 3]: _ = rasl_cross_val_score( pipeline=self._make_rasl_trainable("bagging_monoid"), batches=mockup_data_loader(train_X, train_y, n_batches, "pandas"), scoring=rasl_get_scorer("accuracy"), cv=KFold(3), unique_class_labels=list(train_y.unique()), max_resident=None, prio=PrioBatch(), same_fold=True, verbose=0, ) class TestXGBoost(unittest.TestCase): def test_partial_fit_xgb_classifier(self): X, y = sklearn.datasets.load_iris(return_X_y=True, as_frame=True) est = XGBClassifier(verbosity=0) for bX, by in mockup_data_loader(X, y, 3, "pandas"): est = est.partial_fit(bX, by) _ = est.predict(bX) def test_partial_fit_xgb_regressor(self): X, y = sklearn.datasets.load_diabetes(return_X_y=True, as_frame=True) est = XGBRegressor(verbosity=0) for bX, by in mockup_data_loader(X, y, 3, "pandas"): est = est.partial_fit(bX, by) _ = est.predict(bX) class TestLightGBM(unittest.TestCase): def test_partial_fit_lgbm_classifier(self): X, y = sklearn.datasets.load_iris(return_X_y=True, as_frame=True) est = LGBMClassifier() for bX, by in mockup_data_loader(X, y, 3, "pandas"): est = est.partial_fit(bX, by) _ = est.predict(bX) def test_partial_fit_lgbm_regressor(self): X, y = sklearn.datasets.load_diabetes(return_X_y=True, as_frame=True) est = LGBMRegressor() for bX, by in mockup_data_loader(X, y, 3, "pandas"): est = est.partial_fit(bX, by) _ = est.predict(bX)
103,433
42.29594
162
py
lale
lale-master/test/test_json_pretty_viz.py
# Copyright 2019-2023 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import unittest import lale.operators import lale.pretty_print class TestToGraphviz(unittest.TestCase): def test_with_operator_choice(self): from lale.lib.lale import NoOp from lale.lib.sklearn import ( PCA, KNeighborsClassifier, LogisticRegression, Nystroem, ) from lale.operators import make_choice kernel_tfm_or_not = NoOp | Nystroem tfm = PCA clf = make_choice(LogisticRegression, KNeighborsClassifier) clf.visualize(ipython_display=False) optimizable = kernel_tfm_or_not >> tfm >> clf optimizable.visualize(ipython_display=False) def test_invalid_input(self): from sklearn.linear_model import LogisticRegression as SklearnLR scikit_lr = SklearnLR() from lale.helpers import to_graphviz with self.assertRaises(TypeError): to_graphviz(scikit_lr) class TestPrettyPrint(unittest.TestCase): # pylint:disable=reimported,redefined-outer-name def _roundtrip(self, expected, printed): self.maxDiff = None # sklearn_version_family changes based on the Python as well as sklearn version, # so remove that hyperparameter while comparing if present. import re expected = re.sub(r"""sklearn_version_family=.\d*.,""", "", expected) printed = re.sub(r"""sklearn_version_family=.\d*.,""", "", printed) self.assertEqual(expected, printed) globals2 = {} locals2 = {} try: exec(printed, globals2, locals2) except Exception as e: import pprint print("error during exec(printed, globals2, locals2) where:") print(f'printed = """{printed}"""') print(f"globals2 = {pprint.pformat(globals2)}") print(f"locals2 = {pprint.pformat(locals2)}") raise e pipeline2 = locals2["pipeline"] import sklearn.pipeline self.assertIsInstance( pipeline2, (lale.operators.PlannedOperator, sklearn.pipeline.Pipeline) ) def test_distance_threshold_validation_error(self): import jsonschema from lale.lib.sklearn import FeatureAgglomeration, LogisticRegression with self.assertRaises(jsonschema.ValidationError): _ = ( FeatureAgglomeration( distance_threshold=0.5, n_clusters=None, compute_full_tree=True ) >> LogisticRegression() ) def test_indiv_op_1(self): from lale.lib.sklearn import LogisticRegression pipeline = LogisticRegression(solver=LogisticRegression.enum.solver.saga, C=0.9) expected = """from sklearn.linear_model import LogisticRegression import lale lale.wrap_imported_operators() pipeline = LogisticRegression(solver="saga", C=0.9)""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_indiv_op_2(self): from lale.lib.sklearn import LogisticRegression pipeline = LogisticRegression() expected = """from sklearn.linear_model import LogisticRegression import lale lale.wrap_imported_operators() pipeline = LogisticRegression()""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_reducible(self): from lale.lib.lale import NoOp from lale.lib.rasl import ConcatFeatures from lale.lib.sklearn import ( PCA, KNeighborsClassifier, LogisticRegression, MinMaxScaler, Nystroem, ) from lale.lib.xgboost import XGBClassifier as XGB pca = PCA(copy=False) logistic_regression = LogisticRegression(solver="saga", C=0.9) pipeline = ( (MinMaxScaler | NoOp) >> (pca & Nystroem) >> ConcatFeatures >> (KNeighborsClassifier | logistic_regression | XGB) ) expected = """from sklearn.preprocessing import MinMaxScaler from lale.lib.lale import NoOp from sklearn.decomposition import PCA from sklearn.kernel_approximation import Nystroem from lale.lib.rasl import ConcatFeatures from sklearn.neighbors import KNeighborsClassifier from sklearn.linear_model import LogisticRegression from xgboost import XGBClassifier as XGB import lale lale.wrap_imported_operators() pca = PCA(copy=False) logistic_regression = LogisticRegression(solver="saga", C=0.9) pipeline = ( (MinMaxScaler | NoOp) >> (pca & Nystroem) >> ConcatFeatures >> (KNeighborsClassifier | logistic_regression | XGB) )""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_no_combinators(self): from lale.lib.lale import NoOp from lale.lib.rasl import ConcatFeatures from lale.lib.sklearn import ( PCA, KNeighborsClassifier, LogisticRegression, MinMaxScaler, Nystroem, ) pca = PCA(copy=False) logistic_regression = LogisticRegression(solver="saga", C=0.9) pipeline = ( (MinMaxScaler | NoOp) >> (pca & Nystroem & NoOp) >> ConcatFeatures >> (KNeighborsClassifier | logistic_regression) ) expected = """from sklearn.preprocessing import MinMaxScaler from lale.lib.lale import NoOp from lale.operators import make_choice from sklearn.decomposition import PCA from sklearn.kernel_approximation import Nystroem from lale.operators import make_union from sklearn.neighbors import KNeighborsClassifier from sklearn.linear_model import LogisticRegression from lale.operators import make_pipeline choice_0 = make_choice(MinMaxScaler, NoOp) pca = PCA(copy=False) union = make_union(pca, Nystroem, NoOp) logistic_regression = LogisticRegression(solver="saga", C=0.9) choice_1 = make_choice(KNeighborsClassifier, logistic_regression) pipeline = make_pipeline(choice_0, union, choice_1)""" printed = lale.pretty_print.to_string(pipeline, combinators=False) self._roundtrip(expected, printed) def test_astype_sklearn(self): from lale.lib.rasl import ConcatFeatures from lale.lib.sklearn import PCA, LogisticRegression, MinMaxScaler, Nystroem pca = PCA(copy=False) logistic_regression = LogisticRegression(solver="saga", C=0.9) pipeline = ( MinMaxScaler() >> (pca & Nystroem()) >> ConcatFeatures >> logistic_regression ) expected = """from sklearn.preprocessing import MinMaxScaler from sklearn.decomposition import PCA from sklearn.kernel_approximation import Nystroem from sklearn.pipeline import make_union from sklearn.linear_model import LogisticRegression from sklearn.pipeline import make_pipeline pca = PCA(copy=False) union = make_union(pca, Nystroem()) logistic_regression = LogisticRegression(solver="saga", C=0.9) pipeline = make_pipeline(MinMaxScaler(), union, logistic_regression)""" printed = lale.pretty_print.to_string(pipeline, astype="sklearn") self._roundtrip(expected, printed) def test_import_as_1(self): from lale.lib.sklearn import LogisticRegression as LR pipeline = LR(solver="saga", C=0.9) expected = """from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() pipeline = LR(solver="saga", C=0.9)""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_import_as_2(self): from lale.lib.lale import NoOp from lale.lib.rasl import ConcatFeatures as Concat from lale.lib.sklearn import PCA from lale.lib.sklearn import KNeighborsClassifier as KNN from lale.lib.sklearn import LogisticRegression as LR from lale.lib.sklearn import MinMaxScaler as Scaler from lale.lib.sklearn import Nystroem pca = PCA(copy=False) lr = LR(solver="saga", C=0.9) pipeline = (Scaler | NoOp) >> (pca & Nystroem) >> Concat >> (KNN | lr) expected = """from sklearn.preprocessing import MinMaxScaler as Scaler from lale.lib.lale import NoOp from sklearn.decomposition import PCA from sklearn.kernel_approximation import Nystroem from lale.lib.rasl import ConcatFeatures as Concat from sklearn.neighbors import KNeighborsClassifier as KNN from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() pca = PCA(copy=False) lr = LR(solver="saga", C=0.9) pipeline = (Scaler | NoOp) >> (pca & Nystroem) >> Concat >> (KNN | lr)""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_operator_choice(self): from lale.lib.sklearn import PCA from lale.lib.sklearn import MinMaxScaler as Scl pipeline = PCA | Scl expected = """from sklearn.decomposition import PCA from sklearn.preprocessing import MinMaxScaler as Scl import lale lale.wrap_imported_operators() pipeline = PCA | Scl""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_higher_order(self): from lale.lib.lale import Both from lale.lib.sklearn import PCA, Nystroem pipeline = Both(op1=PCA(n_components=2), op2=Nystroem) expected = """from lale.lib.lale import Both from sklearn.decomposition import PCA from sklearn.kernel_approximation import Nystroem import lale lale.wrap_imported_operators() pca = PCA(n_components=2) pipeline = Both(op1=pca, op2=Nystroem)""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_higher_order_2(self): from lale.lib.sklearn import PCA from lale.lib.sklearn import KNeighborsClassifier as KNN from lale.lib.sklearn import LogisticRegression as LR from lale.lib.sklearn import VotingClassifier as Vote pipeline = Vote( estimators=[("knn", KNN), ("pipeline", PCA() >> LR)], voting="soft" ) expected = """from sklearn.ensemble import VotingClassifier as Vote from sklearn.neighbors import KNeighborsClassifier as KNN from sklearn.decomposition import PCA from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() pipeline = Vote( estimators=[("knn", KNN), ("pipeline", PCA() >> LR)], voting="soft" )""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_multimodal(self): from lale.lib.rasl import ConcatFeatures as Cat from lale.lib.rasl import Project from lale.lib.sklearn import LinearSVC from lale.lib.sklearn import Normalizer as Norm from lale.lib.sklearn import OneHotEncoder as OneHot project_0 = Project(columns={"type": "number"}) project_1 = Project(columns={"type": "string"}) linear_svc = LinearSVC(C=29617.4, dual=False, tol=0.005266) pipeline = ( ((project_0 >> Norm()) & (project_1 >> OneHot())) >> Cat >> linear_svc ) expected = """from lale.lib.rasl import Project from sklearn.preprocessing import Normalizer as Norm from sklearn.preprocessing import OneHotEncoder as OneHot from lale.lib.rasl import ConcatFeatures as Cat from sklearn.svm import LinearSVC import lale lale.wrap_imported_operators() project_0 = Project(columns={"type": "number"}) project_1 = Project(columns={"type": "string"}) linear_svc = LinearSVC(C=29617.4, dual=False, tol=0.005266) pipeline = ( ((project_0 >> Norm()) & (project_1 >> OneHot())) >> Cat >> linear_svc )""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_irreducible_1(self): from lale.lib.sklearn import ( PCA, KNeighborsClassifier, LogisticRegression, MinMaxScaler, Nystroem, ) from lale.operators import make_pipeline_graph choice = PCA | Nystroem pipeline = make_pipeline_graph( steps=[choice, MinMaxScaler, LogisticRegression, KNeighborsClassifier], edges=[ (choice, LogisticRegression), (MinMaxScaler, LogisticRegression), (MinMaxScaler, KNeighborsClassifier), ], ) expected = """from sklearn.decomposition import PCA from sklearn.kernel_approximation import Nystroem from sklearn.preprocessing import MinMaxScaler from sklearn.linear_model import LogisticRegression from sklearn.neighbors import KNeighborsClassifier from lale.operators import make_pipeline_graph import lale lale.wrap_imported_operators() choice = PCA | Nystroem pipeline = make_pipeline_graph( steps=[choice, MinMaxScaler, LogisticRegression, KNeighborsClassifier], edges=[ (choice, LogisticRegression), (MinMaxScaler, LogisticRegression), (MinMaxScaler, KNeighborsClassifier), ], )""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_irreducible_2(self): from lale.lib.rasl import ConcatFeatures as HStack from lale.lib.sklearn import PCA from lale.lib.sklearn import KNeighborsClassifier as KNN from lale.lib.sklearn import LogisticRegression as LR from lale.lib.sklearn import MinMaxScaler as MMS from lale.operators import make_pipeline_graph pipeline_0 = HStack >> LR pipeline = make_pipeline_graph( steps=[PCA, MMS, KNN, pipeline_0], edges=[(PCA, KNN), (PCA, pipeline_0), (MMS, pipeline_0)], ) expected = """from sklearn.decomposition import PCA from sklearn.preprocessing import MinMaxScaler as MMS from sklearn.neighbors import KNeighborsClassifier as KNN from lale.lib.rasl import ConcatFeatures as HStack from sklearn.linear_model import LogisticRegression as LR from lale.operators import make_pipeline_graph import lale lale.wrap_imported_operators() pipeline_0 = HStack >> LR pipeline = make_pipeline_graph( steps=[PCA, MMS, KNN, pipeline_0], edges=[(PCA, KNN), (PCA, pipeline_0), (MMS, pipeline_0)], )""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_nested(self): from lale.lib.lale import NoOp from lale.lib.sklearn import PCA from lale.lib.sklearn import LogisticRegression as LR lr_0 = LR(C=0.09) lr_1 = LR(C=0.19) pipeline = PCA >> (lr_0 | NoOp >> lr_1) expected = """from sklearn.decomposition import PCA from sklearn.linear_model import LogisticRegression as LR from lale.lib.lale import NoOp import lale lale.wrap_imported_operators() lr_0 = LR(C=0.09) lr_1 = LR(C=0.19) pipeline = PCA >> (lr_0 | NoOp >> lr_1)""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_autoai_libs_cat_encoder(self): import numpy as np from autoai_libs.transformers.exportable import CatEncoder from lale.lib.sklearn import LogisticRegression as LR cat_encoder = CatEncoder( encoding="ordinal", categories="auto", dtype=np.float64, handle_unknown="error", ) pipeline = cat_encoder >> LR() expected = """from autoai_libs.transformers.exportable import CatEncoder import numpy as np from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() cat_encoder = CatEncoder( encoding="ordinal", categories="auto", dtype=np.float64, handle_unknown="error", sklearn_version_family="23", ) pipeline = cat_encoder >> LR()""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_autoai_libs_cat_encoder_defaults(self): import numpy as np from lale.lib.autoai_libs import CatEncoder from lale.lib.sklearn import LogisticRegression as LR cat_encoder = CatEncoder( dtype=np.float64, handle_unknown="error", sklearn_version_family="1" ) pipeline = cat_encoder >> LR() expected = """from autoai_libs.transformers.exportable import CatEncoder import numpy as np from sklearn.linear_model import LogisticRegression as LR from sklearn.pipeline import make_pipeline cat_encoder = CatEncoder( dtype=np.float64, handle_unknown="error", sklearn_version_family="1", encoding="ordinal", categories="auto", ) pipeline = make_pipeline(cat_encoder, LR())""" self._roundtrip( expected, lale.pretty_print.to_string(pipeline, astype="sklearn") ) def test_autoai_libs_fs1(self): from autoai_libs.cognito.transforms.transform_utils import FS1 from lale.lib.sklearn import LogisticRegression as LR fs1 = FS1( cols_ids_must_keep=range(0, 7), additional_col_count_to_keep=8, ptype="classification", ) pipeline = fs1 >> LR() expected = """from autoai_libs.cognito.transforms.transform_utils import FS1 from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() fs1 = FS1( cols_ids_must_keep=range(0, 7), additional_col_count_to_keep=8, ptype="classification", ) pipeline = fs1 >> LR()""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_autoai_libs_numpy_replace_missing_values(self): from autoai_libs.transformers.exportable import NumpyReplaceMissingValues from lale.lib.sklearn import LogisticRegression as LR numpy_replace_missing_values = NumpyReplaceMissingValues( filling_values=float("nan"), missing_values=["?"] ) pipeline = numpy_replace_missing_values >> LR() expected = """from autoai_libs.transformers.exportable import NumpyReplaceMissingValues from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() numpy_replace_missing_values = NumpyReplaceMissingValues( missing_values=["?"], filling_values=float("nan") ) pipeline = numpy_replace_missing_values >> LR()""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_autoai_libs_numpy_replace_unknown_values1(self): from autoai_libs.transformers.exportable import NumpyReplaceUnknownValues from lale.lib.sklearn import LogisticRegression as LR numpy_replace_unknown_values = NumpyReplaceUnknownValues( filling_values=float("nan"), filling_values_list=[float("nan")], known_values_list=[[36, 45, 56, 67, 68, 75, 78, 89]], missing_values_reference_list=["", "-", "?", float("nan")], ) pipeline = numpy_replace_unknown_values >> LR() expected = """from autoai_libs.transformers.exportable import NumpyReplaceUnknownValues from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() numpy_replace_unknown_values = NumpyReplaceUnknownValues( filling_values=float("nan"), filling_values_list=[float("nan")], missing_values_reference_list=["", "-", "?", float("nan")], ) pipeline = numpy_replace_unknown_values >> LR()""" try: self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) except BaseException: expected = """from autoai_libs.transformers.exportable import NumpyReplaceUnknownValues from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() numpy_replace_unknown_values = NumpyReplaceUnknownValues( filling_values=float("nan"), filling_values_list=[float("nan")], missing_values_reference_list=("", "-", "?", float("nan")), ) pipeline = numpy_replace_unknown_values >> LR()""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_autoai_libs_numpy_replace_unknown_values2(self): from lale.lib.autoai_libs import NumpyReplaceUnknownValues from lale.lib.sklearn import LogisticRegression as LR CustomOp = NumpyReplaceUnknownValues.customize_schema( known_values_list={ "anyOf": [ {"type": "array", "items": {"laleType": "Any"}}, {"enum": [None]}, ], "default": None, } ) numpy_replace_unknown_values = CustomOp( filling_values=float("nan"), filling_values_list=[float("nan")], known_values_list=[[36, 45, 56, 67, 68, 75, 78, 89]], missing_values_reference_list=["", "-", "?", float("nan")], ) pipeline = numpy_replace_unknown_values >> LR() expected = """from autoai_libs.transformers.exportable import ( NumpyReplaceUnknownValues as CustomOp, ) from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() custom_op = CustomOp( filling_values=float("nan"), filling_values_list=[float("nan")], known_values_list=[[36, 45, 56, 67, 68, 75, 78, 89]], missing_values_reference_list=["", "-", "?", float("nan")], ) pipeline = custom_op >> LR()""" try: self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) except BaseException: expected = """from autoai_libs.transformers.exportable import ( NumpyReplaceUnknownValues as CustomOp, ) from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() custom_op = CustomOp( filling_values=float("nan"), filling_values_list=[float("nan")], known_values_list=[[36, 45, 56, 67, 68, 75, 78, 89]], missing_values_reference_list=("", "-", "?", float("nan")), ) pipeline = custom_op >> LR()""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_autoai_libs_tam_1(self): import autoai_libs.cognito.transforms.transform_extras import numpy as np from autoai_libs.cognito.transforms.transform_utils import TAM from lale.lib.sklearn import LogisticRegression as LR tam = TAM( tans_class=autoai_libs.cognito.transforms.transform_extras.IsolationForestAnomaly, name="isoforestanomaly", col_names=["a", "b", "c"], col_dtypes=[np.dtype("float32"), np.dtype("float32"), np.dtype("float32")], ) pipeline = tam >> LR() expected = """from autoai_libs.cognito.transforms.transform_utils import TAM import autoai_libs.cognito.transforms.transform_extras import numpy as np from sklearn.linear_model import LogisticRegression as LR from sklearn.pipeline import make_pipeline tam = TAM( tans_class=autoai_libs.cognito.transforms.transform_extras.IsolationForestAnomaly, name="isoforestanomaly", col_names=["a", "b", "c"], col_dtypes=[ np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), ], ) pipeline = make_pipeline(tam, LR())""" self._roundtrip( expected, lale.pretty_print.to_string(pipeline, astype="sklearn") ) def test_autoai_libs_tam_2(self): import numpy as np from lightgbm import LGBMClassifier from sklearn.decomposition import PCA from lale.lib.autoai_libs import TAM from lale.operators import make_pipeline pca = PCA(copy=False) tam = TAM( tans_class=pca, name="pca", col_names=["a", "b", "c"], col_dtypes=[np.dtype("float32"), np.dtype("float32"), np.dtype("float32")], ) lgbm_classifier = LGBMClassifier(class_weight="balanced", learning_rate=0.18) pipeline = make_pipeline(tam, lgbm_classifier) expected = """from autoai_libs.cognito.transforms.transform_utils import TAM import sklearn.decomposition import numpy as np from lightgbm import LGBMClassifier from lale.operators import make_pipeline tam = TAM( tans_class=sklearn.decomposition.PCA(copy=False), name="pca", col_names=["a", "b", "c"], col_dtypes=[ np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), ], ) lgbm_classifier = LGBMClassifier( class_weight="balanced", learning_rate=0.18, n_estimators=100 ) pipeline = make_pipeline(tam, lgbm_classifier)""" self._roundtrip( expected, lale.pretty_print.to_string(pipeline, combinators=False) ) def test_autoai_libs_tam_3(self): import autoai_libs.cognito.transforms.transform_utils import numpy as np import sklearn.cluster import sklearn.linear_model import sklearn.pipeline import lale.helpers import lale.operators import lale.pretty_print sklearn_pipeline = sklearn.pipeline.make_pipeline( autoai_libs.cognito.transforms.transform_utils.TAM( tans_class=sklearn.cluster.FeatureAgglomeration( compute_full_tree="auto", connectivity=None, linkage="ward", memory=None, n_clusters=2, pooling_func=np.mean, ), name="featureagglomeration", col_names=["a", "b", "c"], col_dtypes=[ np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), ], ), sklearn.linear_model.LogisticRegression( solver="liblinear", multi_class="ovr" ), ) pipeline = lale.helpers.import_from_sklearn_pipeline(sklearn_pipeline) expected = """from autoai_libs.cognito.transforms.transform_utils import TAM from sklearn.cluster import FeatureAgglomeration import numpy as np from sklearn.linear_model import LogisticRegression import lale lale.wrap_imported_operators() tam = TAM( tans_class=FeatureAgglomeration(), name="featureagglomeration", col_names=["a", "b", "c"], col_dtypes=[ np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), ], ) logistic_regression = LogisticRegression( multi_class="ovr", solver="liblinear" ) pipeline = tam >> logistic_regression""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_autoai_libs_tam_4(self): import autoai_libs.cognito.transforms.transform_utils import numpy as np import sklearn.decomposition import sklearn.linear_model import sklearn.pipeline import lale.helpers import lale.operators import lale.pretty_print sklearn_pipeline = sklearn.pipeline.make_pipeline( autoai_libs.cognito.transforms.transform_utils.TAM( tans_class=sklearn.decomposition.PCA(), name="pca", col_names=["a", "b", "c"], col_dtypes=[ np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), ], ), sklearn.linear_model.LogisticRegression( solver="liblinear", multi_class="ovr" ), ) pipeline = lale.helpers.import_from_sklearn_pipeline( sklearn_pipeline, fitted=False ) assert isinstance(pipeline, lale.operators.TrainableOperator) expected = """from autoai_libs.cognito.transforms.transform_utils import TAM from sklearn.decomposition import PCA import numpy as np from sklearn.linear_model import LogisticRegression import lale lale.wrap_imported_operators() tam = TAM( tans_class=PCA(), name="pca", col_names=["a", "b", "c"], col_dtypes=[ np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), ], ) logistic_regression = LogisticRegression( multi_class="ovr", solver="liblinear" ) pipeline = tam >> logistic_regression""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) import numpy as np import pandas as pd test = pd.DataFrame( np.random.randint(0, 100, size=(15, 3)), columns=["a", "b", "c"], dtype=np.dtype("float32"), ) trained = pipeline.fit( test.to_numpy(), [0, 1, 1, 0, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 1] ) trained.predict(test.to_numpy()) def test_autoai_libs_ta1(self): import autoai_libs.utils.fc_methods import numpy as np from autoai_libs.cognito.transforms.transform_utils import TA1 from lale.lib.sklearn import LogisticRegression as LR ta1 = TA1( fun=np.rint, name="round", datatypes=["numeric"], feat_constraints=[autoai_libs.utils.fc_methods.is_not_categorical], col_names=[ "a____________", "b____________", "c____________", "d____________", "e____________", ], col_dtypes=[ np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), ], ) pipeline = ta1 >> LR() expected = """from autoai_libs.cognito.transforms.transform_utils import TA1 import numpy as np import autoai_libs.utils.fc_methods from sklearn.linear_model import LogisticRegression as LR import lale lale.wrap_imported_operators() ta1 = TA1( fun=np.rint, name="round", datatypes=["numeric"], feat_constraints=[autoai_libs.utils.fc_methods.is_not_categorical], col_names=[ "a____________", "b____________", "c____________", "d____________", "e____________", ], col_dtypes=[ np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), np.dtype("float32"), ], ) pipeline = ta1 >> LR()""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_autoai_libs_t_no_op(self): from lightgbm import LGBMClassifier from lale.lib.autoai_libs import TNoOp from lale.operators import make_pipeline t_no_op = TNoOp( fun="fun", name="no_action", datatypes="x", feat_constraints=[], tgraph="tgraph", ) lgbm_classifier = LGBMClassifier(class_weight="balanced", learning_rate=0.18) pipeline = make_pipeline(t_no_op, lgbm_classifier) expected = """from autoai_libs.cognito.transforms.transform_utils import TNoOp from lightgbm import LGBMClassifier from lale.operators import make_pipeline t_no_op = TNoOp( fun="fun", name="no_action", datatypes="x", feat_constraints=[], tgraph="tgraph", ) lgbm_classifier = LGBMClassifier( class_weight="balanced", learning_rate=0.18, n_estimators=100 ) pipeline = make_pipeline(t_no_op, lgbm_classifier)""" self._roundtrip( expected, lale.pretty_print.to_string(pipeline, combinators=False) ) def test_autoai_libs_two_ops_with_combinator(self): from autoai_libs.transformers.exportable import ( CompressStrings, NumpyColumnSelector, ) numpy_column_selector = NumpyColumnSelector(columns=[0, 2, 3, 5]) compress_strings = CompressStrings( compress_type="hash", dtypes_list=["char_str", "char_str", "char_str", "char_str"], misslist_list=[[], [], [], []], ) pipeline = lale.operators.make_pipeline(numpy_column_selector, compress_strings) expected = """from autoai_libs.transformers.exportable import NumpyColumnSelector from autoai_libs.transformers.exportable import CompressStrings import lale lale.wrap_imported_operators() numpy_column_selector = NumpyColumnSelector(columns=[0, 2, 3, 5]) compress_strings = CompressStrings( compress_type="hash", dtypes_list=["char_str", "char_str", "char_str", "char_str"], missing_values_reference_list=["?", "", "-", float("nan")], misslist_list=[[], [], [], []], ) pipeline = numpy_column_selector >> compress_strings""" printed = lale.pretty_print.to_string(pipeline, combinators=True) try: self._roundtrip(expected, printed) except BaseException: expected = """from autoai_libs.transformers.exportable import NumpyColumnSelector from autoai_libs.transformers.exportable import CompressStrings import lale lale.wrap_imported_operators() numpy_column_selector = NumpyColumnSelector(columns=[0, 2, 3, 5]) compress_strings = CompressStrings( compress_type="hash", dtypes_list=["char_str", "char_str", "char_str", "char_str"], missing_values_reference_list=("?", "", "-", float("nan")), misslist_list=[[], [], [], []], ) pipeline = numpy_column_selector >> compress_strings""" self._roundtrip(expected, printed) def test_expression(self): from lale.expressions import it, mean from lale.lib.rasl import Aggregate, Join, Scan scan1 = Scan(table=it["table1.csv"]) scan2 = Scan(table=it["table2.csv"]) join = Join(pred=[it["table1.csv"].k1 == it["table2.csv"].k2]) aggregate = Aggregate(columns={"talk_time|mean": mean(it.talk_time)}) pipeline = (scan1 & scan2) >> join >> aggregate expected = """from lale.lib.rasl import Scan from lale.expressions import it from lale.lib.rasl import Join from lale.lib.rasl import Aggregate from lale.expressions import mean import lale lale.wrap_imported_operators() scan_0 = Scan(table=it["table1.csv"]) scan_1 = Scan(table=it["table2.csv"]) join = Join(pred=[it["table1.csv"].k1 == it["table2.csv"].k2]) aggregate = Aggregate(columns={"talk_time|mean": mean(it.talk_time)}) pipeline = (scan_0 & scan_1) >> join >> aggregate""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_sklearn_pipeline(self): from lale.lib.sklearn import PCA, LogisticRegression, Pipeline pipeline = Pipeline(steps=[("pca", PCA), ("lr", LogisticRegression(C=0.1))]) expected = """from sklearn.pipeline import Pipeline from sklearn.decomposition import PCA from sklearn.linear_model import LogisticRegression import lale lale.wrap_imported_operators() logistic_regression = LogisticRegression(C=0.1) pipeline = Pipeline(steps=[("pca", PCA), ("lr", logistic_regression)])""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_sklearn_pipeline_2(self): from lale.lib.sklearn import PCA, LogisticRegression, Pipeline pipeline = Pipeline(steps=[("pca", PCA), ("lr", LogisticRegression(C=0.1))]) expected = """from sklearn.pipeline import Pipeline from sklearn.decomposition import PCA from sklearn.linear_model import LogisticRegression logistic_regression = LogisticRegression(C=0.1) pipeline = Pipeline(steps=[("pca", PCA), ("lr", logistic_regression)])""" printed = lale.pretty_print.to_string(pipeline, astype="sklearn") self._roundtrip(expected, printed) def test_customize_schema_enum_and_number(self): from lale.lib.sklearn import LogisticRegression pipeline = LogisticRegression.customize_schema( solver={"enum": ["lbfgs", "liblinear"], "default": "liblinear"}, tol={ "type": "number", "minimum": 0.00001, "maximum": 0.1, "default": 0.0001, }, )(solver="lbfgs") expected = """from sklearn.linear_model import LogisticRegression import lale lale.wrap_imported_operators() pipeline = LogisticRegression.customize_schema( solver={"enum": ["lbfgs", "liblinear"], "default": "liblinear"}, tol={ "type": "number", "minimum": 1e-05, "maximum": 0.1, "default": 0.0001, }, )(solver="lbfgs")""" self._roundtrip(expected, pipeline.pretty_print(customize_schema=True)) def test_customize_schema_none_and_boolean(self): from lale.lib.sklearn import RandomForestRegressor pipeline = RandomForestRegressor.customize_schema( bootstrap={"type": "boolean", "default": True}, random_state={ "anyOf": [ {"laleType": "numpy.random.RandomState"}, { "description": "RandomState used by np.random", "enum": [None], }, {"description": "Explicit seed.", "type": "integer"}, ], "default": 33, }, )(n_estimators=50) expected = """from sklearn.ensemble import RandomForestRegressor import lale lale.wrap_imported_operators() pipeline = RandomForestRegressor.customize_schema( bootstrap={"type": "boolean", "default": True}, random_state={ "anyOf": [ {"laleType": "numpy.random.RandomState"}, {"description": "RandomState used by np.random", "enum": [None]}, {"description": "Explicit seed.", "type": "integer"}, ], "default": 33, }, )(n_estimators=50)""" # this should not include "random_state=33" because that would be # redundant with the schema, and would prevent automated search self._roundtrip(expected, pipeline.pretty_print(customize_schema=True)) def test_customize_schema_print_defaults(self): from lale.lib.sklearn import RandomForestRegressor pipeline = RandomForestRegressor.customize_schema( bootstrap={"type": "boolean", "default": True}, # default unchanged random_state={ "anyOf": [ {"laleType": "numpy.random.RandomState"}, {"enum": [None]}, {"type": "integer"}, ], "default": 33, # default changed }, )(n_estimators=50) expected = """from sklearn.ensemble import RandomForestRegressor import lale lale.wrap_imported_operators() pipeline = RandomForestRegressor(n_estimators=50, random_state=33)""" # print exactly those defaults that changed self._roundtrip(expected, pipeline.pretty_print(customize_schema=False)) def test_user_operator_in_toplevel_module(self): import importlib import os.path import sys import tempfile with tempfile.NamedTemporaryFile(mode="w", suffix=".py") as tmp_py_file: file_contents = """import numpy as np import lale.operators class _MockClassifierImpl: def __init__(self, int_hp=0): self.int_hp = int_hp def fit(self, X, y): self.some_y = list(y)[0] def predict(self, X): return self.some_y MockClassifier = lale.operators.make_operator(_MockClassifierImpl) """ tmp_py_file.write(file_contents) tmp_py_file.flush() dir_name = os.path.dirname(tmp_py_file.name) old_pythonpath = sys.path try: sys.path.append(dir_name) module_name = os.path.basename(tmp_py_file.name)[: -len(".py")] module = importlib.import_module(module_name) MockClf = getattr(module, "MockClassifier") self.assertIsInstance(MockClf, lale.operators.PlannedIndividualOp) self.assertEqual(MockClf.name(), "MockClassifier") pipeline = MockClf(int_hp=42) expected = f"""from {module_name} import MockClassifier as MockClf import lale lale.wrap_imported_operators(["{module_name}"]) pipeline = MockClf(int_hp=42)""" self._roundtrip(expected, pipeline.pretty_print()) finally: sys.path = old_pythonpath def test_nonlib_operator(self): from test.mock_custom_operators import CustomOrigOperator from lale.lib.sklearn import LogisticRegression pipeline = CustomOrigOperator() >> LogisticRegression() expected = """from test.mock_module import CustomOrigOperator from sklearn.linear_model import LogisticRegression import lale lale.wrap_imported_operators(["test.mock_module"]) pipeline = CustomOrigOperator() >> LogisticRegression()""" self._roundtrip(expected, pipeline.pretty_print()) @unittest.skip("TODO: avoid spurious 'name' keys in printed dictionaries") def test_fairness_info(self): from lale.lib.aif360 import DisparateImpactRemover, fetch_creditg_df from lale.lib.lale import Hyperopt, Project from lale.lib.sklearn import KNeighborsClassifier X, y, fairness_info = fetch_creditg_df() disparate_impact_remover = DisparateImpactRemover( **fairness_info, preparation=Project(columns={"type": "number"}), ) planned = disparate_impact_remover >> KNeighborsClassifier() frozen = planned.freeze_trainable() pipeline = frozen.auto_configure(X, y, optimizer=Hyperopt, cv=2, max_evals=1) expected = """from aif360.algorithms.preprocessing import DisparateImpactRemover from lale.lib.rasl import Project from sklearn.neighbors import KNeighborsClassifier import lale lale.wrap_imported_operators() project = Project(columns={"type": "number"}) disparate_impact_remover = DisparateImpactRemover( favorable_labels=["good"], protected_attributes=[ { "reference_group": [ "male div/sep", "male mar/wid", "male single", ], "feature": "personal_status", }, { "reference_group": [[26, 1000]], "feature": "age", }, ], preparation=project, ) pipeline = disparate_impact_remover >> KNeighborsClassifier()""" self._roundtrip(expected, pipeline.pretty_print()) def test_snap_logistic_regression_1(self): # force printing arguments via "transient": "alwaysPrint", case True from lale.lib.snapml import SnapLogisticRegression pipeline = SnapLogisticRegression(normalize=True) expected = """from snapml import SnapLogisticRegression import lale lale.wrap_imported_operators() pipeline = SnapLogisticRegression(fit_intercept=True, normalize=True)""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) def test_snap_logistic_regression_2(self): # force printing arguments via "transient": "alwaysPrint", case False from lale.lib.snapml import SnapLogisticRegression pipeline = SnapLogisticRegression(normalize=False) expected = """from snapml import SnapLogisticRegression import lale lale.wrap_imported_operators() pipeline = SnapLogisticRegression(normalize=False, fit_intercept=True)""" self._roundtrip(expected, lale.pretty_print.to_string(pipeline)) class TestToAndFromJSON(unittest.TestCase): def test_trainable_individual_op(self): self.maxDiff = None from lale.json_operator import from_json, to_json from lale.lib.sklearn import LogisticRegression as LR operator = LR(LR.enum.solver.sag, C=0.1) json_expected = { "class": LR.class_name(), "state": "trainable", "operator": "LogisticRegression", "label": "LR", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.logistic_regression.html", "hyperparams": {"C": 0.1, "solver": "sag"}, "is_frozen_trainable": False, } json = to_json(operator) self.assertEqual(json, json_expected) operator_2 = from_json(json) json_2 = to_json(operator_2) self.assertEqual(json_2, json_expected) def test_operator_choice(self): self.maxDiff = None from lale.json_operator import from_json, to_json from lale.lib.sklearn import PCA from lale.lib.sklearn import MinMaxScaler as Scl operator = PCA | Scl json_expected = { "class": "lale.operators.OperatorChoice", "operator": "OperatorChoice", "state": "planned", "steps": { "pca": { "class": PCA.class_name(), "state": "planned", "operator": "PCA", "label": "PCA", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.pca.html", }, "scl": { "class": Scl.class_name(), "state": "planned", "operator": "MinMaxScaler", "label": "Scl", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.min_max_scaler.html", }, }, } json = to_json(operator) self.assertEqual(json, json_expected) operator_2 = from_json(json) json_2 = to_json(operator_2) self.assertEqual(json_2, json_expected) def test_pipeline_1(self): self.maxDiff = None from lale.json_operator import from_json, to_json from lale.lib.lale import NoOp from lale.lib.rasl import ConcatFeatures from lale.lib.sklearn import PCA from lale.lib.sklearn import LogisticRegression as LR operator = (PCA & NoOp) >> ConcatFeatures >> LR json_expected = { "class": "lale.operators.PlannedPipeline", "state": "planned", "edges": [ ["pca", "concat_features"], ["no_op", "concat_features"], ["concat_features", "lr"], ], "steps": { "pca": { "class": PCA.class_name(), "state": "planned", "operator": "PCA", "label": "PCA", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.pca.html", }, "no_op": { "class": NoOp.class_name(), "state": "trained", "operator": "NoOp", "label": "NoOp", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.lale.no_op.html", "hyperparams": None, "coefs": None, "is_frozen_trainable": True, "is_frozen_trained": True, }, "concat_features": { "class": ConcatFeatures.class_name(), "state": "trained", "operator": "ConcatFeatures", "label": "ConcatFeatures", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.rasl.concat_features.html", "hyperparams": None, "coefs": None, "is_frozen_trainable": True, "is_frozen_trained": True, }, "lr": { "class": LR.class_name(), "state": "planned", "operator": "LogisticRegression", "label": "LR", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.logistic_regression.html", }, }, } json = to_json(operator) self.assertEqual(json, json_expected) operator_2 = from_json(json) json_2 = to_json(operator_2) self.assertEqual(json, json_2) def test_pipeline_2(self): from lale.json_operator import from_json, to_json from lale.lib.lale import NoOp from lale.lib.sklearn import ( PCA, KNeighborsClassifier, LogisticRegression, Nystroem, ) from lale.operators import make_choice, make_pipeline kernel_tfm_or_not = make_choice(NoOp, Nystroem) tfm = PCA clf = make_choice(LogisticRegression, KNeighborsClassifier) operator = make_pipeline(kernel_tfm_or_not, tfm, clf) json = to_json(operator) operator_2 = from_json(json) json_2 = to_json(operator_2) self.assertEqual(json, json_2) def test_higher_order_1(self): from lale.json_operator import from_json from lale.lib.lale import Both from lale.lib.sklearn import PCA, Nystroem operator = Both(op1=PCA(n_components=2), op2=Nystroem) json_expected = { "class": Both.class_name(), "state": "trainable", "operator": "Both", "label": "Both", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.lale.both.html", "hyperparams": { "op1": {"$ref": "../steps/pca"}, "op2": {"$ref": "../steps/nystroem"}, }, "steps": { "pca": { "class": PCA.class_name(), "state": "trainable", "operator": "PCA", "label": "PCA", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.pca.html", "hyperparams": {"n_components": 2}, "is_frozen_trainable": False, }, "nystroem": { "class": Nystroem.class_name(), "state": "planned", "operator": "Nystroem", "label": "Nystroem", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.nystroem.html", }, }, "is_frozen_trainable": False, } json = operator.to_json() self.assertEqual(json, json_expected) operator_2 = from_json(json) json_2 = operator_2.to_json() self.assertEqual(json, json_2) def test_higher_order_2(self): self.maxDiff = None from lale.json_operator import from_json from lale.lib.sklearn import PCA from lale.lib.sklearn import KNeighborsClassifier as KNN from lale.lib.sklearn import LogisticRegression as LR from lale.lib.sklearn import VotingClassifier as Vote operator = Vote( estimators=[("knn", KNN), ("pipeline", PCA() >> LR)], voting="soft" ) json_expected = { "class": Vote.class_name(), "state": "trainable", "operator": "VotingClassifier", "is_frozen_trainable": True, "label": "Vote", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.voting_classifier.html", "hyperparams": { "estimators": [ ("knn", {"$ref": "../steps/knn"}), ("pipeline", {"$ref": "../steps/pipeline"}), ], "voting": "soft", }, "steps": { "knn": { "class": KNN.class_name(), "state": "planned", "operator": "KNeighborsClassifier", "label": "KNN", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.k_neighbors_classifier.html", }, "pipeline": { "class": "lale.operators.PlannedPipeline", "state": "planned", "edges": [["pca", "lr"]], "steps": { "pca": { "class": PCA.class_name(), "state": "trainable", "operator": "PCA", "label": "PCA", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.pca.html", "hyperparams": {}, "is_frozen_trainable": False, }, "lr": { "class": LR.class_name(), "state": "planned", "operator": "LogisticRegression", "label": "LR", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.logistic_regression.html", }, }, }, }, } json = operator.to_json() self.assertEqual(json, json_expected) operator_2 = from_json(json) json_2 = operator_2.to_json() self.assertEqual(json, json_2) def test_nested(self): self.maxDiff = None from lale.json_operator import from_json, to_json from lale.lib.lale import NoOp from lale.lib.sklearn import PCA from lale.lib.sklearn import LogisticRegression as LR operator = PCA >> (LR(C=0.09) | NoOp >> LR(C=0.19)) json_expected = { "class": "lale.operators.PlannedPipeline", "state": "planned", "edges": [["pca", "choice"]], "steps": { "pca": { "class": PCA.class_name(), "state": "planned", "operator": "PCA", "label": "PCA", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.pca.html", }, "choice": { "class": "lale.operators.OperatorChoice", "state": "planned", "operator": "OperatorChoice", "steps": { "lr_0": { "class": LR.class_name(), "state": "trainable", "operator": "LogisticRegression", "label": "LR", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.logistic_regression.html", "hyperparams": {"C": 0.09}, "is_frozen_trainable": False, }, "pipeline_1": { "class": "lale.operators.TrainablePipeline", "state": "trainable", "edges": [["no_op", "lr_1"]], "steps": { "no_op": { "class": NoOp.class_name(), "state": "trained", "operator": "NoOp", "label": "NoOp", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.lale.no_op.html", "hyperparams": None, "coefs": None, "is_frozen_trainable": True, "is_frozen_trained": True, }, "lr_1": { "class": LR.class_name(), "state": "trainable", "operator": "LogisticRegression", "label": "LR", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.logistic_regression.html", "hyperparams": {"C": 0.19}, "is_frozen_trainable": False, }, }, }, }, }, }, } json = to_json(operator) self.assertEqual(json, json_expected) operator_2 = from_json(json) json_2 = to_json(operator_2) self.assertEqual(json, json_2) def test_customize_schema(self): from lale.json_operator import from_json, to_json from lale.lib.sklearn import LogisticRegression as LR operator = LR.customize_schema( solver={"enum": ["lbfgs", "liblinear"], "default": "liblinear"}, tol={ "type": "number", "minimum": 0.00001, "maximum": 0.1, "default": 0.0001, }, ) json_expected = { "class": LR.class_name(), "state": "planned", "operator": "LogisticRegression", "label": "LR", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.sklearn.logistic_regression.html", "customize_schema": { "properties": { "hyperparams": { "allOf": [ { "type": "object", "properties": { "solver": { "default": "liblinear", "enum": ["lbfgs", "liblinear"], }, "tol": { "type": "number", "minimum": 0.00001, "maximum": 0.1, "default": 0.0001, }, }, } ] } } }, } json = to_json(operator) self.maxDiff = None self.assertEqual(json, json_expected) operator_2 = from_json(json) json_2 = to_json(operator_2) self.assertEqual(json, json_2) class TestDiff(unittest.TestCase): def test_single_op(self): from lale.lib.sklearn import LogisticRegression single_op = LogisticRegression() single_op_param = LogisticRegression(solver="saga") expected_diff = ( " from sklearn.linear_model import LogisticRegression\n" " import lale\n" " \n" " lale.wrap_imported_operators()\n" "- pipeline = LogisticRegression()\n" '+ pipeline = LogisticRegression(solver="saga")\n' "? +++++++++++++\n" ) diff_str = single_op.diff(single_op_param, ipython_display=False) self.assertEqual(diff_str, expected_diff) expected_diff_reverse = ( " from sklearn.linear_model import LogisticRegression\n" " import lale\n" " \n" " lale.wrap_imported_operators()\n" '- pipeline = LogisticRegression(solver="saga")\n' "? -------------\n\n" "+ pipeline = LogisticRegression()" ) diff_str_reverse = single_op_param.diff(single_op, ipython_display=False) self.assertEqual(diff_str_reverse, expected_diff_reverse) def test_pipeline(self): from lale.lib.lale import NoOp from lale.lib.sklearn import PCA, LogisticRegression, SelectKBest pipeline_simple = PCA >> SelectKBest >> LogisticRegression pipeline_choice = (PCA | NoOp) >> SelectKBest >> LogisticRegression expected_diff = ( " from sklearn.decomposition import PCA\n" "+ from lale.lib.lale import NoOp\n" " from sklearn.feature_selection import SelectKBest\n" " from sklearn.linear_model import LogisticRegression\n" " import lale\n" " \n" " lale.wrap_imported_operators()\n" "- pipeline = PCA >> SelectKBest >> LogisticRegression\n" "+ pipeline = (PCA | NoOp) >> SelectKBest >> LogisticRegression\n" "? + ++++++++\n" ) diff_str = pipeline_simple.diff(pipeline_choice, ipython_display=False) self.assertEqual(diff_str, expected_diff) expected_diff_reverse = ( " from sklearn.decomposition import PCA\n" "- from lale.lib.lale import NoOp\n" " from sklearn.feature_selection import SelectKBest\n" " from sklearn.linear_model import LogisticRegression\n" " import lale\n" " \n" " lale.wrap_imported_operators()\n" "- pipeline = (PCA | NoOp) >> SelectKBest >> LogisticRegression\n" "? - --------\n\n" "+ pipeline = PCA >> SelectKBest >> LogisticRegression" ) diff_str_reverse = pipeline_choice.diff(pipeline_simple, ipython_display=False) self.assertEqual(diff_str_reverse, expected_diff_reverse) def test_single_op_pipeline(self): from lale.lib.sklearn import PCA, LogisticRegression, SelectKBest single_op = LogisticRegression() pipeline = PCA >> SelectKBest >> LogisticRegression expected_diff = ( "+ from sklearn.decomposition import PCA\n" "+ from sklearn.feature_selection import SelectKBest\n" " from sklearn.linear_model import LogisticRegression\n" " import lale\n" " \n" " lale.wrap_imported_operators()\n" "- pipeline = LogisticRegression()\n" "+ pipeline = PCA >> SelectKBest >> LogisticRegression" ) diff_str = single_op.diff(pipeline, ipython_display=False) self.assertEqual(expected_diff, diff_str) expected_diff_reverse = ( "- from sklearn.decomposition import PCA\n" "- from sklearn.feature_selection import SelectKBest\n" " from sklearn.linear_model import LogisticRegression\n" " import lale\n" " \n" " lale.wrap_imported_operators()\n" "- pipeline = PCA >> SelectKBest >> LogisticRegression\n" "+ pipeline = LogisticRegression()" ) diff_str_reverse = pipeline.diff(single_op, ipython_display=False) self.assertEqual(expected_diff_reverse, diff_str_reverse) def test_options(self): from lale.lib.sklearn import LogisticRegression single_op = LogisticRegression() single_op_schema = single_op.customize_schema(solver={"enum": ["saga"]}) expected_diff_no_imports = " pipeline = LogisticRegression()" diff_str_no_imports = single_op.diff( single_op_schema, show_imports=False, ipython_display=False ) self.assertEqual(diff_str_no_imports, expected_diff_no_imports) expected_diff_no_schema = ( " from sklearn.linear_model import LogisticRegression\n" " import lale\n" " \n" " lale.wrap_imported_operators()\n" "- pipeline = LogisticRegression()\n" '+ pipeline = LogisticRegression.customize_schema(solver={"enum": ["saga"]})()' ) diff_str_no_schema = single_op.diff( single_op_schema, customize_schema=True, ipython_display=False ) self.assertEqual(diff_str_no_schema, expected_diff_no_schema)
64,209
37.106825
147
py
lale
lale-master/lale/operator_wrapper.py
# Copyright 2019-2022 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import inspect import logging import sys from typing import List, Optional, Set from lale.operators import Operator, clone_op, get_op_from_lale_lib if sys.version_info < (3, 9): from typing import Container # raises a mypy error for <3.8 else: from collections.abc import Container logger = logging.getLogger(__name__) def _wrap_operators_in_symtab( symtab, exclude_classes: Optional[Container[str]] = None, wrapper_modules: Optional[List[str]] = None, ) -> None: for name, impl in symtab.items(): if ( inspect.isclass(impl) and not issubclass(impl, Operator) and (hasattr(impl, "predict") or hasattr(impl, "transform")) ): if exclude_classes is not None: if name in exclude_classes: continue operator = get_op_from_lale_lib(impl, wrapper_modules) if operator is None: # symtab[name] = make_operator(impl=impl, name=name) logger.info(f"Lale:Not wrapping unknown operator:{name}") else: symtab[name] = clone_op(operator, name) if operator.class_name().startswith("lale.lib.autogen"): logger.info(f"Lale:Wrapped autogen operator:{name}") else: logger.info(f"Lale:Wrapped known operator:{name}") def wrap_imported_operators( exclude_classes: Optional[Container[str]] = None, wrapper_modules: Optional[List[str]] = None, ) -> None: """Wrap the currently imported operators from the symbol table to their lale wrappers. Parameters ---------- exclude_classes : string, optional, default None List of class names to exclude from wrapping, alias names if they are used while importing. wrapper_modules : set of string, optional, default None Set of Lale modules to use for wrapping operators. """ current_frame = inspect.currentframe() assert ( current_frame is not None ), "Try to use inspect.stack()[1][0] to get the calling frame" calling_frame = current_frame.f_back assert ( calling_frame is not None ), "Try to use inspect.stack()[1][0] to get the calling frame" if wrapper_modules is not None: wrapper_modules.extend(get_lale_wrapper_modules()) else: wrapper_modules = list(get_lale_wrapper_modules()) _wrap_operators_in_symtab( calling_frame.f_globals, exclude_classes, wrapper_modules=wrapper_modules ) if calling_frame.f_code.co_name == "<module>": # for testing with exec() _wrap_operators_in_symtab( calling_frame.f_locals, exclude_classes, wrapper_modules=wrapper_modules ) _lale_wrapper_modules: Set[str] = set() def register_lale_wrapper_modules(m: str) -> None: """Register a module with lale's import system so that :meth:`lale.helpers.import_from_sklearn_pipeline` will look for replacement classes in that module. Example: (in `__init__.py` file for the module): .. code-block:: python from lale import register_lale_wrapper_modules register_lale_wrapper_modules(__name__) Parameters ---------- m : [str] The module name """ _lale_wrapper_modules.add(m) def get_lale_wrapper_modules() -> Set[str]: return _lale_wrapper_modules for builtin_lale_modules in [ "lale.lib.sklearn", "lale.lib.autoai_libs", "lale.lib.xgboost", "lale.lib.lightgbm", "lale.lib.snapml", "autoai_ts_libs.lale", ]: register_lale_wrapper_modules(builtin_lale_modules)
4,203
31.589147
111
py
lale
lale-master/lale/helpers.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import ast import copy import importlib import logging import sys import time import traceback from importlib import util from typing import ( TYPE_CHECKING, Any, Dict, Iterable, List, Mapping, Optional, Set, Tuple, TypeVar, Union, overload, ) import numpy as np import pandas as pd import scipy.sparse import sklearn.pipeline from numpy.random import RandomState from sklearn.metrics import accuracy_score, check_scoring, log_loss from sklearn.model_selection import StratifiedKFold from sklearn.utils.metaestimators import _safe_split import lale.datasets.data_schemas if sys.version_info >= (3, 8): from typing import Literal # raises a mypy error for <3.8 else: from typing_extensions import Literal try: import torch torch_installed = True except ImportError: torch_installed = False spark_loader = util.find_spec("pyspark") spark_installed = spark_loader is not None if spark_installed: from pyspark.sql.dataframe import DataFrame as spark_df logger = logging.getLogger(__name__) LALE_NESTED_SPACE_KEY = "__lale_nested_space" astype_type = Literal["lale", "sklearn"] datatype_param_type = Literal["pandas", "spark"] randomstate_type = Union[RandomState, int, None] def make_nested_hyperopt_space(sub_space): return {LALE_NESTED_SPACE_KEY: sub_space} def assignee_name(level=1) -> Optional[str]: tb = traceback.extract_stack() file_name, _line_number, _function_name, text = tb[-(level + 2)] try: tree = ast.parse(text, file_name) except SyntaxError: return None assert tree is not None and isinstance(tree, ast.Module) if len(tree.body) == 1: stmt = tree.body[0] if isinstance(stmt, ast.Assign): lhs = stmt.targets if len(lhs) == 1: res = lhs[0] if isinstance(res, ast.Name): return res.id return None def arg_name(pos=0, level=1) -> Optional[str]: tb = traceback.extract_stack() file_name, _line_number, _function_name, text = tb[-(level + 2)] try: tree = ast.parse(text, file_name) except SyntaxError: return None assert tree is not None and isinstance(tree, ast.Module) if len(tree.body) == 1: stmt = tree.body[0] if isinstance(stmt, ast.Expr): expr = stmt.value if isinstance(expr, ast.Call): args = expr.args if pos < len(args): res = args[pos] if isinstance(res, ast.Name): return res.id return None def data_to_json(data, subsample_array: bool = True) -> Union[list, dict, int, float]: if isinstance(data, tuple): # convert to list return [data_to_json(elem, subsample_array) for elem in data] if isinstance(data, list): return [data_to_json(elem, subsample_array) for elem in data] elif isinstance(data, dict): return {key: data_to_json(data[key], subsample_array) for key in data} elif isinstance(data, np.ndarray): return ndarray_to_json(data, subsample_array) elif isinstance(data, scipy.sparse.csr_matrix): return ndarray_to_json(data.toarray(), subsample_array) elif isinstance(data, (pd.DataFrame, pd.Series)): np_array = data.values return ndarray_to_json(np_array, subsample_array) elif torch_installed and isinstance(data, torch.Tensor): np_array = data.detach().numpy() return ndarray_to_json(np_array, subsample_array) elif isinstance(data, (np.int64, np.int32, np.int16)): # type: ignore return int(data) elif isinstance(data, (np.float32, np.float64)): # type: ignore return float(data) else: return data def is_empty_dict(val) -> bool: return isinstance(val, dict) and len(val) == 0 def dict_without(orig_dict: Dict[str, Any], key: str) -> Dict[str, Any]: if key not in orig_dict: return orig_dict return {k: v for k, v in orig_dict.items() if k != key} def json_lookup(ptr, jsn, default=None): steps = ptr.split("/") sub_jsn = jsn for s in steps: if s not in sub_jsn: return default sub_jsn = sub_jsn[s] return sub_jsn def ndarray_to_json(arr: np.ndarray, subsample_array: bool = True) -> Union[list, dict]: # sample 10 rows and no limit on columns num_subsamples: List[int] if subsample_array: num_subsamples = [10, np.iinfo(int).max, np.iinfo(int).max] else: num_subsamples = [ np.iinfo(int).max, np.iinfo(int).max, np.iinfo(int).max, ] def subarray_to_json(indices: Tuple[int, ...]) -> Any: if len(indices) == len(arr.shape): if isinstance(arr[indices], (bool, int, float, str)): return arr[indices] elif np.issubdtype(arr.dtype, np.bool_): return bool(arr[indices]) elif np.issubdtype(arr.dtype, np.integer): return int(arr[indices]) elif np.issubdtype(arr.dtype, np.number): return float(arr[indices]) elif arr.dtype.kind in ["U", "S", "O"]: return str(arr[indices]) else: raise ValueError( f"Unexpected dtype {arr.dtype}, " f"kind {arr.dtype.kind}, " f"type {type(arr[indices])}." ) else: assert len(indices) < len(arr.shape) return [ subarray_to_json(indices + (i,)) for i in range( min(num_subsamples[len(indices)], arr.shape[len(indices)]) ) ] return subarray_to_json(()) def split_with_schemas(estimator, all_X, all_y, indices, train_indices=None): subset_X, subset_y = _safe_split(estimator, all_X, all_y, indices, train_indices) if hasattr(all_X, "json_schema"): n_rows = subset_X.shape[0] schema = { "type": "array", "minItems": n_rows, "maxItems": n_rows, "items": all_X.json_schema["items"], } lale.datasets.data_schemas.add_schema(subset_X, schema) if hasattr(all_y, "json_schema"): n_rows = subset_y.shape[0] schema = { "type": "array", "minItems": n_rows, "maxItems": n_rows, "items": all_y.json_schema["items"], } lale.datasets.data_schemas.add_schema(subset_y, schema) return subset_X, subset_y def fold_schema(X, y, cv=1, is_classifier=True): def fold_schema_aux(data, n_rows): orig_schema = lale.datasets.data_schemas._to_schema(data) aux_result = {**orig_schema, "minItems": n_rows, "maxItems": n_rows} return aux_result n_splits = cv if isinstance(cv, int) else cv.get_n_splits() try: n_samples = X.shape[0] if hasattr(X, "shape") else len(X) except TypeError: # raised for Spark dataframes. n_samples = X.count() if hasattr(X, "count") else 0 if n_splits == 1: n_rows_fold = n_samples elif is_classifier: n_classes = len(set(y)) n_rows_unstratified = (n_samples // n_splits) * (n_splits - 1) # in stratified case, fold sizes can differ by up to n_classes n_rows_fold = max(1, n_rows_unstratified - n_classes) else: n_rows_fold = (n_samples // n_splits) * (n_splits - 1) schema_X = fold_schema_aux(X, n_rows_fold) schema_y = fold_schema_aux(y, n_rows_fold) result = {"properties": {"X": schema_X, "y": schema_y}} return result def cross_val_score_track_trials( estimator, X, y=None, scoring: Any = accuracy_score, cv: Any = 5, args_to_scorer: Optional[Dict[str, Any]] = None, args_to_cv: Optional[Dict[str, Any]] = None, **fit_params, ): """ Use the given estimator to perform fit and predict for splits defined by 'cv' and compute the given score on each of the splits. Parameters ---------- estimator: A valid sklearn_wrapper estimator X: Valid data that works with the estimator y: Valid target that works with the estimator scoring: string or a scorer object created using https://scikit-learn.org/stable/modules/generated/sklearn.metrics.make_scorer.html#sklearn.metrics.make_scorer. A string from sklearn.metrics.SCORERS.keys() can be used or a scorer created from one of sklearn.metrics (https://scikit-learn.org/stable/modules/classes.html#module-sklearn.metrics). A completely custom scorer object can be created from a python function following the example at https://scikit-learn.org/stable/modules/model_evaluation.html The metric has to return a scalar value, cv: an integer or an object that has a split function as a generator yielding (train, test) splits as arrays of indices. Integer value is used as number of folds in sklearn.model_selection.StratifiedKFold, default is 5. Note that any of the iterators from https://scikit-learn.org/stable/modules/cross_validation.html#cross-validation-iterators can be used here. args_to_scorer: A dictionary of additional keyword arguments to pass to the scorer. Used for cases where the scorer has a signature such as ``scorer(estimator, X, y, **kwargs)``. args_to_cv: A dictionary of additional keyword arguments to pass to the split method of cv. This is only applicable when cv is not an integer. fit_params: Additional parameters that should be passed when calling fit on the estimator Returns ------- cv_results: a list of scores corresponding to each cross validation fold """ if isinstance(cv, int): cv = StratifiedKFold(cv) if args_to_scorer is None: args_to_scorer = {} if args_to_cv is None: args_to_cv = {} scorer = check_scoring(estimator, scoring=scoring) cv_results: List[float] = [] log_loss_results = [] time_results = [] for train, test in cv.split(X, y, **args_to_cv): X_train, y_train = split_with_schemas(estimator, X, y, train) X_test, y_test = split_with_schemas(estimator, X, y, test, train) start = time.time() # Not calling sklearn.base.clone() here, because: # (1) For Lale pipelines, clone() calls the pipeline constructor # with edges=None, so the resulting topology is incorrect. # (2) For Lale individual operators, the fit() method already # clones the impl object, so cloning again is redundant. trained = estimator.fit(X_train, y_train, **fit_params) score_value = scorer(trained, X_test, y_test, **args_to_scorer) execution_time = time.time() - start # not all estimators have predict probability try: y_pred_proba = trained.predict_proba(X_test) logloss = log_loss(y_true=y_test, y_pred=y_pred_proba) log_loss_results.append(logloss) except BaseException: logger.debug("Warning, log loss cannot be computed") cv_results.append(score_value) time_results.append(execution_time) result = ( np.array(cv_results).mean(), np.array(log_loss_results).mean(), np.array(time_results).mean(), ) return result def cross_val_score(estimator, X, y=None, scoring: Any = accuracy_score, cv: Any = 5): """ Use the given estimator to perform fit and predict for splits defined by 'cv' and compute the given score on each of the splits. Parameters ---------- estimator: A valid sklearn_wrapper estimator X: Valid data value that works with the estimator y: Valid target value that works with the estimator scoring: a scorer object from sklearn.metrics (https://scikit-learn.org/stable/modules/classes.html#module-sklearn.metrics) Default value is accuracy_score. cv: an integer or an object that has a split function as a generator yielding (train, test) splits as arrays of indices. Integer value is used as number of folds in sklearn.model_selection.StratifiedKFold, default is 5. Note that any of the iterators from https://scikit-learn.org/stable/modules/cross_validation.html#cross-validation-iterators can be used here. Returns ------- cv_results: a list of scores corresponding to each cross validation fold """ if isinstance(cv, int): cv = StratifiedKFold(cv) cv_results = [] for train, test in cv.split(X, y): X_train, y_train = split_with_schemas(estimator, X, y, train) X_test, y_test = split_with_schemas(estimator, X, y, test, train) trained_estimator = estimator.fit(X_train, y_train) predicted_values = trained_estimator.predict(X_test) cv_results.append(scoring(y_test, predicted_values)) return cv_results def create_individual_op_using_reflection(class_name, operator_name, param_dict): instance = None if class_name is not None: class_name_parts = class_name.split(".") assert ( len(class_name_parts) ) > 1, ( "The class name needs to be fully qualified, i.e. module name + class name" ) module_name = ".".join(class_name_parts[0:-1]) class_name = class_name_parts[-1] module = importlib.import_module(module_name) class_ = getattr(module, class_name) if param_dict is None: instance = class_() else: instance = class_(**param_dict) return instance if TYPE_CHECKING: import lale.operators def to_graphviz( lale_operator: "lale.operators.Operator", ipython_display: bool = True, call_depth: int = 1, **dot_graph_attr, ): import lale.json_operator import lale.operators import lale.visualize if not isinstance(lale_operator, lale.operators.Operator): raise TypeError("The input to to_graphviz needs to be a valid LALE operator.") jsn = lale.json_operator.to_json(lale_operator, call_depth=call_depth + 1) dot = lale.visualize.json_to_graphviz(jsn, ipython_display, dot_graph_attr) return dot def instantiate_from_hyperopt_search_space(obj_hyperparams, new_hyperparams): if isinstance(new_hyperparams, dict) and LALE_NESTED_SPACE_KEY in new_hyperparams: sub_params = new_hyperparams[LALE_NESTED_SPACE_KEY] sub_op = obj_hyperparams if isinstance(sub_op, list): if len(sub_op) == 1: sub_op = sub_op[0] else: step_index, step_params = list(sub_params)[0] if step_index < len(sub_op): sub_op = sub_op[step_index] sub_params = step_params return create_instance_from_hyperopt_search_space(sub_op, sub_params) elif isinstance(new_hyperparams, (list, tuple)): assert isinstance(obj_hyperparams, (list, tuple)) params_len = len(new_hyperparams) assert params_len == len(obj_hyperparams) res: Optional[List[Any]] = None for i in range(params_len): nhi = new_hyperparams[i] ohi = obj_hyperparams[i] updated_params = instantiate_from_hyperopt_search_space(ohi, nhi) if updated_params is not None: if res is None: res = list(new_hyperparams) res[i] = updated_params if res is not None: if isinstance(obj_hyperparams, tuple): return tuple(res) else: return res # workaround for what seems to be a hyperopt bug # where hyperopt returns a tuple even though the # hyperopt search space specifies a list is_obj_tuple = isinstance(obj_hyperparams, tuple) is_new_tuple = isinstance(new_hyperparams, tuple) if is_obj_tuple != is_new_tuple: if is_obj_tuple: return tuple(new_hyperparams) else: return list(new_hyperparams) return None elif isinstance(new_hyperparams, dict): assert isinstance(obj_hyperparams, dict) for k, sub_params in new_hyperparams.items(): if k in obj_hyperparams: sub_op = obj_hyperparams[k] updated_params = instantiate_from_hyperopt_search_space( sub_op, sub_params ) if updated_params is not None: new_hyperparams[k] = updated_params return None else: return None def create_instance_from_hyperopt_search_space( lale_object, hyperparams ) -> "lale.operators.Operator": """ Hyperparams is a n-tuple of dictionaries of hyper-parameters, each dictionary corresponds to an operator in the pipeline """ # lale_object can either be an individual operator, a pipeline or an operatorchoice # Validate that the number of elements in the n-tuple is the same # as the number of steps in the current pipeline from lale.operators import ( BasePipeline, OperatorChoice, PlannedIndividualOp, TrainableOperator, TrainablePipeline, ) if isinstance(lale_object, PlannedIndividualOp): new_hyperparams: Dict[str, Any] = dict_without(hyperparams, "name") hps = lale_object.hyperparams() if hps: obj_hyperparams = dict(hps) else: obj_hyperparams = {} for k, sub_params in new_hyperparams.items(): if k in obj_hyperparams: sub_op = obj_hyperparams[k] updated_params = instantiate_from_hyperopt_search_space( sub_op, sub_params ) if updated_params is not None: new_hyperparams[k] = updated_params all_hyperparams = {**obj_hyperparams, **new_hyperparams} return lale_object(**all_hyperparams) elif isinstance(lale_object, BasePipeline): steps = lale_object.steps_list() if len(hyperparams) != len(steps): raise ValueError( "The number of steps in the hyper-parameter space does not match the number of steps in the pipeline." ) op_instances = [] edges = lale_object.edges() # op_map:Dict[PlannedOpType, TrainableOperator] = {} op_map = {} for op_index, sub_params in enumerate(hyperparams): sub_op = steps[op_index] op_instance = create_instance_from_hyperopt_search_space(sub_op, sub_params) assert isinstance(op_instance, TrainableOperator) assert ( isinstance(sub_op, OperatorChoice) or sub_op.class_name() == op_instance.class_name() ), f"sub_op {sub_op.class_name()}, op_instance {op_instance.class_name()}" op_instances.append(op_instance) op_map[sub_op] = op_instance # trainable_edges:List[Tuple[TrainableOperator, TrainableOperator]] try: trainable_edges = [(op_map[x], op_map[y]) for (x, y) in edges] except KeyError as e: raise ValueError( "An edge was found with an endpoint that is not a step (" + str(e) + ")" ) from e return TrainablePipeline(op_instances, trainable_edges, ordered=True) # type: ignore elif isinstance(lale_object, OperatorChoice): # Hyperopt search space for an OperatorChoice is generated as a dictionary with a single element # corresponding to the choice made, the only key is the index of the step and the value is # the params corresponding to that step. step_index: int choices = lale_object.steps_list() if len(choices) == 1: step_index = 0 else: step_index_str, hyperparams = list(hyperparams.items())[0] step_index = int(step_index_str) step_object = choices[step_index] return create_instance_from_hyperopt_search_space(step_object, hyperparams) else: assert False, f"Unknown operator type: {type(lale_object)}" def find_lale_wrapper(sklearn_obj: Any) -> Optional[Any]: """ :param sklearn_obj: An sklearn compatible object that may have a lale wrapper :return: The lale wrapper type, or None if one could not be found """ from .operator_wrapper import get_lale_wrapper_modules module_names = get_lale_wrapper_modules() class_name = sklearn_obj.__class__.__name__ for module_name in module_names: try: module = importlib.import_module(module_name) except ModuleNotFoundError: continue try: class_ = getattr(module, class_name) return class_ except AttributeError: continue return None def _import_from_sklearn_inplace_helper( sklearn_obj, fitted: bool = True, is_nested=False ): """ This method take an object and tries to wrap sklearn objects (at the top level or contained within hyperparameters of other sklearn objects). It will modify the object to add in the appropriate lale wrappers. It may also return a wrapper or different object than given. :param sklearn_obj: the object that we are going to try and wrap :param fitted: should we return a TrainedOperator :param is_hyperparams: is this a nested invocation (which allows for returning a Trainable operator even if fitted is set to True) """ @overload def import_nested_params( orig_hyperparams: dict, partial_dict: bool ) -> Optional[dict]: ... @overload def import_nested_params(orig_hyperparams: Any, partial_dict: bool) -> Any: ... def import_nested_params(orig_hyperparams: Any, partial_dict: bool = False): """ look through lists/tuples/dictionaries for sklearn compatible objects to import. :param orig_hyperparams: the input to recursively look through for sklearn compatible objects :param partial_dict: If this is True and the input is a dictionary, the returned dictionary will only have the keys with modified values :return: Either a modified version of the input or None if nothing was changed """ if isinstance(orig_hyperparams, (tuple, list)): new_list: list = [] list_modified: bool = False for e in orig_hyperparams: new_e = import_nested_params(e, partial_dict=False) if new_e is None: new_list.append(e) else: new_list.append(new_e) list_modified = True if not list_modified: return None if isinstance(orig_hyperparams, tuple): return tuple(new_list) else: return new_list if isinstance(orig_hyperparams, dict): new_dict: dict = {} dict_modified: bool = False for k, v in orig_hyperparams.items(): new_v = import_nested_params(v, partial_dict=False) if new_v is None: if not partial_dict: new_dict[k] = v else: new_dict[k] = new_v dict_modified = True if not dict_modified: return None return new_dict if isinstance(orig_hyperparams, object) and hasattr( orig_hyperparams, "get_params" ): newobj = _import_from_sklearn_inplace_helper( orig_hyperparams, fitted=fitted, is_nested=True ) # allow nested_op to be trainable if newobj is orig_hyperparams: return None return newobj return None if sklearn_obj is None: return None if isinstance(sklearn_obj, lale.operators.TrainedIndividualOp): # if fitted=False, we may want to return a TrainedIndidivualOp return sklearn_obj # if the object is a trainable operator, we clean that up if isinstance(sklearn_obj, lale.operators.TrainableIndividualOp) and hasattr( sklearn_obj, "_trained" ): if fitted: # get rid of the indirection, and just return the trained operator directly return sklearn_obj._trained else: # since we are not supposed to be trained, delete the trained part delattr(sklearn_obj, "_trained") # delete _trained before returning return sklearn_obj if isinstance(sklearn_obj, lale.operators.Operator): if ( fitted and is_nested or not hasattr(sklearn_obj._impl_instance(), "fit") ): # Operators such as NoOp do not have a fit, so return them as is. return sklearn_obj if fitted: raise ValueError( f"""The input pipeline has an operator {sklearn_obj} that is not trained and fitted is set to True, please pass fitted=False if you want a trainable pipeline as output.""" ) # the lale operator is not trained and fitted=False return sklearn_obj # special case for FeatureUnion. # An alternative would be to (like for sklearn pipeline) # create a lale wrapper for the sklearn feature union # as a higher order operator # and then the special case would be just to throw away the outer wrapper # Note that lale union does not currently support weights or other features of feature union. if isinstance(sklearn_obj, sklearn.pipeline.FeatureUnion): transformer_list = sklearn_obj.transformer_list concat_predecessors = [ _import_from_sklearn_inplace_helper( transformer[1], fitted=fitted, is_nested=is_nested ) for transformer in transformer_list ] return lale.operators.make_union(*concat_predecessors) if not hasattr(sklearn_obj, "get_params"): # if it does not have a get_params method, # then we just return it without trying to wrap it return sklearn_obj class_ = find_lale_wrapper(sklearn_obj) if not class_: return sklearn_obj # Return the original object # next, we need to figure out what the right hyperparameters are orig_hyperparams = sklearn_obj.get_params(deep=False) hyperparams = import_nested_params(orig_hyperparams, partial_dict=True) if hyperparams: # if we have updated any of the hyperparameters then we modify them in the actual sklearn object try: new_obj = sklearn_obj.set_params(**hyperparams) if new_obj is not None: sklearn_obj = new_obj except NotImplementedError: # if the set_params method does not work, then do our best pass all_new_hyperparams = {**orig_hyperparams, **hyperparams} else: all_new_hyperparams = orig_hyperparams # now, we get the lale operator for the wrapper, with the corresponding hyperparameters if not fitted: # If fitted is False, we do not want to return a Trained operator. lale_op_obj_base = class_ else: lale_op_obj_base = lale.operators.TrainedIndividualOp( class_._name, class_._impl, class_._schemas, None, _lale_trained=True, ) lale_op_obj = lale_op_obj_base(**all_new_hyperparams) from lale.lib.sklearn import Pipeline as LaleSKPipelineWrapper # If this is a scklearn pipeline, then we want to discard the outer wrapper # and just return a lale pipeline if isinstance(lale_op_obj, LaleSKPipelineWrapper): # type: ignore return lale_op_obj.shallow_impl._pipeline # at this point, the object's hyper-parameters are modified as needed # and our wrapper is initialized with the correct hyperparameters. # Now we need to replace the wrapper impl with our (possibly modified) # sklearn object cl_shallow_impl = lale_op_obj.shallow_impl if hasattr(cl_shallow_impl, "_wrapped_model"): cl_shallow_impl._wrapped_model = sklearn_obj else: lale_op_obj._impl = sklearn_obj lale_op_obj._impl_class_ = sklearn_obj.__class__ return lale_op_obj def import_from_sklearn(sklearn_obj: Any, fitted: bool = True, in_place: bool = False): """ This method take an object and tries to wrap sklearn objects (at the top level or contained within hyperparameters of other sklearn objects). It will modify the object to add in the appropriate lale wrappers. It may also return a wrapper or different object than given. :param sklearn_obj: the object that we are going to try and wrap :param fitted: should we return a TrainedOperator :param in_place: should we try to mutate what we can in place, or should we aggressively deepcopy everything :return: The wrapped object (or the input object if we could not wrap it) """ obj = sklearn_obj if in_place: obj = sklearn_obj else: obj = copy.deepcopy(sklearn_obj) return _import_from_sklearn_inplace_helper(obj, fitted=fitted, is_nested=False) def import_from_sklearn_pipeline(sklearn_pipeline: Any, fitted: bool = True): """ Note: Same as import_from_sklearn. This alternative name exists for backwards compatibility. This method take an object and tries to wrap sklearn objects (at the top level or contained within hyperparameters of other sklearn objects). It will modify the object to add in the appropriate lale wrappers. It may also return a wrapper or different object than given. :param sklearn_pipeline: the object that we are going to try and wrap :param fitted: should we return a TrainedOperator :return: The wrapped object (or the input object if we could not wrap it) """ op = import_from_sklearn(sklearn_pipeline, fitted=fitted, in_place=False) from typing import cast from lale.operators import TrainableOperator # simplify using the returned value in the common case return cast(TrainableOperator, op) class val_wrapper: """This is used to wrap values that cause problems for hyper-optimizer backends lale will unwrap these when given them as the value of a hyper-parameter""" def __init__(self, base): self._base = base def unwrap_self(self): return self._base @classmethod def unwrap(cls, obj): if isinstance(obj, cls): return cls.unwrap(obj.unwrap_self()) else: return obj def append_batch(data, batch_data): if data is None: return batch_data elif isinstance(data, np.ndarray): if isinstance(batch_data, np.ndarray): if len(data.shape) == 1 and len(batch_data.shape) == 1: return np.concatenate([data, batch_data]) else: return np.vstack((data, batch_data)) elif isinstance(data, tuple): X, y = data if isinstance(batch_data, tuple): batch_X, batch_y = batch_data X = append_batch(X, batch_X) y = append_batch(y, batch_y) return X, y elif torch_installed and isinstance(data, torch.Tensor): if isinstance(batch_data, torch.Tensor): return torch.cat((data, batch_data)) elif isinstance(data, (pd.Series, pd.DataFrame)): return pd.concat([data, batch_data], axis=0) try: import h5py if isinstance(data, h5py.File): if isinstance(batch_data, tuple): batch_X, batch_y = batch_data except ModuleNotFoundError: pass raise ValueError( f"{type(data)} is unsupported. Supported types are np.ndarray, torch.Tensor and h5py file" ) def create_data_loader( X: Any, y: Any = None, batch_size: int = 1, num_workers: int = 0, shuffle: bool = True, ): """A function that takes a dataset as input and outputs a Pytorch dataloader. Parameters ---------- X : Input data. The formats supported are Pandas DataFrame, Numpy array, a sparse matrix, torch.tensor, torch.utils.data.Dataset, path to a HDF5 file, lale.util.batch_data_dictionary_dataset.BatchDataDict, a Python dictionary of the format `{"dataset": torch.utils.data.Dataset, "collate_fn":collate_fn for torch.utils.data.DataLoader}` y : Labels., optional Supported formats are Numpy array or Pandas series, by default None batch_size : int, optional Number of samples in each batch, by default 1 num_workers : int, optional Number of workers used by the data loader, by default 0 shuffle: boolean, optional, default True Whether to use SequentialSampler or RandomSampler for creating batches Returns ------- torch.utils.data.DataLoader Raises ------ TypeError Raises a TypeError if the input format is not supported. """ from torch.utils.data import DataLoader, Dataset, TensorDataset from lale.util.batch_data_dictionary_dataset import BatchDataDict from lale.util.hdf5_to_torch_dataset import HDF5TorchDataset from lale.util.numpy_torch_dataset import NumpyTorchDataset, numpy_collate_fn from lale.util.pandas_torch_dataset import PandasTorchDataset, pandas_collate_fn collate_fn = None worker_init_fn = None if isinstance(X, Dataset) and not isinstance(X, BatchDataDict): dataset = X elif isinstance(X, pd.DataFrame): dataset = PandasTorchDataset(X, y) collate_fn = pandas_collate_fn elif isinstance(X, scipy.sparse.csr_matrix): # unfortunately, NumpyTorchDataset won't accept a subclass of np.ndarray X = X.toarray() # type: ignore if isinstance(y, lale.datasets.data_schemas.NDArrayWithSchema): y = y.view(np.ndarray) dataset = NumpyTorchDataset(X, y) collate_fn = numpy_collate_fn elif isinstance(X, np.ndarray): # unfortunately, NumpyTorchDataset won't accept a subclass of np.ndarray if isinstance(X, lale.datasets.data_schemas.NDArrayWithSchema): X = X.view(np.ndarray) if isinstance(y, lale.datasets.data_schemas.NDArrayWithSchema): y = y.view(np.ndarray) dataset = NumpyTorchDataset(X, y) collate_fn = numpy_collate_fn elif isinstance(X, str): # Assume that this is path to hdf5 file dataset = HDF5TorchDataset(X) elif isinstance(X, BatchDataDict): dataset = X def my_collate_fn(batch): return batch[ 0 ] # because BatchDataDict's get_item returns a batch, so no collate is required. return DataLoader( dataset, batch_size=1, collate_fn=my_collate_fn, shuffle=shuffle ) elif isinstance(X, dict): # Assumed that it is data indexed by batch number if "dataset" in X: dataset = X["dataset"] collate_fn = X.get("collate_fn", None) worker_init_fn = getattr(dataset, "worker_init_fn", None) else: return [X] elif isinstance(X, torch.Tensor) and y is not None: if isinstance(y, np.ndarray): y = torch.from_numpy(y) dataset = TensorDataset(X, y) elif isinstance(X, torch.Tensor): dataset = TensorDataset(X) else: raise TypeError( f"Can not create a data loader for a dataset with type {type(X)}" ) return DataLoader( dataset, batch_size=batch_size, collate_fn=collate_fn, num_workers=num_workers, worker_init_fn=worker_init_fn, shuffle=shuffle, ) def write_batch_output_to_file( file_obj, file_path, total_len, batch_idx, batch_X, batch_y, batch_out_X, batch_out_y, ): if file_obj is None and file_path is None: raise ValueError("Only one of the file object or file path can be None.") if file_obj is None: import h5py file_obj = h5py.File(file_path, "w") # estimate the size of the dataset based on the first batch output size transform_ratio = int(len(batch_out_X) / len(batch_X)) if len(batch_out_X.shape) == 1: h5_data_shape = (transform_ratio * total_len,) elif len(batch_out_X.shape) == 2: h5_data_shape = (transform_ratio * total_len, batch_out_X.shape[1]) elif len(batch_out_X.shape) == 3: h5_data_shape = ( transform_ratio * total_len, batch_out_X.shape[1], batch_out_X.shape[2], ) else: raise ValueError( "batch_out_X is expected to be a 1-d, 2-d or 3-d array. Any other data types are not handled." ) dataset = file_obj.create_dataset( name="X", shape=h5_data_shape, chunks=True, compression="gzip" ) if batch_out_y is None and batch_y is not None: batch_out_y = batch_y if batch_out_y is not None: if len(batch_out_y.shape) == 1: h5_labels_shape = (transform_ratio * total_len,) elif len(batch_out_y.shape) == 2: h5_labels_shape = (transform_ratio * total_len, batch_out_y.shape[1]) else: raise ValueError( "batch_out_y is expected to be a 1-d or 2-d array. Any other data types are not handled." ) dataset = file_obj.create_dataset( name="y", shape=h5_labels_shape, chunks=True, compression="gzip" ) dataset = file_obj["X"] dataset[ batch_idx * len(batch_out_X) : (batch_idx + 1) * len(batch_out_X) ] = batch_out_X if batch_out_y is not None or batch_y is not None: labels = file_obj["y"] if batch_out_y is not None: labels[ batch_idx * len(batch_out_y) : (batch_idx + 1) * len(batch_out_y) ] = batch_out_y else: labels[batch_idx * len(batch_y) : (batch_idx + 1) * len(batch_y)] = batch_y return file_obj def add_missing_values(orig_X, missing_rate=0.1, seed=None): # see scikit-learn.org/stable/auto_examples/impute/plot_missing_values.html n_samples, n_features = orig_X.shape n_missing_samples = int(n_samples * missing_rate) if seed is None: rng = np.random.RandomState() else: rng = np.random.RandomState(seed) missing_samples = np.zeros(n_samples, dtype=bool) missing_samples[:n_missing_samples] = True rng.shuffle(missing_samples) missing_features = rng.randint(0, n_features, n_missing_samples) missing_X = orig_X.copy() if isinstance(missing_X, np.ndarray): missing_X[missing_samples, missing_features] = np.nan else: assert isinstance(missing_X, pd.DataFrame) i_missing_sample = 0 for i_sample in range(n_samples): if missing_samples[i_sample]: i_feature = missing_features[i_missing_sample] i_missing_sample += 1 missing_X.iloc[i_sample, i_feature] = np.nan return missing_X # helpers for manipulating (extended) sklearn style paths. # documentation of the path format is part of the operators module docstring def partition_sklearn_params( d: Dict[str, Any] ) -> Tuple[Dict[str, Any], Dict[str, Dict[str, Any]]]: sub_parts: Dict[str, Dict[str, Any]] = {} main_parts: Dict[str, Any] = {} for k, v in d.items(): ks = k.split("__", 1) if len(ks) == 1: assert k not in main_parts main_parts[k] = v else: assert len(ks) == 2 bucket: Dict[str, Any] = {} group: str = ks[0] param: str = ks[1] if group in sub_parts: bucket = sub_parts[group] else: sub_parts[group] = bucket assert param not in bucket bucket[param] = v return (main_parts, sub_parts) def partition_sklearn_choice_params(d: Dict[str, Any]) -> Tuple[int, Dict[str, Any]]: discriminant_value: int = -1 choice_parts: Dict[str, Any] = {} for k, v in d.items(): if k == discriminant_name: assert discriminant_value == -1 discriminant_value = int(v) else: k_rest = unnest_choice(k) choice_parts[k_rest] = v assert discriminant_value != -1 return (discriminant_value, choice_parts) DUMMY_SEARCH_SPACE_GRID_PARAM_NAME: str = "$" discriminant_name: str = "?" choice_prefix: str = "?" structure_type_name: str = "#" structure_type_list: str = "list" structure_type_tuple: str = "tuple" structure_type_dict: str = "dict" def get_name_and_index(name: str) -> Tuple[str, int]: """given a name of the form "name@i", returns (name, i) if given a name of the form "name", returns (name, 0) """ splits = name.split("@", 1) if len(splits) == 1: return splits[0], 0 else: return splits[0], int(splits[1]) def make_degen_indexed_name(name, index): return f"{name}@{index}" def make_indexed_name(name, index): if index == 0: return name else: return f"{name}@{index}" def make_array_index_name(index, is_tuple: bool = False): sep = "##" if is_tuple else "#" return f"{sep}{str(index)}" def is_numeric_structure(structure_type: str): if structure_type in ["list", "tuple"]: return True elif structure_type == "dict": return False else: assert False, f"Unknown structure type {structure_type} found" V = TypeVar("V") def nest_HPparam(name: str, key: str): if key == DUMMY_SEARCH_SPACE_GRID_PARAM_NAME: # we can get rid of the dummy now, since we have a name for it return name return name + "__" + key def nest_HPparams(name: str, grid: Mapping[str, V]) -> Dict[str, V]: return {(nest_HPparam(name, k)): v for k, v in grid.items()} def nest_all_HPparams( name: str, grids: Iterable[Mapping[str, V]] ) -> List[Dict[str, V]]: """Given the name of an operator in a pipeline, this transforms every key(parameter name) in the grids to use the operator name as a prefix (separated by __). This is the convention in scikit-learn pipelines. """ return [nest_HPparams(name, grid) for grid in grids] def nest_choice_HPparam(key: str): return choice_prefix + key def nest_choice_HPparams(grid: Mapping[str, V]) -> Dict[str, V]: return {(nest_choice_HPparam(k)): v for k, v in grid.items()} def nest_choice_all_HPparams(grids: Iterable[Mapping[str, V]]) -> List[Dict[str, V]]: """this transforms every key(parameter name) in the grids to be nested under a choice, using a ? as a prefix (separated by __). This is the convention in scikit-learn pipelines. """ return [nest_choice_HPparams(grid) for grid in grids] def unnest_choice(k: str) -> str: assert k.startswith(choice_prefix) return k[len(choice_prefix) :] def unnest_HPparams(k: str) -> List[str]: return k.split("__") def are_hyperparameters_equal(hyperparam1, hyperparam2): if isinstance( hyperparam1, np.ndarray ): # hyperparam2 is from schema default, so it may not always be an array return np.all(hyperparam1 == hyperparam2) else: return hyperparam1 == hyperparam2 def _is_ast_subscript(expr): return isinstance(expr, ast.Subscript) def _is_ast_attribute(expr): return isinstance(expr, ast.Attribute) def _is_ast_constant(expr): return isinstance(expr, ast.Constant) def _is_ast_subs_or_attr(expr): return isinstance(expr, (ast.Subscript, ast.Attribute)) def _is_ast_call(expr): return isinstance(expr, ast.Call) def _is_ast_name(expr): return isinstance(expr, ast.Name) def _ast_func_id(expr): if isinstance(expr, ast.Name): return expr.id else: raise ValueError("function name expected") def _is_df(df): return _is_pandas_df(df) or _is_spark_df(df) def _is_pandas_series(df): return isinstance(df, pd.Series) def _is_pandas_df(df): return isinstance(df, pd.DataFrame) def _is_pandas(df): return isinstance(df, (pd.Series, pd.DataFrame)) def _is_spark_df(df): if spark_installed: return isinstance(df, lale.datasets.data_schemas.SparkDataFrameWithIndex) else: return False def _is_spark_df_without_index(df): if spark_installed: return isinstance(df, spark_df) and not _is_spark_df(df) else: return False def _ensure_pandas(df) -> pd.DataFrame: if _is_spark_df(df): return df.toPandas() assert _is_pandas(df), type(df) return df def _get_subscript_value(subscript_expr): if isinstance(subscript_expr.slice, ast.Constant): # for Python 3.9 subscript_value = subscript_expr.slice.value else: subscript_value = subscript_expr.slice.value.s # type: ignore return subscript_value class GenSym: def __init__(self, names: Set[str]): self._names = names def __call__(self, prefix): if prefix in self._names: suffix = 0 while True: result = f"{prefix}_{suffix}" if result not in self._names: break suffix += 1 else: result = prefix self._names |= {result} return result def get_sklearn_estimator_name() -> str: """Some higher order sklearn operators changed the name of the nested estimatator in later versions. This returns the appropriate version dependent paramater name """ from packaging import version import lale.operators if lale.operators.sklearn_version < version.Version("1.2"): return "base_estimator" else: return "estimator" def get_estimator_param_name_from_hyperparams(hyperparams): be = hyperparams.get("base_estimator", "deprecated") if be == "deprecated" or (be is None and "estimator" in hyperparams): return "estimator" else: return "base_estimator"
46,947
34.459215
150
py
lale
lale-master/lale/util/pandas_torch_dataset.py
# Copyright 2021 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import pandas as pd try: from torch.utils.data import Dataset except ModuleNotFoundError as exc: raise ModuleNotFoundError( """Your Python environment does not have torch installed. You can install it with pip install torch or with pip install 'lale[full]'""" ) from exc class PandasTorchDataset(Dataset): """Pytorch Dataset subclass that takes a pandas DataFrame and an optional label pandas Series.""" def __init__(self, X, y=None): """X and y are the dataset and labels respectively. Parameters ---------- X : pandas DataFrame Two dimensional dataset of input features. y : pandas Series Labels """ self.X = X self.y = y def __len__(self): return self.X.shape[0] def __getitem__(self, idx): if self.y is not None: return self.X.iloc[idx], self.y.iloc[idx] else: return self.X.iloc[idx] def get_data(self): if self.y is None: return self.X else: return self.X, self.y def pandas_collate_fn(batch): return_X = None return_y = None for item in batch: if isinstance(item, tuple): if return_X is None: return_X = [item[0].to_dict()] else: return_X.append(item[0].to_dict()) if return_y is None: return_y = [item[1]] else: return_y.append(item[1]) else: if return_X is None: return_X = [item.to_dict()] else: return_X.append(item.to_dict()) if return_y is not None: return (pd.DataFrame(return_X), pd.Series(return_y)) else: return pd.DataFrame(return_X)
2,475
28.831325
101
py
lale
lale-master/lale/util/hdf5_to_torch_dataset.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. try: from torch.utils.data import Dataset except ModuleNotFoundError as import_exc: raise ModuleNotFoundError( """Your Python environment does not have torch installed. You can install it with pip install torch or with pip install 'lale[full]'""" ) from import_exc try: import h5py except ModuleNotFoundError as import_exc: raise ModuleNotFoundError( """Your Python environment does not have h5py installed. You can install it with pip install h5py or with pip install 'lale[full]'""" ) from import_exc class HDF5TorchDataset(Dataset): """Pytorch Dataset subclass that takes a hdf5 file pointer.""" def __init__(self, file_path): """. Parameters ---------- file : file is an object of class h5py.File """ self.file_path = file_path h5_file = h5py.File(file_path) self.length = h5_file["X"].shape[0] def __len__(self): return self.length def __getitem__(self, idx): with h5py.File(self.file_path) as h5_file: X = h5_file["X"] try: y = h5_file["y"] except KeyError: y = None if y is None: element = X[idx] else: element = X[idx], y[idx] return element def get_data(self): with h5py.File(self.file_path) as h5_file: X = h5_file["X"][:] try: y = h5_file["y"][:] except KeyError: y = None if y is None: return X else: return X, y
2,412
30.75
89
py
lale
lale-master/lale/util/batch_data_dictionary_dataset.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from torch.utils.data import Dataset class BatchDataDict(Dataset): """Pytorch Dataset subclass that takes a dictionary of format {'<batch_idx>': <batch_data>}.""" def __init__(self, X, y=None): """X is the dictionary dataset and y is ignored. Parameters ---------- X : dict Dictionary of format {'<batch_idx>': <batch_data>} y : None Ignored. """ self.data_dict = X def __len__(self): return len(self.data_dict) def __getitem__(self, idx): # This returns the batch at idx instead of a single element. return self.data_dict[idx]
1,227
30.487179
99
py
lale
lale-master/lale/util/numpy_to_torch_dataset.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np try: from torch.utils.data import Dataset except ModuleNotFoundError as exc: raise ModuleNotFoundError( """Your Python environment does not have torch installed. You can install it with pip install torch or with pip install 'lale[full]'""" ) from exc class NumpyTorchDataset(Dataset): """Pytorch Dataset subclass that takes a numpy array and an optional label array.""" def __init__(self, X, y=None): """X and y are the dataset and labels respectively. Parameters ---------- X : numpy array Two dimensional dataset of input features. y : numpy array Labels """ self.X = X self.y = y def __len__(self): return self.X.shape[0] def __getitem__(self, idx): if self.y is not None: return self.X[idx], self.y[idx] else: return self.X[idx] def get_data(self): if self.y is None: return self.X else: return self.X, self.y def numpy_collate_fn(batch): return_X = None return_y = None for item in batch: if isinstance(item, tuple): if return_X is None: return_X = item[0] else: return_X = np.vstack((return_X, item[0])) if return_y is None: return_y = item[1] else: return_y = np.vstack((return_y, item[1])) # type: ignore else: if return_X is None: return_X = item else: return_X = np.vstack((return_X, item)) # type: ignore if return_y is not None: if len(return_y.shape) > 1 and return_y.shape[1] == 1: return_y = np.reshape(return_y, (len(return_y),)) return return_X, return_y else: return return_X
2,557
29.452381
89
py
lale
lale-master/lale/util/numpy_torch_dataset.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np try: from torch.utils.data import Dataset except ModuleNotFoundError as exc: raise ModuleNotFoundError( """Your Python environment does not have torch installed. You can install it with pip install torch or with pip install 'lale[full]'""" ) from exc class NumpyTorchDataset(Dataset): """Pytorch Dataset subclass that takes a numpy array and an optional label array.""" def __init__(self, X, y=None): """X and y are the dataset and labels respectively. Parameters ---------- X : numpy array Two dimensional dataset of input features. y : numpy array Labels """ self.X = X self.y = y def __len__(self): return self.X.shape[0] def __getitem__(self, idx): if self.y is not None: return self.X[idx], self.y[idx] else: return self.X[idx] def get_data(self): if self.y is None: return self.X else: return self.X, self.y def numpy_collate_fn(batch): return_X = None return_y = None for item in batch: if isinstance(item, tuple): if return_X is None: return_X = item[0] else: return_X = np.vstack((return_X, item[0])) if return_y is None: return_y = item[1] else: return_y = np.vstack((return_y, item[1])) # type: ignore else: if return_X is None: return_X = item else: return_X = np.vstack((return_X, item)) # type: ignore if return_y is not None: if len(return_y.shape) > 1 and return_y.shape[1] == 1: return_y = np.reshape(return_y, (len(return_y),)) return return_X, return_y else: return return_X
2,557
29.452381
89
py
lale
lale-master/lale/util/pandas_to_torch_dataset.py
# Copyright 2021 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import pandas as pd try: from torch.utils.data import Dataset except ModuleNotFoundError as exc: raise ModuleNotFoundError( """Your Python environment does not have torch installed. You can install it with pip install torch or with pip install 'lale[full]'""" ) from exc class PandasTorchDataset(Dataset): """Pytorch Dataset subclass that takes a pandas DataFrame and an optional label pandas Series.""" def __init__(self, X, y=None): """X and y are the dataset and labels respectively. Parameters ---------- X : pandas DataFrame Two dimensional dataset of input features. y : pandas Series Labels """ self.X = X self.y = y def __len__(self): return self.X.shape[0] def __getitem__(self, idx): if self.y is not None: return self.X.iloc[idx], self.y.iloc[idx] else: return self.X.iloc[idx] def get_data(self): if self.y is None: return self.X else: return self.X, self.y def pandas_collate_fn(batch): return_X = None return_y = None for item in batch: if isinstance(item, tuple): if return_X is None: return_X = [item[0].to_dict()] else: return_X.append(item[0].to_dict()) if return_y is None: return_y = [item[1]] else: return_y.append(item[1]) else: if return_X is None: return_X = [item[0].to_dict()] else: return_X.append(item[0].to_dict()) if return_y is not None: return (pd.DataFrame(return_X), pd.Series(return_y)) else: return pd.DataFrame(return_X)
2,481
28.903614
101
py
lale
lale-master/lale/datasets/data_schemas.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from typing import Any, List, Optional, Tuple, Type, Union import numpy as np from numpy import issubdtype, ndarray from pandas import DataFrame, Series from pandas.core.groupby import DataFrameGroupBy, SeriesGroupBy from scipy.sparse import csr_matrix import lale.type_checking from lale.helpers import _is_spark_df from lale.type_checking import JSON_TYPE try: import torch from torch import Tensor torch_installed = True except ImportError: torch_installed = False try: import py4j.protocol from pyspark.sql import DataFrame as SparkDataFrame from pyspark.sql import GroupedData as SparkGroupedData spark_installed = True except ImportError: spark_installed = False # See instructions for subclassing numpy ndarray: # https://docs.scipy.org/doc/numpy/user/basics.subclassing.html class NDArrayWithSchema(ndarray): def __new__( cls, shape, dtype=float, buffer=None, offset=0, strides=None, order=None, json_schema=None, table_name=None, ): result = super( # pylint:disable=too-many-function-args NDArrayWithSchema, cls ).__new__( cls, shape, dtype, buffer, offset, strides, order # type: ignore ) result.json_schema = json_schema result.table_name = table_name return result def __array_finalize__(self, obj): if obj is None: return self.json_schema = getattr(obj, "json_schema", None) self.table_name = getattr(obj, "table_name", None) # See instructions for subclassing pandas DataFrame: # https://pandas.pydata.org/pandas-docs/stable/development/extending.html#extending-subclassing-pandas class DataFrameWithSchema(DataFrame): _internal_names = DataFrame._internal_names + ["json_schema", "table_name"] _internal_names_set = set(_internal_names) @property def _constructor(self): return DataFrameWithSchema class SeriesWithSchema(Series): _internal_names = DataFrame._internal_names + [ "json_schema", "table_name", "folds_for_monoid", ] _internal_names_set = set(_internal_names) @property def _constructor(self): return SeriesWithSchema if spark_installed: def _gen_index_name(df, cpt=None): name = f"index{cpt if cpt is not None else ''}" if name in df.columns: return _gen_index_name(df, cpt=cpt + 1 if cpt is not None else 0) else: return name class SparkDataFrameWithIndex(SparkDataFrame): # type: ignore def __init__(self, df, index_names=None): if index_names is not None and len(index_names) == 1: index_name = index_names[0] elif index_names is None or len(index_names) == 0: index_name = _gen_index_name(df) index_names = [index_name] else: index_name = None if index_name is not None and index_name not in df.columns: df_with_index = ( df.rdd.zipWithIndex() .map(lambda row: row[0] + (row[1],)) .toDF(df.columns + [index_name]) ) else: df_with_index = df table_name = get_table_name(df) if table_name is not None: df_with_index = df_with_index.alias(table_name) super().__init__(df_with_index._jdf, df_with_index.sql_ctx) self.index_name = index_name self.index_names = index_names for f in df.schema.fieldNames(): self.schema[f].metadata = df.schema[f].metadata def drop_indexes(self): result = self.drop(*self.index_names) result = add_table_name(result, get_table_name(self)) return result @property def columns_without_indexes(self): cols = list(self.columns) for name in self.index_names: cols.remove(name) return cols def toPandas(self, *args, **kwargs): df = super().toPandas(*args, **kwargs) return df.set_index(self.index_names) else: class SparkDataFrameWithIndex: # type: ignore def __init__(self, df, index_names=None) -> None: raise ValueError("pyspark is not installed") # type: ignore @property def index_name(self) -> Union[str, None]: raise ValueError("pyspark is not installed") # type: ignore @property def index_names(self) -> List[str]: raise ValueError("pyspark is not installed") # type: ignore def toPandas(self, *args, **kwargs) -> DataFrame: raise ValueError("pyspark is not installed") # type: ignore @property def schema(self) -> Any: raise ValueError("pyspark is not installed") # type: ignore def add_schema(obj, schema=None, raise_on_failure=False, recalc=False) -> Any: from lale.settings import disable_data_schema_validation if disable_data_schema_validation: return obj if obj is None: return None if isinstance(obj, NDArrayWithSchema): result = obj elif isinstance(obj, ndarray): result = obj.view(NDArrayWithSchema) elif isinstance(obj, SeriesWithSchema): result = obj elif isinstance(obj, Series): result = SeriesWithSchema(obj) elif isinstance(obj, DataFrameWithSchema): result = obj elif isinstance(obj, DataFrame): result = DataFrameWithSchema(obj) elif is_list_tensor(obj): obj = np.array(obj) result = obj.view(NDArrayWithSchema) elif raise_on_failure: raise ValueError(f"unexpected type(obj) {type(obj)}") else: return obj if recalc: setattr(result, "json_schema", None) if getattr(result, "json_schema", None) is None: if schema is None: setattr(result, "json_schema", to_schema(obj)) else: lale.type_checking.validate_is_schema(schema) setattr(result, "json_schema", schema) return result def add_schema_adjusting_n_rows(obj, schema): assert isinstance(obj, (ndarray, DataFrame, Series)), type(obj) assert schema.get("type", None) == "array", schema n_rows = obj.shape[0] mod_schema = {**schema, "minItems": n_rows, "maxItems": n_rows} result = add_schema(obj, mod_schema) return result def add_table_name(obj, name) -> Any: if obj is None: return None if name is None: return obj if spark_installed and isinstance(obj, SparkDataFrame): # alias method documentation: https://spark.apache.org/docs/latest/api/python/reference/api/pyspark.sql.DataFrame.alias.html # Python class DataFrame with method alias(self, alias): https://github.com/apache/spark/blob/master/python/pyspark/sql/dataframe.py # Scala type DataFrame: https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/package.scala # Scala class DataSet with method as(alias: String): https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/Dataset.scala o = obj.alias(name) for f in obj.schema.fieldNames(): o.schema[f].metadata = obj.schema[f].metadata if isinstance(obj, SparkDataFrameWithIndex): o = SparkDataFrameWithIndex(o, obj.index_names) return o if isinstance(obj, NDArrayWithSchema): result = obj.view(NDArrayWithSchema) if hasattr(obj, "json_schema"): result.json_schema = obj.json_schema elif isinstance(obj, ndarray): result = obj.view(NDArrayWithSchema) elif isinstance(obj, SeriesWithSchema): result = obj.copy(deep=False) if hasattr(obj, "json_schema"): result.json_schema = obj.json_schema elif isinstance(obj, Series): result = SeriesWithSchema(obj) elif isinstance(obj, DataFrameWithSchema): result = obj.copy(deep=False) if hasattr(obj, "json_schema"): result.json_schema = obj.json_schema elif isinstance(obj, DataFrame): result = DataFrameWithSchema(obj) elif is_list_tensor(obj): obj = np.array(obj) result = obj.view(NDArrayWithSchema) elif isinstance(obj, (DataFrameGroupBy, SeriesGroupBy)): result = obj elif spark_installed and isinstance(obj, SparkGroupedData): result = obj else: raise ValueError(f"unexpected type(obj) {type(obj)}") setattr(result, "table_name", name) return result def get_table_name(obj): if spark_installed and isinstance(obj, SparkDataFrame): # Python class DataFrame with field self._jdf: https://github.com/apache/spark/blob/master/python/pyspark/sql/dataframe.py # Scala type DataFrame: https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/package.scala # Scala class DataSet with field queryExecution: https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/Dataset.scala # Scala fields turn into Java nullary methods # Py4J exposes Java methods as Python methods # Scala class QueryExecution with field analyzed: LogicalPlan: https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/QueryExecution.scala spark_query = obj._jdf.queryExecution().analyzed() # type: ignore try: # calling spark_df.explain("extended") shows the analyzed contents # after spark_df.alias("foo"), analyzed contents should be SubqueryAlias # Scala class SuqueryAlias with field identifier: https://github.com/apache/spark/blob/master/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/logical/basicLogicalOperators.scala # str(..) converts the Java string into a Python string result = str(spark_query.identifier()) except py4j.protocol.Py4JError: result = None return result if isinstance( obj, ( NDArrayWithSchema, SeriesWithSchema, DataFrameWithSchema, DataFrameGroupBy, SeriesGroupBy, ), ) or (spark_installed and isinstance(obj, SparkGroupedData)): return getattr(obj, "table_name", None) return None def get_index_name(obj): result = None if spark_installed and isinstance(obj, SparkDataFrameWithIndex): result = obj.index_name elif isinstance( obj, ( SeriesWithSchema, DataFrameWithSchema, DataFrameGroupBy, SeriesGroupBy, ), ): result = obj.index.name return result def get_index_names(obj): result = None if spark_installed and isinstance(obj, SparkDataFrameWithIndex): result = obj.index_names elif isinstance( obj, ( SeriesWithSchema, DataFrameWithSchema, DataFrameGroupBy, SeriesGroupBy, ), ): result = obj.index.names return result def forward_metadata(old, new): new = add_table_name(new, get_table_name(old)) if isinstance(old, SparkDataFrameWithIndex): new = SparkDataFrameWithIndex(new, index_names=get_index_names(old)) return new def strip_schema(obj): if isinstance(obj, NDArrayWithSchema): result = np.array(obj) assert type(result) == ndarray # pylint:disable=unidiomatic-typecheck elif isinstance(obj, SeriesWithSchema): result = Series(obj) assert type(result) == Series # pylint:disable=unidiomatic-typecheck elif isinstance(obj, DataFrameWithSchema): result = DataFrame(obj) assert type(result) == DataFrame # pylint:disable=unidiomatic-typecheck else: result = obj return result def _dtype_to_schema(typ) -> JSON_TYPE: result: JSON_TYPE if typ is bool or issubdtype(typ, np.bool_): result = {"type": "boolean"} elif issubdtype(typ, np.unsignedinteger): result = {"type": "integer", "minimum": 0} elif issubdtype(typ, np.integer): result = {"type": "integer"} elif issubdtype(typ, np.number): result = {"type": "number"} elif issubdtype(typ, np.string_) or issubdtype(typ, np.unicode_): result = {"type": "string"} elif isinstance(typ, np.dtype): if typ.fields: props = {k: _dtype_to_schema(t) for k, t in typ.fields.items()} result = {"type": "object", "properties": props} elif typ.shape: result = _shape_and_dtype_to_schema(typ.shape, typ.subdtype) elif issubdtype(typ, np.object_): result = {"type": "string"} else: assert False, f"unexpected dtype {typ}" else: assert False, f"unexpected non-dtype {typ}" return result def dtype_to_schema(typ) -> JSON_TYPE: result = _dtype_to_schema(typ) lale.type_checking.validate_is_schema(result) return result def _shape_and_dtype_to_schema(shape, dtype) -> JSON_TYPE: result = _dtype_to_schema(dtype) for dim in reversed(shape): result = {"type": "array", "minItems": dim, "maxItems": dim, "items": result} return result def shape_and_dtype_to_schema(shape, dtype) -> JSON_TYPE: result = _shape_and_dtype_to_schema(shape, dtype) lale.type_checking.validate_is_schema(result) return result def list_tensor_to_shape_and_dtype(ls) -> Optional[Tuple[Tuple[int, ...], Type]]: if isinstance(ls, (int, float, str)): return ((), type(ls)) if isinstance(ls, list): sub_result: Any = "Any" for item in ls: item_result = list_tensor_to_shape_and_dtype(item) if item_result is None: return None if sub_result == "Any": sub_result = item_result elif sub_result != item_result: return None if sub_result == "Any" and len(ls) == 0: return ((len(ls),) + (), int) sub_shape, sub_dtype = sub_result return ((len(ls),) + sub_shape, sub_dtype) return None def is_list_tensor(obj) -> bool: if isinstance(obj, list): shape_and_dtype = list_tensor_to_shape_and_dtype(obj) return shape_and_dtype is not None return False def _list_tensor_to_schema(ls) -> Optional[JSON_TYPE]: shape_and_dtype = list_tensor_to_shape_and_dtype(ls) if shape_and_dtype is None: return None result = _shape_and_dtype_to_schema(*shape_and_dtype) return result def list_tensor_to_schema(ls) -> Optional[JSON_TYPE]: result = _list_tensor_to_schema(ls) if result is None: return None lale.type_checking.validate_is_schema(result) return result def _ndarray_to_schema(array) -> JSON_TYPE: assert isinstance(array, ndarray) if ( isinstance(array, NDArrayWithSchema) and hasattr(array, "json_schema") and array.json_schema is not None ): return array.json_schema return _shape_and_dtype_to_schema(array.shape, array.dtype) def ndarray_to_schema(array) -> JSON_TYPE: result = _ndarray_to_schema(array) lale.type_checking.validate_is_schema(result) return result def _csr_matrix_to_schema(matrix) -> JSON_TYPE: assert isinstance(matrix, csr_matrix) result = _shape_and_dtype_to_schema(matrix.shape, matrix.dtype) result["isSparse"] = {} # true schema return result def csr_matrix_to_schema(matrix) -> JSON_TYPE: result = _csr_matrix_to_schema(matrix) lale.type_checking.validate_is_schema(result) return result def _dataframe_to_schema(df) -> JSON_TYPE: assert isinstance(df, DataFrame) if ( isinstance(df, DataFrameWithSchema) and hasattr(df, "json_schema") and df.json_schema is not None ): return df.json_schema n_rows, n_columns = df.shape df_dtypes = df.dtypes assert n_columns == len(df.columns) and n_columns == len(df_dtypes) items = [ {"description": str(col), **_dtype_to_schema(df_dtypes[col])} for col in df.columns ] result = { "type": "array", "minItems": n_rows, "maxItems": n_rows, "items": { "type": "array", "minItems": n_columns, "maxItems": n_columns, "items": items, }, } return result def dataframe_to_schema(df) -> JSON_TYPE: result = _dataframe_to_schema(df) lale.type_checking.validate_is_schema(result) return result def _series_to_schema(series) -> JSON_TYPE: assert isinstance(series, Series) if ( isinstance(series, SeriesWithSchema) and hasattr(series, "json_schema") and series.json_schema is not None ): return series.json_schema (n_rows,) = series.shape result = { "type": "array", "minItems": n_rows, "maxItems": n_rows, "items": {"description": str(series.name), **_dtype_to_schema(series.dtype)}, } return result def series_to_schema(series) -> JSON_TYPE: result = _series_to_schema(series) lale.type_checking.validate_is_schema(result) return result def _torch_tensor_to_schema(tensor) -> JSON_TYPE: assert torch_installed, """Your Python environment does not have torch installed. You can install it with pip install torch or with pip install 'lale[full]'""" assert isinstance(tensor, Tensor) result: JSON_TYPE # https://pytorch.org/docs/stable/tensor_attributes.html#torch-dtype if tensor.dtype == torch.bool: result = {"type": "boolean"} elif tensor.dtype == torch.uint8: result = {"type": "integer", "minimum": 0, "maximum": 255} elif torch.is_floating_point(tensor): result = {"type": "number"} else: result = {"type": "integer"} for dim in reversed(tensor.shape): result = {"type": "array", "minItems": dim, "maxItems": dim, "items": result} return result def torch_tensor_to_schema(tensor) -> JSON_TYPE: result = _torch_tensor_to_schema(tensor) lale.type_checking.validate_is_schema(result) return result def is_liac_arff(obj) -> bool: expected_types = { "description": str, "relation": str, "attributes": list, "data": list, } if not isinstance(obj, dict): return False for k, t in expected_types.items(): if k not in obj or not isinstance(obj[k], t): return False return True def _liac_arff_to_schema(larff) -> JSON_TYPE: assert is_liac_arff( larff ), """Your Python environment might contain an 'arff' package different from 'liac-arff'. You can install it with pip install 'liac-arff>=2.4.0' or with pip install 'lale[full]'""" n_rows, n_columns = len(larff["data"]), len(larff["attributes"]) def larff_type_to_schema(larff_type) -> JSON_TYPE: if isinstance(larff_type, str): a2j = { "numeric": "number", "real": "number", "integer": "integer", "string": "string", } return {"type": a2j[larff_type.lower()]} assert isinstance(larff_type, list) return {"enum": [*larff_type]} items = [ {"description": attr[0], **larff_type_to_schema(attr[1])} for attr in larff["attributes"] ] result = { "type": "array", "minItems": n_rows, "maxItems": n_rows, "items": { "type": "array", "minItems": n_columns, "maxItems": n_columns, "items": items, }, } return result def liac_arff_to_schema(larff) -> JSON_TYPE: result = _liac_arff_to_schema(larff) lale.type_checking.validate_is_schema(result) return result def make_optional_schema(schema: JSON_TYPE) -> JSON_TYPE: return {"anyOf": [schema, {"enum": [None]}]} def _spark_df_to_schema(df) -> JSON_TYPE: assert spark_installed, """Your Python environment does not have spark installed. You can install it with pip install pyspark """ assert isinstance(df, SparkDataFrameWithIndex) import pyspark.sql.types as stypes from pyspark.sql.types import StructField, StructType def maybe_make_optional(schema: JSON_TYPE, is_option: bool) -> JSON_TYPE: if is_option: return make_optional_schema(schema) return schema def spark_datatype_to_json_schema(dtype: stypes.DataType) -> JSON_TYPE: if isinstance(dtype, stypes.ArrayType): return { "type": "array", "items": maybe_make_optional( spark_datatype_to_json_schema(dtype.elementType), dtype.containsNull ), } if isinstance(dtype, stypes.BooleanType): return {"type": "boolean"} if isinstance(dtype, stypes.DoubleType): return {"type": "number"} if isinstance(dtype, stypes.FloatType): return {"type": "number"} if isinstance(dtype, stypes.IntegerType): return {"type": "integer"} if isinstance(dtype, stypes.LongType): return {"type": "integer"} if isinstance(dtype, stypes.ShortType): return {"type": "integer"} if isinstance(dtype, stypes.NullType): return {"enum": [None]} if isinstance(dtype, stypes.StringType): return {"type": "string"} return {} def spark_struct_field_to_json_schema(f: StructField) -> JSON_TYPE: type_schema = spark_datatype_to_json_schema(f.dataType) result = maybe_make_optional(type_schema, f.nullable) if f.name is not None: result["description"] = f.name return result def spark_struct_to_json_schema( s: StructType, index_names, table_name: Optional[str] = None ) -> JSON_TYPE: items = [ spark_struct_field_to_json_schema(f) for f in s if f.name not in index_names ] num_items = len(items) result = { "type": "array", "items": { "type": "array", "description": "rows", "minItems": num_items, "maxItems": num_items, "items": items, }, } if table_name is not None: result["description"] = table_name return result return spark_struct_to_json_schema(df.schema, df.index_names, get_table_name(df)) def spark_df_to_schema(df) -> JSON_TYPE: result = _spark_df_to_schema(df) lale.type_checking.validate_is_schema(result) return result def _to_schema(obj) -> JSON_TYPE: result = None if obj is None: result = {"enum": [None]} elif isinstance(obj, ndarray): result = _ndarray_to_schema(obj) elif isinstance(obj, csr_matrix): result = _csr_matrix_to_schema(obj) elif isinstance(obj, DataFrame): result = _dataframe_to_schema(obj) elif isinstance(obj, Series): result = _series_to_schema(obj) elif torch_installed and isinstance(obj, Tensor): result = _torch_tensor_to_schema(obj) elif is_liac_arff(obj): result = _liac_arff_to_schema(obj) elif isinstance(obj, list): result = _list_tensor_to_schema(obj) elif _is_spark_df(obj): result = _spark_df_to_schema(obj) elif lale.type_checking.is_schema(obj): result = obj # Does not need to validate again the schema return result # type: ignore if result is None: raise ValueError(f"to_schema(obj), type {type(obj)}, value {obj}") return result def to_schema(obj) -> JSON_TYPE: result = _to_schema(obj) lale.type_checking.validate_is_schema(result) return result
24,607
32.389417
205
py
lale
lale-master/lale/lib/rasl/batching.py
# Copyright 2019-2022 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np import lale.docstrings import lale.helpers import lale.operators from .task_graphs import PrioBatch, PrioResourceAware, PrioStep, fit_with_batches class _BatchingImpl: def __init__( self, operator=None, batch_size=32, shuffle=True, num_workers=0, inmemory=False, num_epochs=None, max_resident=None, scoring=None, progress_callback=None, partial_transform=False, priority="resource_aware", verbose=0, ): self.operator = operator self.batch_size = batch_size self.shuffle = shuffle self.num_workers = num_workers self.inmemory = inmemory self.num_epochs = num_epochs self.max_resident = max_resident self.scoring = scoring self.progress_callback = progress_callback self.partial_transform = partial_transform self.priority = priority self.verbose = verbose def fit(self, X, y=None, classes=None): if self.operator is None: raise ValueError("The pipeline object can't be None at the time of fit.") if hasattr(X, "__next__") and hasattr( X, "__iter__" ): # allow an iterable that is not a torch data loader assert y is None, "When X is an Iterable, y should be None" data_loader = X else: try: from torch.utils.data import DataLoader except ImportError as exc: raise ImportError( """Batching uses Pytorch for data loading. It is not installed in the current environment, please install the package and try again.""" ) from exc if isinstance(X, DataLoader): assert ( y is None ), "When X is a torch.utils.data.DataLoader, y should be None" data_loader = X else: data_loader = lale.helpers.create_data_loader( X=X, y=y, batch_size=self.batch_size, num_workers=self.num_workers, shuffle=self.shuffle, ) if y is not None and classes is None: classes = np.unique(y) if self.priority == "batch": prio = PrioBatch() elif self.priority == "step": prio = PrioStep() else: prio = PrioResourceAware() self.operator = fit_with_batches( pipeline=self.operator, batches_train=data_loader, # type:ignore batches_valid=None, unique_class_labels=classes, # type:ignore max_resident=self.max_resident, prio=prio, partial_transform=self.partial_transform, scoring=self.scoring, progress_callback=self.progress_callback, verbose=self.verbose, ) return self def transform(self, X, y=None): if hasattr(X, "__next__") and hasattr( X, "__iter__" ): # allow an iterable that is not a torch data loader assert y is None, "When X is an Iterable, y should be None" data_loader = X else: try: from torch.utils.data import DataLoader except ImportError as exc: raise ImportError( """Batching uses Pytorch for data loading. It is not installed in the current environment, please install the package and try again.""" ) from exc if isinstance(X, DataLoader): assert ( y is None ), "When X is a torch.utils.data.DataLoader, y should be None" data_loader = X else: data_loader = lale.helpers.create_data_loader( X=X, y=y, batch_size=self.batch_size, num_workers=self.num_workers, shuffle=self.shuffle, ) op = self.operator assert op is not None transformed_data = op.transform_with_batches( data_loader, serialize=self.inmemory ) return transformed_data def predict(self, X, y=None): return self.transform(X, y) _input_fit_schema = { "description": "Input data schema for fit.", "type": "object", "required": ["X", "y"], "additionalProperties": False, "properties": { "X": { "description": "Features; the outer array is over samples.", "anyOf": [ { "type": "array", "items": { "anyOf": [ {"type": "number"}, {"type": "string"}, {"type": "boolean"}, ] }, }, { "type": "array", "items": { "type": "array", "items": { "anyOf": [ {"type": "number"}, {"type": "string"}, {"type": "boolean"}, ] }, }, }, {"type": "object"}, ], }, "y": { "anyOf": [ { "type": "array", "items": { "anyOf": [ {"type": "integer"}, {"type": "number"}, {"type": "string"}, ] }, }, {"enum": [None]}, ], }, "classes": { "anyOf": [ { "type": "array", "items": { "anyOf": [ {"type": "number"}, {"type": "string"}, {"type": "boolean"}, ] }, }, {"enum": [None]}, ], "description": """The total number of classes in the entire training dataset.""", }, }, } _input_predict_transform_schema = { # TODO: separate predict vs. transform "description": "Input data schema for predictions.", "type": "object", "required": ["X"], "additionalProperties": False, "properties": { "X": { "description": "Features; the outer array is over samples.", "anyOf": [ { "type": "array", "items": { "anyOf": [ {"type": "number"}, {"type": "string"}, {"type": "boolean"}, ] }, }, { "type": "array", "items": { "type": "array", "items": { "anyOf": [ {"type": "number"}, {"type": "string"}, {"type": "boolean"}, ] }, }, }, {}, ], }, "y": { "type": "array", "items": {"anyOf": [{"type": "integer"}, {"type": "number"}]}, }, }, } _output_schema = { # TODO: separate predict vs. transform "description": "Output data schema for transformed data.", "laleType": "Any", } _hyperparams_schema = { "description": "Hyperparameter schema.", "allOf": [ { "description": "This first sub-object lists all constructor arguments with their " "types, one at a time, omitting cross-argument constraints.", "type": "object", "additionalProperties": False, "relevantToOptimizer": ["batch_size"], "properties": { "operator": { "description": "A lale pipeline object to be used inside of batching", "laleType": "operator", }, "batch_size": { "description": "Batch size used for transform.", "type": "integer", "default": 64, "minimum": 1, "distribution": "uniform", "minimumForOptimizer": 32, "maximumForOptimizer": 128, }, "shuffle": { "type": "boolean", "default": False, "description": "Shuffle dataset before batching or not.", }, "num_workers": { "type": "integer", "default": 0, "description": "Number of workers for pytorch dataloader.", }, "inmemory": { "type": "boolean", "default": False, "description": """Whether all the computations are done in memory or intermediate outputs are serialized. Only applies to transform/predict. For fit, use the `max_resident` argument.""", }, "num_epochs": { "anyOf": [{"type": "integer"}, {"enum": [None]}], "default": None, "description": "Number of epochs. If the operator has `num_epochs` as a parameter, that takes precedence.", }, "max_resident": { "anyOf": [{"type": "integer"}, {"enum": [None]}], "default": None, "description": "Amount of memory to be used in bytes.", }, "scoring": { "anyOf": [{"laleType": "callable"}, {"enum": [None]}], "default": None, "description": "Batch-wise scoring metrics from `lale.lib.rasl`.", }, "progress_callback": { "anyOf": [{"laleType": "callable"}, {"enum": [None]}], "default": None, "description": "Callback function to get performance metrics per batch.", }, "partial_transform": { "type": "boolean", "default": False, "description": """Whether to allow partially-trained upstream operators to transform data for training downstream operators even before the upstream operator has been fully trained.""", }, "priority": { "description": """Scheduling priority in task graphs. "batch" will execute tasks from earlier batches first. "step" will execute tasks from earlier steps first, like nested-loop algorithm. And "resource_aware" will execute tasks with less non-resident data first.""", "enum": ["batch", "step", "resource_aware"], "default": "resource_aware", }, "verbose": { "type": "integer", "default": 0, "description": "Verbosity level, higher values mean more information.", }, }, } ], } _combined_schemas = { "$schema": "http://json-schema.org/draft-04/schema#", "description": """Batching trains the given pipeline using batches. The batch_size is used across all steps of the pipeline, serializing the intermediate outputs if specified.""", "type": "object", "tags": {"pre": [], "op": ["estimator", "transformer"], "post": []}, "properties": { "hyperparams": _hyperparams_schema, "input_fit": _input_fit_schema, "input_predict": _input_predict_transform_schema, "output_predict": _output_schema, "input_transform": _input_predict_transform_schema, "output_transform": _output_schema, }, } Batching = lale.operators.make_operator(_BatchingImpl, _combined_schemas) lale.docstrings.set_docstrings(Batching)
13,230
34.953804
127
py
lale
lale-master/lale/lib/rasl/concat_features.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import logging from functools import reduce from typing import Optional import numpy as np import pandas as pd import scipy.sparse import lale.docstrings import lale.operators import lale.pretty_print from lale.datasets.data_schemas import add_table_name, get_index_names, get_table_name from lale.expressions import it from lale.helpers import _is_spark_df from lale.json_operator import JSON_TYPE from lale.lib.rasl.alias import Alias from lale.lib.rasl.join import Join from lale.type_checking import is_subschema, join_schemas, validate_is_schema logger = logging.getLogger(__name__) logger.setLevel(logging.WARNING) try: import torch torch_installed = True except ImportError: torch_installed = False def _is_pandas_df(d): return isinstance(d, pd.DataFrame) def _is_pandas_series(d): return isinstance(d, pd.Series) def _is_pandas(d): return _is_pandas_df(d) or _is_pandas_series(d) def _gen_table_name(avoid, cpt=0): name = f"tbl{cpt}" if name in avoid: return _gen_table_name(avoid, cpt=cpt + 1) else: return name class _ConcatFeaturesImpl: def transform(self, X): if all(_is_pandas(d) for d in X): name2series = {} for dataset in X: if _is_pandas_df(dataset): for name in dataset.columns: name2series[name] = name2series.get(name, []) + [dataset[name]] elif _is_pandas_series(dataset): name = dataset.name name2series[name] = name2series.get(name, []) + [dataset] else: assert False duplicates = [name for name, ls in name2series.items() if len(ls) > 1] if len(duplicates) == 0: result = pd.concat(X, axis=1) else: logger.info(f"ConcatFeatures duplicate column names {duplicates}") deduplicated = [ls[-1] for _, ls in name2series.items()] result = pd.concat(deduplicated, axis=1) elif all(_is_spark_df(d) for d in X): def join(d1, d2): n1 = get_table_name(d1) n2 = get_table_name(d2) if n1 is None: n1 = _gen_table_name([n2]) d1 = Alias(name=n1).transform(d1) if n2 is None: n2 = _gen_table_name([n1]) d2 = Alias(name=n2).transform(d2) indexes_col1 = get_index_names(d1) indexes_col2 = get_index_names(d2) if indexes_col1 is None or indexes_col2 is None: raise ValueError( "Index columns are required to concatenate features of Spark dataframes (see SparkDataFrameWithIndex)" ) transformer = Join( pred=[ it[n1][index_col1] == it[n2][index_col2] for index_col1, index_col2 in zip(indexes_col1, indexes_col2) ] ) return transformer.transform([d1, d2]) result = reduce(join, X) elif all(_is_pandas(d) or _is_spark_df(d) for d in X): X = [d.toPandas() if _is_spark_df(d) else d for d in X] result = self.transform(X) else: np_datasets = [] # Preprocess the datasets to convert them to 2-d numpy arrays for dataset in X: if _is_pandas(dataset): np_dataset = dataset.values elif _is_spark_df(dataset): np_dataset = dataset.toPandas().values elif isinstance(dataset, scipy.sparse.csr_matrix): np_dataset = dataset.toarray() elif torch_installed and isinstance(dataset, torch.Tensor): np_dataset = dataset.detach().cpu().numpy() else: np_dataset = dataset if hasattr(np_dataset, "shape"): if len(np_dataset.shape) == 1: # To handle numpy column vectors np_dataset = np.reshape(np_dataset, (np_dataset.shape[0], 1)) np_datasets.append(np_dataset) result = np.concatenate(np_datasets, axis=1) name = reduce( ( lambda x, y: get_table_name(x) if get_table_name(x) == get_table_name(y) else None ), X, ) return add_table_name(result, name) def transform_schema(self, s_X): """Used internally by Lale for type-checking downstream operators.""" min_cols, max_cols, elem_schema = 0, 0, None def add_ranges(min_a, max_a, min_b, max_b): min_ab = min_a + min_b if max_a == "unbounded" or max_b == "unbounded": max_ab = "unbounded" else: max_ab = max_a + max_b return min_ab, max_ab elem_schema: Optional[JSON_TYPE] = None for s_dataset in s_X["items"]: if s_dataset.get("laleType", None) == "Any": return {"laleType": "Any"} arr_1d_num = {"type": "array", "items": {"type": "number"}} arr_2d_num = {"type": "array", "items": arr_1d_num} s_decision_func = {"anyOf": [arr_1d_num, arr_2d_num]} if is_subschema(s_decision_func, s_dataset): s_dataset = arr_2d_num assert "items" in s_dataset, lale.pretty_print.to_string(s_dataset) s_rows = s_dataset["items"] if "type" in s_rows and "array" == s_rows["type"]: s_cols = s_rows["items"] if isinstance(s_cols, dict): min_c = s_rows["minItems"] if "minItems" in s_rows else 1 max_c = s_rows["maxItems"] if "maxItems" in s_rows else "unbounded" if elem_schema is None: elem_schema = s_cols else: elem_schema = join_schemas(elem_schema, s_cols) else: min_c, max_c = len(s_cols), len(s_cols) for s_col in s_cols: if elem_schema is None: elem_schema = s_col else: elem_schema = join_schemas(elem_schema, s_col) min_cols, max_cols = add_ranges(min_cols, max_cols, min_c, max_c) else: if elem_schema is None: elem_schema = s_rows else: elem_schema = join_schemas(elem_schema, s_rows) min_cols, max_cols = add_ranges(min_cols, max_cols, 1, 1) s_result = { "$schema": "http://json-schema.org/draft-04/schema#", "type": "array", "items": {"type": "array", "minItems": min_cols, "items": elem_schema}, } if max_cols != "unbounded": s_result["items"]["maxItems"] = max_cols validate_is_schema(s_result) return s_result _hyperparams_schema = { "allOf": [ { "description": "This first sub-object lists all constructor arguments with their " "types, one at a time, omitting cross-argument constraints, if any.", "type": "object", "additionalProperties": False, "relevantToOptimizer": [], "properties": {}, } ] } _input_transform_schema = { "type": "object", "required": ["X"], "additionalProperties": False, "properties": { "X": { "description": "Outermost array dimension is over datasets.", "type": "array", "items": { "description": "Middle array dimension is over samples (aka rows).", "type": "array", "items": { "description": "Innermost array dimension is over features (aka columns).", "anyOf": [ { "type": "array", "items": {"type": "number"}, }, {"type": "number"}, ], }, }, } }, } _output_transform_schema = { "description": "Features; the outer array is over samples.", "type": "array", "items": { "type": "array", "description": "Outer array dimension is over samples (aka rows).", "items": { "description": "Inner array dimension is over features (aka columns).", "laleType": "Any", }, }, } _combined_schemas = { "$schema": "http://json-schema.org/draft-04/schema#", "description": """Horizontal stacking concatenates features (aka columns) of input datasets. Examples -------- >>> A = [ [11, 12, 13], ... [21, 22, 23], ... [31, 32, 33] ] >>> B = [ [14, 15], ... [24, 25], ... [34, 35] ] >>> ConcatFeatures.transform([A, B]) NDArrayWithSchema([[11, 12, 13, 14, 15], [21, 22, 23, 24, 25], [31, 32, 33, 34, 35]])""", "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.rasl.concat_features.html", "import_from": "lale.lib.rasl", "type": "object", "tags": {"pre": [], "op": ["transformer"], "post": []}, "properties": { "hyperparams": _hyperparams_schema, "input_transform": _input_transform_schema, "output_transform": _output_transform_schema, }, } ConcatFeatures = lale.operators.make_pretrained_operator( _ConcatFeaturesImpl, _combined_schemas ) lale.docstrings.set_docstrings(ConcatFeatures)
10,375
35.407018
126
py
lale
lale-master/lale/lib/xgboost/xgb_regressor.py
# Copyright 2019-2022 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from typing import TYPE_CHECKING import numpy as np import pandas as pd from packaging import version import lale.docstrings import lale.helpers import lale.operators import lale.schemas from ._common_schemas import schema_silent try: import xgboost # type: ignore xgboost_version = version.parse(getattr(xgboost, "__version__")) except ImportError: xgboost_version = None if TYPE_CHECKING: import xgboost # type: ignore # xgboost does not like column names with some characters (which are legal in pandas) # so we encode them def _rename_one_feature(name): mapping = {"[": "&#91;", "]": "&#93;", "<": "&lt;"} for old, new in mapping.items(): name = name.replace(old, new) return name def _rename_all_features(X): if not isinstance(X, pd.DataFrame): return X mapped = [_rename_one_feature(f) for f in X.columns] if list(X.columns) == mapped: return X return pd.DataFrame(data=X, columns=mapped) class _XGBRegressorImpl: _wrapped_model: xgboost.XGBRegressor @classmethod def validate_hyperparams(cls, **hyperparams): assert ( xgboost_version is not None ), """Your Python environment does not have xgboost installed. You can install it with pip install xgboost or with pip install 'lale[full]'""" def __init__(self, **hyperparams): self.validate_hyperparams(**hyperparams) self._wrapped_model = xgboost.XGBRegressor(**hyperparams) def fit(self, X, y, **fit_params): renamed_X = _rename_all_features(X) self._wrapped_model.fit(renamed_X, y, **fit_params) return self def partial_fit(self, X, y, **fit_params): fit_params = lale.helpers.dict_without(fit_params, "classes") if self._wrapped_model.__sklearn_is_fitted__(): booster = self._wrapped_model.get_booster() fit_params = {**fit_params, "xgb_model": booster} return self.fit(X, y, **fit_params) def predict(self, X, **predict_params): renamed_X = _rename_all_features(X) result = self._wrapped_model.predict(renamed_X, **predict_params) return result def score(self, X, y): from sklearn.metrics import r2_score y_pred = self.predict(X) return r2_score(y, y_pred) _hyperparams_schema = { "description": "Hyperparameter schema for a Lale wrapper for XGBoost.", "allOf": [ { "description": "This first sub-object lists all constructor arguments with their " "types, one at a time, omitting cross-argument constraints.", "type": "object", "additionalProperties": False, "required": [ "max_depth", "learning_rate", "n_estimators", "verbosity", "objective", "booster", "tree_method", "n_jobs", "gamma", "min_child_weight", "max_delta_step", "subsample", "colsample_bytree", "colsample_bylevel", "colsample_bynode", "reg_alpha", "reg_lambda", "scale_pos_weight", "base_score", "random_state", "missing", ], "relevantToOptimizer": [ "max_depth", "learning_rate", "n_estimators", "gamma", "min_child_weight", "subsample", "reg_alpha", "reg_lambda", ], "properties": { "max_depth": { "description": "Maximum tree depth for base learners.", "type": "integer", "default": 4, "minimum": 0, "distribution": "uniform", "minimumForOptimizer": 1, "maximumForOptimizer": 7, }, "learning_rate": { "description": """Boosting learning rate (xgb's "eta")""", "type": "number", "default": 0.1, "distribution": "loguniform", "minimumForOptimizer": 0.02, "maximumForOptimizer": 1, }, "n_estimators": { "description": "Number of trees to fit.", "type": "integer", "default": 200, "minimumForOptimizer": 50, "maximumForOptimizer": 1000, }, "verbosity": { "description": "The degree of verbosity.", "type": "integer", "default": 1, "minimum": 0, "maximum": 3, }, "silent": schema_silent, "objective": { "description": "Specify the learning task and the corresponding " "learning objective or a custom objective function to be used.", "anyOf": [ { "enum": [ "reg:linear", "reg:logistic", "reg:gamma", "reg:tweedie", ] }, {"laleType": "callable"}, ], "default": "reg:linear", }, "booster": { "description": "Specify which booster to use.", "enum": ["gbtree", "gblinear", "dart"], "default": "gbtree", }, "tree_method": { "description": """Specify which tree method to use. Default to auto. If this parameter is set to default, XGBoost will choose the most conservative option available. Refer to https://xgboost.readthedocs.io/en/latest/parameter.html. """, "enum": ["auto", "exact", "approx", "hist", "gpu_hist"], "default": "auto", }, "n_jobs": { "type": "integer", "description": "Number of parallel threads used to run xgboost. (replaces ``nthread``)", "default": 1, }, "nthread": { "anyOf": [{"type": "integer"}, {"enum": [None]}], "default": None, "description": "Number of parallel threads used to run xgboost. Deprecated, please use n_jobs", }, "gamma": { "type": "number", "description": "Minimum loss reduction required to make a further partition on a leaf node of the tree.", "default": 0, "minimum": 0, "maximumForOptimizer": 1.0, }, "min_child_weight": { "type": "integer", "description": "Minimum sum of instance weight(hessian) needed in a child.", "default": 10, "distribution": "uniform", "minimumForOptimizer": 2, "maximumForOptimizer": 20, }, "max_delta_step": { "type": "integer", "description": "Maximum delta step we allow each tree's weight estimation to be.", "default": 0, }, "subsample": { "type": "number", "description": "Subsample ratio of the training instance.", "default": 1, "minimum": 0, "exclusiveMinimum": True, "distribution": "uniform", "minimumForOptimizer": 0.01, "maximumForOptimizer": 1.0, }, "colsample_bytree": { "type": "number", "description": "Subsample ratio of columns when constructing each tree.", "default": 1, "minimum": 0, "exclusiveMinimum": True, "maximum": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1.0, }, "colsample_bylevel": { "type": "number", "description": "Subsample ratio of columns for each split, in each level.", "default": 1, "minimum": 0, "exclusiveMinimum": True, "maximum": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1.0, }, "colsample_bynode": { "type": "number", "description": "Subsample ratio of columns for each split.", "default": 1, "minimum": 0, "exclusiveMinimum": True, "maximum": 1, }, "reg_alpha": { "type": "number", "description": "L1 regularization term on weights", "default": 0, "distribution": "uniform", "minimumForOptimizer": 0, "maximumForOptimizer": 1, }, "reg_lambda": { "type": "number", "description": "L2 regularization term on weights", "default": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1, }, "scale_pos_weight": { "type": "number", "description": "Balancing of positive and negative weights.", "default": 1, }, "base_score": { "type": "number", "description": "The initial prediction score of all instances, global bias.", "default": 0.5, }, "random_state": { "type": "integer", "description": "Random number seed. (replaces seed)", "default": 0, }, "missing": { "anyOf": [ { "type": "number", }, { "enum": [None], }, ], "default": None, "description": "Value in the data which needs to be present as a missing value. If" " If None, defaults to np.nan.", }, "importance_type": { "enum": [ "gain", "weight", "cover", "total_gain", "total_cover", None, ], "default": "gain", "description": "The feature importance type for the `feature_importances_` property.", }, "seed": { "default": None, "description": "deprecated and replaced with random_state, but adding to be backward compatible. ", }, }, } ], } _input_fit_schema = { "description": "Fit gradient boosting classifier", "type": "object", "required": ["X", "y"], "properties": { "X": { "type": "array", "items": { "type": "array", "items": {"type": "number"}, }, "description": "Feature matrix", }, "y": { "type": "array", "items": {"type": "number"}, "description": "Labels", }, "sample_weight": { "anyOf": [ { "type": "array", "items": {"type": "number"}, }, {"enum": [None]}, ], "description": "Weight for each instance", "default": None, }, "eval_set": { "anyOf": [ { "type": "array", }, { "enum": [None], }, ], "default": None, "description": "A list of (X, y) pairs to use as a validation set for", }, "sample_weight_eval_set": { "anyOf": [ { "type": "array", }, { "enum": [None], }, ], "default": None, "description": "A list of the form [L_1, L_2, ..., L_n], where each L_i is a list of", }, "eval_metric": { "anyOf": [ {"type": "array", "items": {"type": "string"}}, {"type": "string"}, {"enum": [None]}, {"type": "object"}, ], "default": None, "description": "If a str, should be a built-in evaluation metric to use. See", }, "early_stopping_rounds": { "anyOf": [ { "type": "integer", }, { "enum": [None], }, ], "default": None, "description": "Activates early stopping. Validation error needs to decrease at", }, "verbose": { "type": "boolean", "description": "If `verbose` and an evaluation set is used, writes the evaluation", "default": True, }, "xgb_model": { "anyOf": [{"type": "string"}, {"enum": [None]}], "description": "file name of stored xgb model or 'Booster' instance Xgb model to be", "default": None, }, "callbacks": { "anyOf": [{"type": "array", "items": {"type": "object"}}, {"enum": [None]}], "default": None, "description": "List of callback functions that are applied at each iteration. ", }, }, } _input_predict_schema = { "description": "Predict with `data`.", "type": "object", "required": ["X"], "properties": { "X": { "type": "array", "items": { "type": "array", "items": {"type": "number"}, }, "description": "The dmatrix storing the input.", }, "output_margin": { "type": "boolean", "default": False, "description": "Whether to output the raw untransformed margin value.", }, "ntree_limit": { "anyOf": [{"type": "integer"}, {"enum": [None]}], "description": "Limit number of trees in the prediction; defaults to best_ntree_limit if defined", }, "validate_features": { "type": "boolean", "default": True, "description": "When this is True, validate that the Booster's and data's feature_names are identical.", }, }, } _output_predict_schema = { "description": "Output data schema for predictions (target class labels).", "type": "array", "items": {"type": "number"}, } _combined_schemas = { "$schema": "http://json-schema.org/draft-04/schema#", "description": """`XGBRegressor`_ gradient boosted decision trees. .. _`XGBRegressor`: https://xgboost.readthedocs.io/en/latest/python/python_api.html#xgboost.XGBRegressor """, "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.xgboost.xgb_regressor.html", "import_from": "xgboost", "tags": {"pre": [], "op": ["estimator", "regressor"], "post": []}, "properties": { "hyperparams": _hyperparams_schema, "input_fit": _input_fit_schema, "input_partial_fit": _input_fit_schema, "input_predict": _input_predict_schema, "output_predict": _output_predict_schema, }, } XGBRegressor: lale.operators.PlannedIndividualOp XGBRegressor = lale.operators.make_operator(_XGBRegressorImpl, _combined_schemas) if xgboost_version is not None and xgboost_version >= version.Version("0.90"): # page 58 of https://readthedocs.org/projects/xgboost/downloads/pdf/release_0.90/ XGBRegressor = XGBRegressor.customize_schema( objective=lale.schemas.JSON( { "description": "Specify the learning task and the corresponding learning objective or a custom objective function to be used.", "anyOf": [ { "enum": [ "reg:linear", "reg:logistic", "reg:gamma", "reg:tweedie", "reg:squarederror", ] }, {"laleType": "callable"}, ], "default": "reg:linear", } ), set_as_available=True, ) if xgboost_version is not None and xgboost_version >= version.Version("1.3"): # https://xgboost.readthedocs.io/en/latest/python/python_api.html#module-xgboost.sklearn XGBRegressor = XGBRegressor.customize_schema( monotone_constraints={ "description": "Constraint of variable monotonicity.", "anyOf": [{"enum": [None]}, {"type": "string"}], "default": None, }, interaction_constraints={ "description": "Constraints for interaction representing permitted interactions. The constraints must be specified in the form of a nest list, e.g. [[0, 1], [2, 3, 4]], where each inner list is a group of indices of features that are allowed to interact with each other.", "anyOf": [{"enum": [None]}, {"type": "string"}], "default": None, }, num_parallel_tree={ "description": "Used for boosting random forest.", "anyOf": [{"enum": [None]}, {"type": "integer"}], "default": None, }, validate_parameters={ "description": "Give warnings for unknown parameter.", "anyOf": [{"enum": [None]}, {"type": "boolean"}, {"type": "integer"}], "default": None, }, gpu_id={ "description": "Device ordinal.", "anyOf": [ {"type": "integer"}, {"enum": [None]}, ], "default": None, }, max_depth={ "description": "Maximum tree depth for base learners.", "anyOf": [ { "type": "integer", "minimum": 0, "distribution": "uniform", "minimumForOptimizer": 1, "maximumForOptimizer": 7, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, learning_rate={ "description": """Boosting learning rate (xgb's "eta")""", "anyOf": [ { "type": "number", "distribution": "loguniform", "minimumForOptimizer": 0.02, "maximumForOptimizer": 1, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, booster={ "description": "Specify which booster to use.", "enum": ["gbtree", "gblinear", "dart", None], "default": None, }, tree_method={ "description": """Specify which tree method to use. Default to auto. If this parameter is set to default, XGBoost will choose the most conservative option available. Refer to https://xgboost.readthedocs.io/en/latest/parameter.html. """, "enum": ["auto", "exact", "approx", "hist", "gpu_hist", None], "default": None, }, gamma={ "description": "Minimum loss reduction required to make a further partition on a leaf node of the tree.", "anyOf": [ { "type": "number", "minimum": 0, "maximumForOptimizer": 1.0, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, min_child_weight={ "description": "Minimum sum of instance weight(hessian) needed in a child.", "anyOf": [ { "type": "integer", "distribution": "uniform", "minimumForOptimizer": 2, "maximumForOptimizer": 20, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, max_delta_step={ "description": "Maximum delta step we allow each tree's weight estimation to be.", "anyOf": [{"enum": [None]}, {"type": "integer"}], "default": None, }, subsample={ "description": "Subsample ratio of the training instance.", "anyOf": [ { "type": "number", "minimum": 0, "exclusiveMinimum": True, "distribution": "uniform", "minimumForOptimizer": 0.01, "maximumForOptimizer": 1.0, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, colsample_bytree={ "description": "Subsample ratio of columns when constructing each tree.", "anyOf": [ { "type": "number", "minimum": 0, "exclusiveMinimum": True, "maximum": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1.0, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, colsample_bylevel={ "description": "Subsample ratio of columns for each split, in each level.", "anyOf": [ { "type": "number", "minimum": 0, "exclusiveMinimum": True, "maximum": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1.0, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, colsample_bynode={ "description": "Subsample ratio of columns for each split.", "anyOf": [ { "type": "number", "minimum": 0, "exclusiveMinimum": True, "maximum": 1, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, reg_alpha={ "description": "L1 regularization term on weights", "anyOf": [ { "type": "number", "distribution": "uniform", "minimumForOptimizer": 0, "maximumForOptimizer": 1, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, reg_lambda={ "description": "L2 regularization term on weights", "anyOf": [ { "type": "number", "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, scale_pos_weight={ "description": "Balancing of positive and negative weights.", "anyOf": [ {"type": "number"}, {"enum": [None], "forOptimizer": False}, ], "default": None, }, base_score={ "description": "The initial prediction score of all instances, global bias.", "anyOf": [ {"type": "number"}, {"enum": [None], "forOptimizer": False}, ], "default": None, }, missing={ "anyOf": [ { "type": "number", }, { "enum": [None, np.NaN], }, ], "default": np.NaN, "description": "Value in the data which needs to be present as a missing value. If" " If None, defaults to np.nan.", }, verbosity={ "description": "The degree of verbosity.", "anyOf": [{"type": "integer"}, {"enum": [None]}], "default": None, "minimum": 0, "maximum": 3, }, set_as_available=True, ) if xgboost_version is not None and xgboost_version >= version.Version("1.5"): # https://xgboost.readthedocs.io/en/latest/python/python_api.html#module-xgboost.sklearn XGBRegressor = XGBRegressor.customize_schema( enable_categorical={ "type": "boolean", "description": """Experimental support for categorical data. Do not set to true unless you are interested in development. Only valid when gpu_hist and dataframe are used.""", "default": False, }, predictor={ "anyOf": [{"type": "string"}, {"enum": [None]}], "description": """Force XGBoost to use specific predictor, available choices are [cpu_predictor, gpu_predictor].""", "default": None, }, set_as_available=True, ) if xgboost_version is not None and xgboost_version >= version.Version("1.6"): # https://xgboost.readthedocs.io/en/latest/python/python_api.html#module-xgboost.sklearn XGBRegressor = XGBRegressor.customize_schema( max_leaves={ "description": """Maximum number of leaves; 0 indicates no limit.""", "anyOf": [ {"type": "integer"}, {"enum": [None], "forOptimizer": False}, ], "default": None, }, max_bin={ "description": """If using histogram-based algorithm, maximum number of bins per feature.""", "anyOf": [ {"type": "integer"}, {"enum": [None], "forOptimizer": False}, ], "default": None, }, grow_policy={ "description": """Tree growing policy. 0 or depthwise: favor splitting at nodes closest to the node, i.e. grow depth-wise. 1 or lossguide: favor splitting at nodes with highest loss change.""", "enum": [0, 1, "depthwise", "lossguide", None], "default": None, }, sampling_method={ "description": """Sampling method. Used only by gpu_hist tree method. - uniform: select random training instances uniformly. - gradient_based select random training instances with higher probability when the gradient and hessian are larger. (cf. CatBoost)""", "enum": ["uniform", "gadient_based", None], "default": None, }, max_cat_to_onehot={ "description": """A threshold for deciding whether XGBoost should use one-hot encoding based split for categorical data.""", "anyOf": [ {"type": "integer"}, {"enum": [None]}, ], "default": None, }, eval_metric={ "description": """Metric used for monitoring the training result and early stopping.""", "anyOf": [ {"type": "string"}, {"type": "array", "items": {"type": "string"}}, {"type": "array", "items": {"laleType": "callable"}}, {"enum": [None]}, ], "default": None, }, early_stopping_rounds={ "description": """Activates early stopping. Validation metric needs to improve at least once in every early_stopping_rounds round(s) to continue training.""", "anyOf": [ {"type": "integer"}, {"enum": [None]}, ], "default": None, }, callbacks={ "description": """List of callback functions that are applied at end of each iteration. It is possible to use predefined callbacks by using Callback API.""", "anyOf": [ {"type": "array", "items": {"laleType": "callable"}}, {"enum": [None]}, ], "default": None, }, n_jobs={ "anyOf": [{"type": "integer"}, {"enum": [None]}], "description": "Number of parallel threads used to run xgboost. (replaces ``nthread``)", "default": 1, }, random_state={ "anyOf": [{"type": "integer"}, {"enum": [None]}], "description": "Random number seed. (replaces seed)", "default": 0, }, ) lale.docstrings.set_docstrings(XGBRegressor)
30,974
35.963007
284
py
lale
lale-master/lale/lib/xgboost/xgb_classifier.py
# Copyright 2019-2022 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import warnings from typing import TYPE_CHECKING import numpy as np import pandas as pd from packaging import version import lale.docstrings import lale.helpers import lale.operators import lale.schemas from ._common_schemas import schema_silent with warnings.catch_warnings(): warnings.filterwarnings("ignore", category=FutureWarning) try: import xgboost # type: ignore xgboost_version = version.parse(getattr(xgboost, "__version__")) except ImportError: xgboost_version = None if TYPE_CHECKING: import xgboost # type: ignore # xgboost does not like column names with some characters (which are legal in pandas) # so we encode them def _rename_one_feature(name): mapping = {"[": "&#91;", "]": "&#93;", "<": "&lt;"} for old, new in mapping.items(): name = name.replace(old, new) return name def _rename_all_features(X): if not isinstance(X, pd.DataFrame): return X mapped = [_rename_one_feature(f) for f in X.columns] if list(X.columns) == mapped: return X return pd.DataFrame(data=X, columns=mapped) class _XGBClassifierImpl: _wrapped_model: xgboost.XGBClassifier @classmethod def validate_hyperparams(cls, **hyperparams): assert ( xgboost_version is not None ), """Your Python environment does not have xgboost installed. You can install it with pip install xgboost or with pip install 'lale[full]'""" def __init__(self, **hyperparams): self.validate_hyperparams(**hyperparams) self._wrapped_model = xgboost.XGBClassifier(**hyperparams) def fit(self, X, y, **fit_params): renamed_X = _rename_all_features(X) assert xgboost_version is not None if ( xgboost_version >= version.Version("1.3.0") and "eval_metric" not in fit_params ): # set eval_metric explicitly to avoid spurious warning fit_params = {"eval_metric": "logloss", **fit_params} with warnings.catch_warnings(): if fit_params.get("use_label_encoder", True): warnings.filterwarnings("ignore", category=UserWarning) warnings.filterwarnings("ignore", category=FutureWarning) self._wrapped_model.fit(renamed_X, y, **fit_params) return self def partial_fit(self, X, y, **fit_params): fit_params = lale.helpers.dict_without(fit_params, "classes") if self._wrapped_model.__sklearn_is_fitted__(): booster = self._wrapped_model.get_booster() fit_params = {**fit_params, "xgb_model": booster} return self.fit(X, y, **fit_params) def predict(self, X, **predict_params): renamed_X = _rename_all_features(X) with warnings.catch_warnings(): warnings.filterwarnings("ignore", category=FutureWarning) result = self._wrapped_model.predict(renamed_X, **predict_params) return result def predict_proba(self, X): return self._wrapped_model.predict_proba(X) def score(self, X, y): from sklearn.metrics import accuracy_score y_pred = self.predict(X) return accuracy_score(y, y_pred) _hyperparams_schema = { "description": "Hyperparameter schema for a Lale wrapper for XGBoost.", "allOf": [ { "description": "This first sub-object lists all constructor arguments with their " "types, one at a time, omitting cross-argument constraints.", "type": "object", "additionalProperties": False, "required": [ "max_depth", "learning_rate", "n_estimators", "verbosity", "objective", "booster", "tree_method", "n_jobs", "gamma", "min_child_weight", "max_delta_step", "subsample", "colsample_bytree", "colsample_bylevel", "colsample_bynode", "reg_alpha", "reg_lambda", "scale_pos_weight", "base_score", "random_state", "missing", ], "relevantToOptimizer": [ "gamma", "max_depth", "learning_rate", "n_estimators", "min_child_weight", "subsample", "reg_alpha", "reg_lambda", ], "properties": { "max_depth": { "description": "Maximum tree depth for base learners.", "type": "integer", "default": 4, "minimum": 0, "distribution": "uniform", "minimumForOptimizer": 1, "maximumForOptimizer": 7, }, "learning_rate": { "description": "Boosting learning rate (xgb’s “eta”)", "type": "number", "default": 0.1, "distribution": "loguniform", "minimumForOptimizer": 0.02, "maximumForOptimizer": 1, }, "n_estimators": { "description": "Number of trees to fit.", "type": "integer", "default": 100, "minimumForOptimizer": 50, "maximumForOptimizer": 1000, }, "verbosity": { "description": "The degree of verbosity.", "anyOf": [{"type": "integer"}, {"enum": [None]}], "default": 1, "minimum": 0, "maximum": 3, }, "objective": { "description": "Specify the learning task and the corresponding " "learning objective or a custom objective function to be used.", "anyOf": [ { "enum": [ "binary:logistic", "binary:logitraw", "binary:hinge", "multi:softprob", "multi:softmax", ] }, {"laleType": "callable"}, ], "default": "binary:logistic", }, "booster": { "description": "Specify which booster to use.", "enum": ["gbtree", "gblinear", "dart"], "default": "gbtree", }, "tree_method": { "description": """Specify which tree method to use. Default to auto. If this parameter is set to default, XGBoost will choose the most conservative option available. Refer to https://xgboost.readthedocs.io/en/latest/parameter.html. """, "enum": ["auto", "exact", "approx", "hist", "gpu_hist"], "default": "auto", }, "n_jobs": { "anyOf": [{"type": "integer"}, {"enum": [None]}], "description": "Number of parallel threads used to run xgboost. (replaces ``nthread``)", "default": 1, }, "nthread": { "anyOf": [{"type": "integer"}, {"enum": [None]}], "default": None, "description": "Number of parallel threads used to run xgboost. Deprecated, please use n_jobs", }, "gamma": { "type": "number", "description": "Minimum loss reduction required to make a further partition on a leaf node of the tree.", "default": 0, "minimum": 0, "maximumForOptimizer": 1.0, }, "min_child_weight": { "type": "integer", "description": "Minimum sum of instance weight(hessian) needed in a child.", "default": 10, "distribution": "uniform", "minimumForOptimizer": 2, "maximumForOptimizer": 20, }, "max_delta_step": { "type": "integer", "description": "Maximum delta step we allow each tree's weight estimation to be.", "default": 0, }, "subsample": { "type": "number", "description": "Subsample ratio of the training instance.", "default": 1, "minimum": 0, "exclusiveMinimum": True, "distribution": "uniform", "minimumForOptimizer": 0.01, "maximumForOptimizer": 1.0, }, "colsample_bytree": { "type": "number", "description": "Subsample ratio of columns when constructing each tree.", "default": 1, "minimum": 0, "exclusiveMinimum": True, "maximum": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1.0, }, "colsample_bylevel": { "type": "number", "description": "Subsample ratio of columns for each split, in each level.", "default": 1, "minimum": 0, "exclusiveMinimum": True, "maximum": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1.0, }, "colsample_bynode": { "type": "number", "description": "Subsample ratio of columns for each split.", "default": 1, "minimum": 0, "exclusiveMinimum": True, "maximum": 1, }, "reg_alpha": { "type": "number", "description": "L1 regularization term on weights", "default": 0, "distribution": "uniform", "minimumForOptimizer": 0.0, "maximumForOptimizer": 1.0, }, "reg_lambda": { "type": "number", "description": "L2 regularization term on weights", "default": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1.0, }, "scale_pos_weight": { "anyOf": [ { "type": "number", }, { "enum": [None], }, ], "description": "Balancing of positive and negative weights.", "default": 1, }, "base_score": { "anyOf": [ { "type": "number", }, { "enum": [None], }, ], "description": "The initial prediction score of all instances, global bias.", "default": 0.5, }, "random_state": { "anyOf": [ { "type": "integer", }, { "enum": [None], }, ], "description": "Random number seed. (replaces seed)", "default": 0, }, "missing": { "anyOf": [ { "type": "number", }, { "enum": [None], }, ], "default": None, "description": "Value in the data which needs to be present as a missing value. If" " If None, defaults to np.nan.", }, "silent": schema_silent, "seed": { "default": None, "description": "deprecated and replaced with random_state, but adding to be backward compatible. ", }, }, } ], } _input_fit_schema = { "description": "Fit gradient boosting classifier", "type": "object", "required": ["X", "y"], "properties": { "X": { "type": "array", "items": { "type": "array", "items": {"type": "number"}, }, "description": "Feature matrix", }, "y": { "anyOf": [ {"type": "array", "items": {"type": "number"}}, {"type": "array", "items": {"type": "string"}}, {"type": "array", "items": {"type": "boolean"}}, ], "description": "Labels", }, "sample_weight": { "anyOf": [ { "type": "array", "items": {"type": "number"}, }, {"enum": [None]}, ], "description": "Weight for each instance", "default": None, }, "eval_set": { "anyOf": [ { "type": "array", }, { "enum": [None], }, ], "default": None, "description": "A list of (X, y) pairs to use as a validation set for", }, "sample_weight_eval_set": { "anyOf": [ { "type": "array", }, { "enum": [None], }, ], "default": None, "description": "A list of the form [L_1, L_2, ..., L_n], where each L_i is a list of", }, "eval_metric": { "anyOf": [ {"type": "array", "items": {"type": "string"}}, {"type": "string"}, {"enum": [None]}, {"type": "object"}, ], "default": None, "description": "If a str, should be a built-in evaluation metric to use. See", }, "early_stopping_rounds": { "anyOf": [ { "type": "integer", }, { "enum": [None], }, ], "default": None, "description": "Activates early stopping. Validation error needs to decrease at", }, "verbose": { "type": "boolean", "description": "If `verbose` and an evaluation set is used, writes the evaluation", "default": True, }, "xgb_model": { "anyOf": [{"type": "string"}, {"enum": [None]}], "description": "file name of stored xgb model or 'Booster' instance Xgb model to be", "default": None, }, "callbacks": { "anyOf": [{"type": "array", "items": {"type": "object"}}, {"enum": [None]}], "default": None, "description": "List of callback functions that are applied at each iteration. ", }, }, } _input_predict_schema = { "description": "Predict with `data`.", "type": "object", "required": ["X"], "properties": { "X": { "type": "array", "items": {"type": "array", "items": {"type": "number"}}, "description": "The dmatrix storing the input.", }, "output_margin": { "type": "boolean", "default": False, "description": "Whether to output the raw untransformed margin value.", }, "ntree_limit": { "anyOf": [{"type": "integer"}, {"enum": [None]}], "description": "Limit number of trees in the prediction; defaults to best_ntree_limit if defined", }, "validate_features": { "type": "boolean", "default": True, "description": "When this is True, validate that the Booster's and data's feature_names are identical.", }, }, } _output_predict_schema = { "description": "Predicted class label per sample.", "anyOf": [ {"type": "array", "items": {"type": "number"}}, {"type": "array", "items": {"type": "string"}}, {"type": "array", "items": {"type": "boolean"}}, ], } _input_predict_proba_schema = { "type": "object", "required": ["X"], "properties": { "X": {"type": "array", "items": {"type": "array", "items": {"type": "number"}}} }, } _output_predict_proba_schema = { "description": "Probability of the sample for each class in the model.", "type": "array", "items": {"type": "array", "items": {"type": "number"}}, } _combined_schemas = { "$schema": "http://json-schema.org/draft-04/schema#", "description": """`XGBClassifier`_ gradient boosted decision trees. .. _`XGBClassifier`: https://xgboost.readthedocs.io/en/latest/python/python_api.html#xgboost.XGBClassifier """, "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.xgboost.xgb_classifier.html", "import_from": "xgboost", "tags": {"pre": [], "op": ["estimator", "classifier"], "post": []}, "properties": { "hyperparams": _hyperparams_schema, "input_fit": _input_fit_schema, "input_partial_fit": _input_fit_schema, "input_predict": _input_predict_schema, "output_predict": _output_predict_schema, "input_predict_proba": _input_predict_proba_schema, "output_predict_proba": _output_predict_proba_schema, }, } XGBClassifier: lale.operators.PlannedIndividualOp XGBClassifier = lale.operators.make_operator(_XGBClassifierImpl, _combined_schemas) if xgboost_version is not None and xgboost_version >= version.Version("0.90"): # page 58 of https://readthedocs.org/projects/xgboost/downloads/pdf/release_0.90/ XGBClassifier = XGBClassifier.customize_schema( objective=lale.schemas.JSON( { "description": "Specify the learning task and the corresponding learning objective or a custom objective function to be used.", "anyOf": [ { "enum": [ "binary:hinge", "binary:logistic", "binary:logitraw", "multi:softmax", "multi:softprob", ] }, {"laleType": "callable"}, ], "default": "binary:logistic", } ), set_as_available=True, ) if xgboost_version is not None and xgboost_version >= version.Version("1.3"): # https://xgboost.readthedocs.io/en/latest/python/python_api.html#module-xgboost.sklearn XGBClassifier = XGBClassifier.customize_schema( monotone_constraints={ "description": "Constraint of variable monotonicity.", "anyOf": [{"enum": [None]}, {"type": "string"}], "default": None, }, interaction_constraints={ "description": "Constraints for interaction representing permitted interactions. The constraints must be specified in the form of a nest list, e.g. [[0, 1], [2, 3, 4]], where each inner list is a group of indices of features that are allowed to interact with each other.", "anyOf": [{"enum": [None]}, {"type": "string"}], "default": None, }, num_parallel_tree={ "description": "Used for boosting random forest.", "anyOf": [{"enum": [None]}, {"type": "integer"}], "default": None, }, validate_parameters={ "description": "Give warnings for unknown parameter.", "anyOf": [{"enum": [None]}, {"type": "boolean"}, {"type": "integer"}], "default": None, }, gpu_id={ "description": "Device ordinal.", "anyOf": [ {"type": "integer"}, {"enum": [None]}, ], "default": None, }, max_depth={ "description": "Maximum tree depth for base learners.", "anyOf": [ { "type": "integer", "minimum": 0, "distribution": "uniform", "minimumForOptimizer": 1, "maximumForOptimizer": 7, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, learning_rate={ "description": """Boosting learning rate (xgb's "eta")""", "anyOf": [ { "type": "number", "distribution": "loguniform", "minimumForOptimizer": 0.02, "maximumForOptimizer": 1, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, booster={ "description": "Specify which booster to use.", "enum": ["gbtree", "gblinear", "dart", None], "default": None, }, tree_method={ "description": """Specify which tree method to use. Default to auto. If this parameter is set to default, XGBoost will choose the most conservative option available. Refer to https://xgboost.readthedocs.io/en/latest/parameter.html. """, "enum": ["auto", "exact", "approx", "hist", "gpu_hist", None], "default": None, }, gamma={ "description": "Minimum loss reduction required to make a further partition on a leaf node of the tree.", "anyOf": [ { "type": "number", "minimum": 0, "maximumForOptimizer": 1.0, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, min_child_weight={ "description": "Minimum sum of instance weight(hessian) needed in a child.", "anyOf": [ { "type": "integer", "distribution": "uniform", "minimumForOptimizer": 2, "maximumForOptimizer": 20, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, max_delta_step={ "description": "Maximum delta step we allow each tree's weight estimation to be.", "anyOf": [{"enum": [None]}, {"type": "integer"}], "default": None, }, subsample={ "description": "Subsample ratio of the training instance.", "anyOf": [ { "type": "number", "minimum": 0, "exclusiveMinimum": True, "distribution": "uniform", "minimumForOptimizer": 0.01, "maximumForOptimizer": 1.0, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, colsample_bytree={ "description": "Subsample ratio of columns when constructing each tree.", "anyOf": [ { "type": "number", "minimum": 0, "exclusiveMinimum": True, "maximum": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1.0, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, colsample_bylevel={ "description": "Subsample ratio of columns for each split, in each level.", "anyOf": [ { "type": "number", "minimum": 0, "exclusiveMinimum": True, "maximum": 1, "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1.0, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, colsample_bynode={ "description": "Subsample ratio of columns for each split.", "anyOf": [ { "type": "number", "minimum": 0, "exclusiveMinimum": True, "maximum": 1, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, reg_alpha={ "description": "L1 regularization term on weights", "anyOf": [ { "type": "number", "distribution": "uniform", "minimumForOptimizer": 0, "maximumForOptimizer": 1, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, reg_lambda={ "description": "L2 regularization term on weights", "anyOf": [ { "type": "number", "distribution": "uniform", "minimumForOptimizer": 0.1, "maximumForOptimizer": 1, }, {"enum": [None], "forOptimizer": False}, ], "default": None, }, scale_pos_weight={ "description": "Balancing of positive and negative weights.", "anyOf": [ {"type": "number"}, {"enum": [None], "forOptimizer": False}, ], "default": None, }, base_score={ "description": "The initial prediction score of all instances, global bias.", "anyOf": [ {"type": "number"}, {"enum": [None], "forOptimizer": False}, ], "default": None, }, importance_type={ "description": "The feature importance type for the `feature_importances_` property.", "enum": ["gain", "weight", "cover", "total_gain", "total_cover", None], "default": "gain", }, use_label_encoder={ "description": """(Deprecated) Use the label encoder from scikit-learn to encode the labels. For new code, we recommend that you set this parameter to False.""", "type": "boolean", "default": True, }, missing={ "anyOf": [ { "type": "number", }, { "enum": [None, np.NaN], }, ], "default": np.NaN, "description": "Value in the data which needs to be present as a missing value. If" " If None, defaults to np.nan.", }, verbosity={ "description": "The degree of verbosity.", "anyOf": [{"type": "integer"}, {"enum": [None]}], "default": None, "minimum": 0, "maximum": 3, }, set_as_available=True, ) if xgboost_version is not None and xgboost_version >= version.Version("1.5"): # https://xgboost.readthedocs.io/en/latest/python/python_api.html#module-xgboost.sklearn XGBClassifier = XGBClassifier.customize_schema( enable_categorical={ "type": "boolean", "description": """Experimental support for categorical data. Do not set to true unless you are interested in development. Only valid when gpu_hist and dataframe are used.""", "default": False, }, predictor={ "anyOf": [{"type": "string"}, {"enum": [None]}], "description": """Force XGBoost to use specific predictor, available choices are [cpu_predictor, gpu_predictor].""", "default": None, }, set_as_available=True, ) if xgboost_version is not None and xgboost_version >= version.Version("1.6"): # https://xgboost.readthedocs.io/en/latest/python/python_api.html#module-xgboost.sklearn XGBClassifier = XGBClassifier.customize_schema( use_label_encoder={ "description": """(Deprecated) Use the label encoder from scikit-learn to encode the labels. For new code, we recommend that you set this parameter to False.""", "type": "boolean", "default": False, }, max_leaves={ "description": """Maximum number of leaves; 0 indicates no limit.""", "anyOf": [ {"type": "integer"}, {"enum": [None], "forOptimizer": False}, ], "default": None, }, max_bin={ "description": """If using histogram-based algorithm, maximum number of bins per feature.""", "anyOf": [ {"type": "integer"}, {"enum": [None], "forOptimizer": False}, ], "default": None, }, grow_policy={ "description": """Tree growing policy. 0 or depthwise: favor splitting at nodes closest to the node, i.e. grow depth-wise. 1 or lossguide: favor splitting at nodes with highest loss change.""", "enum": [0, 1, "depthwise", "lossguide", None], "default": None, }, sampling_method={ "description": """Sampling method. Used only by gpu_hist tree method. - uniform: select random training instances uniformly. - gradient_based select random training instances with higher probability when the gradient and hessian are larger. (cf. CatBoost)""", "enum": ["uniform", "gadient_based", None], "default": None, }, max_cat_to_onehot={ "description": """A threshold for deciding whether XGBoost should use one-hot encoding based split for categorical data.""", "anyOf": [ {"type": "integer"}, {"enum": [None]}, ], "default": None, }, eval_metric={ "description": """Metric used for monitoring the training result and early stopping.""", "anyOf": [ {"type": "string"}, {"type": "array", "items": {"type": "string"}}, {"type": "array", "items": {"laleType": "callable"}}, {"enum": [None]}, ], "default": None, }, early_stopping_rounds={ "description": """Activates early stopping. Validation metric needs to improve at least once in every early_stopping_rounds round(s) to continue training.""", "anyOf": [ {"type": "integer"}, {"enum": [None]}, ], "default": None, }, callbacks={ "description": """List of callback functions that are applied at end of each iteration. It is possible to use predefined callbacks by using Callback API.""", "anyOf": [ {"type": "array", "items": {"laleType": "callable"}}, {"enum": [None]}, ], "default": None, }, ) lale.docstrings.set_docstrings(XGBClassifier)
33,369
36.326622
284
py
lale
lale-master/lale/lib/xgboost/__init__.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Scikit-learn compatible wrappers for XGBoost_ along with schemas to enable hyperparameter tuning. .. _XGBoost: https://xgboost.readthedocs.io/en/latest/ Operators: ========== * `XGBClassifier`_ * `XGBRegressor`_ .. _`XGBClassifier`: lale.lib.xgboost.xgb_classifier.html .. _`XGBRegressor`: lale.lib.xgboost.xgb_regressor.html """ from lale import register_lale_wrapper_modules from .xgb_classifier import XGBClassifier as XGBClassifier from .xgb_regressor import XGBRegressor as XGBRegressor # Note: all imports should be done as # from .xxx import XXX as XXX # this ensures that pyright considers them to be publicly available # and not private imports (this affects lale users that use pyright) register_lale_wrapper_modules(__name__)
1,325
31.341463
97
py
lale
lale-master/lale/lib/lale/auto_pipeline.py
# Copyright 2020 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import time import warnings from typing import Optional import hyperopt import pandas as pd import sklearn.metrics import sklearn.model_selection import lale.docstrings import lale.helpers import lale.operators from lale.lib._common_schemas import ( schema_best_score_single, schema_cv, schema_max_opt_time, schema_scoring_single, ) with warnings.catch_warnings(): warnings.filterwarnings("ignore", category=FutureWarning) try: import xgboost # noqa: F401 xgboost_installed = True except ImportError: xgboost_installed = False try: import lightgbm.sklearn # noqa: F401 lightgbm_installed = True except ImportError: lightgbm_installed = False def auto_prep(X): from lale.lib.lale import ConcatFeatures, Project, categorical from lale.lib.sklearn import OneHotEncoder, SimpleImputer n_cols = X.shape[1] n_cats = len(categorical()(X)) prep_num = SimpleImputer(strategy="mean") prep_cat = SimpleImputer(strategy="most_frequent") >> OneHotEncoder( handle_unknown="ignore" ) if n_cats == 0: result = prep_num elif n_cats == n_cols: result = prep_cat else: result = ( ( Project(columns={"type": "number"}, drop_columns=categorical()) >> prep_num ) & (Project(columns=categorical()) >> prep_cat) ) >> ConcatFeatures return result def auto_gbt(prediction_type): if prediction_type == "regression": if xgboost_installed: from lale.lib.xgboost import XGBRegressor return XGBRegressor(verbosity=0) elif lightgbm_installed: from lale.lib.lightgbm import LGBMRegressor return LGBMRegressor() else: from lale.lib.sklearn import GradientBoostingRegressor return GradientBoostingRegressor() else: assert prediction_type in ["binary", "multiclass", "classification"] if xgboost_installed: from lale.lib.xgboost import XGBClassifier return XGBClassifier(verbosity=0) elif lightgbm_installed: from lale.lib.lightgbm import LGBMClassifier return LGBMClassifier() else: from lale.lib.sklearn import GradientBoostingClassifier return GradientBoostingClassifier() class _AutoPipelineImpl: _summary: Optional[pd.DataFrame] def __init__( self, *, prediction_type="classification", scoring=None, best_score=0.0, verbose=False, max_evals=100, max_opt_time=600.0, max_eval_time=120.0, cv=5, ): self.prediction_type = prediction_type self.max_opt_time = max_opt_time self.max_eval_time = max_eval_time self.max_evals = max_evals self.verbose = verbose if scoring is None: scoring = "r2" if prediction_type == "regression" else "accuracy" self.scoring = scoring self._scorer = sklearn.metrics.get_scorer(scoring) self.best_score = best_score self._summary = None self.cv = cv def _try_and_add(self, name, trainable, X, y): assert name not in self._pipelines if self._name_of_best is not None: if time.time() > self._start_fit + self.max_opt_time: return with warnings.catch_warnings(): warnings.simplefilter("ignore") cv = sklearn.model_selection.check_cv( cv=self.cv, classifier=(self.prediction_type != "regression") ) ( cv_score, logloss, execution_time, ) = lale.helpers.cross_val_score_track_trials( trainable, X, y, self.scoring, cv ) loss = self.best_score - cv_score if self._name_of_best is None or ( self._summary is None or loss < self._summary.at[self._name_of_best, "loss"] ): self._name_of_best = name record = { "name": name, "loss": loss, "time": execution_time, "log_loss": logloss, "status": hyperopt.STATUS_OK, } singleton_summary = pd.DataFrame.from_records([record], index="name") if self._summary is None: self._summary = singleton_summary else: self._summary = pd.concat([self._summary, singleton_summary]) if name == self._name_of_best: self._pipelines[name] = trainable.fit(X, y) else: self._pipelines[name] = trainable def _fit_dummy(self, X, y): from lale.lib.sklearn import DummyClassifier, DummyRegressor if self.prediction_type == "regression": trainable = DummyRegressor() else: trainable = DummyClassifier() self._try_and_add("dummy", trainable, X, y) def _fit_gbt_num(self, X, y): from lale.lib.lale import Project from lale.lib.sklearn import SimpleImputer gbt = auto_gbt(self.prediction_type) trainable = ( Project(columns={"type": "number"}) >> SimpleImputer(strategy="mean") >> gbt ) self._try_and_add("gbt_num", trainable, X, y) def _fit_gbt_all(self, X, y): prep = auto_prep(X) gbt = auto_gbt(self.prediction_type) trainable = prep >> gbt self._try_and_add("gbt_all", trainable, X, y) def _fit_hyperopt(self, X, y): from lale.lib.lale import Hyperopt, NoOp from lale.lib.sklearn import ( PCA, DecisionTreeClassifier, DecisionTreeRegressor, KNeighborsClassifier, KNeighborsRegressor, MinMaxScaler, RandomForestClassifier, RandomForestRegressor, RobustScaler, SelectKBest, SGDClassifier, SGDRegressor, StandardScaler, ) remaining_time = self.max_opt_time - (time.time() - self._start_fit) if remaining_time <= 0: return prep = auto_prep(X) scale = MinMaxScaler | StandardScaler | RobustScaler | NoOp reduce_dims = PCA | SelectKBest | NoOp gbt = auto_gbt(self.prediction_type) if self.prediction_type == "regression": estim_trees = gbt | DecisionTreeRegressor | RandomForestRegressor estim_notree = SGDRegressor | KNeighborsRegressor else: estim_trees = gbt | DecisionTreeClassifier | RandomForestClassifier estim_notree = SGDClassifier | KNeighborsClassifier model_trees = reduce_dims >> estim_trees model_notree = scale >> reduce_dims >> estim_notree planned = prep >> (model_trees | model_notree) prior_evals = self._summary.shape[0] if self._summary is not None else 0 trainable = Hyperopt( estimator=planned, max_evals=self.max_evals - prior_evals, scoring=self.scoring, best_score=self.best_score, max_opt_time=remaining_time, max_eval_time=self.max_eval_time, verbose=self.verbose, show_progressbar=False, cv=self.cv, ) trained = trainable.fit(X, y) # The static types are not currently smart enough to verify # that the conditionally defined summary method is actually present # But it must be, since the hyperopt impl type provides it summary: pd.DataFrame = trained.summary() # type: ignore if list(summary.status) == ["new"]: return # only one trial and that one timed out best_trial = trained._impl._trials.best_trial if "loss" in best_trial["result"]: if ( self._summary is None or best_trial["result"]["loss"] < self._summary.at[self._name_of_best, "loss"] ): self._name_of_best = f'p{best_trial["tid"]}' if self._summary is None: self._summary = summary else: self._summary = pd.concat([self._summary, summary]) for name in summary.index: assert name not in self._pipelines if summary.at[name, "status"] == hyperopt.STATUS_OK: self._pipelines[name] = trained.get_pipeline(name) def fit(self, X, y): self._start_fit = time.time() self._name_of_best = None self._summary = None self._pipelines = {} self._fit_dummy(X, y) self._fit_gbt_num(X, y) self._fit_gbt_all(X, y) self._fit_hyperopt(X, y) return self def predict(self, X, **predict_params): best_pipeline = self._pipelines[self._name_of_best] result = best_pipeline.predict(X, **predict_params) return result def summary(self): """Table summarizing the trial results (name, tid, loss, time, log_loss, status). Returns ------- result : DataFrame""" if self._summary is not None: self._summary.sort_values(by="loss", inplace=True) return self._summary def get_pipeline( self, pipeline_name: Optional[str] = None, astype: lale.helpers.astype_type = "lale", ): """Retrieve one of the trials. Parameters ---------- pipeline_name : union type, default None - string Key for table returned by summary(), return a trainable pipeline. - None When not specified, return the best trained pipeline found. astype : 'lale' or 'sklearn', default 'lale' Type of resulting pipeline. Returns ------- result : Trained operator if best, trainable operator otherwise.""" if pipeline_name is None: pipeline_name = self._name_of_best result = self._pipelines[pipeline_name] if result is None or astype == "lale": return result assert astype == "sklearn", astype return result.export_to_sklearn_pipeline() _hyperparams_schema = { "allOf": [ { "type": "object", "required": [ "prediction_type", "scoring", "max_evals", "max_opt_time", "max_eval_time", "cv", ], "relevantToOptimizer": [], "additionalProperties": False, "properties": { "prediction_type": { "description": "The kind of learning problem.", "enum": ["binary", "multiclass", "classification", "regression"], "default": "classification", }, "scoring": schema_scoring_single, "best_score": schema_best_score_single, "verbose": { "description": """Whether to print errors from each of the trials if any. This is also logged using logger.warning in Hyperopt.""", "type": "boolean", "default": False, }, "max_evals": { "description": "Number of trials of Hyperopt search.", "type": "integer", "minimum": 1, "default": 100, }, "max_opt_time": { **schema_max_opt_time, "default": 600.0, }, "max_eval_time": { "description": "Maximum time in seconds for each evaluation.", "anyOf": [ {"type": "number", "minimum": 0.0, "exclusiveMinimum": True}, {"description": "No runtime bound.", "enum": [None]}, ], "default": 120.0, }, "cv": schema_cv, }, } ] } _input_fit_schema = { "type": "object", "required": ["X", "y"], "properties": { "X": { "type": "array", "items": {"type": "array", "items": {"laleType": "Any"}}, }, "y": { "anyOf": [ {"type": "array", "items": {"type": "number"}}, {"type": "array", "items": {"type": "string"}}, {"type": "array", "items": {"type": "boolean"}}, ] }, }, } _input_predict_schema = { "type": "object", "required": ["X"], "properties": { "X": {"type": "array", "items": {"type": "array", "items": {"laleType": "Any"}}} }, } _output_predict_schema = { "anyOf": [ {"type": "array", "items": {"type": "number"}}, {"type": "array", "items": {"type": "string"}}, {"type": "array", "items": {"type": "boolean"}}, ] } _combined_schemas = { "description": """Automatically find a pipeline for a dataset. This is a high-level entry point to get an initial trained pipeline without having to specify your own planned pipeline first. It is designed to be simple at the expense of not offering much control. For an example, see `demo_auto_pipeline.ipynb`_. .. _`demo_auto_pipeline.ipynb`: https://nbviewer.jupyter.org/github/IBM/lale/blob/master/examples/demo_auto_pipeline.ipynb """, "documentation_url": "https://lale.readthedocs.io/en/latest/modules/lale.lib.lale.auto_pipeline.html", "import_from": "lale.lib.lale", "type": "object", "tags": {"pre": [], "op": ["estimator"], "post": []}, "properties": { "hyperparams": _hyperparams_schema, "input_fit": _input_fit_schema, "input_predict": _input_predict_schema, "output_predict": _output_predict_schema, }, } AutoPipeline = lale.operators.make_operator(_AutoPipelineImpl, _combined_schemas) lale.docstrings.set_docstrings(AutoPipeline)
14,599
32.87471
122
py
lale
lale-master/docs/conf.py
# Copyright 2019 IBM Corporation # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # -*- coding: utf-8 -*- # # Configuration file for the Sphinx documentation builder. # # This file does only contain a selection of the most common options. For a # full list see the documentation: # http://www.sphinx-doc.org/en/master/config # -- Path setup -------------------------------------------------------------- # If extensions (or modules to document with autodoc) are in another directory, # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. # # import os # import sys # sys.path.insert(0, os.path.abspath('.')) import builtins import os import sys from typing import Dict, List import lale from lale.settings import set_disable_hyperparams_schema_validation # -- Project information ----------------------------------------------------- project = "LALE" project_copyright = "2019-2022, IBM AI Research" author = "IBM AI Research" # The short X.Y version version = lale.__version__ # The full version, including alpha/beta/rc tags release = f"{lale.__version__}-dev" sys.path.append(os.path.join(os.path.dirname(__name__), "../lale")) import sphinx_rtd_theme # isort:skip # noqa:E402 # pylint:disable=wrong-import-position,wrong-import-order # For packages with mock imports, if we have wrappers without our impl classes, # schema validation fails as the mocking adds methods such as `transform`, `predict` etc. # when the schema may not have those tags. So we disable schema validation during doc generation. set_disable_hyperparams_schema_validation(True) # This is so that we can detect if we are running a sphinx build # and so generate pseudo-classes for documentation setattr(builtins, "__sphinx_build__", True) # -- General configuration --------------------------------------------------- # If your documentation needs a minimal Sphinx version, state it here. # # needs_sphinx = '1.0' # Add any Sphinx extension module names here, as strings. They can be # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom # ones. extensions = [ "sphinx.ext.autodoc", "sphinx.ext.doctest", "sphinx.ext.napoleon", "sphinx.ext.intersphinx", "sphinx.ext.todo", "sphinx.ext.coverage", "sphinx.ext.imgmath", "sphinx.ext.ifconfig", "sphinx.ext.viewcode", "sphinxcontrib.rsvgconverter", "m2r2", "sphinxcontrib.apidoc", ] apidoc_module_dir = "../lale" apidoc_output_dir = "modules" apidoc_separate_modules = True autoclass_content = "both" # Mock requirements to save resources during doc build machine setup autodoc_mock_imports = [ "aif360", "autoai_libs", "fairlearn", "pytorch", "tensorflow", "torch", ] # Add any paths that contain templates here, relative to this directory. templates_path = ["_templates"] # The suffix(es) of source filenames. # You can specify multiple suffix as a list of string: # source_suffix = [".rst", ".md"] # The master toctree document. master_doc = "index" # The language for content autogenerated by Sphinx. Refer to documentation # for a list of supported languages. # # This is also used if you do content translation via gettext catalogs. # Usually you set "language" from the command line for these cases. language = "en" # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. # This pattern also affects html_static_path and html_extra_path. exclude_patterns = ["_build", "Thumbs.db", ".DS_Store", "README-*.md"] # The name of the Pygments (syntax highlighting) style to use. pygments_style = None # -- Options for HTML output ------------------------------------------------- # The theme to use for HTML and HTML Help pages. See the documentation for # a list of builtin themes. # html_theme = "sphinx_rtd_theme" html_theme_path = [sphinx_rtd_theme.get_html_theme_path()] # Theme options are theme-specific and customize the look and feel of a theme # further. For a list of options available for each theme, see the # documentation. # # html_theme_options = {} # Add any paths that contain custom static files (such as style sheets) here, # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". # html_static_path = ["_static"] html_static_path: List[str] = [] # Custom sidebar templates, must be a dictionary that maps document names # to template names. # # The default sidebars (for documents that don't match any pattern) are # defined by theme itself. Builtin themes are using these templates by # default: ``['localtoc.html', 'relations.html', 'sourcelink.html', # 'searchbox.html']``. # # html_sidebars = {} # -- Options for HTMLHelp output --------------------------------------------- # Output file base name for HTML help builder. htmlhelp_basename = "LALEdoc" # -- Options for LaTeX output ------------------------------------------------ latex_elements: Dict[str, str] = { # The paper size ('letterpaper' or 'a4paper'). # # 'papersize': 'letterpaper', # The font size ('10pt', '11pt' or '12pt'). # # 'pointsize': '10pt', # Additional stuff for the LaTeX preamble. # # 'preamble': '', # Latex figure (float) alignment # # 'figure_align': 'htbp', } # Grouping the document tree into LaTeX files. List of tuples # (source start file, target name, title, # author, documentclass [howto, manual, or own class]). latex_documents = [ (master_doc, "LALE.tex", "LALE Documentation", "IBM AI Research", "manual"), ] # -- Options for manual page output ------------------------------------------ # One entry per manual page. List of tuples # (source start file, name, description, authors, manual section). man_pages = [(master_doc, "lale", "LALE Documentation", [author], 1)] # -- Options for Texinfo output ---------------------------------------------- # Grouping the document tree into Texinfo files. List of tuples # (source start file, target name, title, author, # dir menu entry, description, category) texinfo_documents = [ ( master_doc, "LALE", "LALE Documentation", author, "LALE", "One line description of project.", "Miscellaneous", ), ] # -- Options for Epub output ------------------------------------------------- # Bibliographic Dublin Core info. epub_title = project # The unique identifier of the text. This can be a ISBN number # or the project homepage. # # epub_identifier = '' # A unique identification for the text. # # epub_uid = '' # A list of files that should not be packed into the epub file. epub_exclude_files = ["search.html"] # -- Extension configuration ------------------------------------------------- # -- Options for intersphinx extension --------------------------------------- # Example configuration for intersphinx: refer to the Python standard library. intersphinx_mapping = {"https://docs.python.org/": None} # -- Options for todo extension ---------------------------------------------- # If true, `todo` and `todoList` produce output, else they produce nothing. todo_include_todos = True
7,789
30.538462
108
py
PoolNet
PoolNet-master/main.py
import argparse import os from dataset.dataset import get_loader from solver import Solver def get_test_info(sal_mode='e'): if sal_mode == 'e': image_root = './data/ECSSD/Imgs/' image_source = './data/ECSSD/test.lst' elif sal_mode == 'p': image_root = './data/PASCALS/Imgs/' image_source = './data/PASCALS/test.lst' elif sal_mode == 'd': image_root = './data/DUTOMRON/Imgs/' image_source = './data/DUTOMRON/test.lst' elif sal_mode == 'h': image_root = './data/HKU-IS/Imgs/' image_source = './data/HKU-IS/test.lst' elif sal_mode == 's': image_root = './data/SOD/Imgs/' image_source = './data/SOD/test.lst' elif sal_mode == 't': image_root = './data/DUTS-TE/Imgs/' image_source = './data/DUTS-TE/test.lst' elif sal_mode == 'm_r': # for speed test image_root = './data/MSRA/Imgs_resized/' image_source = './data/MSRA/test_resized.lst' return image_root, image_source def main(config): if config.mode == 'train': train_loader = get_loader(config) run = 0 while os.path.exists("%s/run-%d" % (config.save_folder, run)): run += 1 os.mkdir("%s/run-%d" % (config.save_folder, run)) os.mkdir("%s/run-%d/models" % (config.save_folder, run)) config.save_folder = "%s/run-%d" % (config.save_folder, run) train = Solver(train_loader, None, config) train.train() elif config.mode == 'test': config.test_root, config.test_list = get_test_info(config.sal_mode) test_loader = get_loader(config, mode='test') if not os.path.exists(config.test_fold): os.mkdir(config.test_fold) test = Solver(None, test_loader, config) test.test() else: raise IOError("illegal input!!!") if __name__ == '__main__': vgg_path = './dataset/pretrained/vgg16_20M.pth' resnet_path = './dataset/pretrained/resnet50_caffe.pth' parser = argparse.ArgumentParser() # Hyper-parameters parser.add_argument('--n_color', type=int, default=3) parser.add_argument('--lr', type=float, default=5e-5) # Learning rate resnet:5e-5, vgg:1e-4 parser.add_argument('--wd', type=float, default=0.0005) # Weight decay parser.add_argument('--no-cuda', dest='cuda', action='store_false') # Training settings parser.add_argument('--arch', type=str, default='resnet') # resnet or vgg parser.add_argument('--pretrained_model', type=str, default=resnet_path) parser.add_argument('--epoch', type=int, default=24) parser.add_argument('--batch_size', type=int, default=1) # only support 1 now parser.add_argument('--num_thread', type=int, default=1) parser.add_argument('--load', type=str, default='') parser.add_argument('--save_folder', type=str, default='./results') parser.add_argument('--epoch_save', type=int, default=3) parser.add_argument('--iter_size', type=int, default=10) parser.add_argument('--show_every', type=int, default=50) # Train data parser.add_argument('--train_root', type=str, default='') parser.add_argument('--train_list', type=str, default='') # Testing settings parser.add_argument('--model', type=str, default=None) # Snapshot parser.add_argument('--test_fold', type=str, default=None) # Test results saving folder parser.add_argument('--sal_mode', type=str, default='e') # Test image dataset # Misc parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) config = parser.parse_args() if not os.path.exists(config.save_folder): os.mkdir(config.save_folder) # Get test set info test_root, test_list = get_test_info(config.sal_mode) config.test_root = test_root config.test_list = test_list main(config)
3,827
38.061224
95
py
PoolNet
PoolNet-master/joint_main.py
import argparse import os from dataset.joint_dataset import get_loader from joint_solver import Solver def get_test_info(sal_mode='e'): if sal_mode == 'e': image_root = './data/ECSSD/Imgs/' image_source = './data/ECSSD/test.lst' elif sal_mode == 'p': image_root = './data/PASCALS/Imgs/' image_source = './data/PASCALS/test.lst' elif sal_mode == 'd': image_root = './data/DUTOMRON/Imgs/' image_source = './data/DUTOMRON/test.lst' elif sal_mode == 'h': image_root = './data/HKU-IS/Imgs/' image_source = './data/HKU-IS/test.lst' elif sal_mode == 's': image_root = './data/SOD/Imgs/' image_source = './data/SOD/test.lst' elif sal_mode == 't': image_root = './data/DUTS-TE/Imgs/' image_source = './data/DUTS-TE/test.lst' elif sal_mode == 'm_r': # for speed test image_root = './data/MSRA/Imgs_resized/' image_source = './data/MSRA/test_resized.lst' elif sal_mode == 'b': # BSDS dataset for edge evaluation image_root = './data/HED-BSDS_PASCAL/HED-BSDS/test/' image_source = './data/HED-BSDS_PASCAL/HED-BSDS/test.lst' return image_root, image_source def main(config): if config.mode == 'train': train_loader = get_loader(config) run = 0 while os.path.exists("%s/run-%d" % (config.save_folder, run)): run += 1 os.mkdir("%s/run-%d" % (config.save_folder, run)) os.mkdir("%s/run-%d/models" % (config.save_folder, run)) config.save_folder = "%s/run-%d" % (config.save_folder, run) train = Solver(train_loader, None, config) train.train() elif config.mode == 'test': config.test_root, config.test_list = get_test_info(config.sal_mode) test_loader = get_loader(config, mode='test') if not os.path.exists(config.test_fold): os.mkdir(config.test_fold) test = Solver(None, test_loader, config) test.test(test_mode=config.test_mode) else: raise IOError("illegal input!!!") if __name__ == '__main__': vgg_path = './dataset/pretrained/vgg16_20M.pth' resnet_path = './dataset/pretrained/resnet50_caffe.pth' parser = argparse.ArgumentParser() # Hyper-parameters parser.add_argument('--n_color', type=int, default=3) parser.add_argument('--lr', type=float, default=5e-5) # Learning rate resnet:5e-5, vgg:1e-4 parser.add_argument('--wd', type=float, default=0.0005) # Weight decay parser.add_argument('--no-cuda', dest='cuda', action='store_false') # Training settings parser.add_argument('--arch', type=str, default='resnet') # resnet or vgg parser.add_argument('--pretrained_model', type=str, default=resnet_path) parser.add_argument('--epoch', type=int, default=11) parser.add_argument('--batch_size', type=int, default=1) # only support 1 now parser.add_argument('--num_thread', type=int, default=1) parser.add_argument('--load', type=str, default='') parser.add_argument('--save_folder', type=str, default='./results') parser.add_argument('--epoch_save', type=int, default=3) parser.add_argument('--iter_size', type=int, default=10) parser.add_argument('--show_every', type=int, default=50) # Train data parser.add_argument('--train_root', type=str, default='') parser.add_argument('--train_list', type=str, default='') parser.add_argument('--train_edge_root', type=str, default='') # path for edge data parser.add_argument('--train_edge_list', type=str, default='') # list file for edge data # Testing settings parser.add_argument('--model', type=str, default=None) # Snapshot parser.add_argument('--test_fold', type=str, default=None) # Test results saving folder parser.add_argument('--test_mode', type=int, default=1) # 0->edge, 1->saliency parser.add_argument('--sal_mode', type=str, default='e') # Test image dataset # Misc parser.add_argument('--mode', type=str, default='train', choices=['train', 'test']) config = parser.parse_args() if not os.path.exists(config.save_folder): os.mkdir(config.save_folder) # Get test set info test_root, test_list = get_test_info(config.sal_mode) config.test_root = test_root config.test_list = test_list main(config)
4,316
40.912621
95
py
PoolNet
PoolNet-master/solver.py
import torch from collections import OrderedDict from torch.nn import utils, functional as F from torch.optim import Adam from torch.autograd import Variable from torch.backends import cudnn from networks.poolnet import build_model, weights_init import scipy.misc as sm import numpy as np import os import torchvision.utils as vutils import cv2 import math import time class Solver(object): def __init__(self, train_loader, test_loader, config): self.train_loader = train_loader self.test_loader = test_loader self.config = config self.iter_size = config.iter_size self.show_every = config.show_every self.lr_decay_epoch = [15,] self.build_model() if config.mode == 'test': print('Loading pre-trained model from %s...' % self.config.model) if self.config.cuda: self.net.load_state_dict(torch.load(self.config.model)) else: self.net.load_state_dict(torch.load(self.config.model, map_location='cpu')) self.net.eval() # print the network information and parameter numbers def print_network(self, model, name): num_params = 0 for p in model.parameters(): num_params += p.numel() print(name) print(model) print("The number of parameters: {}".format(num_params)) # build the network def build_model(self): self.net = build_model(self.config.arch) if self.config.cuda: self.net = self.net.cuda() # self.net.train() self.net.eval() # use_global_stats = True self.net.apply(weights_init) if self.config.load == '': self.net.base.load_pretrained_model(torch.load(self.config.pretrained_model)) else: self.net.load_state_dict(torch.load(self.config.load)) self.lr = self.config.lr self.wd = self.config.wd self.optimizer = Adam(filter(lambda p: p.requires_grad, self.net.parameters()), lr=self.lr, weight_decay=self.wd) self.print_network(self.net, 'PoolNet Structure') def test(self): mode_name = 'sal_fuse' time_s = time.time() img_num = len(self.test_loader) for i, data_batch in enumerate(self.test_loader): images, name, im_size = data_batch['image'], data_batch['name'][0], np.asarray(data_batch['size']) with torch.no_grad(): images = Variable(images) if self.config.cuda: images = images.cuda() preds = self.net(images) pred = np.squeeze(torch.sigmoid(preds).cpu().data.numpy()) multi_fuse = 255 * pred cv2.imwrite(os.path.join(self.config.test_fold, name[:-4] + '_' + mode_name + '.png'), multi_fuse) time_e = time.time() print('Speed: %f FPS' % (img_num/(time_e-time_s))) print('Test Done!') # training phase def train(self): iter_num = len(self.train_loader.dataset) // self.config.batch_size aveGrad = 0 for epoch in range(self.config.epoch): r_sal_loss= 0 self.net.zero_grad() for i, data_batch in enumerate(self.train_loader): sal_image, sal_label = data_batch['sal_image'], data_batch['sal_label'] if (sal_image.size(2) != sal_label.size(2)) or (sal_image.size(3) != sal_label.size(3)): print('IMAGE ERROR, PASSING```') continue sal_image, sal_label= Variable(sal_image), Variable(sal_label) if self.config.cuda: # cudnn.benchmark = True sal_image, sal_label = sal_image.cuda(), sal_label.cuda() sal_pred = self.net(sal_image) sal_loss_fuse = F.binary_cross_entropy_with_logits(sal_pred, sal_label, reduction='sum') sal_loss = sal_loss_fuse / (self.iter_size * self.config.batch_size) r_sal_loss += sal_loss.data sal_loss.backward() aveGrad += 1 # accumulate gradients as done in DSS if aveGrad % self.iter_size == 0: self.optimizer.step() self.optimizer.zero_grad() aveGrad = 0 if i % (self.show_every // self.config.batch_size) == 0: if i == 0: x_showEvery = 1 print('epoch: [%2d/%2d], iter: [%5d/%5d] || Sal : %10.4f' % ( epoch, self.config.epoch, i, iter_num, r_sal_loss/x_showEvery)) print('Learning rate: ' + str(self.lr)) r_sal_loss= 0 if (epoch + 1) % self.config.epoch_save == 0: torch.save(self.net.state_dict(), '%s/models/epoch_%d.pth' % (self.config.save_folder, epoch + 1)) if epoch in self.lr_decay_epoch: self.lr = self.lr * 0.1 self.optimizer = Adam(filter(lambda p: p.requires_grad, self.net.parameters()), lr=self.lr, weight_decay=self.wd) torch.save(self.net.state_dict(), '%s/models/final.pth' % self.config.save_folder) def bce2d(input, target, reduction=None): assert(input.size() == target.size()) pos = torch.eq(target, 1).float() neg = torch.eq(target, 0).float() num_pos = torch.sum(pos) num_neg = torch.sum(neg) num_total = num_pos + num_neg alpha = num_neg / num_total beta = 1.1 * num_pos / num_total # target pixel = 1 -> weight beta # target pixel = 0 -> weight 1-beta weights = alpha * pos + beta * neg return F.binary_cross_entropy_with_logits(input, target, weights, reduction=reduction)
5,765
38.493151
129
py
PoolNet
PoolNet-master/joint_solver.py
import torch from collections import OrderedDict from torch.nn import utils, functional as F from torch.optim import Adam from torch.autograd import Variable from torch.backends import cudnn from networks.joint_poolnet import build_model, weights_init import scipy.misc as sm import numpy as np import os import torchvision.utils as vutils import cv2 import math import time class Solver(object): def __init__(self, train_loader, test_loader, config): self.train_loader = train_loader self.test_loader = test_loader self.config = config self.iter_size = config.iter_size self.show_every = config.show_every self.lr_decay_epoch = [8,] self.build_model() if config.mode == 'test': print('Loading pre-trained model from %s...' % self.config.model) if self.config.cuda: self.net.load_state_dict(torch.load(self.config.model)) else: self.net.load_state_dict(torch.load(self.config.model, map_location='cpu')) self.net.eval() # print the network information and parameter numbers def print_network(self, model, name): num_params = 0 for p in model.parameters(): num_params += p.numel() print(name) print(model) print("The number of parameters: {}".format(num_params)) # build the network def build_model(self): self.net = build_model(self.config.arch) if self.config.cuda: self.net = self.net.cuda() # self.net.train() self.net.eval() # use_global_stats = True self.net.apply(weights_init) if self.config.load == '': self.net.base.load_pretrained_model(torch.load(self.config.pretrained_model)) else: self.net.load_state_dict(torch.load(self.config.load)) self.lr = self.config.lr self.wd = self.config.wd self.optimizer = Adam(filter(lambda p: p.requires_grad, self.net.parameters()), lr=self.lr, weight_decay=self.wd) self.print_network(self.net, 'PoolNet Structure') def test(self, test_mode=1): mode_name = ['edge_fuse', 'sal_fuse'] EPSILON = 1e-8 time_s = time.time() img_num = len(self.test_loader) for i, data_batch in enumerate(self.test_loader): images, name, im_size = data_batch['image'], data_batch['name'][0], np.asarray(data_batch['size']) if test_mode == 0: images = images.numpy()[0].transpose((1,2,0)) scale = [0.5, 1, 1.5, 2] # uncomment for multi-scale testing # scale = [1] multi_fuse = np.zeros(im_size, np.float32) for k in range(0, len(scale)): im_ = cv2.resize(images, None, fx=scale[k], fy=scale[k], interpolation=cv2.INTER_LINEAR) im_ = im_.transpose((2, 0, 1)) im_ = torch.Tensor(im_[np.newaxis, ...]) with torch.no_grad(): im_ = Variable(im_) if self.config.cuda: im_ = im_.cuda() preds = self.net(im_, mode=test_mode) pred_0 = np.squeeze(torch.sigmoid(preds[1][0]).cpu().data.numpy()) pred_1 = np.squeeze(torch.sigmoid(preds[1][1]).cpu().data.numpy()) pred_2 = np.squeeze(torch.sigmoid(preds[1][2]).cpu().data.numpy()) pred_fuse = np.squeeze(torch.sigmoid(preds[0]).cpu().data.numpy()) pred = (pred_0 + pred_1 + pred_2 + pred_fuse) / 4 pred = (pred - np.min(pred) + EPSILON) / (np.max(pred) - np.min(pred) + EPSILON) pred = cv2.resize(pred, (im_size[1], im_size[0]), interpolation=cv2.INTER_LINEAR) multi_fuse += pred multi_fuse /= len(scale) multi_fuse = 255 * (1 - multi_fuse) cv2.imwrite(os.path.join(self.config.test_fold, name[:-4] + '_' + mode_name[test_mode] + '.png'), multi_fuse) elif test_mode == 1: with torch.no_grad(): images = Variable(images) if self.config.cuda: images = images.cuda() preds = self.net(images, mode=test_mode) pred = np.squeeze(torch.sigmoid(preds).cpu().data.numpy()) multi_fuse = 255 * pred cv2.imwrite(os.path.join(self.config.test_fold, name[:-4] + '_' + mode_name[test_mode] + '.png'), multi_fuse) time_e = time.time() print('Speed: %f FPS' % (img_num/(time_e-time_s))) print('Test Done!') # training phase def train(self): iter_num = 30000 # each batch only train 30000 iters.(This number is just a random choice...) aveGrad = 0 for epoch in range(self.config.epoch): r_edge_loss, r_sal_loss, r_sum_loss= 0,0,0 self.net.zero_grad() for i, data_batch in enumerate(self.train_loader): if (i + 1) == iter_num: break edge_image, edge_label, sal_image, sal_label = data_batch['edge_image'], data_batch['edge_label'], data_batch['sal_image'], data_batch['sal_label'] if (sal_image.size(2) != sal_label.size(2)) or (sal_image.size(3) != sal_label.size(3)): print('IMAGE ERROR, PASSING```') continue edge_image, edge_label, sal_image, sal_label= Variable(edge_image), Variable(edge_label), Variable(sal_image), Variable(sal_label) if self.config.cuda: edge_image, edge_label, sal_image, sal_label = edge_image.cuda(), edge_label.cuda(), sal_image.cuda(), sal_label.cuda() # edge part edge_pred = self.net(edge_image, mode=0) edge_loss_fuse = bce2d(edge_pred[0], edge_label, reduction='sum') edge_loss_part = [] for ix in edge_pred[1]: edge_loss_part.append(bce2d(ix, edge_label, reduction='sum')) edge_loss = (edge_loss_fuse + sum(edge_loss_part)) / (self.iter_size * self.config.batch_size) r_edge_loss += edge_loss.data # sal part sal_pred = self.net(sal_image, mode=1) sal_loss_fuse = F.binary_cross_entropy_with_logits(sal_pred, sal_label, reduction='sum') sal_loss = sal_loss_fuse / (self.iter_size * self.config.batch_size) r_sal_loss += sal_loss.data loss = sal_loss + edge_loss r_sum_loss += loss.data loss.backward() aveGrad += 1 # accumulate gradients as done in DSS if aveGrad % self.iter_size == 0: self.optimizer.step() self.optimizer.zero_grad() aveGrad = 0 if i % (self.show_every // self.config.batch_size) == 0: if i == 0: x_showEvery = 1 print('epoch: [%2d/%2d], iter: [%5d/%5d] || Edge : %10.4f || Sal : %10.4f || Sum : %10.4f' % ( epoch, self.config.epoch, i, iter_num, r_edge_loss/x_showEvery, r_sal_loss/x_showEvery, r_sum_loss/x_showEvery)) print('Learning rate: ' + str(self.lr)) r_edge_loss, r_sal_loss, r_sum_loss= 0,0,0 if (epoch + 1) % self.config.epoch_save == 0: torch.save(self.net.state_dict(), '%s/models/epoch_%d.pth' % (self.config.save_folder, epoch + 1)) if epoch in self.lr_decay_epoch: self.lr = self.lr * 0.1 self.optimizer = Adam(filter(lambda p: p.requires_grad, self.net.parameters()), lr=self.lr, weight_decay=self.wd) torch.save(self.net.state_dict(), '%s/models/final.pth' % self.config.save_folder) def bce2d(input, target, reduction=None): assert(input.size() == target.size()) pos = torch.eq(target, 1).float() neg = torch.eq(target, 0).float() num_pos = torch.sum(pos) num_neg = torch.sum(neg) num_total = num_pos + num_neg alpha = num_neg / num_total beta = 1.1 * num_pos / num_total # target pixel = 1 -> weight beta # target pixel = 0 -> weight 1-beta weights = alpha * pos + beta * neg return F.binary_cross_entropy_with_logits(input, target, weights, reduction=reduction)
8,569
44.105263
163
py
PoolNet
PoolNet-master/networks/joint_poolnet.py
import torch from torch import nn from torch.nn import init import torch.nn.functional as F import math from torch.autograd import Variable import numpy as np from .deeplab_resnet import resnet50_locate from .vgg import vgg16_locate config_vgg = {'convert': [[128,256,512,512,512],[64,128,256,512,512]], 'deep_pool': [[512, 512, 256, 128], [512, 256, 128, 128], [True, True, True, False], [True, True, True, False]], 'score': 256, 'edgeinfoc':[48,128], 'block': [[512, [16]], [256, [16]], [128, [16]]], 'fuse': [[16, 16, 16], True]} # no convert layer, no conv6 config_resnet = {'convert': [[64,256,512,1024,2048],[128,256,256,512,512]], 'deep_pool': [[512, 512, 256, 256, 128], [512, 256, 256, 128, 128], [False, True, True, True, False], [True, True, True, True, False]], 'score': 256, 'edgeinfoc':[64,128], 'block': [[512, [16]], [256, [16]], [256, [16]], [128, [16]]], 'fuse': [[16, 16, 16, 16], True]} class ConvertLayer(nn.Module): def __init__(self, list_k): super(ConvertLayer, self).__init__() up = [] for i in range(len(list_k[0])): up.append(nn.Sequential(nn.Conv2d(list_k[0][i], list_k[1][i], 1, 1, bias=False), nn.ReLU(inplace=True))) self.convert0 = nn.ModuleList(up) def forward(self, list_x): resl = [] for i in range(len(list_x)): resl.append(self.convert0[i](list_x[i])) return resl class DeepPoolLayer(nn.Module): def __init__(self, k, k_out, need_x2, need_fuse): super(DeepPoolLayer, self).__init__() self.pools_sizes = [2,4,8] self.need_x2 = need_x2 self.need_fuse = need_fuse pools, convs = [],[] for i in self.pools_sizes: pools.append(nn.AvgPool2d(kernel_size=i, stride=i)) convs.append(nn.Conv2d(k, k, 3, 1, 1, bias=False)) self.pools = nn.ModuleList(pools) self.convs = nn.ModuleList(convs) self.relu = nn.ReLU() self.conv_sum = nn.Conv2d(k, k_out, 3, 1, 1, bias=False) if self.need_fuse: self.conv_sum_c = nn.Conv2d(k_out, k_out, 3, 1, 1, bias=False) def forward(self, x, x2=None, x3=None): x_size = x.size() resl = x for i in range(len(self.pools_sizes)): y = self.convs[i](self.pools[i](x)) resl = torch.add(resl, F.interpolate(y, x_size[2:], mode='bilinear', align_corners=True)) resl = self.relu(resl) if self.need_x2: resl = F.interpolate(resl, x2.size()[2:], mode='bilinear', align_corners=True) resl = self.conv_sum(resl) if self.need_fuse: resl = self.conv_sum_c(torch.add(torch.add(resl, x2), x3)) return resl class BlockLayer(nn.Module): def __init__(self, k_in, k_out_list): super(BlockLayer, self).__init__() up_in1, up_mid1, up_in2, up_mid2, up_out = [], [], [], [], [] for k in k_out_list: up_in1.append(nn.Conv2d(k_in, k_in//4, 1, 1, bias=False)) up_mid1.append(nn.Sequential(nn.Conv2d(k_in//4, k_in//4, 3, 1, 1, bias=False), nn.Conv2d(k_in//4, k_in, 1, 1, bias=False))) up_in2.append(nn.Conv2d(k_in, k_in//4, 1, 1, bias=False)) up_mid2.append(nn.Sequential(nn.Conv2d(k_in//4, k_in//4, 3, 1, 1, bias=False), nn.Conv2d(k_in//4, k_in, 1, 1, bias=False))) up_out.append(nn.Conv2d(k_in, k, 1, 1, bias=False)) self.block_in1 = nn.ModuleList(up_in1) self.block_in2 = nn.ModuleList(up_in2) self.block_mid1 = nn.ModuleList(up_mid1) self.block_mid2 = nn.ModuleList(up_mid2) self.block_out = nn.ModuleList(up_out) self.relu = nn.ReLU() def forward(self, x, mode=0): x_tmp = self.relu(x + self.block_mid1[mode](self.block_in1[mode](x))) # x_tmp = self.block_mid2[mode](self.block_in2[mode](self.relu(x + x_tmp))) x_tmp = self.relu(x_tmp + self.block_mid2[mode](self.block_in2[mode](x_tmp))) x_tmp = self.block_out[mode](x_tmp) return x_tmp class EdgeInfoLayerC(nn.Module): def __init__(self, k_in, k_out): super(EdgeInfoLayerC, self).__init__() self.trans = nn.Sequential(nn.Conv2d(k_in, k_in, 3, 1, 1, bias=False), nn.ReLU(inplace=True), nn.Conv2d(k_in, k_out, 3, 1, 1, bias=False), nn.ReLU(inplace=True), nn.Conv2d(k_out, k_out, 3, 1, 1, bias=False), nn.ReLU(inplace=True), nn.Conv2d(k_out, k_out, 3, 1, 1, bias=False), nn.ReLU(inplace=True)) def forward(self, x, x_size): tmp_x = [] for i_x in x: tmp_x.append(F.interpolate(i_x, x_size[2:], mode='bilinear', align_corners=True)) x = self.trans(torch.cat(tmp_x, dim=1)) return x class FuseLayer1(nn.Module): def __init__(self, list_k, deep_sup): super(FuseLayer1, self).__init__() up = [] for i in range(len(list_k)): up.append(nn.Conv2d(list_k[i], 1, 1, 1)) self.trans = nn.ModuleList(up) self.fuse = nn.Conv2d(len(list_k), 1, 1, 1) self.deep_sup = deep_sup def forward(self, list_x, x_size): up_x = [] for i, i_x in enumerate(list_x): up_x.append(F.interpolate(self.trans[i](i_x), x_size[2:], mode='bilinear', align_corners=True)) out_fuse = self.fuse(torch.cat(up_x, dim = 1)) if self.deep_sup: out_all = [] for up_i in up_x: out_all.append(up_i) return [out_fuse, out_all] else: return [out_fuse] class ScoreLayer(nn.Module): def __init__(self, k): super(ScoreLayer, self).__init__() self.score = nn.Conv2d(k ,1, 3, 1, 1) def forward(self, x, x_size=None): x = self.score(x) if x_size is not None: x = F.interpolate(x, x_size[2:], mode='bilinear', align_corners=True) return x def extra_layer(base_model_cfg, base): if base_model_cfg == 'vgg': config = config_vgg elif base_model_cfg == 'resnet': config = config_resnet convert_layers, deep_pool_layers, block_layers, fuse_layers, edgeinfo_layers, score_layers = [], [], [], [], [], [] convert_layers = ConvertLayer(config['convert']) for k in config['block']: block_layers += [BlockLayer(k[0], k[1])] for i in range(len(config['deep_pool'][0])): deep_pool_layers += [DeepPoolLayer(config['deep_pool'][0][i], config['deep_pool'][1][i], config['deep_pool'][2][i], config['deep_pool'][3][i])] fuse_layers = FuseLayer1(config['fuse'][0], config['fuse'][1]) edgeinfo_layers = EdgeInfoLayerC(config['edgeinfoc'][0], config['edgeinfoc'][1]) score_layers = ScoreLayer(config['score']) return base, convert_layers, deep_pool_layers, block_layers, fuse_layers, edgeinfo_layers, score_layers class PoolNet(nn.Module): def __init__(self, base_model_cfg, base, convert_layers, deep_pool_layers, block_layers, fuse_layers, edgeinfo_layers, score_layers): super(PoolNet, self).__init__() self.base_model_cfg = base_model_cfg self.base = base self.block = nn.ModuleList(block_layers) self.deep_pool = nn.ModuleList(deep_pool_layers) self.fuse = fuse_layers self.edgeinfo = edgeinfo_layers self.score = score_layers if self.base_model_cfg == 'resnet': self.convert = convert_layers def forward(self, x, mode): x_size = x.size() conv2merge, infos = self.base(x) if self.base_model_cfg == 'resnet': conv2merge = self.convert(conv2merge) conv2merge = conv2merge[::-1] edge_merge = [] merge = self.deep_pool[0](conv2merge[0], conv2merge[1], infos[0]) edge_merge.append(merge) for k in range(1, len(conv2merge)-1): merge = self.deep_pool[k](merge, conv2merge[k+1], infos[k]) edge_merge.append(merge) if mode == 0: edge_merge = [self.block[i](kk) for i, kk in enumerate(edge_merge)] merge = self.fuse(edge_merge, x_size) elif mode == 1: merge = self.deep_pool[-1](merge) edge_merge = [self.block[i](kk).detach() for i, kk in enumerate(edge_merge)] edge_merge = self.edgeinfo(edge_merge, merge.size()) merge = self.score(torch.cat([merge, edge_merge], dim=1), x_size) return merge def build_model(base_model_cfg='vgg'): if base_model_cfg == 'vgg': return PoolNet(base_model_cfg, *extra_layer(base_model_cfg, vgg16_locate())) elif base_model_cfg == 'resnet': return PoolNet(base_model_cfg, *extra_layer(base_model_cfg, resnet50_locate())) def weights_init(m): if isinstance(m, nn.Conv2d): m.weight.data.normal_(0, 0.01) if m.bias is not None: m.bias.data.zero_()
8,853
41.772947
344
py
PoolNet
PoolNet-master/networks/vgg.py
import torch.nn as nn import math import torch import numpy as np import torch.nn.functional as F # vgg16 def vgg(cfg, i, batch_norm=False): layers = [] in_channels = i stage = 1 for v in cfg: if v == 'M': stage += 1 if stage == 6: layers += [nn.MaxPool2d(kernel_size=3, stride=1, padding=1)] else: layers += [nn.MaxPool2d(kernel_size=3, stride=2, padding=1)] else: if stage == 6: conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1) else: conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1) if batch_norm: layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)] else: layers += [conv2d, nn.ReLU(inplace=True)] in_channels = v return layers class vgg16(nn.Module): def __init__(self): super(vgg16, self).__init__() self.cfg = {'tun': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'], 'tun_ex': [512, 512, 512]} self.extract = [8, 15, 22, 29] # [3, 8, 15, 22, 29] self.base = nn.ModuleList(vgg(self.cfg['tun'], 3)) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, 0.01) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def load_pretrained_model(self, model): self.base.load_state_dict(model, strict=False) def forward(self, x): tmp_x = [] for k in range(len(self.base)): x = self.base[k](x) if k in self.extract: tmp_x.append(x) return tmp_x class vgg16_locate(nn.Module): def __init__(self): super(vgg16_locate,self).__init__() self.vgg16 = vgg16() self.in_planes = 512 self.out_planes = [512, 256, 128] ppms, infos = [], [] for ii in [1, 3, 5]: ppms.append(nn.Sequential(nn.AdaptiveAvgPool2d(ii), nn.Conv2d(self.in_planes, self.in_planes, 1, 1, bias=False), nn.ReLU(inplace=True))) self.ppms = nn.ModuleList(ppms) self.ppm_cat = nn.Sequential(nn.Conv2d(self.in_planes * 4, self.in_planes, 3, 1, 1, bias=False), nn.ReLU(inplace=True)) for ii in self.out_planes: infos.append(nn.Sequential(nn.Conv2d(self.in_planes, ii, 3, 1, 1, bias=False), nn.ReLU(inplace=True))) self.infos = nn.ModuleList(infos) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, 0.01) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def load_pretrained_model(self, model): self.vgg16.load_pretrained_model(model) def forward(self, x): x_size = x.size()[2:] xs = self.vgg16(x) xls = [xs[-1]] for k in range(len(self.ppms)): xls.append(F.interpolate(self.ppms[k](xs[-1]), xs[-1].size()[2:], mode='bilinear', align_corners=True)) xls = self.ppm_cat(torch.cat(xls, dim=1)) infos = [] for k in range(len(self.infos)): infos.append(self.infos[k](F.interpolate(xls, xs[len(self.infos) - 1 - k].size()[2:], mode='bilinear', align_corners=True))) return xs, infos
3,581
35.927835
148
py
PoolNet
PoolNet-master/networks/poolnet.py
import torch from torch import nn from torch.nn import init import torch.nn.functional as F import math from torch.autograd import Variable import numpy as np from .deeplab_resnet import resnet50_locate from .vgg import vgg16_locate config_vgg = {'convert': [[128,256,512,512,512],[64,128,256,512,512]], 'deep_pool': [[512, 512, 256, 128], [512, 256, 128, 128], [True, True, True, False], [True, True, True, False]], 'score': 128} # no convert layer, no conv6 config_resnet = {'convert': [[64,256,512,1024,2048],[128,256,256,512,512]], 'deep_pool': [[512, 512, 256, 256, 128], [512, 256, 256, 128, 128], [False, True, True, True, False], [True, True, True, True, False]], 'score': 128} class ConvertLayer(nn.Module): def __init__(self, list_k): super(ConvertLayer, self).__init__() up = [] for i in range(len(list_k[0])): up.append(nn.Sequential(nn.Conv2d(list_k[0][i], list_k[1][i], 1, 1, bias=False), nn.ReLU(inplace=True))) self.convert0 = nn.ModuleList(up) def forward(self, list_x): resl = [] for i in range(len(list_x)): resl.append(self.convert0[i](list_x[i])) return resl class DeepPoolLayer(nn.Module): def __init__(self, k, k_out, need_x2, need_fuse): super(DeepPoolLayer, self).__init__() self.pools_sizes = [2,4,8] self.need_x2 = need_x2 self.need_fuse = need_fuse pools, convs = [],[] for i in self.pools_sizes: pools.append(nn.AvgPool2d(kernel_size=i, stride=i)) convs.append(nn.Conv2d(k, k, 3, 1, 1, bias=False)) self.pools = nn.ModuleList(pools) self.convs = nn.ModuleList(convs) self.relu = nn.ReLU() self.conv_sum = nn.Conv2d(k, k_out, 3, 1, 1, bias=False) if self.need_fuse: self.conv_sum_c = nn.Conv2d(k_out, k_out, 3, 1, 1, bias=False) def forward(self, x, x2=None, x3=None): x_size = x.size() resl = x for i in range(len(self.pools_sizes)): y = self.convs[i](self.pools[i](x)) resl = torch.add(resl, F.interpolate(y, x_size[2:], mode='bilinear', align_corners=True)) resl = self.relu(resl) if self.need_x2: resl = F.interpolate(resl, x2.size()[2:], mode='bilinear', align_corners=True) resl = self.conv_sum(resl) if self.need_fuse: resl = self.conv_sum_c(torch.add(torch.add(resl, x2), x3)) return resl class ScoreLayer(nn.Module): def __init__(self, k): super(ScoreLayer, self).__init__() self.score = nn.Conv2d(k ,1, 1, 1) def forward(self, x, x_size=None): x = self.score(x) if x_size is not None: x = F.interpolate(x, x_size[2:], mode='bilinear', align_corners=True) return x def extra_layer(base_model_cfg, vgg): if base_model_cfg == 'vgg': config = config_vgg elif base_model_cfg == 'resnet': config = config_resnet convert_layers, deep_pool_layers, score_layers = [], [], [] convert_layers = ConvertLayer(config['convert']) for i in range(len(config['deep_pool'][0])): deep_pool_layers += [DeepPoolLayer(config['deep_pool'][0][i], config['deep_pool'][1][i], config['deep_pool'][2][i], config['deep_pool'][3][i])] score_layers = ScoreLayer(config['score']) return vgg, convert_layers, deep_pool_layers, score_layers class PoolNet(nn.Module): def __init__(self, base_model_cfg, base, convert_layers, deep_pool_layers, score_layers): super(PoolNet, self).__init__() self.base_model_cfg = base_model_cfg self.base = base self.deep_pool = nn.ModuleList(deep_pool_layers) self.score = score_layers if self.base_model_cfg == 'resnet': self.convert = convert_layers def forward(self, x): x_size = x.size() conv2merge, infos = self.base(x) if self.base_model_cfg == 'resnet': conv2merge = self.convert(conv2merge) conv2merge = conv2merge[::-1] edge_merge = [] merge = self.deep_pool[0](conv2merge[0], conv2merge[1], infos[0]) for k in range(1, len(conv2merge)-1): merge = self.deep_pool[k](merge, conv2merge[k+1], infos[k]) merge = self.deep_pool[-1](merge) merge = self.score(merge, x_size) return merge def build_model(base_model_cfg='vgg'): if base_model_cfg == 'vgg': return PoolNet(base_model_cfg, *extra_layer(base_model_cfg, vgg16_locate())) elif base_model_cfg == 'resnet': return PoolNet(base_model_cfg, *extra_layer(base_model_cfg, resnet50_locate())) def weights_init(m): if isinstance(m, nn.Conv2d): m.weight.data.normal_(0, 0.01) if m.bias is not None: m.bias.data.zero_()
4,800
37.103175
227
py
PoolNet
PoolNet-master/networks/deeplab_resnet.py
import torch.nn as nn import math import torch import numpy as np import torch.nn.functional as F affine_par = True def conv3x3(in_planes, out_planes, stride=1): "3x3 convolution with padding" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes, affine = affine_par) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes, affine = affine_par) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, dilation_ = 1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, stride=stride, bias=False) # change self.bn1 = nn.BatchNorm2d(planes,affine = affine_par) for i in self.bn1.parameters(): i.requires_grad = False padding = 1 if dilation_ == 2: padding = 2 elif dilation_ == 4: padding = 4 self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, # change padding=padding, bias=False, dilation = dilation_) self.bn2 = nn.BatchNorm2d(planes,affine = affine_par) for i in self.bn2.parameters(): i.requires_grad = False self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * 4, affine = affine_par) for i in self.bn3.parameters(): i.requires_grad = False self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class ResNet(nn.Module): def __init__(self, block, layers): self.inplanes = 64 super(ResNet, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64,affine = affine_par) for i in self.bn1.parameters(): i.requires_grad = False self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=True) # change self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=1, dilation__ = 2) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, 0.01) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, stride=1,dilation__ = 1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion or dilation__ == 2 or dilation__ == 4: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion,affine = affine_par), ) for i in downsample._modules['1'].parameters(): i.requires_grad = False layers = [] layers.append(block(self.inplanes, planes, stride,dilation_=dilation__, downsample = downsample )) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes,dilation_=dilation__)) return nn.Sequential(*layers) def forward(self, x): tmp_x = [] x = self.conv1(x) x = self.bn1(x) x = self.relu(x) tmp_x.append(x) x = self.maxpool(x) x = self.layer1(x) tmp_x.append(x) x = self.layer2(x) tmp_x.append(x) x = self.layer3(x) tmp_x.append(x) x = self.layer4(x) tmp_x.append(x) return tmp_x class ResNet_locate(nn.Module): def __init__(self, block, layers): super(ResNet_locate,self).__init__() self.resnet = ResNet(block, layers) self.in_planes = 512 self.out_planes = [512, 256, 256, 128] self.ppms_pre = nn.Conv2d(2048, self.in_planes, 1, 1, bias=False) ppms, infos = [], [] for ii in [1, 3, 5]: ppms.append(nn.Sequential(nn.AdaptiveAvgPool2d(ii), nn.Conv2d(self.in_planes, self.in_planes, 1, 1, bias=False), nn.ReLU(inplace=True))) self.ppms = nn.ModuleList(ppms) self.ppm_cat = nn.Sequential(nn.Conv2d(self.in_planes * 4, self.in_planes, 3, 1, 1, bias=False), nn.ReLU(inplace=True)) for ii in self.out_planes: infos.append(nn.Sequential(nn.Conv2d(self.in_planes, ii, 3, 1, 1, bias=False), nn.ReLU(inplace=True))) self.infos = nn.ModuleList(infos) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, 0.01) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def load_pretrained_model(self, model): self.resnet.load_state_dict(model, strict=False) def forward(self, x): x_size = x.size()[2:] xs = self.resnet(x) xs_1 = self.ppms_pre(xs[-1]) xls = [xs_1] for k in range(len(self.ppms)): xls.append(F.interpolate(self.ppms[k](xs_1), xs_1.size()[2:], mode='bilinear', align_corners=True)) xls = self.ppm_cat(torch.cat(xls, dim=1)) infos = [] for k in range(len(self.infos)): infos.append(self.infos[k](F.interpolate(xls, xs[len(self.infos) - 1 - k].size()[2:], mode='bilinear', align_corners=True))) return xs, infos def resnet50_locate(): model = ResNet_locate(Bottleneck, [3, 4, 6, 3]) return model
7,161
34.107843
148
py
PoolNet
PoolNet-master/dataset/dataset.py
import os from PIL import Image import cv2 import torch from torch.utils import data from torchvision import transforms from torchvision.transforms import functional as F import numbers import numpy as np import random class ImageDataTrain(data.Dataset): def __init__(self, data_root, data_list): self.sal_root = data_root self.sal_source = data_list with open(self.sal_source, 'r') as f: self.sal_list = [x.strip() for x in f.readlines()] self.sal_num = len(self.sal_list) def __getitem__(self, item): # sal data loading im_name = self.sal_list[item % self.sal_num].split()[0] gt_name = self.sal_list[item % self.sal_num].split()[1] sal_image = load_image(os.path.join(self.sal_root, im_name)) sal_label = load_sal_label(os.path.join(self.sal_root, gt_name)) sal_image, sal_label = cv_random_flip(sal_image, sal_label) sal_image = torch.Tensor(sal_image) sal_label = torch.Tensor(sal_label) sample = {'sal_image': sal_image, 'sal_label': sal_label} return sample def __len__(self): return self.sal_num class ImageDataTest(data.Dataset): def __init__(self, data_root, data_list): self.data_root = data_root self.data_list = data_list with open(self.data_list, 'r') as f: self.image_list = [x.strip() for x in f.readlines()] self.image_num = len(self.image_list) def __getitem__(self, item): image, im_size = load_image_test(os.path.join(self.data_root, self.image_list[item])) image = torch.Tensor(image) return {'image': image, 'name': self.image_list[item % self.image_num], 'size': im_size} def __len__(self): return self.image_num def get_loader(config, mode='train', pin=False): shuffle = False if mode == 'train': shuffle = True dataset = ImageDataTrain(config.train_root, config.train_list) data_loader = data.DataLoader(dataset=dataset, batch_size=config.batch_size, shuffle=shuffle, num_workers=config.num_thread, pin_memory=pin) else: dataset = ImageDataTest(config.test_root, config.test_list) data_loader = data.DataLoader(dataset=dataset, batch_size=config.batch_size, shuffle=shuffle, num_workers=config.num_thread, pin_memory=pin) return data_loader def load_image(path): if not os.path.exists(path): print('File {} not exists'.format(path)) im = cv2.imread(path) in_ = np.array(im, dtype=np.float32) in_ -= np.array((104.00699, 116.66877, 122.67892)) in_ = in_.transpose((2,0,1)) return in_ def load_image_test(path): if not os.path.exists(path): print('File {} not exists'.format(path)) im = cv2.imread(path) in_ = np.array(im, dtype=np.float32) im_size = tuple(in_.shape[:2]) in_ -= np.array((104.00699, 116.66877, 122.67892)) in_ = in_.transpose((2,0,1)) return in_, im_size def load_sal_label(path): if not os.path.exists(path): print('File {} not exists'.format(path)) im = Image.open(path) label = np.array(im, dtype=np.float32) if len(label.shape) == 3: label = label[:,:,0] label = label / 255. label = label[np.newaxis, ...] return label def cv_random_flip(img, label): flip_flag = random.randint(0, 1) if flip_flag == 1: img = img[:,:,::-1].copy() label = label[:,:,::-1].copy() return img, label
3,469
32.047619
148
py
PoolNet
PoolNet-master/dataset/joint_dataset.py
import os from PIL import Image import cv2 import torch from torch.utils import data from torchvision import transforms from torchvision.transforms import functional as F import numbers import numpy as np import random class ImageDataTrain(data.Dataset): def __init__(self, sal_data_root, sal_data_list, edge_data_root, edge_data_list): self.sal_root = sal_data_root self.sal_source = sal_data_list self.edge_root = edge_data_root self.edge_source = edge_data_list with open(self.sal_source, 'r') as f: self.sal_list = [x.strip() for x in f.readlines()] with open(self.edge_source, 'r') as f: self.edge_list = [x.strip() for x in f.readlines()] self.sal_num = len(self.sal_list) self.edge_num = len(self.edge_list) def __getitem__(self, item): # edge data loading edge_im_name = self.edge_list[item % self.edge_num].split()[0] edge_gt_name = self.edge_list[item % self.edge_num].split()[1] edge_image = load_image(os.path.join(self.edge_root, edge_im_name)) edge_label = load_edge_label(os.path.join(self.edge_root, edge_gt_name)) edge_image = torch.Tensor(edge_image) edge_label = torch.Tensor(edge_label) # sal data loading sal_im_name = self.sal_list[item % self.sal_num].split()[0] sal_gt_name = self.sal_list[item % self.sal_num].split()[1] sal_image = load_image(os.path.join(self.sal_root, sal_im_name)) sal_label = load_sal_label(os.path.join(self.sal_root, sal_gt_name)) sal_image, sal_label = cv_random_flip(sal_image, sal_label) sal_image = torch.Tensor(sal_image) sal_label = torch.Tensor(sal_label) sample = {'edge_image': edge_image, 'edge_label': edge_label, 'sal_image': sal_image, 'sal_label': sal_label} return sample def __len__(self): return max(self.sal_num, self.edge_num) class ImageDataTest(data.Dataset): def __init__(self, data_root, data_list): self.data_root = data_root self.data_list = data_list with open(self.data_list, 'r') as f: self.image_list = [x.strip() for x in f.readlines()] self.image_num = len(self.image_list) def __getitem__(self, item): image, im_size = load_image_test(os.path.join(self.data_root, self.image_list[item])) image = torch.Tensor(image) return {'image': image, 'name': self.image_list[item % self.image_num], 'size': im_size} def __len__(self): return self.image_num def get_loader(config, mode='train', pin=False): shuffle = False if mode == 'train': shuffle = True dataset = ImageDataTrain(config.train_root, config.train_list, config.train_edge_root, config.train_edge_list) data_loader = data.DataLoader(dataset=dataset, batch_size=config.batch_size, shuffle=shuffle, num_workers=config.num_thread, pin_memory=pin) else: dataset = ImageDataTest(config.test_root, config.test_list) data_loader = data.DataLoader(dataset=dataset, batch_size=config.batch_size, shuffle=shuffle, num_workers=config.num_thread, pin_memory=pin) return data_loader def load_image(path): if not os.path.exists(path): print('File {} not exists'.format(path)) im = cv2.imread(path) in_ = np.array(im, dtype=np.float32) in_ -= np.array((104.00699, 116.66877, 122.67892)) in_ = in_.transpose((2,0,1)) return in_ def load_image_test(path): if not os.path.exists(path): print('File {} not exists'.format(path)) im = cv2.imread(path) in_ = np.array(im, dtype=np.float32) im_size = tuple(in_.shape[:2]) in_ -= np.array((104.00699, 116.66877, 122.67892)) in_ = in_.transpose((2,0,1)) return in_, im_size def load_sal_label(path): if not os.path.exists(path): print('File {} not exists'.format(path)) im = Image.open(path) label = np.array(im, dtype=np.float32) if len(label.shape) == 3: label = label[:,:,0] label = label / 255. label = label[np.newaxis, ...] return label def load_edge_label(path): """ pixels > 0.5 -> 1. """ if not os.path.exists(path): print('File {} not exists'.format(path)) im = Image.open(path) label = np.array(im, dtype=np.float32) if len(label.shape) == 3: label = label[:,:,0] label = label / 255. label[np.where(label > 0.5)] = 1. label = label[np.newaxis, ...] return label def cv_random_flip(img, label): flip_flag = random.randint(0, 1) if flip_flag == 1: img = img[:,:,::-1].copy() label = label[:,:,::-1].copy() return img, label
4,702
34.360902
148
py
GNNDelete
GNNDelete-main/train_node.py
import os import wandb import pickle import torch from torch_geometric.seed import seed_everything from torch_geometric.utils import to_undirected, is_undirected import torch_geometric.transforms as T from torch_geometric.datasets import CitationFull, Coauthor, Flickr, RelLinkPredDataset, WordNet18, WordNet18RR from torch_geometric.seed import seed_everything from framework import get_model, get_trainer from framework.training_args import parse_args from framework.trainer.base import NodeClassificationTrainer from framework.utils import negative_sampling_kg device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') def main(): args = parse_args() args.checkpoint_dir = 'checkpoint_node' args.dataset = 'DBLP' args.unlearning_model = 'original' args.checkpoint_dir = os.path.join(args.checkpoint_dir, args.dataset, args.gnn, args.unlearning_model, str(args.random_seed)) os.makedirs(args.checkpoint_dir, exist_ok=True) seed_everything(args.random_seed) # Dataset dataset = CitationFull(os.path.join(args.data_dir, args.dataset), args.dataset, transform=T.NormalizeFeatures()) data = dataset[0] print('Original data', data) split = T.RandomNodeSplit() data = split(data) assert is_undirected(data.edge_index) print('Split data', data) args.in_dim = data.x.shape[1] args.out_dim = dataset.num_classes wandb.init(config=args) # Model model = get_model(args, num_nodes=data.num_nodes, num_edge_type=args.num_edge_type).to(device) wandb.watch(model, log_freq=100) optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)#, weight_decay=args.weight_decay) # Train trainer = NodeClassificationTrainer(args) trainer.train(model, data, optimizer, args) # Test trainer.test(model, data) trainer.save_log() if __name__ == "__main__": main()
1,881
30.898305
129
py
GNNDelete
GNNDelete-main/graph_stat.py
import os from torch_geometric.data import Data import torch_geometric.transforms as T from torch_geometric.datasets import CitationFull, Coauthor, Flickr, RelLinkPredDataset, WordNet18RR from ogb.linkproppred import PygLinkPropPredDataset data_dir = './data' datasets = ['Cora', 'PubMed', 'DBLP', 'CS', 'Physics', 'ogbl-citation2', 'ogbl-collab', 'FB15k-237', 'WordNet18RR', 'ogbl-biokg', 'ogbl-wikikg2'][-2:] def get_stat(d): if d in ['Cora', 'PubMed', 'DBLP']: dataset = CitationFull(os.path.join(data_dir, d), d, transform=T.NormalizeFeatures()) if d in ['CS', 'Physics']: dataset = Coauthor(os.path.join(data_dir, d), d, transform=T.NormalizeFeatures()) if d in ['Flickr']: dataset = Flickr(os.path.join(data_dir, d), transform=T.NormalizeFeatures()) if 'ogbl' in d: dataset = PygLinkPropPredDataset(root=os.path.join(data_dir, d), name=d) data = dataset[0] print(d) print('Number of nodes:', data.num_nodes) print('Number of edges:', data.num_edges) print('Number of max deleted edges:', int(0.05 * data.num_edges)) if hasattr(data, 'edge_type'): print('Number of nodes:', data.edge_type.unique().shape) def main(): for d in datasets: get_stat(d) if __name__ == "__main__": main()
1,292
35.942857
150
py
GNNDelete
GNNDelete-main/delete_node_feature.py
import os import copy import json import wandb import pickle import argparse import torch import torch.nn as nn from torch_geometric.utils import to_undirected, to_networkx, k_hop_subgraph, is_undirected from torch_geometric.data import Data import torch_geometric.transforms as T from torch_geometric.datasets import CitationFull, Coauthor, Flickr, RelLinkPredDataset, WordNet18, WordNet18RR from torch_geometric.loader import GraphSAINTRandomWalkSampler from torch_geometric.seed import seed_everything from framework import get_model, get_trainer from framework.models.gcn import GCN from framework.models.deletion import GCNDelete from framework.training_args import parse_args from framework.utils import * from framework.trainer.gnndelete_nodeemb import GNNDeleteNodeClassificationTrainer from train_mi import MLPAttacker device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') torch.autograd.set_detect_anomaly(True) def to_directed(edge_index): row, col = edge_index mask = row < col return torch.cat([row[mask], col[mask]], dim=0) def main(): args = parse_args() args.checkpoint_dir = 'checkpoint_node_feature' args.dataset = 'DBLP' original_path = os.path.join(args.checkpoint_dir, args.dataset, args.gnn, 'original', str(args.random_seed)) attack_path_all = os.path.join(args.checkpoint_dir, args.dataset, 'member_infer_all', str(args.random_seed)) attack_path_sub = os.path.join(args.checkpoint_dir, args.dataset, 'member_infer_sub', str(args.random_seed)) seed_everything(args.random_seed) if 'gnndelete' in args.unlearning_model: args.checkpoint_dir = os.path.join( args.checkpoint_dir, args.dataset, args.gnn, f'{args.unlearning_model}-node_deletion', '-'.join([str(i) for i in [args.loss_fct, args.loss_type, args.alpha, args.neg_sample_random]]), '-'.join([str(i) for i in [args.df, args.df_size, args.random_seed]])) else: args.checkpoint_dir = os.path.join( args.checkpoint_dir, args.dataset, args.gnn, f'{args.unlearning_model}-node_deletion', '-'.join([str(i) for i in [args.df, args.df_size, args.random_seed]])) os.makedirs(args.checkpoint_dir, exist_ok=True) # Dataset dataset = CitationFull(os.path.join(args.data_dir, args.dataset), args.dataset, transform=T.NormalizeFeatures()) data = dataset[0] print('Original data', data) split = T.RandomNodeSplit() data = split(data) assert is_undirected(data.edge_index) print('Split data', data) args.in_dim = data.x.shape[1] args.out_dim = dataset.num_classes wandb.init(config=args) # Df and Dr if args.df_size >= 100: # df_size is number of nodes/edges to be deleted df_size = int(args.df_size) else: # df_size is the ratio df_size = int(args.df_size / 100 * data.train_pos_edge_index.shape[1]) print(f'Original size: {data.num_nodes:,}') print(f'Df size: {df_size:,}') # Delete node feature df_nodes = torch.randperm(data.num_nodes)[:df_size] global_node_mask = torch.ones(data.num_nodes, dtype=torch.bool) # global_node_mask[df_nodes] = False data.x[df_nodes] = 0 assert data.x[df_nodes].sum() == 0 dr_mask_node = torch.ones(data.num_nodes, dtype=torch.bool) df_mask_node = ~global_node_mask # assert df_mask_node.sum() == df_size # Delete edges associated with deleted nodes from training set res = [torch.eq(data.edge_index, aelem).logical_or_(torch.eq(data.edge_index, aelem)) for aelem in df_nodes] df_mask_edge = torch.any(torch.stack(res, dim=0), dim = 0) df_mask_edge = df_mask_edge.sum(0).bool() dr_mask_edge = ~df_mask_edge df_edge = data.edge_index[:, df_mask_edge] data.directed_df_edge_index = to_directed(df_edge) # print(df_edge.shape, directed_df_edge_index.shape) # raise print('Deleting the following nodes:', df_nodes) # # Delete edges associated with deleted nodes from valid and test set # res = [torch.eq(data.val_pos_edge_index, aelem).logical_or_(torch.eq(data.val_pos_edge_index, aelem)) for aelem in df_nodes] # mask = torch.any(torch.stack(res, dim=0), dim = 0) # mask = mask.sum(0).bool() # mask = ~mask # data.val_pos_edge_index = data.val_pos_edge_index[:, mask] # data.val_neg_edge_index = data.val_neg_edge_index[:, :data.val_pos_edge_index.shape[1]] # res = [torch.eq(data.test_pos_edge_index, aelem).logical_or_(torch.eq(data.test_pos_edge_index, aelem)) for aelem in df_nodes] # mask = torch.any(torch.stack(res, dim=0), dim = 0) # mask = mask.sum(0).bool() # mask = ~mask # data.test_pos_edge_index = data.test_pos_edge_index[:, mask] # data.test_neg_edge_index = data.test_neg_edge_index[:, :data.test_pos_edge_index.shape[1]] # For testing # data.directed_df_edge_index = data.train_pos_edge_index[:, df_mask_edge] # if args.gnn in ['rgcn', 'rgat']: # data.directed_df_edge_type = data.train_edge_type[df_mask] # Edges in S_Df _, two_hop_edge, _, two_hop_mask = k_hop_subgraph( data.edge_index[:, df_mask_edge].flatten().unique(), 2, data.edge_index, num_nodes=data.num_nodes) # Nodes in S_Df _, one_hop_edge, _, one_hop_mask = k_hop_subgraph( data.edge_index[:, df_mask_edge].flatten().unique(), 1, data.edge_index, num_nodes=data.num_nodes) sdf_node_1hop = torch.zeros(data.num_nodes, dtype=torch.bool) sdf_node_2hop = torch.zeros(data.num_nodes, dtype=torch.bool) sdf_node_1hop[one_hop_edge.flatten().unique()] = True sdf_node_2hop[two_hop_edge.flatten().unique()] = True assert sdf_node_1hop.sum() == len(one_hop_edge.flatten().unique()) assert sdf_node_2hop.sum() == len(two_hop_edge.flatten().unique()) data.sdf_node_1hop_mask = sdf_node_1hop data.sdf_node_2hop_mask = sdf_node_2hop # To undirected for message passing # print(is_undir0.0175ected(data.train_pos_edge_index), data.train_pos_edge_index.shape, two_hop_mask.shape, df_mask.shape, two_hop_mask.shape) # assert not is_undirected(data.edge_index) print(is_undirected(data.edge_index)) if args.gnn in ['rgcn', 'rgat']: r, c = data.train_pos_edge_index rev_edge_index = torch.stack([c, r], dim=0) rev_edge_type = data.train_edge_type + args.num_edge_type data.edge_index = torch.cat((data.train_pos_edge_index, rev_edge_index), dim=1) data.edge_type = torch.cat([data.train_edge_type, rev_edge_type], dim=0) # data.train_mask = data.train_mask.repeat(2) two_hop_mask = two_hop_mask.repeat(2).view(-1) df_mask = df_mask.repeat(2).view(-1) dr_mask = dr_mask.repeat(2).view(-1) assert is_undirected(data.edge_index) else: # train_pos_edge_index, [df_mask, two_hop_mask] = to_undirected(data.train_pos_edge_index, [df_mask.int(), two_hop_mask.int()]) two_hop_mask = two_hop_mask.bool() df_mask_edge = df_mask_edge.bool() dr_mask_edge = ~df_mask_edge # data.train_pos_edge_index = train_pos_edge_index # assert is_undirected(data.train_pos_edge_index) print('Undirected dataset:', data) # print(is_undirected(train_pos_edge_index), train_pos_edge_index.shape, two_hop_mask.shape, df_mask.shape, two_hop_mask.shape) data.sdf_mask = two_hop_mask data.df_mask = df_mask_edge data.dr_mask = dr_mask_edge data.dtrain_mask = dr_mask_edge # print(is_undirected(data.train_pos_edge_index), data.train_pos_edge_index.shape, data.two_hop_mask.shape, data.df_mask.shape, data.two_hop_mask.shape) # raise # Model model = GCNDelete(args) # model = get_model(args, sdf_node_1hop, sdf_node_2hop, num_nodes=data.num_nodes, num_edge_type=args.num_edge_type) if args.unlearning_model != 'retrain': # Start from trained GNN model if os.path.exists(os.path.join(original_path, 'pred_proba.pt')): logits_ori = torch.load(os.path.join(original_path, 'pred_proba.pt')) if logits_ori is not None: logits_ori = logits_ori.to(device) else: logits_ori = None model_ckpt = torch.load(os.path.join(original_path, 'model_best.pt'), map_location=device) model.load_state_dict(model_ckpt['model_state'], strict=False) else: # Initialize a new GNN model retrain = None logits_ori = None model = model.to(device) if 'gnndelete' in args.unlearning_model and 'nodeemb' in args.unlearning_model: parameters_to_optimize = [ {'params': [p for n, p in model.named_parameters() if 'del' in n], 'weight_decay': 0.0} ] print('parameters_to_optimize', [n for n, p in model.named_parameters() if 'del' in n]) if 'layerwise' in args.loss_type: optimizer1 = torch.optim.Adam(model.deletion1.parameters(), lr=args.lr) optimizer2 = torch.optim.Adam(model.deletion2.parameters(), lr=args.lr) optimizer = [optimizer1, optimizer2] else: optimizer = torch.optim.Adam(parameters_to_optimize, lr=args.lr) else: if 'gnndelete' in args.unlearning_model: parameters_to_optimize = [ {'params': [p for n, p in model.named_parameters() if 'del' in n], 'weight_decay': 0.0} ] print('parameters_to_optimize', [n for n, p in model.named_parameters() if 'del' in n]) else: parameters_to_optimize = [ {'params': [p for n, p in model.named_parameters()], 'weight_decay': 0.0} ] print('parameters_to_optimize', [n for n, p in model.named_parameters()]) optimizer = torch.optim.Adam(parameters_to_optimize, lr=args.lr)#, weight_decay=args.weight_decay) wandb.watch(model, log_freq=100) # MI attack model attack_model_all = None # attack_model_all = MLPAttacker(args) # attack_ckpt = torch.load(os.path.join(attack_path_all, 'attack_model_best.pt')) # attack_model_all.load_state_dict(attack_ckpt['model_state']) # attack_model_all = attack_model_all.to(device) attack_model_sub = None # attack_model_sub = MLPAttacker(args) # attack_ckpt = torch.load(os.path.join(attack_path_sub, 'attack_model_best.pt')) # attack_model_sub.load_state_dict(attack_ckpt['model_state']) # attack_model_sub = attack_model_sub.to(device) # Train trainer = GNNDeleteNodeClassificationTrainer(args) trainer.train(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) # Test if args.unlearning_model != 'retrain': retrain_path = os.path.join( 'checkpoint', args.dataset, args.gnn, 'retrain', '-'.join([str(i) for i in [args.df, args.df_size, args.random_seed]])) retrain_ckpt = torch.load(os.path.join(retrain_path, 'model_best.pt'), map_location=device) retrain_args = copy.deepcopy(args) retrain_args.unlearning_model = 'retrain' retrain = get_model(retrain_args, num_nodes=data.num_nodes, num_edge_type=args.num_edge_type) retrain.load_state_dict(retrain_ckpt['model_state']) retrain = retrain.to(device) retrain.eval() else: retrain = None trainer.test(model, data, model_retrain=retrain, attack_model_all=attack_model_all, attack_model_sub=attack_model_sub) trainer.save_log() if __name__ == "__main__": main()
11,564
40.902174
156
py
GNNDelete
GNNDelete-main/delete_gnn.py
import os import copy import json import wandb import pickle import argparse import torch import torch.nn as nn from torch_geometric.utils import to_undirected, to_networkx, k_hop_subgraph, is_undirected from torch_geometric.data import Data from torch_geometric.loader import GraphSAINTRandomWalkSampler from torch_geometric.seed import seed_everything from framework import get_model, get_trainer from framework.models.gcn import GCN from framework.training_args import parse_args from framework.utils import * from train_mi import MLPAttacker device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') def load_args(path): with open(path, 'r') as f: d = json.load(f) parser = argparse.ArgumentParser() for k, v in d.items(): parser.add_argument('--' + k, default=v) try: parser.add_argument('--df_size', default=0.5) except: pass args = parser.parse_args() for k, v in d.items(): setattr(args, k, v) return args @torch.no_grad() def get_node_embedding(model, data): model.eval() node_embedding = model(data.x.to(device), data.edge_index.to(device)) return node_embedding @torch.no_grad() def get_output(model, node_embedding, data): model.eval() node_embedding = node_embedding.to(device) edge = data.edge_index.to(device) output = model.decode(node_embedding, edge, edge_type) return output torch.autograd.set_detect_anomaly(True) def main(): args = parse_args() original_path = os.path.join(args.checkpoint_dir, args.dataset, args.gnn, 'original', str(args.random_seed)) attack_path_all = os.path.join(args.checkpoint_dir, args.dataset, 'member_infer_all', str(args.random_seed)) attack_path_sub = os.path.join(args.checkpoint_dir, args.dataset, 'member_infer_sub', str(args.random_seed)) seed_everything(args.random_seed) if 'gnndelete' in args.unlearning_model: args.checkpoint_dir = os.path.join( args.checkpoint_dir, args.dataset, args.gnn, args.unlearning_model, '-'.join([str(i) for i in [args.loss_fct, args.loss_type, args.alpha, args.neg_sample_random]]), '-'.join([str(i) for i in [args.df, args.df_size, args.random_seed]])) else: args.checkpoint_dir = os.path.join( args.checkpoint_dir, args.dataset, args.gnn, args.unlearning_model, '-'.join([str(i) for i in [args.df, args.df_size, args.random_seed]])) os.makedirs(args.checkpoint_dir, exist_ok=True) # Dataset with open(os.path.join(args.data_dir, args.dataset, f'd_{args.random_seed}.pkl'), 'rb') as f: dataset, data = pickle.load(f) print('Directed dataset:', dataset, data) if args.gnn not in ['rgcn', 'rgat']: args.in_dim = dataset.num_features print('Training args', args) wandb.init(config=args) # Df and Dr assert args.df != 'none' if args.df_size >= 100: # df_size is number of nodes/edges to be deleted df_size = int(args.df_size) else: # df_size is the ratio df_size = int(args.df_size / 100 * data.train_pos_edge_index.shape[1]) print(f'Original size: {data.train_pos_edge_index.shape[1]:,}') print(f'Df size: {df_size:,}') df_mask_all = torch.load(os.path.join(args.data_dir, args.dataset, f'df_{args.random_seed}.pt'))[args.df] df_nonzero = df_mask_all.nonzero().squeeze() idx = torch.randperm(df_nonzero.shape[0])[:df_size] df_global_idx = df_nonzero[idx] print('Deleting the following edges:', df_global_idx) # df_idx = [int(i) for i in args.df_idx.split(',')] # df_idx_global = df_mask.nonzero()[df_idx] dr_mask = torch.ones(data.train_pos_edge_index.shape[1], dtype=torch.bool) dr_mask[df_global_idx] = False df_mask = torch.zeros(data.train_pos_edge_index.shape[1], dtype=torch.bool) df_mask[df_global_idx] = True # For testing data.directed_df_edge_index = data.train_pos_edge_index[:, df_mask] if args.gnn in ['rgcn', 'rgat']: data.directed_df_edge_type = data.train_edge_type[df_mask] # data.dr_mask = dr_mask # data.df_mask = df_mask # data.edge_index = data.train_pos_edge_index[:, dr_mask] # assert df_mask.sum() == len(df_global_idx) # assert dr_mask.shape[0] - len(df_global_idx) == data.train_pos_edge_index[:, dr_mask].shape[1] # data.dtrain_mask = dr_mask # Edges in S_Df _, two_hop_edge, _, two_hop_mask = k_hop_subgraph( data.train_pos_edge_index[:, df_mask].flatten().unique(), 2, data.train_pos_edge_index, num_nodes=data.num_nodes) data.sdf_mask = two_hop_mask # Nodes in S_Df _, one_hop_edge, _, one_hop_mask = k_hop_subgraph( data.train_pos_edge_index[:, df_mask].flatten().unique(), 1, data.train_pos_edge_index, num_nodes=data.num_nodes) sdf_node_1hop = torch.zeros(data.num_nodes, dtype=torch.bool) sdf_node_2hop = torch.zeros(data.num_nodes, dtype=torch.bool) sdf_node_1hop[one_hop_edge.flatten().unique()] = True sdf_node_2hop[two_hop_edge.flatten().unique()] = True assert sdf_node_1hop.sum() == len(one_hop_edge.flatten().unique()) assert sdf_node_2hop.sum() == len(two_hop_edge.flatten().unique()) data.sdf_node_1hop_mask = sdf_node_1hop data.sdf_node_2hop_mask = sdf_node_2hop # To undirected for message passing # print(is_undir0.0175ected(data.train_pos_edge_index), data.train_pos_edge_index.shape, two_hop_mask.shape, df_mask.shape, two_hop_mask.shape) assert not is_undirected(data.train_pos_edge_index) if args.gnn in ['rgcn', 'rgat']: r, c = data.train_pos_edge_index rev_edge_index = torch.stack([c, r], dim=0) rev_edge_type = data.train_edge_type + args.num_edge_type data.edge_index = torch.cat((data.train_pos_edge_index, rev_edge_index), dim=1) data.edge_type = torch.cat([data.train_edge_type, rev_edge_type], dim=0) if hasattr(data, 'train_mask'): data.train_mask = data.train_mask.repeat(2).view(-1) two_hop_mask = two_hop_mask.repeat(2).view(-1) df_mask = df_mask.repeat(2).view(-1) dr_mask = dr_mask.repeat(2).view(-1) assert is_undirected(data.edge_index) else: train_pos_edge_index, [df_mask, two_hop_mask] = to_undirected(data.train_pos_edge_index, [df_mask.int(), two_hop_mask.int()]) two_hop_mask = two_hop_mask.bool() df_mask = df_mask.bool() dr_mask = ~df_mask data.train_pos_edge_index = train_pos_edge_index data.edge_index = train_pos_edge_index assert is_undirected(data.train_pos_edge_index) print('Undirected dataset:', data) data.sdf_mask = two_hop_mask data.df_mask = df_mask data.dr_mask = dr_mask # data.dtrain_mask = dr_mask # print(is_undirected(train_pos_edge_index), train_pos_edge_index.shape, two_hop_mask.shape, df_mask.shape, two_hop_mask.shape) # print(is_undirected(data.train_pos_edge_index), data.train_pos_edge_index.shape, data.df_mask.shape, ) # raise # Model model = get_model(args, sdf_node_1hop, sdf_node_2hop, num_nodes=data.num_nodes, num_edge_type=args.num_edge_type) if args.unlearning_model != 'retrain': # Start from trained GNN model if os.path.exists(os.path.join(original_path, 'pred_proba.pt')): logits_ori = torch.load(os.path.join(original_path, 'pred_proba.pt')) if logits_ori is not None: logits_ori = logits_ori.to(device) else: logits_ori = None model_ckpt = torch.load(os.path.join(original_path, 'model_best.pt'), map_location=device) model.load_state_dict(model_ckpt['model_state'], strict=False) else: # Initialize a new GNN model retrain = None logits_ori = None model = model.to(device) if 'gnndelete' in args.unlearning_model and 'nodeemb' in args.unlearning_model: parameters_to_optimize = [ {'params': [p for n, p in model.named_parameters() if 'del' in n], 'weight_decay': 0.0} ] print('parameters_to_optimize', [n for n, p in model.named_parameters() if 'del' in n]) if 'layerwise' in args.loss_type: optimizer1 = torch.optim.Adam(model.deletion1.parameters(), lr=args.lr) optimizer2 = torch.optim.Adam(model.deletion2.parameters(), lr=args.lr) optimizer = [optimizer1, optimizer2] else: optimizer = torch.optim.Adam(parameters_to_optimize, lr=args.lr) else: if 'gnndelete' in args.unlearning_model: parameters_to_optimize = [ {'params': [p for n, p in model.named_parameters() if 'del' in n], 'weight_decay': 0.0} ] print('parameters_to_optimize', [n for n, p in model.named_parameters() if 'del' in n]) else: parameters_to_optimize = [ {'params': [p for n, p in model.named_parameters()], 'weight_decay': 0.0} ] print('parameters_to_optimize', [n for n, p in model.named_parameters()]) optimizer = torch.optim.Adam(parameters_to_optimize, lr=args.lr)#, weight_decay=args.weight_decay) wandb.watch(model, log_freq=100) # MI attack model attack_model_all = None # attack_model_all = MLPAttacker(args) # attack_ckpt = torch.load(os.path.join(attack_path_all, 'attack_model_best.pt')) # attack_model_all.load_state_dict(attack_ckpt['model_state']) # attack_model_all = attack_model_all.to(device) attack_model_sub = None # attack_model_sub = MLPAttacker(args) # attack_ckpt = torch.load(os.path.join(attack_path_sub, 'attack_model_best.pt')) # attack_model_sub.load_state_dict(attack_ckpt['model_state']) # attack_model_sub = attack_model_sub.to(device) # Train trainer = get_trainer(args) trainer.train(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) # Test if args.unlearning_model != 'retrain': retrain_path = os.path.join( 'checkpoint', args.dataset, args.gnn, 'retrain', '-'.join([str(i) for i in [args.df, args.df_size, args.random_seed]]), 'model_best.pt') if os.path.exists(retrain_path): retrain_ckpt = torch.load(retrain_path, map_location=device) retrain_args = copy.deepcopy(args) retrain_args.unlearning_model = 'retrain' retrain = get_model(retrain_args, num_nodes=data.num_nodes, num_edge_type=args.num_edge_type) retrain.load_state_dict(retrain_ckpt['model_state']) retrain = retrain.to(device) retrain.eval() else: retrain = None else: retrain = None test_results = trainer.test(model, data, model_retrain=retrain, attack_model_all=attack_model_all, attack_model_sub=attack_model_sub) print(test_results[-1]) trainer.save_log() if __name__ == "__main__": main()
11,069
37.4375
147
py
GNNDelete
GNNDelete-main/train_gnn.py
import os import wandb import pickle import torch from torch_geometric.seed import seed_everything from torch_geometric.utils import to_undirected, is_undirected from torch_geometric.datasets import RelLinkPredDataset, WordNet18 from torch_geometric.seed import seed_everything from framework import get_model, get_trainer from framework.training_args import parse_args from framework.trainer.base import Trainer from framework.utils import negative_sampling_kg device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') def main(): args = parse_args() args.unlearning_model = 'original' args.checkpoint_dir = os.path.join(args.checkpoint_dir, args.dataset, args.gnn, args.unlearning_model, str(args.random_seed)) os.makedirs(args.checkpoint_dir, exist_ok=True) seed_everything(args.random_seed) # Dataset with open(os.path.join(args.data_dir, args.dataset, f'd_{args.random_seed}.pkl'), 'rb') as f: dataset, data = pickle.load(f) print('Directed dataset:', dataset, data) if args.gnn not in ['rgcn', 'rgat']: args.in_dim = dataset.num_features wandb.init(config=args) # Use proper training data for original and Dr if args.gnn in ['rgcn', 'rgat']: if not hasattr(data, 'train_mask'): data.train_mask = torch.ones(data.edge_index.shape[1], dtype=torch.bool) # data.dtrain_mask = torch.ones(data.train_pos_edge_index.shape[1], dtype=torch.bool) # data.edge_index_mask = data.dtrain_mask.repeat(2) else: data.dtrain_mask = torch.ones(data.train_pos_edge_index.shape[1], dtype=torch.bool) # To undirected if args.gnn in ['rgcn', 'rgat']: r, c = data.train_pos_edge_index rev_edge_index = torch.stack([c, r], dim=0) rev_edge_type = data.train_edge_type + args.num_edge_type data.edge_index = torch.cat((data.train_pos_edge_index, rev_edge_index), dim=1) data.edge_type = torch.cat([data.train_edge_type, rev_edge_type], dim=0) # data.train_mask = data.train_mask.repeat(2) data.dr_mask = torch.ones(data.edge_index.shape[1], dtype=torch.bool) assert is_undirected(data.edge_index) else: train_pos_edge_index = to_undirected(data.train_pos_edge_index) data.train_pos_edge_index = train_pos_edge_index data.dtrain_mask = torch.ones(data.train_pos_edge_index.shape[1], dtype=torch.bool) assert is_undirected(data.train_pos_edge_index) print('Undirected dataset:', data) # Model model = get_model(args, num_nodes=data.num_nodes, num_edge_type=args.num_edge_type).to(device) wandb.watch(model, log_freq=100) optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)#, weight_decay=args.weight_decay) # Train trainer = get_trainer(args) trainer.train(model, data, optimizer, args) # Test trainer.test(model, data) trainer.save_log() if __name__ == "__main__": main()
2,977
34.035294
129
py
GNNDelete
GNNDelete-main/prepare_dataset.py
import os import math import pickle import torch import pandas as pd import networkx as nx from tqdm import tqdm from torch_geometric.seed import seed_everything import torch_geometric.transforms as T from torch_geometric.data import Data from torch_geometric.datasets import CitationFull, Coauthor, Flickr, RelLinkPredDataset, WordNet18, WordNet18RR from torch_geometric.utils import train_test_split_edges, k_hop_subgraph, negative_sampling, to_undirected, is_undirected, to_networkx from ogb.linkproppred import PygLinkPropPredDataset from framework.utils import * data_dir = './data' df_size = [i / 100 for i in range(10)] + [i / 10 for i in range(10)] + [i for i in range(10)] # Df_size in percentage seeds = [42, 21, 13, 87, 100] graph_datasets = ['Cora', 'PubMed', 'DBLP', 'CS', 'ogbl-citation2', 'ogbl-collab'][4:] kg_datasets = ['FB15k-237', 'WordNet18', 'WordNet18RR', 'ogbl-biokg'][-1:] os.makedirs(data_dir, exist_ok=True) num_edge_type_mapping = { 'FB15k-237': 237, 'WordNet18': 18, 'WordNet18RR': 11 } def train_test_split_edges_no_neg_adj_mask(data, val_ratio: float = 0.05, test_ratio: float = 0.1, two_hop_degree=None, kg=False): '''Avoid adding neg_adj_mask''' num_nodes = data.num_nodes row, col = data.edge_index edge_attr = data.edge_attr if kg: edge_type = data.edge_type data.edge_index = data.edge_attr = data.edge_weight = data.edge_year = data.edge_type = None if not kg: # Return upper triangular portion. mask = row < col row, col = row[mask], col[mask] if edge_attr is not None: edge_attr = edge_attr[mask] n_v = int(math.floor(val_ratio * row.size(0))) n_t = int(math.floor(test_ratio * row.size(0))) if two_hop_degree is not None: # Use low degree edges for test sets low_degree_mask = two_hop_degree < 50 low = low_degree_mask.nonzero().squeeze() high = (~low_degree_mask).nonzero().squeeze() low = low[torch.randperm(low.size(0))] high = high[torch.randperm(high.size(0))] perm = torch.cat([low, high]) else: perm = torch.randperm(row.size(0)) row = row[perm] col = col[perm] # Train r, c = row[n_v + n_t:], col[n_v + n_t:] if kg: # data.edge_index and data.edge_type has reverse edges and edge types for message passing pos_edge_index = torch.stack([r, c], dim=0) # rev_pos_edge_index = torch.stack([r, c], dim=0) train_edge_type = edge_type[n_v + n_t:] # train_rev_edge_type = edge_type[n_v + n_t:] + edge_type.unique().shape[0] # data.edge_index = torch.cat((torch.stack([r, c], dim=0), torch.stack([r, c], dim=0)), dim=1) # data.edge_type = torch.cat([train_edge_type, train_rev_edge_type], dim=0) data.edge_index = pos_edge_index data.edge_type = train_edge_type # data.train_pos_edge_index and data.train_edge_type only has one direction edges and edge types for decoding data.train_pos_edge_index = torch.stack([r, c], dim=0) data.train_edge_type = train_edge_type else: data.train_pos_edge_index = torch.stack([r, c], dim=0) if edge_attr is not None: # out = to_undirected(data.train_pos_edge_index, edge_attr[n_v + n_t:]) data.train_pos_edge_index, data.train_pos_edge_attr = out else: data.train_pos_edge_index = data.train_pos_edge_index # data.train_pos_edge_index = to_undirected(data.train_pos_edge_index) assert not is_undirected(data.train_pos_edge_index) # Test r, c = row[:n_t], col[:n_t] data.test_pos_edge_index = torch.stack([r, c], dim=0) if kg: data.test_edge_type = edge_type[:n_t] neg_edge_index = negative_sampling_kg( edge_index=data.test_pos_edge_index, edge_type=data.test_edge_type) else: neg_edge_index = negative_sampling( edge_index=data.test_pos_edge_index, num_nodes=data.num_nodes, num_neg_samples=data.test_pos_edge_index.shape[1]) data.test_neg_edge_index = neg_edge_index # Valid r, c = row[n_t:n_t+n_v], col[n_t:n_t+n_v] data.val_pos_edge_index = torch.stack([r, c], dim=0) if kg: data.val_edge_type = edge_type[n_t:n_t+n_v] neg_edge_index = negative_sampling_kg( edge_index=data.val_pos_edge_index, edge_type=data.val_edge_type) else: neg_edge_index = negative_sampling( edge_index=data.val_pos_edge_index, num_nodes=data.num_nodes, num_neg_samples=data.val_pos_edge_index.shape[1]) data.val_neg_edge_index = neg_edge_index return data def process_graph(): for d in graph_datasets: if d in ['Cora', 'PubMed', 'DBLP']: dataset = CitationFull(os.path.join(data_dir, d), d, transform=T.NormalizeFeatures()) elif d in ['CS', 'Physics']: dataset = Coauthor(os.path.join(data_dir, d), d, transform=T.NormalizeFeatures()) elif d in ['Flickr']: dataset = Flickr(os.path.join(data_dir, d), transform=T.NormalizeFeatures()) elif 'ogbl' in d: dataset = PygLinkPropPredDataset(root=os.path.join(data_dir, d), name=d) else: raise NotImplementedError print('Processing:', d) print(dataset) data = dataset[0] data.train_mask = data.val_mask = data.test_mask = None graph = to_networkx(data) # Get two hop degree for all nodes node_to_neighbors = {} for n in tqdm(graph.nodes(), desc='Two hop neighbors'): neighbor_1 = set(graph.neighbors(n)) neighbor_2 = sum([list(graph.neighbors(i)) for i in neighbor_1], []) neighbor_2 = set(neighbor_2) neighbor = neighbor_1 | neighbor_2 node_to_neighbors[n] = neighbor two_hop_degree = [] row, col = data.edge_index mask = row < col row, col = row[mask], col[mask] for r, c in tqdm(zip(row, col), total=len(row)): neighbor_row = node_to_neighbors[r.item()] neighbor_col = node_to_neighbors[c.item()] neighbor = neighbor_row | neighbor_col num = len(neighbor) two_hop_degree.append(num) two_hop_degree = torch.tensor(two_hop_degree) for s in seeds: seed_everything(s) # D data = dataset[0] if 'ogbl' in d: data = train_test_split_edges_no_neg_adj_mask(data, test_ratio=0.05, two_hop_degree=two_hop_degree) else: data = train_test_split_edges_no_neg_adj_mask(data, test_ratio=0.05) print(s, data) with open(os.path.join(data_dir, d, f'd_{s}.pkl'), 'wb') as f: pickle.dump((dataset, data), f) # Two ways to sample Df from the training set ## 1. Df is within 2 hop local enclosing subgraph of Dtest ## 2. Df is outside of 2 hop local enclosing subgraph of Dtest # All the candidate edges (train edges) # graph = to_networkx(Data(edge_index=data.train_pos_edge_index, x=data.x)) # Get the 2 hop local enclosing subgraph for all test edges _, local_edges, _, mask = k_hop_subgraph( data.test_pos_edge_index.flatten().unique(), 2, data.train_pos_edge_index, num_nodes=dataset[0].num_nodes) distant_edges = data.train_pos_edge_index[:, ~mask] print('Number of edges. Local: ', local_edges.shape[1], 'Distant:', distant_edges.shape[1]) in_mask = mask out_mask = ~mask # df_in_mask = torch.zeros_like(mask) # df_out_mask = torch.zeros_like(mask) # df_in_all_idx = in_mask.nonzero().squeeze() # df_out_all_idx = out_mask.nonzero().squeeze() # df_in_selected_idx = df_in_all_idx[torch.randperm(df_in_all_idx.shape[0])[:df_size]] # df_out_selected_idx = df_out_all_idx[torch.randperm(df_out_all_idx.shape[0])[:df_size]] # df_in_mask[df_in_selected_idx] = True # df_out_mask[df_out_selected_idx] = True # assert (in_mask & out_mask).sum() == 0 # assert (df_in_mask & df_out_mask).sum() == 0 # local_edges = set() # for i in range(data.test_pos_edge_index.shape[1]): # edge = data.test_pos_edge_index[:, i].tolist() # subgraph = get_enclosing_subgraph(graph, edge) # local_edges = local_edges | set(subgraph[2]) # distant_edges = graph.edges() - local_edges # print('aaaaaaa', len(local_edges), len(distant_edges)) # local_edges = torch.tensor(sorted(list([i for i in local_edges if i[0] < i[1]]))) # distant_edges = torch.tensor(sorted(list([i for i in distant_edges if i[0] < i[1]]))) # df_in = torch.randperm(local_edges.shape[1])[:df_size] # df_out = torch.randperm(distant_edges.shape[1])[:df_size] # df_in = local_edges[:, df_in] # df_out = distant_edges[:, df_out] # df_in_mask = torch.zeros(data.train_pos_edge_index.shape[1], dtype=torch.bool) # df_out_mask = torch.zeros(data.train_pos_edge_index.shape[1], dtype=torch.bool) # for row in df_in: # i = (data.train_pos_edge_index.T == row).all(axis=1).nonzero() # df_in_mask[i] = True # for row in df_out: # i = (data.train_pos_edge_index.T == row).all(axis=1).nonzero() # df_out_mask[i] = True torch.save( {'out': out_mask, 'in': in_mask}, os.path.join(data_dir, d, f'df_{s}.pt') ) def process_kg(): for d in kg_datasets: # Create the dataset to calculate node degrees if d in ['FB15k-237']: dataset = RelLinkPredDataset(os.path.join(data_dir, d), d, transform=T.NormalizeFeatures()) data = dataset[0] data.x = torch.arange(data.num_nodes) edge_index = torch.cat([data.train_edge_index, data.valid_edge_index, data.test_edge_index], dim=1) edge_type = torch.cat([data.train_edge_type, data.valid_edge_type, data.test_edge_type]) data = Data(edge_index=edge_index, edge_type=edge_type) elif d in ['WordNet18RR']: dataset = WordNet18RR(os.path.join(data_dir, d), transform=T.NormalizeFeatures()) data = dataset[0] data.x = torch.arange(data.num_nodes) data.train_mask = data.val_mask = data.test_mask = None elif d in ['WordNet18']: dataset = WordNet18(os.path.join(data_dir, d), transform=T.NormalizeFeatures()) data = dataset[0] data.x = torch.arange(data.num_nodes) # Use original split data.train_pos_edge_index = data.edge_index[:, data.train_mask] data.train_edge_type = data.edge_type[data.train_mask] data.val_pos_edge_index = data.edge_index[:, data.val_mask] data.val_edge_type = data.edge_type[data.val_mask] data.val_neg_edge_index = negative_sampling_kg(data.val_pos_edge_index, data.val_edge_type) data.test_pos_edge_index = data.edge_index[:, data.test_mask] data.test_edge_type = data.edge_type[data.test_mask] data.test_neg_edge_index = negative_sampling_kg(data.test_pos_edge_index, data.test_edge_type) elif 'ogbl' in d: dataset = PygLinkPropPredDataset(root=os.path.join(data_dir, d), name=d) split_edge = dataset.get_edge_split() train_edge, valid_edge, test_edge = split_edge["train"], split_edge["valid"], split_edge["test"] entity_dict = dict() cur_idx = 0 for key in dataset[0]['num_nodes_dict']: entity_dict[key] = (cur_idx, cur_idx + dataset[0]['num_nodes_dict'][key]) cur_idx += dataset[0]['num_nodes_dict'][key] nentity = sum(dataset[0]['num_nodes_dict'].values()) valid_head_neg = valid_edge.pop('head_neg') valid_tail_neg = valid_edge.pop('tail_neg') test_head_neg = test_edge.pop('head_neg') test_tail_neg = test_edge.pop('tail_neg') train = pd.DataFrame(train_edge) valid = pd.DataFrame(valid_edge) test = pd.DataFrame(test_edge) # Convert to global index train['head'] = [idx + entity_dict[tp][0] for idx, tp in zip(train['head'], train['head_type'])] train['tail'] = [idx + entity_dict[tp][0] for idx, tp in zip(train['tail'], train['tail_type'])] valid['head'] = [idx + entity_dict[tp][0] for idx, tp in zip(valid['head'], valid['head_type'])] valid['tail'] = [idx + entity_dict[tp][0] for idx, tp in zip(valid['tail'], valid['tail_type'])] test['head'] = [idx + entity_dict[tp][0] for idx, tp in zip(test['head'], test['head_type'])] test['tail'] = [idx + entity_dict[tp][0] for idx, tp in zip(test['tail'], test['tail_type'])] valid_pos_edge_index = torch.tensor([valid['head'], valid['tail']]) valid_edge_type = torch.tensor(valid.relation) valid_neg_edge_index = torch.stack([valid_pos_edge_index[0], valid_tail_neg[:, 0]]) test_pos_edge_index = torch.tensor([test['head'], test['tail']]) test_edge_type = torch.tensor(test.relation) test_neg_edge_index = torch.stack([test_pos_edge_index[0], test_tail_neg[:, 0]]) train_directed = train[train.head_type != train.tail_type] train_undirected = train[train.head_type == train.tail_type] train_undirected_uni = train_undirected[train_undirected['head'] < train_undirected['tail']] train_uni = pd.concat([train_directed, train_undirected_uni], ignore_index=True) train_pos_edge_index = torch.tensor([train_uni['head'], train_uni['tail']]) train_edge_type = torch.tensor(train_uni.relation) r, c = train_pos_edge_index rev_edge_index = torch.stack([c, r]) rev_edge_type = train_edge_type + 51 edge_index = torch.cat([train_pos_edge_index, rev_edge_index], dim=1) edge_type = torch.cat([train_edge_type, rev_edge_type], dim=0) data = Data( x=torch.arange(nentity), edge_index=edge_index, edge_type=edge_type, train_pos_edge_index=train_pos_edge_index, train_edge_type=train_edge_type, val_pos_edge_index=valid_pos_edge_index, val_edge_type=valid_edge_type, val_neg_edge_index=valid_neg_edge_index, test_pos_edge_index=test_pos_edge_index, test_edge_type=test_edge_type, test_neg_edge_index=test_neg_edge_index) else: raise NotImplementedError print('Processing:', d) print(dataset) for s in seeds: seed_everything(s) # D # data = train_test_split_edges_no_neg_adj_mask(data, test_ratio=0.05, two_hop_degree=two_hop_degree, kg=True) print(s, data) with open(os.path.join(data_dir, d, f'd_{s}.pkl'), 'wb') as f: pickle.dump((dataset, data), f) # Two ways to sample Df from the training set ## 1. Df is within 2 hop local enclosing subgraph of Dtest ## 2. Df is outside of 2 hop local enclosing subgraph of Dtest # All the candidate edges (train edges) # graph = to_networkx(Data(edge_index=data.train_pos_edge_index, x=data.x)) # Get the 2 hop local enclosing subgraph for all test edges _, local_edges, _, mask = k_hop_subgraph( data.test_pos_edge_index.flatten().unique(), 2, data.train_pos_edge_index, num_nodes=dataset[0].num_nodes) distant_edges = data.train_pos_edge_index[:, ~mask] print('Number of edges. Local: ', local_edges.shape[1], 'Distant:', distant_edges.shape[1]) in_mask = mask out_mask = ~mask torch.save( {'out': out_mask, 'in': in_mask}, os.path.join(data_dir, d, f'df_{s}.pt') ) def main(): process_graph() # process_kg() if __name__ == "__main__": main()
16,717
39.97549
134
py
GNNDelete
GNNDelete-main/delete_node.py
import os import copy import json import wandb import pickle import argparse import torch import torch.nn as nn from torch_geometric.utils import to_undirected, to_networkx, k_hop_subgraph, is_undirected from torch_geometric.data import Data import torch_geometric.transforms as T from torch_geometric.datasets import CitationFull, Coauthor, Flickr, RelLinkPredDataset, WordNet18, WordNet18RR from torch_geometric.loader import GraphSAINTRandomWalkSampler from torch_geometric.seed import seed_everything from framework import get_model, get_trainer from framework.models.gcn import GCN from framework.models.deletion import GCNDelete from framework.training_args import parse_args from framework.utils import * from framework.trainer.gnndelete_nodeemb import GNNDeleteNodeClassificationTrainer from train_mi import MLPAttacker device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') torch.autograd.set_detect_anomaly(True) def to_directed(edge_index): row, col = edge_index mask = row < col return torch.cat([row[mask], col[mask]], dim=0) def main(): args = parse_args() args.checkpoint_dir = 'checkpoint_node' args.dataset = 'DBLP' original_path = os.path.join(args.checkpoint_dir, args.dataset, args.gnn, 'original', str(args.random_seed)) attack_path_all = os.path.join(args.checkpoint_dir, args.dataset, 'member_infer_all', str(args.random_seed)) attack_path_sub = os.path.join(args.checkpoint_dir, args.dataset, 'member_infer_sub', str(args.random_seed)) seed_everything(args.random_seed) if 'gnndelete' in args.unlearning_model: args.checkpoint_dir = os.path.join( args.checkpoint_dir, args.dataset, args.gnn, f'{args.unlearning_model}-node_deletion', '-'.join([str(i) for i in [args.loss_fct, args.loss_type, args.alpha, args.neg_sample_random]]), '-'.join([str(i) for i in [args.df, args.df_size, args.random_seed]])) else: args.checkpoint_dir = os.path.join( args.checkpoint_dir, args.dataset, args.gnn, f'{args.unlearning_model}-node_deletion', '-'.join([str(i) for i in [args.df, args.df_size, args.random_seed]])) os.makedirs(args.checkpoint_dir, exist_ok=True) # Dataset dataset = CitationFull(os.path.join(args.data_dir, args.dataset), args.dataset, transform=T.NormalizeFeatures()) data = dataset[0] print('Original data', data) split = T.RandomNodeSplit() data = split(data) assert is_undirected(data.edge_index) print('Split data', data) args.in_dim = data.x.shape[1] args.out_dim = dataset.num_classes wandb.init(config=args) # Df and Dr if args.df_size >= 100: # df_size is number of nodes/edges to be deleted df_size = int(args.df_size) else: # df_size is the ratio df_size = int(args.df_size / 100 * data.train_pos_edge_index.shape[1]) print(f'Original size: {data.num_nodes:,}') print(f'Df size: {df_size:,}') # Delete nodes df_nodes = torch.randperm(data.num_nodes)[:df_size] global_node_mask = torch.ones(data.num_nodes, dtype=torch.bool) global_node_mask[df_nodes] = False dr_mask_node = global_node_mask df_mask_node = ~global_node_mask assert df_mask_node.sum() == df_size # Delete edges associated with deleted nodes from training set res = [torch.eq(data.edge_index, aelem).logical_or_(torch.eq(data.edge_index, aelem)) for aelem in df_nodes] df_mask_edge = torch.any(torch.stack(res, dim=0), dim = 0) df_mask_edge = df_mask_edge.sum(0).bool() dr_mask_edge = ~df_mask_edge df_edge = data.edge_index[:, df_mask_edge] data.directed_df_edge_index = to_directed(df_edge) # print(df_edge.shape, directed_df_edge_index.shape) # raise print('Deleting the following nodes:', df_nodes) # # Delete edges associated with deleted nodes from valid and test set # res = [torch.eq(data.val_pos_edge_index, aelem).logical_or_(torch.eq(data.val_pos_edge_index, aelem)) for aelem in df_nodes] # mask = torch.any(torch.stack(res, dim=0), dim = 0) # mask = mask.sum(0).bool() # mask = ~mask # data.val_pos_edge_index = data.val_pos_edge_index[:, mask] # data.val_neg_edge_index = data.val_neg_edge_index[:, :data.val_pos_edge_index.shape[1]] # res = [torch.eq(data.test_pos_edge_index, aelem).logical_or_(torch.eq(data.test_pos_edge_index, aelem)) for aelem in df_nodes] # mask = torch.any(torch.stack(res, dim=0), dim = 0) # mask = mask.sum(0).bool() # mask = ~mask # data.test_pos_edge_index = data.test_pos_edge_index[:, mask] # data.test_neg_edge_index = data.test_neg_edge_index[:, :data.test_pos_edge_index.shape[1]] # For testing # data.directed_df_edge_index = data.train_pos_edge_index[:, df_mask_edge] # if args.gnn in ['rgcn', 'rgat']: # data.directed_df_edge_type = data.train_edge_type[df_mask] # Edges in S_Df _, two_hop_edge, _, two_hop_mask = k_hop_subgraph( data.edge_index[:, df_mask_edge].flatten().unique(), 2, data.edge_index, num_nodes=data.num_nodes) # Nodes in S_Df _, one_hop_edge, _, one_hop_mask = k_hop_subgraph( data.edge_index[:, df_mask_edge].flatten().unique(), 1, data.edge_index, num_nodes=data.num_nodes) sdf_node_1hop = torch.zeros(data.num_nodes, dtype=torch.bool) sdf_node_2hop = torch.zeros(data.num_nodes, dtype=torch.bool) sdf_node_1hop[one_hop_edge.flatten().unique()] = True sdf_node_2hop[two_hop_edge.flatten().unique()] = True assert sdf_node_1hop.sum() == len(one_hop_edge.flatten().unique()) assert sdf_node_2hop.sum() == len(two_hop_edge.flatten().unique()) data.sdf_node_1hop_mask = sdf_node_1hop data.sdf_node_2hop_mask = sdf_node_2hop # To undirected for message passing # print(is_undir0.0175ected(data.train_pos_edge_index), data.train_pos_edge_index.shape, two_hop_mask.shape, df_mask.shape, two_hop_mask.shape) # assert not is_undirected(data.edge_index) print(is_undirected(data.edge_index)) if args.gnn in ['rgcn', 'rgat']: r, c = data.train_pos_edge_index rev_edge_index = torch.stack([c, r], dim=0) rev_edge_type = data.train_edge_type + args.num_edge_type data.edge_index = torch.cat((data.train_pos_edge_index, rev_edge_index), dim=1) data.edge_type = torch.cat([data.train_edge_type, rev_edge_type], dim=0) # data.train_mask = data.train_mask.repeat(2) two_hop_mask = two_hop_mask.repeat(2).view(-1) df_mask = df_mask.repeat(2).view(-1) dr_mask = dr_mask.repeat(2).view(-1) assert is_undirected(data.edge_index) else: # train_pos_edge_index, [df_mask, two_hop_mask] = to_undirected(data.train_pos_edge_index, [df_mask.int(), two_hop_mask.int()]) two_hop_mask = two_hop_mask.bool() df_mask_edge = df_mask_edge.bool() dr_mask_edge = ~df_mask_edge # data.train_pos_edge_index = train_pos_edge_index # assert is_undirected(data.train_pos_edge_index) print('Undirected dataset:', data) # print(is_undirected(train_pos_edge_index), train_pos_edge_index.shape, two_hop_mask.shape, df_mask.shape, two_hop_mask.shape) data.sdf_mask = two_hop_mask data.df_mask = df_mask_edge data.dr_mask = dr_mask_edge data.dtrain_mask = dr_mask_edge # print(is_undirected(data.train_pos_edge_index), data.train_pos_edge_index.shape, data.two_hop_mask.shape, data.df_mask.shape, data.two_hop_mask.shape) # raise # Model model = GCNDelete(args) # model = get_model(args, sdf_node_1hop, sdf_node_2hop, num_nodes=data.num_nodes, num_edge_type=args.num_edge_type) if args.unlearning_model != 'retrain': # Start from trained GNN model if os.path.exists(os.path.join(original_path, 'pred_proba.pt')): logits_ori = torch.load(os.path.join(original_path, 'pred_proba.pt')) if logits_ori is not None: logits_ori = logits_ori.to(device) else: logits_ori = None model_ckpt = torch.load(os.path.join(original_path, 'model_best.pt'), map_location=device) model.load_state_dict(model_ckpt['model_state'], strict=False) else: # Initialize a new GNN model retrain = None logits_ori = None model = model.to(device) if 'gnndelete' in args.unlearning_model and 'nodeemb' in args.unlearning_model: parameters_to_optimize = [ {'params': [p for n, p in model.named_parameters() if 'del' in n], 'weight_decay': 0.0} ] print('parameters_to_optimize', [n for n, p in model.named_parameters() if 'del' in n]) if 'layerwise' in args.loss_type: optimizer1 = torch.optim.Adam(model.deletion1.parameters(), lr=args.lr) optimizer2 = torch.optim.Adam(model.deletion2.parameters(), lr=args.lr) optimizer = [optimizer1, optimizer2] else: optimizer = torch.optim.Adam(parameters_to_optimize, lr=args.lr) else: if 'gnndelete' in args.unlearning_model: parameters_to_optimize = [ {'params': [p for n, p in model.named_parameters() if 'del' in n], 'weight_decay': 0.0} ] print('parameters_to_optimize', [n for n, p in model.named_parameters() if 'del' in n]) else: parameters_to_optimize = [ {'params': [p for n, p in model.named_parameters()], 'weight_decay': 0.0} ] print('parameters_to_optimize', [n for n, p in model.named_parameters()]) optimizer = torch.optim.Adam(parameters_to_optimize, lr=args.lr)#, weight_decay=args.weight_decay) wandb.watch(model, log_freq=100) # MI attack model attack_model_all = None # attack_model_all = MLPAttacker(args) # attack_ckpt = torch.load(os.path.join(attack_path_all, 'attack_model_best.pt')) # attack_model_all.load_state_dict(attack_ckpt['model_state']) # attack_model_all = attack_model_all.to(device) attack_model_sub = None # attack_model_sub = MLPAttacker(args) # attack_ckpt = torch.load(os.path.join(attack_path_sub, 'attack_model_best.pt')) # attack_model_sub.load_state_dict(attack_ckpt['model_state']) # attack_model_sub = attack_model_sub.to(device) # Train trainer = GNNDeleteNodeClassificationTrainer(args) trainer.train(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) # Test if args.unlearning_model != 'retrain': retrain_path = os.path.join( 'checkpoint', args.dataset, args.gnn, 'retrain', '-'.join([str(i) for i in [args.df, args.df_size, args.random_seed]])) retrain_ckpt = torch.load(os.path.join(retrain_path, 'model_best.pt'), map_location=device) retrain_args = copy.deepcopy(args) retrain_args.unlearning_model = 'retrain' retrain = get_model(retrain_args, num_nodes=data.num_nodes, num_edge_type=args.num_edge_type) retrain.load_state_dict(retrain_ckpt['model_state']) retrain = retrain.to(device) retrain.eval() else: retrain = None trainer.test(model, data, model_retrain=retrain, attack_model_all=attack_model_all, attack_model_sub=attack_model_sub) trainer.save_log() if __name__ == "__main__": main()
11,453
40.80292
156
py
GNNDelete
GNNDelete-main/framework/data_loader.py
import os import torch from torch_geometric.data import Data, GraphSAINTRandomWalkSampler def load_dict(filename): '''Load entity and relation to id mapping''' mapping = {} with open(filename, 'r') as f: for l in f: l = l.strip().split('\t') mapping[l[0]] = l[1] return mapping def load_edges(filename): with open(filename, 'r') as f: r = f.readlines() r = [i.strip().split('\t') for i in r] return r def generate_true_dict(all_triples): heads = {(r, t) : [] for _, r, t in all_triples} tails = {(h, r) : [] for h, r, _ in all_triples} for h, r, t in all_triples: heads[r, t].append(h) tails[h, r].append(t) return heads, tails def get_loader(args, delete=[]): prefix = os.path.join('./data', args.dataset) # Edges train = load_edges(os.path.join(prefix, 'train.txt')) valid = load_edges(os.path.join(prefix, 'valid.txt')) test = load_edges(os.path.join(prefix, 'test.txt')) train = [(int(i[0]), int(i[1]), int(i[2])) for i in train] valid = [(int(i[0]), int(i[1]), int(i[2])) for i in valid] test = [(int(i[0]), int(i[1]), int(i[2])) for i in test] train_rev = [(int(i[2]), int(i[1]), int(i[0])) for i in train] valid_rev = [(int(i[2]), int(i[1]), int(i[0])) for i in valid] test_rev = [(int(i[2]), int(i[1]), int(i[0])) for i in test] train = train + train_rev valid = valid + valid_rev test = test + test_rev all_edge = train + valid + test true_triples = generate_true_dict(all_edge) edge = torch.tensor([(int(i[0]), int(i[2])) for i in all_edge], dtype=torch.long).t() edge_type = torch.tensor([int(i[1]) for i in all_edge], dtype=torch.long)#.view(-1, 1) # Masks train_size = len(train) valid_size = len(valid) test_size = len(test) total_size = train_size + valid_size + test_size train_mask = torch.zeros((total_size,)).bool() train_mask[:train_size] = True valid_mask = torch.zeros((total_size,)).bool() valid_mask[train_size:train_size + valid_size] = True test_mask = torch.zeros((total_size,)).bool() test_mask[-test_size:] = True # Graph size num_nodes = edge.flatten().unique().shape[0] num_edges = edge.shape[1] num_edge_type = edge_type.unique().shape[0] # Node feature x = torch.rand((num_nodes, args.in_dim)) # Delete edges if len(delete) > 0: delete_idx = torch.tensor(delete, dtype=torch.long) num_train_edges = train_size // 2 train_mask[delete_idx] = False train_mask[delete_idx + num_train_edges] = False train_size -= 2 * len(delete) node_id = torch.arange(num_nodes) dataset = Data( edge_index=edge, edge_type=edge_type, x=x, node_id=node_id, train_mask=train_mask, valid_mask=valid_mask, test_mask=test_mask) dataloader = GraphSAINTRandomWalkSampler( dataset, batch_size=args.batch_size, walk_length=args.walk_length, num_steps=args.num_steps) print(f'Dataset: {args.dataset}, Num nodes: {num_nodes}, Num edges: {num_edges//2}, Num relation types: {num_edge_type}') print(f'Train edges: {train_size//2}, Valid edges: {valid_size//2}, Test edges: {test_size//2}') return dataloader, valid, test, true_triples, num_nodes, num_edges, num_edge_type
3,344
32.45
125
py
GNNDelete
GNNDelete-main/framework/utils.py
import numpy as np import torch import networkx as nx def get_node_edge(graph): degree_sorted_ascend = sorted(graph.degree, key=lambda x: x[1]) return degree_sorted_ascend[-1][0] def h_hop_neighbor(G, node, h): path_lengths = nx.single_source_dijkstra_path_length(G, node) return [node for node, length in path_lengths.items() if length == h] def get_enclosing_subgraph(graph, edge_to_delete): subgraph = {0: [edge_to_delete]} s, t = edge_to_delete neighbor_s = [] neighbor_t = [] for h in range(1, 2+1): neighbor_s += h_hop_neighbor(graph, s, h) neighbor_t += h_hop_neighbor(graph, t, h) nodes = neighbor_s + neighbor_t + [s, t] subgraph[h] = list(graph.subgraph(nodes).edges()) return subgraph @torch.no_grad() def get_link_labels(pos_edge_index, neg_edge_index): E = pos_edge_index.size(1) + neg_edge_index.size(1) link_labels = torch.zeros(E, dtype=torch.float, device=pos_edge_index.device) link_labels[:pos_edge_index.size(1)] = 1. return link_labels @torch.no_grad() def get_link_labels_kg(pos_edge_index, neg_edge_index): E = pos_edge_index.size(1) + neg_edge_index.size(1) link_labels = torch.zeros(E, dtype=torch.float, device=pos_edge_index.device) link_labels[:pos_edge_index.size(1)] = 1. return link_labels @torch.no_grad() def negative_sampling_kg(edge_index, edge_type): '''Generate negative samples but keep the node type the same''' edge_index_copy = edge_index.clone() for et in edge_type.unique(): mask = (edge_type == et) old_source = edge_index_copy[0, mask] new_index = torch.randperm(old_source.shape[0]) new_source = old_source[new_index] edge_index_copy[0, mask] = new_source return edge_index_copy
1,852
30.40678
81
py
GNNDelete
GNNDelete-main/framework/evaluation.py
import torch import torch.nn.functional as F from sklearn.metrics import roc_auc_score, average_precision_score from .utils import get_link_labels device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') @torch.no_grad() def eval_lp(model, stage, data=None, loader=None): model.eval() # For full batch if data is not None: pos_edge_index = data[f'{stage}_pos_edge_index'] neg_edge_index = data[f'{stage}_neg_edge_index'] if hasattr(data, 'dtrain_mask') and data.dtrain_mask is not None: embedding = model(data.x.to(device), data.train_pos_edge_index[:, data.dtrain_mask].to(device)) else: embedding = model(data.x.to(device), data.train_pos_edge_index.to(device)) logits = model.decode(embedding, pos_edge_index, neg_edge_index).sigmoid() label = get_link_labels(pos_edge_index, neg_edge_index) # For mini batch if loader is not None: logits = [] label = [] for batch in loader: edge_index = batch.edge_index.to(device) if hasattr(batch, 'edge_type'): edge_type = batch.edge_type.to(device) embedding1 = model1(edge_index, edge_type) embedding2 = model2(edge_index, edge_type) s1 = model.decode(embedding1, edge_index, edge_type) s2 = model.decode(embedding2, edge_index, edge_type) else: embedding1 = model1(edge_index) embedding2 = model2(edge_index) s1 = model.decode(embedding1, edge_index) s2 = model.decode(embedding2, edge_index) embedding = model(data.train_pos_edge_index.to(device)) lg = model.decode(embedding, pos_edge_index, neg_edge_index).sigmoid() lb = get_link_labels(pos_edge_index, neg_edge_index) logits.append(lg) label.append(lb) loss = F.binary_cross_entropy_with_logits(logits, label) auc = roc_auc_score(label.cpu(), logits.cpu()) aup = average_precision_score(label.cpu(), logits.cpu()) return loss, auc, aup @torch.no_grad() def verification_error(model1, model2): '''L2 distance between aproximate model and re-trained model''' model1 = model1.to('cpu') model2 = model2.to('cpu') modules1 = {n: p for n, p in model1.named_parameters()} modules2 = {n: p for n, p in model2.named_parameters()} all_names = set(modules1.keys()) & set(modules2.keys()) print(all_names) diff = torch.tensor(0.0).float() for n in all_names: diff += torch.norm(modules1[n] - modules2[n]) return diff @torch.no_grad() def member_infer_attack(target_model, attack_model, data, logits=None): '''Membership inference attack''' edge = data.train_pos_edge_index[:, data.df_mask] z = target_model(data.x, data.train_pos_edge_index[:, data.dr_mask]) feature1 = target_model.decode(z, edge).sigmoid() feature0 = 1 - feature1 feature = torch.stack([feature0, feature1], dim=1) # feature = torch.cat([z[edge[0]], z[edge][1]], dim=-1) logits = attack_model(feature) _, pred = torch.max(logits, 1) suc_rate = 1 - pred.float().mean() return torch.softmax(logits, dim=-1).squeeze().tolist(), suc_rate.cpu().item() @torch.no_grad() def member_infer_attack_node(target_model, attack_model, data, logits=None): '''Membership inference attack''' edge = data.train_pos_edge_index[:, data.df_mask] z = target_model(data.x, data.train_pos_edge_index[:, data.dr_mask]) feature = torch.cat([z[edge[0]], z[edge][1]], dim=-1) logits = attack_model(feature) _, pred = torch.max(logits, 1) suc_rate = 1 - pred.float().mean() return torch.softmax(logits, dim=-1).squeeze().tolist(), suc_rate.cpu().item() @torch.no_grad() def get_node_embedding_data(model, data): model.eval() if hasattr(data, 'dtrain_mask') and data.dtrain_mask is not None: node_embedding = model(data.x.to(device), data.train_pos_edge_index[:, data.dtrain_mask].to(device)) else: node_embedding = model(data.x.to(device), data.train_pos_edge_index.to(device)) return node_embedding @torch.no_grad() def output_kldiv(model1, model2, data=None, loader=None): '''KL-Divergence between output distribution of model and re-trained model''' model1.eval() model2.eval() # For full batch if data is not None: embedding1 = get_node_embedding_data(model1, data).to(device) embedding2 = get_node_embedding_data(model2, data).to(device) if data.edge_index is not None: edge_index = data.edge_index.to(device) if data.train_pos_edge_index is not None: edge_index = data.train_pos_edge_index.to(device) if hasattr(data, 'edge_type'): edge_type = data.edge_type.to(device) score1 = model1.decode(embedding1, edge_index, edge_type) score2 = model2.decode(embedding2, edge_index, edge_type) else: score1 = model1.decode(embedding1, edge_index) score2 = model2.decode(embedding2, edge_index) # For mini batch if loader is not None: score1 = [] score2 = [] for batch in loader: edge_index = batch.edge_index.to(device) if hasattr(batch, 'edge_type'): edge_type = batch.edge_type.to(device) embedding1 = model1(edge, edge_type) embedding2 = model2(edge, edge_type) s1 = model.decode(embedding1, edge, edge_type) s2 = model.decode(embedding2, edge, edge_type) else: embedding1 = model1(edge) embedding2 = model2(edge) s1 = model.decode(embedding1, edge) s2 = model.decode(embedding2, edge) score1.append(s1) score2.append(s2) score1 = torch.hstack(score1) score2 = torch.hstack(score2) kldiv = F.kl_div( F.log_softmax(score1, dim=-1), F.softmax(score2, dim=-1) ) return kldiv
6,151
32.254054
108
py
GNNDelete
GNNDelete-main/framework/trainer/base.py
import os import time import json import wandb import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from tqdm import trange, tqdm from ogb.graphproppred import Evaluator from torch_geometric.data import DataLoader from torch_geometric.utils import negative_sampling from torch_geometric.loader import GraphSAINTRandomWalkSampler from sklearn.metrics import roc_auc_score, average_precision_score, accuracy_score, f1_score from ..evaluation import * from ..training_args import parse_args from ..utils import * device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # device = 'cpu' class Trainer: def __init__(self, args): self.args = args self.trainer_log = { 'unlearning_model': args.unlearning_model, 'dataset': args.dataset, 'log': []} self.logit_all_pair = None self.df_pos_edge = [] with open(os.path.join(self.args.checkpoint_dir, 'training_args.json'), 'w') as f: json.dump(vars(args), f) def freeze_unused_weights(self, model, mask): grad_mask = torch.zeros_like(mask) grad_mask[mask] = 1 model.deletion1.deletion_weight.register_hook(lambda grad: grad.mul_(grad_mask)) model.deletion2.deletion_weight.register_hook(lambda grad: grad.mul_(grad_mask)) @torch.no_grad() def get_link_labels(self, pos_edge_index, neg_edge_index): E = pos_edge_index.size(1) + neg_edge_index.size(1) link_labels = torch.zeros(E, dtype=torch.float, device=pos_edge_index.device) link_labels[:pos_edge_index.size(1)] = 1. return link_labels @torch.no_grad() def get_embedding(self, model, data, on_cpu=False): original_device = next(model.parameters()).device if on_cpu: model = model.cpu() data = data.cpu() z = model(data.x, data.train_pos_edge_index[:, data.dtrain_mask]) model = model.to(original_device) return z def train(self, model, data, optimizer, args): if self.args.dataset in ['Cora', 'PubMed', 'DBLP', 'CS']: return self.train_fullbatch(model, data, optimizer, args) if self.args.dataset in ['Physics']: return self.train_minibatch(model, data, optimizer, args) if 'ogbl' in self.args.dataset: return self.train_minibatch(model, data, optimizer, args) def train_fullbatch(self, model, data, optimizer, args): start_time = time.time() best_valid_loss = 1000000 data = data.to(device) for epoch in trange(args.epochs, desc='Epoch'): model.train() # Positive and negative sample neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index, num_nodes=data.num_nodes, num_neg_samples=data.dtrain_mask.sum()) z = model(data.x, data.train_pos_edge_index) # edge = torch.cat([train_pos_edge_index, neg_edge_index], dim=-1) # logits = model.decode(z, edge[0], edge[1]) logits = model.decode(z, data.train_pos_edge_index, neg_edge_index) label = get_link_labels(data.train_pos_edge_index, neg_edge_index) loss = F.binary_cross_entropy_with_logits(logits, label) loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() if (epoch+1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': loss.item() } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(train_log) self.trainer_log['log'].append(valid_log) if valid_loss < best_valid_loss: best_valid_loss = valid_loss best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) torch.save(z, os.path.join(args.checkpoint_dir, 'node_embeddings.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save models and node embeddings print('Saving final checkpoint') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best valid loss = {best_valid_loss:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_valid_loss'] = best_valid_loss def train_minibatch(self, model, data, optimizer, args): start_time = time.time() best_valid_loss = 1000000 data.edge_index = data.train_pos_edge_index loader = GraphSAINTRandomWalkSampler( data, batch_size=args.batch_size, walk_length=2, num_steps=args.num_steps, ) for epoch in trange(args.epochs, desc='Epoch'): model.train() epoch_loss = 0 for step, batch in enumerate(tqdm(loader, desc='Step', leave=False)): # Positive and negative sample train_pos_edge_index = batch.edge_index.to(device) z = model(batch.x.to(device), train_pos_edge_index) neg_edge_index = negative_sampling( edge_index=train_pos_edge_index, num_nodes=z.size(0)) logits = model.decode(z, train_pos_edge_index, neg_edge_index) label = get_link_labels(train_pos_edge_index, neg_edge_index) loss = F.binary_cross_entropy_with_logits(logits, label) loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() log = { 'epoch': epoch, 'step': step, 'train_loss': loss.item(), } wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) epoch_loss += loss.item() if (epoch+1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': epoch_loss / step } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(train_log) self.trainer_log['log'].append(valid_log) if valid_loss < best_valid_loss: best_valid_loss = valid_loss best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) torch.save(z, os.path.join(args.checkpoint_dir, 'node_embeddings.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save models and node embeddings print('Saving final checkpoint') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best valid loss = {best_valid_loss:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_valid_loss'] = best_valid_loss self.trainer_log['training_time'] = np.mean([i['epoch_time'] for i in self.trainer_log['log'] if 'epoch_time' in i]) @torch.no_grad() def eval(self, model, data, stage='val', pred_all=False): model.eval() pos_edge_index = data[f'{stage}_pos_edge_index'] neg_edge_index = data[f'{stage}_neg_edge_index'] if self.args.eval_on_cpu: model = model.to('cpu') if hasattr(data, 'dtrain_mask'): mask = data.dtrain_mask else: mask = data.dr_mask z = model(data.x, data.train_pos_edge_index[:, mask]) logits = model.decode(z, pos_edge_index, neg_edge_index).sigmoid() label = self.get_link_labels(pos_edge_index, neg_edge_index) # DT AUC AUP loss = F.binary_cross_entropy_with_logits(logits, label).cpu().item() dt_auc = roc_auc_score(label.cpu(), logits.cpu()) dt_aup = average_precision_score(label.cpu(), logits.cpu()) # DF AUC AUP if self.args.unlearning_model in ['original']: df_logit = [] else: # df_logit = model.decode(z, data.train_pos_edge_index[:, data.df_mask]).sigmoid().tolist() df_logit = model.decode(z, data.directed_df_edge_index).sigmoid().tolist() if len(df_logit) > 0: df_auc = [] df_aup = [] # Sample pos samples if len(self.df_pos_edge) == 0: for i in range(500): mask = torch.zeros(data.train_pos_edge_index[:, data.dr_mask].shape[1], dtype=torch.bool) idx = torch.randperm(data.train_pos_edge_index[:, data.dr_mask].shape[1])[:len(df_logit)] mask[idx] = True self.df_pos_edge.append(mask) # Use cached pos samples for mask in self.df_pos_edge: pos_logit = model.decode(z, data.train_pos_edge_index[:, data.dr_mask][:, mask]).sigmoid().tolist() logit = df_logit + pos_logit label = [0] * len(df_logit) + [1] * len(df_logit) df_auc.append(roc_auc_score(label, logit)) df_aup.append(average_precision_score(label, logit)) df_auc = np.mean(df_auc) df_aup = np.mean(df_aup) else: df_auc = np.nan df_aup = np.nan # Logits for all node pairs if pred_all: logit_all_pair = (z @ z.t()).cpu() else: logit_all_pair = None log = { f'{stage}_loss': loss, f'{stage}_dt_auc': dt_auc, f'{stage}_dt_aup': dt_aup, f'{stage}_df_auc': df_auc, f'{stage}_df_aup': df_aup, f'{stage}_df_logit_mean': np.mean(df_logit) if len(df_logit) > 0 else np.nan, f'{stage}_df_logit_std': np.std(df_logit) if len(df_logit) > 0 else np.nan } if self.args.eval_on_cpu: model = model.to(device) return loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, log @torch.no_grad() def test(self, model, data, model_retrain=None, attack_model_all=None, attack_model_sub=None, ckpt='best'): if ckpt == 'best': # Load best ckpt ckpt = torch.load(os.path.join(self.args.checkpoint_dir, 'model_best.pt')) model.load_state_dict(ckpt['model_state']) if 'ogbl' in self.args.dataset: pred_all = False else: pred_all = True loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, test_log = self.eval(model, data, 'test', pred_all) self.trainer_log['dt_loss'] = loss self.trainer_log['dt_auc'] = dt_auc self.trainer_log['dt_aup'] = dt_aup self.trainer_log['df_logit'] = df_logit self.logit_all_pair = logit_all_pair self.trainer_log['df_auc'] = df_auc self.trainer_log['df_aup'] = df_aup self.trainer_log['auc_sum'] = dt_auc + df_auc self.trainer_log['aup_sum'] = dt_aup + df_aup self.trainer_log['auc_gap'] = abs(dt_auc - df_auc) self.trainer_log['aup_gap'] = abs(dt_aup - df_aup) # # AUC AUP on Df # if len(df_logit) > 0: # auc = [] # aup = [] # if self.args.eval_on_cpu: # model = model.to('cpu') # z = model(data.x, data.train_pos_edge_index[:, data.dtrain_mask]) # for i in range(500): # mask = torch.zeros(data.train_pos_edge_index[:, data.dr_mask].shape[1], dtype=torch.bool) # idx = torch.randperm(data.train_pos_edge_index[:, data.dr_mask].shape[1])[:len(df_logit)] # mask[idx] = True # pos_logit = model.decode(z, data.train_pos_edge_index[:, data.dr_mask][:, mask]).sigmoid().tolist() # logit = df_logit + pos_logit # label = [0] * len(df_logit) + [1] * len(df_logit) # auc.append(roc_auc_score(label, logit)) # aup.append(average_precision_score(label, logit)) # self.trainer_log['df_auc'] = np.mean(auc) # self.trainer_log['df_aup'] = np.mean(aup) if model_retrain is not None: # Deletion self.trainer_log['ve'] = verification_error(model, model_retrain).cpu().item() # self.trainer_log['dr_kld'] = output_kldiv(model, model_retrain, data=data).cpu().item() # MI Attack after unlearning if attack_model_all is not None: mi_logit_all_after, mi_sucrate_all_after = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_after'] = mi_logit_all_after self.trainer_log['mi_sucrate_all_after'] = mi_sucrate_all_after if attack_model_sub is not None: mi_logit_sub_after, mi_sucrate_sub_after = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_after'] = mi_logit_sub_after self.trainer_log['mi_sucrate_sub_after'] = mi_sucrate_sub_after self.trainer_log['mi_ratio_all'] = np.mean([i[1] / j[1] for i, j in zip(self.trainer_log['mi_logit_all_after'], self.trainer_log['mi_logit_all_before'])]) self.trainer_log['mi_ratio_sub'] = np.mean([i[1] / j[1] for i, j in zip(self.trainer_log['mi_logit_sub_after'], self.trainer_log['mi_logit_sub_before'])]) print(self.trainer_log['mi_ratio_all'], self.trainer_log['mi_ratio_sub'], self.trainer_log['mi_sucrate_all_after'], self.trainer_log['mi_sucrate_sub_after']) print(self.trainer_log['df_auc'], self.trainer_log['df_aup']) return loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, test_log @torch.no_grad() def get_output(self, model, node_embedding, data): model.eval() node_embedding = node_embedding.to(device) edge = data.edge_index.to(device) output = model.decode(node_embedding, edge, edge_type) return output def save_log(self): # print(self.trainer_log) with open(os.path.join(self.args.checkpoint_dir, 'trainer_log.json'), 'w') as f: json.dump(self.trainer_log, f) torch.save(self.logit_all_pair, os.path.join(self.args.checkpoint_dir, 'pred_proba.pt')) class KGTrainer(Trainer): def train(self, model, data, optimizer, args): model = model.to(device) start_time = time.time() best_metric = 0 print('Num workers:', len(os.sched_getaffinity(0))) loader = GraphSAINTRandomWalkSampler( data, batch_size=128, walk_length=2, num_steps=args.num_steps, num_workers=len(os.sched_getaffinity(0)) ) for epoch in trange(args.epochs, desc='Epoch'): model.train() epoch_loss = 0 epoch_time = 0 for step, batch in enumerate(tqdm(loader, desc='Step', leave=False)): start_time = time.time() batch = batch.to(device) # Message passing edge_index = batch.edge_index#[:, batch.train_mask] edge_type = batch.edge_type#[batch.train_mask] z = model(batch.x, edge_index, edge_type) # Positive and negative sample decoding_mask = (edge_type < args.num_edge_type) # Only select directed edges for link prediction decoding_edge_index = edge_index[:, decoding_mask] decoding_edge_type = edge_type[decoding_mask] neg_edge_index = negative_sampling_kg( edge_index=decoding_edge_index, edge_type=decoding_edge_type) pos_logits = model.decode(z, decoding_edge_index, decoding_edge_type) neg_logits = model.decode(z, neg_edge_index, decoding_edge_type) logits = torch.cat([pos_logits, neg_logits], dim=-1) label = get_link_labels(decoding_edge_index, neg_edge_index) # reg_loss = z.pow(2).mean() + model.W.pow(2).mean() loss = F.binary_cross_entropy_with_logits(logits, label)# + 1e-2 * reg_loss loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() log = { 'epoch': epoch, 'step': step, 'train_loss': loss.item(), } wandb.log(log) # msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] # tqdm.write(' | '.join(msg)) epoch_loss += loss.item() epoch_time += time.time() - start_time if (epoch + 1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': epoch_loss / step, 'epoch_time': epoch_time } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(train_log) self.trainer_log['log'].append(valid_log) if dt_aup > best_metric: best_metric = dt_aup best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save models and node embeddings print('Saving final checkpoint') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best valid aup = {best_metric:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_metric'] = best_metric self.trainer_log['training_time'] = np.mean([i['epoch_time'] for i in self.trainer_log['log'] if 'epoch_time' in i]) @torch.no_grad() def eval(self, model, data, stage='val', pred_all=False): model.eval() pos_edge_index = data[f'{stage}_pos_edge_index'] neg_edge_index = data[f'{stage}_neg_edge_index'] pos_edge_type = data[f'{stage}_edge_type'] neg_edge_type = data[f'{stage}_edge_type'] if self.args.eval_on_cpu: model = model.to('cpu') z = model(data.x, data.edge_index[:, data.dr_mask], data.edge_type[data.dr_mask]) decoding_edge_index = torch.cat([pos_edge_index, neg_edge_index], dim=-1) decoding_edge_type = torch.cat([pos_edge_type, neg_edge_type], dim=-1) logits = model.decode(z, decoding_edge_index, decoding_edge_type) label = get_link_labels(pos_edge_index, neg_edge_index) # DT AUC AUP loss = F.binary_cross_entropy_with_logits(logits, label).cpu().item() dt_auc = roc_auc_score(label.cpu(), logits.cpu()) dt_aup = average_precision_score(label.cpu(), logits.cpu()) # DF AUC AUP if self.args.unlearning_model in ['original']: df_logit = [] else: # df_logit = model.decode(z, data.train_pos_edge_index[:, data.df_mask], data.train_edge_type[data.df_mask]).sigmoid().tolist() df_logit = model.decode(z, data.directed_df_edge_index, data.directed_df_edge_type).sigmoid().tolist() dr_mask = data.dr_mask[:data.dr_mask.shape[0] // 2] if len(df_logit) > 0: df_auc = [] df_aup = [] for i in range(500): mask = torch.zeros(data.train_pos_edge_index[:, dr_mask].shape[1], dtype=torch.bool) idx = torch.randperm(data.train_pos_edge_index[:, dr_mask].shape[1])[:len(df_logit)] mask[idx] = True pos_logit = model.decode(z, data.train_pos_edge_index[:, dr_mask][:, mask], data.train_edge_type[dr_mask][mask]).sigmoid().tolist() logit = df_logit + pos_logit label = [0] * len(df_logit) + [1] * len(df_logit) df_auc.append(roc_auc_score(label, logit)) df_aup.append(average_precision_score(label, logit)) df_auc = np.mean(df_auc) df_aup = np.mean(df_aup) else: df_auc = np.nan df_aup = np.nan # Logits for all node pairs if pred_all: logit_all_pair = (z @ z.t()).cpu() else: logit_all_pair = None log = { f'{stage}_loss': loss, f'{stage}_dt_auc': dt_auc, f'{stage}_dt_aup': dt_aup, f'{stage}_df_auc': df_auc, f'{stage}_df_aup': df_aup, f'{stage}_df_logit_mean': np.mean(df_logit) if len(df_logit) > 0 else np.nan, f'{stage}_df_logit_std': np.std(df_logit) if len(df_logit) > 0 else np.nan } if self.args.eval_on_cpu: model = model.to(device) return loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, log @torch.no_grad() def test(self, model, data, model_retrain=None, attack_model_all=None, attack_model_sub=None, ckpt='ckpt'): if ckpt is 'best': # Load best ckpt ckpt = torch.load(os.path.join(self.args.checkpoint_dir, 'model_best.pt')) model.load_state_dict(ckpt['model_state']) if 'ogbl' in self.args.dataset: pred_all = False else: pred_all = True loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, test_log = self.eval(model, data, 'test', pred_all) self.trainer_log['dt_loss'] = loss self.trainer_log['dt_auc'] = dt_auc self.trainer_log['dt_aup'] = dt_aup self.trainer_log['df_logit'] = df_logit self.logit_all_pair = logit_all_pair self.trainer_log['df_auc'] = df_auc self.trainer_log['df_aup'] = df_aup # if model_retrain is not None: # Deletion # self.trainer_log['ve'] = verification_error(model, model_retrain).cpu().item() # self.trainer_log['dr_kld'] = output_kldiv(model, model_retrain, data=data).cpu().item() # MI Attack after unlearning if attack_model_all is not None: mi_logit_all_after, mi_sucrate_all_after = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_after'] = mi_logit_all_after self.trainer_log['mi_sucrate_all_after'] = mi_sucrate_all_after if attack_model_sub is not None: mi_logit_sub_after, mi_sucrate_sub_after = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_after'] = mi_logit_sub_after self.trainer_log['mi_sucrate_sub_after'] = mi_sucrate_sub_after self.trainer_log['mi_ratio_all'] = np.mean([i[1] / j[1] for i, j in zip(self.trainer_log['mi_logit_all_after'], self.trainer_log['mi_logit_all_before'])]) self.trainer_log['mi_ratio_sub'] = np.mean([i[1] / j[1] for i, j in zip(self.trainer_log['mi_logit_sub_after'], self.trainer_log['mi_logit_sub_before'])]) print(self.trainer_log['mi_ratio_all'], self.trainer_log['mi_ratio_sub'], self.trainer_log['mi_sucrate_all_after'], self.trainer_log['mi_sucrate_sub_after']) print(self.trainer_log['df_auc'], self.trainer_log['df_aup']) return loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, test_log def _train(self, model, data, optimizer, args): model = model.to(device) data = data.to(device) start_time = time.time() best_valid_loss = 1000000 for epoch in trange(args.epochs, desc='Epoch'): model.train() # Message passing z = model(data.x, data.edge_index, data.edge_type) # Positive and negative sample mask = (data.edge_type < args.num_edge_type) # Only select directed edges for link prediction neg_edge_index = negative_sampling_kg( edge_index=data.train_pos_edge_index, edge_type=data.train_edge_type) pos_logits = model.decode(z, data.train_pos_edge_index, data.train_edge_type) neg_logits = model.decode(z, neg_edge_index, data.train_edge_type) logits = torch.cat([pos_logits, neg_logits], dim=-1) label = get_link_labels(data.train_pos_edge_index, neg_edge_index) reg_loss = z.pow(2).mean() + model.W.pow(2).mean() loss = F.binary_cross_entropy_with_logits(logits, label) + 1e-2 * reg_loss loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() log = { 'epoch': epoch, 'train_loss': loss.item(), } wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) if (epoch + 1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': epoch_loss / step } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(train_log) self.trainer_log['log'].append(valid_log) if valid_loss < best_valid_loss: best_valid_loss = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) torch.save(z, os.path.join(args.checkpoint_dir, 'node_embeddings.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save models and node embeddings print('Saving final checkpoint') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best valid loss = {best_valid_loss:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_valid_loss'] = best_valid_loss self.trainer_log['training_time'] = np.mean([i['epoch_time'] for i in self.trainer_log['log'] if 'epoch_time' in i]) class NodeClassificationTrainer(Trainer): def train(self, model, data, optimizer, args): start_time = time.time() best_epoch = 0 best_valid_acc = 0 data = data.to(device) for epoch in trange(args.epochs, desc='Epoch'): model.train() z = F.log_softmax(model(data.x, data.edge_index), dim=1) loss = F.nll_loss(z[data.train_mask], data.y[data.train_mask]) loss.backward() optimizer.step() optimizer.zero_grad() if (epoch+1) % args.valid_freq == 0: valid_loss, dt_acc, dt_f1, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': loss.item() } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(train_log) self.trainer_log['log'].append(valid_log) if dt_acc > best_valid_acc: best_valid_acc = dt_acc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid Acc = {dt_acc:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) torch.save(z, os.path.join(args.checkpoint_dir, 'node_embeddings.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save models and node embeddings print('Saving final checkpoint') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best valid acc = {best_valid_acc:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_valid_acc'] = best_valid_acc @torch.no_grad() def eval(self, model, data, stage='val', pred_all=False): model.eval() if self.args.eval_on_cpu: model = model.to('cpu') # if hasattr(data, 'dtrain_mask'): # mask = data.dtrain_mask # else: # mask = data.dr_mask z = F.log_softmax(model(data.x, data.edge_index), dim=1) # DT AUC AUP loss = F.nll_loss(z[data.val_mask], data.y[data.val_mask]).cpu().item() pred = torch.argmax(z[data.val_mask], dim=1).cpu() dt_acc = accuracy_score(data.y[data.val_mask].cpu(), pred) dt_f1 = f1_score(data.y[data.val_mask].cpu(), pred, average='micro') # DF AUC AUP # if self.args.unlearning_model in ['original', 'original_node']: # df_logit = [] # else: # df_logit = model.decode(z, data.directed_df_edge_index).sigmoid().tolist() # if len(df_logit) > 0: # df_auc = [] # df_aup = [] # # Sample pos samples # if len(self.df_pos_edge) == 0: # for i in range(500): # mask = torch.zeros(data.train_pos_edge_index[:, data.dr_mask].shape[1], dtype=torch.bool) # idx = torch.randperm(data.train_pos_edge_index[:, data.dr_mask].shape[1])[:len(df_logit)] # mask[idx] = True # self.df_pos_edge.append(mask) # # Use cached pos samples # for mask in self.df_pos_edge: # pos_logit = model.decode(z, data.train_pos_edge_index[:, data.dr_mask][:, mask]).sigmoid().tolist() # logit = df_logit + pos_logit # label = [0] * len(df_logit) + [1] * len(df_logit) # df_auc.append(roc_auc_score(label, logit)) # df_aup.append(average_precision_score(label, logit)) # df_auc = np.mean(df_auc) # df_aup = np.mean(df_aup) # else: # df_auc = np.nan # df_aup = np.nan # Logits for all node pairs if pred_all: logit_all_pair = (z @ z.t()).cpu() else: logit_all_pair = None log = { f'{stage}_loss': loss, f'{stage}_dt_acc': dt_acc, f'{stage}_dt_f1': dt_f1, } if self.args.eval_on_cpu: model = model.to(device) return loss, dt_acc, dt_f1, log @torch.no_grad() def test(self, model, data, model_retrain=None, attack_model_all=None, attack_model_sub=None, ckpt='best'): if ckpt == 'best': # Load best ckpt ckpt = torch.load(os.path.join(self.args.checkpoint_dir, 'model_best.pt')) model.load_state_dict(ckpt['model_state']) if 'ogbl' in self.args.dataset: pred_all = False else: pred_all = True loss, dt_acc, dt_f1, test_log = self.eval(model, data, 'test', pred_all) self.trainer_log['dt_loss'] = loss self.trainer_log['dt_acc'] = dt_acc self.trainer_log['dt_f1'] = dt_f1 # self.trainer_log['df_logit'] = df_logit # self.logit_all_pair = logit_all_pair # self.trainer_log['df_auc'] = df_auc # self.trainer_log['df_aup'] = df_aup if model_retrain is not None: # Deletion self.trainer_log['ve'] = verification_error(model, model_retrain).cpu().item() # self.trainer_log['dr_kld'] = output_kldiv(model, model_retrain, data=data).cpu().item() # MI Attack after unlearning if attack_model_all is not None: mi_logit_all_after, mi_sucrate_all_after = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_after'] = mi_logit_all_after self.trainer_log['mi_sucrate_all_after'] = mi_sucrate_all_after if attack_model_sub is not None: mi_logit_sub_after, mi_sucrate_sub_after = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_after'] = mi_logit_sub_after self.trainer_log['mi_sucrate_sub_after'] = mi_sucrate_sub_after self.trainer_log['mi_ratio_all'] = np.mean([i[1] / j[1] for i, j in zip(self.trainer_log['mi_logit_all_after'], self.trainer_log['mi_logit_all_before'])]) self.trainer_log['mi_ratio_sub'] = np.mean([i[1] / j[1] for i, j in zip(self.trainer_log['mi_logit_sub_after'], self.trainer_log['mi_logit_sub_before'])]) print(self.trainer_log['mi_ratio_all'], self.trainer_log['mi_ratio_sub'], self.trainer_log['mi_sucrate_all_after'], self.trainer_log['mi_sucrate_sub_after']) print(self.trainer_log['df_auc'], self.trainer_log['df_aup']) return loss, dt_acc, dt_f1, test_log class GraphTrainer(Trainer): def train(self, model, dataset, split_idx, optimizer, args): self.train_loader = DataLoader(dataset[split_idx["train"]], batch_size=32, shuffle=True) self.valid_loader = DataLoader(dataset[split_idx["valid"]], batch_size=32, shuffle=False) self.test_loader = DataLoader(dataset[split_idx["test"]], batch_size=32, shuffle=False) start_time = time.time() best_epoch = 0 best_valid_auc = 0 for epoch in trange(args.epochs, desc='Epoch'): model.train() for batch in tqdm(self.train_loader, desc="Iteration", leave=False): batch = batch.to(device) pred = model(batch) optimizer.zero_grad() ## ignore nan targets (unlabeled) when computing training loss. is_labeled = batch.y == batch.y loss = F.binary_cross_entropy_with_logits(pred.to(torch.float32)[is_labeled], batch.y.to(torch.float32)[is_labeled]) loss.backward() optimizer.step() if (epoch+1) % args.valid_freq == 0: valid_auc, valid_log = self.eval(model, dataset, 'val') train_log = { 'epoch': epoch, 'train_loss': loss.item() } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(train_log) self.trainer_log['log'].append(valid_log) if valid_auc > best_valid_auc: best_valid_auc = valid_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid auc = {valid_auc:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save models and node embeddings print('Saving final checkpoint') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best valid auc = {best_valid_auc:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_valid_auc'] = best_valid_auc @torch.no_grad() def eval(self, model, data, stage='val', pred_all=False): model.eval() y_true = [] y_pred = [] if stage == 'val': loader = self.valid_loader else: loader = self.test_loader if self.args.eval_on_cpu: model = model.to('cpu') for batch in tqdm(loader): batch = batch.to(device) pred = model(batch) y_true.append(batch.y.view(pred.shape).detach().cpu()) y_pred.append(pred.detach().cpu()) y_true = torch.cat(y_true, dim = 0).numpy() y_pred = torch.cat(y_pred, dim = 0).numpy() evaluator = Evaluator('ogbg-molhiv') auc = evaluator.eval({"y_true": y_true, "y_pred": y_pred})['rocauc'] log = { f'val_auc': auc, } if self.args.eval_on_cpu: model = model.to(device) return auc, log @torch.no_grad() def test(self, model, data, model_retrain=None, attack_model_all=None, attack_model_sub=None, ckpt='best'): if ckpt == 'best': # Load best ckpt ckpt = torch.load(os.path.join(self.args.checkpoint_dir, 'model_best.pt')) model.load_state_dict(ckpt['model_state']) dt_auc, test_log = self.eval(model, data, 'test') self.trainer_log['dt_auc'] = dt_auc if model_retrain is not None: # Deletion self.trainer_log['ve'] = verification_error(model, model_retrain).cpu().item() # self.trainer_log['dr_kld'] = output_kldiv(model, model_retrain, data=data).cpu().item() return dt_auc, test_log
41,518
41.366327
169
py
GNNDelete
GNNDelete-main/framework/trainer/member_infer.py
import os import json import wandb import numpy as np import torch import torch.nn as nn from tqdm import trange, tqdm from torch_geometric.utils import negative_sampling from sklearn.metrics import accuracy_score, roc_auc_score, average_precision_score, f1_score from .base import Trainer from ..evaluation import * from ..utils import * device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') class MIAttackTrainer(Trainer): '''This code is adapted from https://github.com/iyempissy/rebMIGraph''' def __init__(self, args): self.args = args self.trainer_log = { 'unlearning_model': 'member_infer', 'dataset': args.dataset, 'seed': args.random_seed, 'shadow_log': [], 'attack_log': []} self.logit_all_pair = None with open(os.path.join(self.args.checkpoint_dir, 'training_args.json'), 'w') as f: json.dump(vars(args), f) def train_shadow(self, model, data, optimizer, args): best_valid_loss = 1000000 all_neg = [] # Train shadow model using the test data for epoch in trange(args.epochs, desc='Train shadow model'): model.train() # Positive and negative sample neg_edge_index = negative_sampling( edge_index=data.test_pos_edge_index, num_nodes=data.num_nodes, num_neg_samples=data.test_pos_edge_index.shape[1]) z = model(data.x, data.test_pos_edge_index) logits = model.decode(z, data.test_pos_edge_index, neg_edge_index) label = get_link_labels(data.test_pos_edge_index, neg_edge_index) loss = F.binary_cross_entropy_with_logits(logits, label) loss.backward() optimizer.step() optimizer.zero_grad() all_neg.append(neg_edge_index.cpu()) if (epoch+1) % args.valid_freq == 0: valid_loss, auc, aup, df_logit, logit_all_pair = self.eval_shadow(model, data, 'val') log = { 'shadow_epoch': epoch, 'shadow_train_loss': loss.item(), 'shadow_valid_loss': valid_loss, 'shadow_valid_auc': auc, 'shadow_valid_aup': aup, 'shadow_df_logit': df_logit } wandb.log(log) self.trainer_log['shadow_log'].append(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) if valid_loss < best_valid_loss: best_valid_loss = valid_loss best_epoch = epoch self.trainer_log['shadow_best_epoch'] = best_epoch self.trainer_log['shadow_best_valid_loss'] = best_valid_loss print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'shadow_model_best.pt')) return torch.cat(all_neg, dim=-1) @torch.no_grad() def eval_shadow(self, model, data, stage='val'): model.eval() pos_edge_index = data[f'{stage}_pos_edge_index'] neg_edge_index = data[f'{stage}_neg_edge_index'] z = model(data.x, data.val_pos_edge_index) logits = model.decode(z, pos_edge_index, neg_edge_index).sigmoid() label = self.get_link_labels(pos_edge_index, neg_edge_index) loss = F.binary_cross_entropy_with_logits(logits, label).cpu().item() auc = roc_auc_score(label.cpu(), logits.cpu()) aup = average_precision_score(label.cpu(), logits.cpu()) df_logit = float('nan') logit_all_pair = (z @ z.t()).cpu() log = { f'{stage}_loss': loss, f'{stage}_auc': auc, f'{stage}_aup': aup, f'{stage}_df_logit': df_logit, } wandb.log(log) msg = [f'{i}: {j:.4f}' if isinstance(j, (np.floating, float)) else f'{i}: {j:>4d}' for i, j in log.items()] tqdm.write(' | '.join(msg)) return loss, auc, aup, df_logit, logit_all_pair def train_attack(self, model, train_loader, valid_loader, optimizer, args): loss_fct = nn.CrossEntropyLoss() best_auc = 0 best_epoch = 0 for epoch in trange(50, desc='Train attack model'): model.train() train_loss = 0 for x, y in train_loader: logits = model(x.to(device)) loss = loss_fct(logits, y.to(device)) loss.backward() optimizer.step() optimizer.zero_grad() train_loss += loss.item() valid_loss, valid_acc, valid_auc, valid_f1 = self.eval_attack(model, valid_loader) log = { 'attack_train_loss': train_loss / len(train_loader), 'attack_valid_loss': valid_loss, 'attack_valid_acc': valid_acc, 'attack_valid_auc': valid_auc, 'attack_valid_f1': valid_f1} wandb.log(log) self.trainer_log['attack_log'].append(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) if valid_auc > best_auc: best_auc = valid_auc best_epoch = epoch self.trainer_log['attack_best_auc'] = valid_auc self.trainer_log['attack_best_epoch'] = epoch ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'attack_model_best.pt')) @torch.no_grad() def eval_attack(self, model, eval_loader): loss_fct = nn.CrossEntropyLoss() pred = [] label = [] for x, y in eval_loader: logits = model(x.to(device)) loss = loss_fct(logits, y.to(device)) _, p = torch.max(logits, 1) pred.extend(p.cpu()) label.extend(y) pred = torch.stack(pred) label = torch.stack(label) return loss.item(), accuracy_score(label.numpy(), pred.numpy()), roc_auc_score(label.numpy(), pred.numpy()), f1_score(label.numpy(), pred.numpy(), average='macro') @torch.no_grad() def prepare_attack_training_data(self, model, data, all_neg=None): '''Prepare the training data of attack model (Present vs. Absent) Present edges (label = 1): training data of shadow model (Test pos and neg edges) Absent edges (label = 0): validation data of shadow model (Valid pos and neg edges) ''' z = model(data.x, data.test_pos_edge_index) # Sample same size of neg as pos sample_idx = torch.randperm(all_neg.shape[1])[:data.test_pos_edge_index.shape[1]] neg_subset = all_neg[:, sample_idx] present_edge_index = torch.cat([data.test_pos_edge_index, data.test_neg_edge_index], dim=-1) if 'sub' in self.args.unlearning_model: absent_edge_index = torch.cat([data.val_pos_edge_index, data.val_neg_edge_index], dim=-1) else: #if 'all' in self.args.unlearning_model: absent_edge_index = torch.cat([data.val_pos_edge_index, data.val_neg_edge_index, data.train_pos_edge_index, neg_subset.to(device)], dim=-1) edge_index = torch.cat([present_edge_index, absent_edge_index], dim=-1) feature = torch.cat([z[edge_index[0]], z[edge_index[1]]], dim=-1).cpu() label = get_link_labels(present_edge_index, absent_edge_index).long().cpu() return feature, label
8,132
37.728571
171
py
GNNDelete
GNNDelete-main/framework/trainer/gradient_ascent_with_mp.py
import os import json from tqdm import tqdm, trange import torch import torch.nn.functional as F from torch_geometric.utils import negative_sampling from .base import Trainer from ..evaluation import * from ..utils import * class GradientAscentWithMessagePassingTrainer(Trainer): def __init__(self,): self.trainer_log = {'unlearning_model': 'gradient_ascent_with_mp', 'log': []} def freeze_unused_mask(self, model, edge_to_delete, subgraph, h): gradient_mask = torch.zeros_like(delete_model.operator) edges = subgraph[h] for s, t in edges: if s < t: gradient_mask[s, t] = 1 gradient_mask = gradient_mask.to(device) model.operator.register_hook(lambda grad: grad.mul_(gradient_mask)) def train(self, model_retrain, model, data, optimizer, args): best_loss = 100000 for epoch in trange(args.epochs, desc='Unlerning'): model.train() total_step = 0 total_loss = 0 ## Gradient Ascent neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index[:, data.ga_mask], num_nodes=data.num_nodes, num_neg_samples=data.ga_mask.sum()) # print('data train to unlearn', data.train_pos_edge_index[:, data.ga_mask]) z = model(data.x, data.train_pos_edge_index[:, data.ga_mask]) logits = model.decode(z, data.train_pos_edge_index[:, data.ga_mask]) label = torch.tensor([1], dtype=torch.float, device='cuda') loss_ga = -F.binary_cross_entropy_with_logits(logits, label) ## Message Passing neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index[:, data.mp_mask], num_nodes=data.num_nodes, num_neg_samples=data.mp_mask.sum()) z = model(data.x, data.train_pos_edge_index[:, data.mp_mask]) logits = model.decode(z, data.train_pos_edge_index[:, data.mp_mask]) label = self.get_link_labels(data.train_pos_edge_index[:, data.mp_mask], dtype=torch.float, device='cuda') loss_mp = F.binary_cross_entropy_with_logits(logits, label) loss = loss_ga + loss_mp loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() total_step += 1 total_loss += loss.item() msg = [ f'Epoch: {epoch:>4d}', f'train loss: {total_loss / total_step:.6f}' ] tqdm.write(' | '.join(msg)) valid_loss, auc, aup = self.eval(model, data, 'val') self.trainer_log['log'].append({ 'dt_loss': valid_loss, 'dt_auc': auc, 'dt_aup': aup }) # Eval unlearn loss, auc, aup = self.test(model, data) self.trainer_log['dt_loss'] = loss self.trainer_log['dt_auc'] = auc self.trainer_log['dt_aup'] = aup self.trainer_log['ve'] = verification_error(model, model_retrain).cpu().item() self.trainer_log['dr_kld'] = output_kldiv(model, model_retrain, data=data).cpu().item() embedding = get_node_embedding_data(model, data) logits = model.decode(embedding, data.train_pos_edge_index[:, data.dtrain_mask]).sigmoid().detach().cpu() self.trainer_log['df_score'] = logits[:1].cpu().item() # Save ckpt = { 'model_state': model.state_dict(), 'node_emb': z, 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model.pt')) print(self.trainer_log) with open(os.path.join(args.checkpoint_dir, 'trainer_log.json'), 'w') as f: json.dump(self.trainer_log, f)
3,937
36.865385
118
py
GNNDelete
GNNDelete-main/framework/trainer/retrain.py
import os import time import wandb from tqdm import tqdm, trange import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from torch_geometric.utils import negative_sampling from torch_geometric.loader import GraphSAINTRandomWalkSampler from .base import Trainer, KGTrainer from ..evaluation import * from ..utils import * device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') class RetrainTrainer(Trainer): def freeze_unused_mask(self, model, edge_to_delete, subgraph, h): gradient_mask = torch.zeros_like(delete_model.operator) edges = subgraph[h] for s, t in edges: if s < t: gradient_mask[s, t] = 1 gradient_mask = gradient_mask.to(device) model.operator.register_hook(lambda grad: grad.mul_(gradient_mask)) def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): if 'ogbl' in self.args.dataset: return self.train_minibatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) else: return self.train_fullbatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) def train_fullbatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): model = model.to('cuda') data = data.to('cuda') best_metric = 0 loss_fct = nn.MSELoss() # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before for epoch in trange(args.epochs, desc='Unlearning'): model.train() start_time = time.time() total_step = 0 total_loss = 0 neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index[:, data.dr_mask], num_nodes=data.num_nodes, num_neg_samples=data.dr_mask.sum()) z = model(data.x, data.train_pos_edge_index[:, data.dr_mask]) logits = model.decode(z, data.train_pos_edge_index[:, data.dr_mask], neg_edge_index) label = self.get_link_labels(data.train_pos_edge_index[:, data.dr_mask], neg_edge_index) loss = F.binary_cross_entropy_with_logits(logits, label) loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() total_step += 1 total_loss += loss.item() end_time = time.time() epoch_time = end_time - start_time step_log = { 'Epoch': epoch, 'train_loss': loss.item(), 'train_time': epoch_time } wandb.log(step_log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in step_log.items()] tqdm.write(' | '.join(msg)) if (epoch + 1) % self.args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') valid_log['epoch'] = epoch train_log = { 'epoch': epoch, 'train_loss': loss.item(), 'train_time': epoch_time, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best metric = {best_metric:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_metric'] = best_metric def train_minibatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): start_time = time.time() best_metric = 0 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before data.edge_index = data.train_pos_edge_index loader = GraphSAINTRandomWalkSampler( data, batch_size=args.batch_size, walk_length=2, num_steps=args.num_steps, ) for epoch in trange(args.epochs, desc='Epoch'): model.train() start_time = time.time() epoch_loss = 0 for step, batch in enumerate(tqdm(loader, desc='Step', leave=False)): batch = batch.to(device) # Positive and negative sample train_pos_edge_index = batch.edge_index[:, batch.dr_mask] z = model(batch.x, train_pos_edge_index) neg_edge_index = negative_sampling( edge_index=train_pos_edge_index, num_nodes=z.size(0)) logits = model.decode(z, train_pos_edge_index, neg_edge_index) label = get_link_labels(train_pos_edge_index, neg_edge_index) loss = F.binary_cross_entropy_with_logits(logits, label) loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() step_log = { 'epoch': epoch, 'step': step, 'train_loss': loss.item(), } wandb.log(step_log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in step_log.items()] tqdm.write(' | '.join(msg)) epoch_loss += loss.item() end_time = time.time() epoch_time = end_time - start_time if (epoch+1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') valid_log['epoch'] = epoch train_log = { 'epoch': epoch, 'train_loss': loss.item(), 'train_time': epoch_time, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) torch.save(z, os.path.join(args.checkpoint_dir, 'node_embeddings.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save models and node embeddings print('Saving final checkpoint') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best metric = {best_metric:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_metric'] = best_metric class KGRetrainTrainer(KGTrainer): def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): model = model.to(device) start_time = time.time() best_metric = 0 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before loader = GraphSAINTRandomWalkSampler( data, batch_size=128, walk_length=2, num_steps=args.num_steps, ) for epoch in trange(args.epochs, desc='Epoch'): model.train() epoch_loss = 0 for step, batch in enumerate(tqdm(loader, desc='Step', leave=False)): batch = batch.to(device) # Message passing edge_index = batch.edge_index[:, batch.dr_mask] edge_type = batch.edge_type[batch.dr_mask] z = model(batch.x, edge_index, edge_type) # Positive and negative sample decoding_mask = (edge_type < args.num_edge_type) # Only select directed edges for link prediction decoding_edge_index = edge_index[:, decoding_mask] decoding_edge_type = edge_type[decoding_mask] neg_edge_index = negative_sampling_kg( edge_index=decoding_edge_index, edge_type=decoding_edge_type) pos_logits = model.decode(z, decoding_edge_index, decoding_edge_type) neg_logits = model.decode(z, neg_edge_index, decoding_edge_type) logits = torch.cat([pos_logits, neg_logits], dim=-1) label = get_link_labels(decoding_edge_index, neg_edge_index) # reg_loss = z.pow(2).mean() + model.W.pow(2).mean() loss = F.binary_cross_entropy_with_logits(logits, label)# + 1e-2 * reg_loss loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() log = { 'epoch': epoch, 'step': step, 'train_loss': loss.item(), } wandb.log(log) # msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] # tqdm.write(' | '.join(msg)) epoch_loss += loss.item() if (epoch + 1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': epoch_loss / step } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(train_log) self.trainer_log['log'].append(valid_log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save models and node embeddings print('Saving final checkpoint') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best valid loss = {best_metric:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_metric'] = best_metric self.trainer_log['training_time'] = np.mean([i['epoch_time'] for i in self.trainer_log['log'] if 'epoch_time' in i])
14,611
41.976471
127
py
GNNDelete
GNNDelete-main/framework/trainer/gnndelete_nodeemb.py
import os import copy import time import wandb from tqdm import tqdm, trange import torch import torch.nn as nn from torch_geometric.utils import negative_sampling, k_hop_subgraph from torch_geometric.loader import GraphSAINTRandomWalkSampler from .base import Trainer, KGTrainer, NodeClassificationTrainer from ..evaluation import * from ..utils import * device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') def BoundedKLDMean(logits, truth): return 1 - torch.exp(-F.kl_div(F.log_softmax(logits, -1), truth.softmax(-1), None, None, 'batchmean')) def BoundedKLDSum(logits, truth): return 1 - torch.exp(-F.kl_div(F.log_softmax(logits, -1), truth.softmax(-1), None, None, 'sum')) def CosineDistanceMean(logits, truth): return (1 - F.cosine_similarity(logits, truth)).mean() def CosineDistanceSum(logits, truth): return (1 - F.cosine_similarity(logits, truth)).sum() def centering(K): n = K.shape[0] unit = torch.ones([n, n], device=K.device) I = torch.eye(n, device=K.device) H = I - unit / n return torch.matmul(torch.matmul(H, K), H) def rbf(X, sigma=None): GX = torch.matmul(X, X.T) KX = torch.diag(GX) - GX + (torch.diag(GX) - GX).T if sigma is None: mdist = torch.median(KX[KX != 0]) sigma = math.sqrt(mdist) KX *= - 0.5 / (sigma * sigma) KX = torch.exp(KX) return KX def kernel_HSIC(X, Y, sigma=None): return torch.sum(centering(rbf(X, sigma)) * centering(rbf(Y, sigma))) def linear_HSIC(X, Y): L_X = torch.matmul(X, X.T) L_Y = torch.matmul(Y, Y.T) return torch.sum(centering(L_X) * centering(L_Y)) def LinearCKA(X, Y): hsic = linear_HSIC(X, Y) var1 = torch.sqrt(linear_HSIC(X, X)) var2 = torch.sqrt(linear_HSIC(Y, Y)) return hsic / (var1 * var2) def RBFCKA(X, Y, sigma=None): hsic = kernel_HSIC(X, Y, sigma) var1 = torch.sqrt(kernel_HSIC(X, X, sigma)) var2 = torch.sqrt(kernel_HSIC(Y, Y, sigma)) return hsic / (var1 * var2) def get_loss_fct(name): # if name == 'mse': # loss_fct = nn.MSELoss(reduction='mean') # elif name == 'kld': # loss_fct = BoundedKLDMean # elif name == 'cosine': # loss_fct = CosineDistanceMean if name == 'kld_mean': loss_fct = BoundedKLDMean elif name == 'kld_sum': loss_fct = BoundedKLDSum elif name == 'mse_mean': loss_fct = nn.MSELoss(reduction='mean') elif name == 'mse_sum': loss_fct = nn.MSELoss(reduction='sum') elif name == 'cosine_mean': loss_fct = CosineDistanceMean elif name == 'cosine_sum': loss_fct = CosineDistanceSum elif name == 'linear_cka': loss_fct = LinearCKA elif name == 'rbf_cka': loss_fct = RBFCKA else: raise NotImplementedError return loss_fct class GNNDeleteNodeembTrainer(Trainer): def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): if 'ogbl' in self.args.dataset: return self.train_minibatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) else: return self.train_fullbatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) def train_fullbatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): model = model.to('cuda') data = data.to('cuda') best_metric = 0 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before # All node paris in S_Df without Df ## S_Df 1 hop all pair mask sdf1_all_pair_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) idx = torch.combinations(torch.arange(data.num_nodes)[data.sdf_node_1hop_mask], with_replacement=True).t() sdf1_all_pair_mask[idx[0], idx[1]] = True sdf1_all_pair_mask[idx[1], idx[0]] = True assert sdf1_all_pair_mask.sum().cpu() == data.sdf_node_1hop_mask.sum().cpu() * data.sdf_node_1hop_mask.sum().cpu() ## Remove Df itself sdf1_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]] = False sdf1_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][1], data.train_pos_edge_index[:, data.df_mask][0]] = False ## S_Df 2 hop all pair mask sdf2_all_pair_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) idx = torch.combinations(torch.arange(data.num_nodes)[data.sdf_node_2hop_mask], with_replacement=True).t() sdf2_all_pair_mask[idx[0], idx[1]] = True sdf2_all_pair_mask[idx[1], idx[0]] = True assert sdf2_all_pair_mask.sum().cpu() == data.sdf_node_2hop_mask.sum().cpu() * data.sdf_node_2hop_mask.sum().cpu() ## Remove Df itself sdf2_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]] = False sdf2_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][1], data.train_pos_edge_index[:, data.df_mask][0]] = False ## Lower triangular mask idx = torch.tril_indices(data.num_nodes, data.num_nodes, -1) lower_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) lower_mask[idx[0], idx[1]] = True ## The final mask is the intersection sdf1_all_pair_without_df_mask = sdf1_all_pair_mask & lower_mask sdf2_all_pair_without_df_mask = sdf2_all_pair_mask & lower_mask # print(data.sdf_node_2hop_mask.sum()) # print(sdf_all_pair_mask.nonzero()) # print(data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]) # print('aaaaaaaaaaaa', data.sdf_node_2hop_mask.sum(), a, sdf_all_pair_mask.sum()) # print('aaaaaaaaaaaa', lower_mask.sum()) # print('aaaaaaaaaaaa', sdf_all_pair_without_df_mask.sum()) # print('aaaaaaaaaaaa', data.sdf_node_2hop_mask.sum()) # assert sdf_all_pair_without_df_mask.sum() == \ # data.sdf_node_2hop_mask.sum().cpu() * (data.sdf_node_2hop_mask.sum().cpu() - 1) // 2 - data.df_mask.sum().cpu() # non_df_node_mask = torch.ones(data.x.shape[0], dtype=torch.bool, device=data.x.device) non_df_node_mask[data.directed_df_edge_index.flatten().unique()] = False data.sdf_node_1hop_mask_non_df_mask = data.sdf_node_1hop_mask & non_df_node_mask data.sdf_node_2hop_mask_non_df_mask = data.sdf_node_2hop_mask & non_df_node_mask # Original node embeddings with torch.no_grad(): z1_ori, z2_ori = model.get_original_embeddings(data.x, data.train_pos_edge_index[:, data.dr_mask], return_all_emb=True) loss_fct = get_loss_fct(self.args.loss_fct) neg_edge = neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index, num_nodes=data.num_nodes, num_neg_samples=data.df_mask.sum()) for epoch in trange(args.epochs, desc='Unlerning'): model.train() start_time = time.time() z1, z2 = model(data.x, data.train_pos_edge_index[:, data.sdf_mask], return_all_emb=True) # print('current deletion weight', model.deletion1.deletion_weight.sum(), model.deletion2.deletion_weight.sum()) # print('aaaaaa', z[data.sdf_node_2hop_mask].sum()) # Randomness pos_edge = data.train_pos_edge_index[:, data.df_mask] # neg_edge = torch.randperm(data.num_nodes)[:pos_edge.view(-1).shape[0]].view(2, -1) embed1 = torch.cat([z1[pos_edge[0]], z1[pos_edge[1]]], dim=0) embed1_ori = torch.cat([z1_ori[neg_edge[0]], z1_ori[neg_edge[1]]], dim=0) embed2 = torch.cat([z2[pos_edge[0]], z2[pos_edge[1]]], dim=0) embed2_ori = torch.cat([z2_ori[neg_edge[0]], z2_ori[neg_edge[1]]], dim=0) loss_r1 = loss_fct(embed1, embed1_ori) loss_r2 = loss_fct(embed2, embed2_ori) # Local causality loss_l1 = loss_fct(z1[data.sdf_node_1hop_mask_non_df_mask], z1_ori[data.sdf_node_1hop_mask_non_df_mask]) loss_l2 = loss_fct(z2[data.sdf_node_2hop_mask_non_df_mask], z2_ori[data.sdf_node_2hop_mask_non_df_mask]) # Total loss '''both_all, both_layerwise, only2_layerwise, only2_all, only1''' if self.args.loss_type == 'both_all': loss_l = loss_l1 + loss_l2 loss_r = loss_r1 + loss_r2 #### alpha * loss_r + (1 - alpha) * loss_l loss = self.args.alpha * loss_r + (1 - self.args.alpha) * loss_l #### loss_r + lambda * loss_l # loss = loss_l + self.args.alpha * loss_r loss.backward() optimizer.step() elif self.args.loss_type == 'both_layerwise': #### alpha * loss_r + (1 - alpha) * loss_l loss_l = loss_l1 + loss_l2 loss_r = loss_r1 + loss_r2 loss1 = self.args.alpha * loss_r1 + (1 - self.args.alpha) * loss_l1 loss1.backward(retain_graph=True) optimizer[0].step() optimizer[0].zero_grad() loss2 = self.args.alpha * loss_r2 + (1 - self.args.alpha) * loss_l2 loss2.backward(retain_graph=True) optimizer[1].step() optimizer[1].zero_grad() loss = loss1 + loss2 #### loss_r + lambda * loss_l # loss_l = loss_l1 + loss_l2 # loss_r = loss_r1 + loss_r2 # loss1 = loss_r1 + self.args.alpha * loss_l1 # loss1.backward(retain_graph=True) # optimizer[0].step() # optimizer[0].zero_grad() # loss2 = loss_r2 + self.args.alpha * loss_l2 # loss2.backward() # optimizer[1].step() # optimizer[1].zero_grad() # loss = loss1 + loss2 elif self.args.loss_type == 'only2_layerwise': loss_l = loss_l1 + loss_l2 loss_r = loss_r1 + loss_r2 optimizer[0].zero_grad() #### alpha * loss_r + (1 - alpha) * loss_l loss2 = self.args.alpha * loss_r2 + (1 - self.args.alpha) * loss_l2 #### loss_r + lambda * loss_l # loss2 = loss_r2 + self.args.alpha * loss_l2 loss2.backward() optimizer[1].step() optimizer[1].zero_grad() loss = loss2 elif self.args.loss_type == 'only2_all': loss_l = loss_l2 loss_r = loss_r2 loss = loss_l + self.args.alpha * loss_r loss.backward() optimizer.step() optimizer.zero_grad() elif self.args.loss_type == 'only1': loss_l = loss_l1 loss_r = loss_r1 loss = loss_l + self.args.alpha * loss_r loss.backward() optimizer.step() optimizer.zero_grad() else: raise NotImplementedError end_time = time.time() epoch_time = end_time - start_time step_log = { 'Epoch': epoch, 'train_loss': loss.item(), 'loss_r': loss_r.item(), 'loss_l': loss_l.item(), 'train_time': epoch_time } wandb.log(step_log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in step_log.items()] tqdm.write(' | '.join(msg)) if (epoch + 1) % self.args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') valid_log['epoch'] = epoch train_log = { 'epoch': epoch, 'train_loss': loss.item(), 'loss_r': loss_r.item(), 'loss_l': loss_l.item(), 'train_time': epoch_time, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), # 'optimizer_state': [optimizer[0].state_dict(), optimizer[1].state_dict()], } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, # 'optimizer_state': [optimizer[0].state_dict(), optimizer[1].state_dict()], } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) def train_minibatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): best_metric = 0 if 'kld' in args.unlearning_model: loss_fct = BoundedKLD else: loss_fct = nn.MSELoss() # neg_size = 10 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before non_df_node_mask = torch.ones(data.x.shape[0], dtype=torch.bool, device=data.x.device) non_df_node_mask[data.directed_df_edge_index.flatten().unique()] = False data.sdf_node_1hop_mask_non_df_mask = data.sdf_node_1hop_mask & non_df_node_mask data.sdf_node_2hop_mask_non_df_mask = data.sdf_node_2hop_mask & non_df_node_mask data.edge_index = data.train_pos_edge_index data.node_id = torch.arange(data.x.shape[0]) loader = GraphSAINTRandomWalkSampler( data, batch_size=args.batch_size, walk_length=2, num_steps=args.num_steps, ) # all_neg_edge = negative_sampling( # edge_index=data.train_pos_edge_index, # num_nodes=data.num_nodes, # num_neg_samples=100000 # ) for epoch in trange(args.epochs, desc='Unlerning'): model.train() epoch_loss_l = 0 epoch_loss_r = 0 epoch_loss = 0 epoch_time = 0 for step, batch in enumerate(tqdm(loader, leave=False)): batch = batch.to(device) start_time = time.time() # Original embedding with torch.no_grad(): z1_ori, z2_ori = model.get_original_embeddings(batch.x, batch.edge_index, return_all_emb=True) # z1_ori = z1_ori[batch.sdf_node_2hop_mask] # z2_ori = z2_ori[batch.sdf_node_2hop_mask] z1, z2 = model(batch.x, batch.edge_index[:, batch.sdf_mask], batch.sdf_node_1hop_mask, batch.sdf_node_2hop_mask, return_all_emb=True) # Randomness pos_edge = batch.edge_index[:, batch.df_mask] neg_edge = negative_sampling( edge_index=batch.edge_index, num_nodes=batch.x.shape[0], num_neg_samples=pos_edge.shape[1] ) # neg_edge = all_neg_edge[:, :pos_edge.shape[1]] embed1 = torch.cat([z1[pos_edge[0]], z1[pos_edge[1]]], dim=0) embed1_ori = torch.cat([z1_ori[neg_edge[0]], z1_ori[neg_edge[1]]], dim=0) embed2 = torch.cat([z2[pos_edge[0]], z2[pos_edge[1]]], dim=0) embed2_ori = torch.cat([z2_ori[neg_edge[0]], z2_ori[neg_edge[1]]], dim=0) loss_r1 = loss_fct(embed1, embed1_ori) loss_r2 = loss_fct(embed2, embed2_ori) # Local causality loss_l1 = loss_fct(z1[batch.sdf_node_1hop_mask_non_df_mask], z1_ori[batch.sdf_node_1hop_mask_non_df_mask]) loss_l2 = loss_fct(z2[batch.sdf_node_2hop_mask_non_df_mask], z2_ori[batch.sdf_node_2hop_mask_non_df_mask]) # Total loss loss_l = loss_l1 + loss_l2 loss_r = loss_r1 + loss_r2 loss1 = self.args.alpha * loss_r1 + (1 - self.args.alpha) * loss_l1 loss1.backward(retain_graph=True) optimizer[0].step() optimizer[0].zero_grad() loss2 = self.args.alpha * loss_r2 + (1 - self.args.alpha) * loss_l2 loss2.backward(retain_graph=True) optimizer[1].step() optimizer[1].zero_grad() loss = loss1 + loss2 end_time = time.time() epoch_loss_l += loss_l.item() epoch_loss_r += loss_r.item() epoch_loss += loss.item() epoch_time += end_time - start_time step_log = { 'Epoch': epoch, 'train_loss': loss.item(), 'train_loss_l': loss_l.item(), 'train_loss_r': loss_r.item(), 'train_time': end_time - start_time } wandb.log(step_log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in step_log.items()] tqdm.write(' | '.join(msg)) if (epoch+1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': epoch_loss / step, 'train_loss_l': epoch_loss_l / step, 'train_loss_r': epoch_loss_r / step, 'train_time': epoch_time / step, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), # 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, # 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) class GNNDeleteNodeClassificationTrainer(NodeClassificationTrainer): def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): model = model.to('cuda') data = data.to('cuda') best_metric = 0 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before # All node paris in S_Df without Df ## S_Df 1 hop all pair mask # sdf1_all_pair_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) # idx = torch.combinations(torch.arange(data.num_nodes)[data.sdf_node_1hop_mask], with_replacement=True).t() # sdf1_all_pair_mask[idx[0], idx[1]] = True # sdf1_all_pair_mask[idx[1], idx[0]] = True # assert sdf1_all_pair_mask.sum().cpu() == data.sdf_node_1hop_mask.sum().cpu() * data.sdf_node_1hop_mask.sum().cpu() # ## Remove Df itself # sdf1_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]] = False # sdf1_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][1], data.train_pos_edge_index[:, data.df_mask][0]] = False # ## S_Df 2 hop all pair mask # sdf2_all_pair_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) # idx = torch.combinations(torch.arange(data.num_nodes)[data.sdf_node_2hop_mask], with_replacement=True).t() # sdf2_all_pair_mask[idx[0], idx[1]] = True # sdf2_all_pair_mask[idx[1], idx[0]] = True # assert sdf2_all_pair_mask.sum().cpu() == data.sdf_node_2hop_mask.sum().cpu() * data.sdf_node_2hop_mask.sum().cpu() # ## Remove Df itself # sdf2_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]] = False # sdf2_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][1], data.train_pos_edge_index[:, data.df_mask][0]] = False # ## Lower triangular mask # idx = torch.tril_indices(data.num_nodes, data.num_nodes, -1) # lower_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) # lower_mask[idx[0], idx[1]] = True # ## The final mask is the intersection # sdf1_all_pair_without_df_mask = sdf1_all_pair_mask & lower_mask # sdf2_all_pair_without_df_mask = sdf2_all_pair_mask & lower_mask non_df_node_mask = torch.ones(data.x.shape[0], dtype=torch.bool, device=data.x.device) non_df_node_mask[data.directed_df_edge_index.flatten().unique()] = False data.sdf_node_1hop_mask_non_df_mask = data.sdf_node_1hop_mask & non_df_node_mask data.sdf_node_2hop_mask_non_df_mask = data.sdf_node_2hop_mask & non_df_node_mask # Original node embeddings with torch.no_grad(): z1_ori, z2_ori = model.get_original_embeddings(data.x, data.edge_index[:, data.dr_mask], return_all_emb=True) loss_fct = get_loss_fct(self.args.loss_fct) neg_edge = neg_edge_index = negative_sampling( edge_index=data.edge_index, num_nodes=data.num_nodes, num_neg_samples=data.df_mask.sum()) for epoch in trange(args.epochs, desc='Unlerning'): model.train() start_time = time.time() z1, z2 = model(data.x, data.edge_index[:, data.sdf_mask], return_all_emb=True) # print('current deletion weight', model.deletion1.deletion_weight.sum(), model.deletion2.deletion_weight.sum()) # print('aaaaaa', z[data.sdf_node_2hop_mask].sum()) # Randomness pos_edge = data.edge_index[:, data.df_mask] # neg_edge = torch.randperm(data.num_nodes)[:pos_edge.view(-1).shape[0]].view(2, -1) embed1 = torch.cat([z1[pos_edge[0]], z1[pos_edge[1]]], dim=0) embed1_ori = torch.cat([z1_ori[neg_edge[0]], z1_ori[neg_edge[1]]], dim=0) embed2 = torch.cat([z2[pos_edge[0]], z2[pos_edge[1]]], dim=0) embed2_ori = torch.cat([z2_ori[neg_edge[0]], z2_ori[neg_edge[1]]], dim=0) loss_r1 = loss_fct(embed1, embed1_ori) loss_r2 = loss_fct(embed2, embed2_ori) # Local causality loss_l1 = loss_fct(z1[data.sdf_node_1hop_mask_non_df_mask], z1_ori[data.sdf_node_1hop_mask_non_df_mask]) loss_l2 = loss_fct(z2[data.sdf_node_2hop_mask_non_df_mask], z2_ori[data.sdf_node_2hop_mask_non_df_mask]) # Total loss '''both_all, both_layerwise, only2_layerwise, only2_all, only1''' loss_l = loss_l1 + loss_l2 loss_r = loss_r1 + loss_r2 loss1 = self.args.alpha * loss_r1 + (1 - self.args.alpha) * loss_l1 loss1.backward(retain_graph=True) optimizer[0].step() optimizer[0].zero_grad() loss2 = self.args.alpha * loss_r2 + (1 - self.args.alpha) * loss_l2 loss2.backward(retain_graph=True) optimizer[1].step() optimizer[1].zero_grad() loss = loss1 + loss2 end_time = time.time() epoch_time = end_time - start_time step_log = { 'Epoch': epoch, 'train_loss': loss.item(), 'loss_r': loss_r.item(), 'loss_l': loss_l.item(), 'train_time': epoch_time } wandb.log(step_log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in step_log.items()] tqdm.write(' | '.join(msg)) if (epoch + 1) % self.args.valid_freq == 0: valid_loss, dt_acc, dt_f1, valid_log = self.eval(model, data, 'val') valid_log['epoch'] = epoch train_log = { 'epoch': epoch, 'train_loss': loss.item(), 'loss_r': loss_r.item(), 'loss_l': loss_l.item(), 'train_time': epoch_time, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_acc + dt_f1 > best_metric: best_metric = dt_acc + dt_f1 best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), # 'optimizer_state': [optimizer[0].state_dict(), optimizer[1].state_dict()], } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, # 'optimizer_state': [optimizer[0].state_dict(), optimizer[1].state_dict()], } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) class KGGNNDeleteNodeembTrainer(KGTrainer): def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): best_metric = 0 if 'kld' in args.unlearning_model: loss_fct = BoundedKLD else: loss_fct = nn.MSELoss() # neg_size = 10 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before # All node paris in S_Df without Df ## S_Df 1 hop all pair mask # sdf1_all_pair_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) # idx = torch.combinations(torch.arange(data.num_nodes)[data.sdf_node_1hop_mask], with_replacement=True).t() # sdf1_all_pair_mask[idx[0], idx[1]] = True # sdf1_all_pair_mask[idx[1], idx[0]] = True # assert sdf1_all_pair_mask.sum().cpu() == data.sdf_node_1hop_mask.sum().cpu() * data.sdf_node_1hop_mask.sum().cpu() # ## Remove Df itself # sdf1_all_pair_mask[data.edge_index[:, data.df_mask][0], data.edge_index[:, data.df_mask][1]] = False # sdf1_all_pair_mask[data.edge_index[:, data.df_mask][1], data.edge_index[:, data.df_mask][0]] = False # ## S_Df 2 hop all pair mask # sdf2_all_pair_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) # idx = torch.combinations(torch.arange(data.num_nodes)[data.sdf_node_2hop_mask], with_replacement=True).t() # sdf2_all_pair_mask[idx[0], idx[1]] = True # sdf2_all_pair_mask[idx[1], idx[0]] = True # assert sdf2_all_pair_mask.sum().cpu() == data.sdf_node_2hop_mask.sum().cpu() * data.sdf_node_2hop_mask.sum().cpu() # ## Remove Df itself # sdf2_all_pair_mask[data.edge_index[:, data.df_mask][0], data.edge_index[:, data.df_mask][1]] = False # sdf2_all_pair_mask[data.edge_index[:, data.df_mask][1], data.edge_index[:, data.df_mask][0]] = False # ## Lower triangular mask # idx = torch.tril_indices(data.num_nodes, data.num_nodes, -1) # lower_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) # lower_mask[idx[0], idx[1]] = True # ## The final mask is the intersection # sdf1_all_pair_without_df_mask = sdf1_all_pair_mask & lower_mask # sdf2_all_pair_without_df_mask = sdf2_all_pair_mask & lower_mask # print(data.sdf_node_2hop_mask.sum()) # print(sdf_all_pair_mask.nonzero()) # print(data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]) # print('aaaaaaaaaaaa', data.sdf_node_2hop_mask.sum(), a, sdf_all_pair_mask.sum()) # print('aaaaaaaaaaaa', lower_mask.sum()) # print('aaaaaaaaaaaa', sdf_all_pair_without_df_mask.sum()) # print('aaaaaaaaaaaa', data.sdf_node_2hop_mask.sum()) # assert sdf_all_pair_without_df_mask.sum() == \ # data.sdf_node_2hop_mask.sum().cpu() * (data.sdf_node_2hop_mask.sum().cpu() - 1) // 2 - data.df_mask.sum().cpu() # non_df_node_mask = torch.ones(data.x.shape[0], dtype=torch.bool, device=data.x.device) non_df_node_mask[data.directed_df_edge_index.flatten().unique()] = False data.sdf_node_1hop_mask_non_df_mask = data.sdf_node_1hop_mask & non_df_node_mask data.sdf_node_2hop_mask_non_df_mask = data.sdf_node_2hop_mask & non_df_node_mask model_ori = copy.deepcopy(model) loss_fct = get_loss_fct(self.args.loss_fct) loader = GraphSAINTRandomWalkSampler( data, batch_size=args.batch_size, walk_length=2, num_steps=args.num_steps, ) for epoch in trange(args.epochs, desc='Unlerning'): model.train() epoch_loss_e = 0 epoch_loss_l = 0 epoch_loss = 0 epoch_time = 0 for step, batch in enumerate(tqdm(loader, leave=False)): start_time = time.time() batch = batch.to(device) # Message passing edge_index = batch.edge_index[:, batch.dr_mask] edge_type = batch.edge_type[batch.dr_mask] z1, z2 = model(batch.x, edge_index, edge_type, batch.sdf_node_1hop_mask_non_df_mask, batch.sdf_node_2hop_mask_non_df_mask, return_all_emb=True) # Original node embeddings with torch.no_grad(): z1_ori, z2_ori = model.get_original_embeddings(batch.x, edge_index, edge_type, return_all_emb=True) # Randomness pos_edge_index = batch.edge_index[:, batch.df_mask] pos_edge_type = batch.edge_type[batch.df_mask] decoding_mask = pos_edge_type < self.args.num_edge_type decoding_edge_index = pos_edge_index[:, decoding_mask] decoding_edge_type = pos_edge_type[decoding_mask] # print(pos_edge_type.max(), decoding_edge_type.max(), self.args.num_edge_type) # raise neg_edge_index = negative_sampling_kg( edge_index=decoding_edge_index, edge_type=decoding_edge_type) embed1 = torch.cat([z1[decoding_edge_index[0]], z1[decoding_edge_index[1]]], dim=0) embed1_ori = torch.cat([z1_ori[neg_edge_index[0]], z1_ori[neg_edge_index[1]]], dim=0) embed2 = torch.cat([z2[decoding_edge_index[0]], z2[decoding_edge_index[1]]], dim=0) embed2_ori = torch.cat([z2_ori[neg_edge_index[0]], z2_ori[neg_edge_index[1]]], dim=0) loss_r1 = loss_fct(embed1, embed1_ori) loss_r2 = loss_fct(embed2, embed2_ori) # Local causality loss_l1 = loss_fct(z1[batch.sdf_node_1hop_mask_non_df_mask], z1_ori[batch.sdf_node_1hop_mask_non_df_mask]) loss_l2 = loss_fct(z2[batch.sdf_node_2hop_mask_non_df_mask], z2_ori[batch.sdf_node_2hop_mask_non_df_mask]) # Loss loss_l = loss_l1 + loss_l2 loss_r = loss_r1 + loss_r2 loss1 = self.args.alpha * loss_r1 + (1 - self.args.alpha) * loss_l1 loss1.backward(retain_graph=True) optimizer[0].step() optimizer[0].zero_grad() loss2 = self.args.alpha * loss_r2 + (1 - self.args.alpha) * loss_l2 loss2.backward(retain_graph=True) optimizer[1].step() optimizer[1].zero_grad() loss = loss1 + loss2 end_time = time.time() epoch_time = end_time - start_time step_log = { 'Epoch': epoch, 'train_loss': loss.item(), 'loss_r': loss_r.item(), 'loss_l': loss_l.item(), 'train_time': epoch_time } wandb.log(step_log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in step_log.items()] tqdm.write(' | '.join(msg)) if (epoch + 1) % self.args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') valid_log['epoch'] = epoch train_log = { 'epoch': epoch, 'train_loss': loss.item(), 'loss_r': loss_r.item(), 'loss_l': loss_l.item(), 'train_time': epoch_time, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt'))
37,356
43.105077
159
py
GNNDelete
GNNDelete-main/framework/trainer/gnndelete.py
import os import time import wandb from tqdm import tqdm, trange import torch import torch.nn as nn from torch_geometric.utils import negative_sampling, k_hop_subgraph from torch_geometric.loader import GraphSAINTRandomWalkSampler from .base import Trainer from ..evaluation import * from ..utils import * def BoundedKLD(logits, truth): # print('aaaaaaaaa', truth.shape, truth) return 1 - torch.exp(-F.kl_div(F.log_softmax(logits, -1), truth.softmax(-1), None, None, 'batchmean')) def CosineDistance(logits, truth): if len(logits.shape) == 1: return 1 - F.cosine_similarity(logits.view(1, -1), truth.view(1, -1)) else: return 1 - F.cosine_similarity(logits, truth) def get_loss_fct(name): if name == 'kld': loss_fct = BoundedKLD elif name == 'mse': loss_fct = nn.MSELoss() elif name == 'cosine': loss_fct = CosineDistance else: raise NotImplementedError return loss_fct class GNNDeleteTrainer(Trainer): def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): if 'ogbl' in self.args.dataset: return self.train_minibatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) else: return self.train_fullbatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) def compute_loss(self, model, data, random_loss_fct, compute_random_on, random_layer, local_loss_fct, compute_local_on, local_layer, z1=None, z2=None, z1_ori=None, z2_ori=None, logits_ori=None, sdf1_all_pair_without_df_mask=None, sdf2_all_pair_without_df_mask=None): # Randomness loss_r = 0 if random_layer == '1': all_z = [z1] elif random_layer == '2': all_z = [z2] elif random_layer == 'both': all_z = [z1, z2] else: raise NotImplementedError neg_size = data.df_mask.sum() neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index, num_nodes=data.num_nodes, num_neg_samples=neg_size) if compute_random_on == 'edgeprob': # Compute Randomness on edge probability for z in all_z: df_logits = model.decode(z, data.train_pos_edge_index[:, data.df_mask], neg_edge_index) loss_r += random_loss_fct(df_logits[:neg_size], df_logits[neg_size:]) elif compute_random_on == 'nodeemb': for z in all_z: z_random_source, z_random_target = z[neg_edge_index[0]], z[neg_edge_index[1]] z_source, z_target = z[data.train_pos_edge_index[:, data.df_mask][0]], z[data.train_pos_edge_index[:, data.df_mask][1]] loss_r += (random_loss_fct(z_source, z_random_source) + random_loss_fct(z_target, z_random_target)) elif compute_random_on == 'none': loss_r = None else: raise NotImplementedError # Local causality loss_l = 0 if local_layer == '1': all_z = [z1] all_z_ori = [z1_ori] all_sdf_lower_triangular_mask = [sdf1_all_pair_without_df_mask] all_sdf_node_mask = [data.sdf_node_1hop_mask] elif local_layer == '2': all_z = [z2] all_z_ori = [z2_ori] all_sdf_lower_triangular_mask = [sdf2_all_pair_without_df_mask] all_sdf_node_mask = [data.sdf_node_2hop_mask] elif local_layer == 'both': all_z = [z1, z2] all_z_ori = [z1_ori, z2_ori] all_sdf_lower_triangular_mask = [sdf1_all_pair_without_df_mask, sdf2_all_pair_without_df_mask] all_sdf_node_mask = [data.sdf_node_1hop_mask, data.sdf_node_2hop_mask] else: raise NotImplementedError if compute_local_on == 'edgeprob': for z_ori, z, sdf_lower_triangular_mask in zip(all_z_ori, all_z, all_sdf_lower_triangular_mask): logits = (z @ z.t())[sdf_lower_triangular_mask].sigmoid() logits_ori = (z_ori @ z_ori.t())[sdf_lower_triangular_mask].sigmoid() loss_l += local_loss_fct(logits, logits_ori) elif compute_local_on == 'nodeemb': for z_ori, z, sdf_node_mask in zip(all_z_ori, all_z, all_sdf_node_mask): print(z_ori.shape, z.shape, sdf_node_mask.shape, sdf_node_mask.sum()) loss_l += local_loss_fct(z_ori[sdf_node_mask], z[sdf_node_mask]) elif compute_local_on == 'none': loss_l = None else: raise NotImplementedError if compute_random_on == 'none': loss = loss_l elif compute_local_on == 'none': loss = loss_r else: alpha = 0.5 loss = alpha * loss_r + (1 - alpha) * loss_l return loss, loss_r, loss_l def train_fullbatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): model = model.to('cuda') data = data.to('cuda') best_metric = 0 # '''Model naming convention: "gnndelete_random_mse_edgeprob_1_local_mse_edgeprob_1" ''' # _, _, random_loss_fct, compute_random_on, random_layer, _, local_loss_fct, compute_local_on, local_layer = self.args.unlearning_model.split('_') # random_loss_fct = get_loss_fct(random_loss_fct) # local_loss_fct = get_loss_fct(local_loss_fct) # neg_size = 10 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before # All node paris in S_Df without Df ## S_Df 1 hop all pair mask sdf1_all_pair_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) idx = torch.combinations(torch.arange(data.num_nodes)[data.sdf_node_1hop_mask], with_replacement=True).t() sdf1_all_pair_mask[idx[0], idx[1]] = True sdf1_all_pair_mask[idx[1], idx[0]] = True assert sdf1_all_pair_mask.sum().cpu() == data.sdf_node_1hop_mask.sum().cpu() * data.sdf_node_1hop_mask.sum().cpu() ## Remove Df itself sdf1_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]] = False sdf1_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][1], data.train_pos_edge_index[:, data.df_mask][0]] = False ## S_Df 2 hop all pair mask sdf2_all_pair_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) idx = torch.combinations(torch.arange(data.num_nodes)[data.sdf_node_2hop_mask], with_replacement=True).t() sdf2_all_pair_mask[idx[0], idx[1]] = True sdf2_all_pair_mask[idx[1], idx[0]] = True assert sdf2_all_pair_mask.sum().cpu() == data.sdf_node_2hop_mask.sum().cpu() * data.sdf_node_2hop_mask.sum().cpu() ## Remove Df itself sdf2_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]] = False sdf2_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][1], data.train_pos_edge_index[:, data.df_mask][0]] = False ## Lower triangular mask idx = torch.tril_indices(data.num_nodes, data.num_nodes, -1) lower_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) lower_mask[idx[0], idx[1]] = True ## The final mask is the intersection sdf1_all_pair_without_df_mask = sdf1_all_pair_mask & lower_mask sdf2_all_pair_without_df_mask = sdf2_all_pair_mask & lower_mask # print(data.sdf_node_2hop_mask.sum()) # print(sdf_all_pair_mask.nonzero()) # print(data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]) # print('aaaaaaaaaaaa', data.sdf_node_2hop_mask.sum(), a, sdf_all_pair_mask.sum()) # print('aaaaaaaaaaaa', lower_mask.sum()) # print('aaaaaaaaaaaa', sdf_all_pair_without_df_mask.sum()) # print('aaaaaaaaaaaa', data.sdf_node_2hop_mask.sum()) # assert sdf_all_pair_without_df_mask.sum() == \ # data.sdf_node_2hop_mask.sum().cpu() * (data.sdf_node_2hop_mask.sum().cpu() - 1) // 2 - data.df_mask.sum().cpu() # Original node embeddings # with torch.no_grad(): # z1_ori, z2_ori = model.get_original_embeddings(data.x, data.train_pos_edge_index[:, data.dtrain_mask], return_all_emb=True) loss_fct = nn.MSELoss() for epoch in trange(args.epochs, desc='Unlerning'): model.train() start_time = time.time() z = model(data.x, data.train_pos_edge_index[:, data.sdf_mask]) # z1, z2 = model(data.x, data.train_pos_edge_index[:, data.sdf_mask], return_all_emb=True) # print('current deletion weight', model.deletion1.deletion_weight.sum(), model.deletion2.deletion_weight.sum()) # print('aaaaaa', z[data.sdf_node_2hop_mask].sum()) # Effectiveness and Randomness neg_size = data.df_mask.sum() neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index, num_nodes=data.num_nodes, num_neg_samples=neg_size) df_logits = model.decode(z, data.train_pos_edge_index[:, data.df_mask], neg_edge_index) loss_r = loss_fct(df_logits[:neg_size], df_logits[neg_size:]) # df_logits = model.decode( # z, # data.train_pos_edge_index[:, data.df_mask].repeat(1, neg_size), # neg_edge_index).sigmoid() # loss_e = loss_fct(df_logits[:neg_size], df_logits[neg_size:]) # print('df_logits', df_logits) # raise # Local causality if sdf2_all_pair_without_df_mask.sum() != 0: logits_sdf = (z @ z.t())[sdf2_all_pair_without_df_mask].sigmoid() loss_l = loss_fct(logits_sdf, logits_ori[sdf2_all_pair_without_df_mask].sigmoid()) # print('local proba', logits_sdf.shape, logits_sdf, logits_ori[sdf2_all_pair_without_df_mask].sigmoid()) else: loss_l = torch.tensor(0) print('local proba', 0) alpha = 0.5 loss = alpha * loss_r + (1 - alpha) * loss_l # loss, loss_r, loss_l = self.compute_loss( # model, data, random_loss_fct, compute_random_on, random_layer, local_loss_fct, compute_local_on, local_layer, # z1, z2, z1_ori, z2_ori, logits_ori, sdf1_all_pair_without_df_mask, sdf2_all_pair_without_df_mask) loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() end_time = time.time() epoch_time = end_time - start_time step_log = { 'Epoch': epoch, 'train_loss': loss.item(), 'loss_r': loss_r.item(), 'loss_l': loss_l.item(), 'train_time': epoch_time } wandb.log(step_log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in step_log.items()] tqdm.write(' | '.join(msg)) if (epoch + 1) % self.args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') valid_log['epoch'] = epoch train_log = { 'epoch': epoch, 'train_loss': loss.item(), 'train_loss_l': loss_l.item(), 'train_loss_r': loss_r.item(), 'train_time': epoch_time, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) def train_minibatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): best_metric = 0 if 'kld' in args.unlearning_model: loss_fct = BoundedKLD else: loss_fct = nn.MSELoss() # neg_size = 10 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before z_ori = self.get_embedding(model, data, on_cpu=True) z_ori_two_hop = z_ori[data.sdf_node_2hop_mask] data.edge_index = data.train_pos_edge_index data.node_id = torch.arange(data.x.shape[0]) loader = GraphSAINTRandomWalkSampler( data, batch_size=args.batch_size, walk_length=2, num_steps=args.num_steps, ) for epoch in trange(args.epochs, desc='Unlerning'): model.train() # print('current deletion weight', model.deletion1.deletion_weight.sum(), model.deletion2.deletion_weight.sum()) epoch_loss_e = 0 epoch_loss_l = 0 epoch_loss = 0 epoch_time = 0 for step, batch in enumerate(tqdm(loader, leave=False)): start_time = time.time() batch = batch.to('cuda') train_pos_edge_index = batch.edge_index z = model(batch.x, train_pos_edge_index[:, batch.sdf_mask], batch.sdf_node_1hop_mask, batch.sdf_node_2hop_mask) z_two_hop = z[batch.sdf_node_2hop_mask] # Effectiveness and Randomness neg_size = batch.df_mask.sum() neg_edge_index = negative_sampling( edge_index=train_pos_edge_index, num_nodes=z.size(0), num_neg_samples=neg_size) df_logits = model.decode(z, train_pos_edge_index[:, batch.df_mask], neg_edge_index) loss_e = loss_fct(df_logits[:neg_size], df_logits[neg_size:]) # Local causality # Only take the lower triangular part # mask = torch.zeros(data.x.shape[0], dtype=torch.bool) # mask[batch.node_id[batch.sdf_node_2hop_mask]] = True # z_ori_subset = z_ori[mask].to('cuda') # num_nodes = z_ori_subset.shape[0] # idx = torch.tril_indices(num_nodes, num_nodes, -1) # local_lower_mask = torch.zeros(num_nodes, num_nodes, dtype=torch.bool) # local_lower_mask[idx[0], idx[1]] = True # logits_ori = (z_ori_subset @ z_ori_subset.t())[local_lower_mask]#.sigmoid() # logits = (z_two_hop @ z_two_hop.t())[local_lower_mask]#.sigmoid() edge = batch.edge_index[:, batch.sdf_mask] lower_mask = edge[0] < edge[1] row, col = edge[0][lower_mask], edge[1][lower_mask] logits_ori = (z_ori[row] * z_ori[col]).sum(dim=-1).to('cuda') logits = (z[row] * z[col]).sum(dim=-1) loss_l = loss_fct(logits, logits_ori) # print(loss_e, loss_l, z_ori.device, z.device) alpha = 0.5 if 'ablation_random' in self.args.unlearning_model: loss_l = torch.tensor(0) loss = loss_e elif 'ablation_locality' in self.args.unlearning_model: loss_e = torch.tensor(0) loss = loss_l else: loss = alpha * loss_e + (1 - alpha) * loss_l loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() end_time = time.time() epoch_loss_e += loss_e.item() epoch_loss_l += loss_l.item() epoch_loss += loss.item() epoch_time += end_time - start_time epoch_loss_e /= step epoch_loss_l /= step epoch_loss /= step epoch_time /= step if (epoch+1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': epoch_loss / step, 'train_loss_l': epoch_loss_e / step, 'train_loss_e': epoch_loss_l / step, 'train_time': epoch_time / step, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) torch.save(z, os.path.join(args.checkpoint_dir, 'node_embeddings.pt')) # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt'))
19,850
43.015521
154
py
GNNDelete
GNNDelete-main/framework/trainer/gradient_ascent.py
import os import time import wandb from tqdm import tqdm, trange import torch import torch.nn as nn import torch.nn.functional as F from torch_geometric.utils import negative_sampling from torch_geometric.loader import GraphSAINTRandomWalkSampler from .base import Trainer, KGTrainer from ..evaluation import * from ..utils import * device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') def weight(model): t = 0 for p in model.parameters(): t += torch.norm(p) return t class GradientAscentTrainer(Trainer): def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): if 'ogbl' in self.args.dataset: return self.train_minibatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) else: return self.train_fullbatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) def train_fullbatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): model = model.to('cuda') data = data.to('cuda') start_time = time.time() best_metric = 0 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before for epoch in trange(args.epochs, desc='Unlerning'): model.train() start_time = time.time() # Positive and negative sample neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index[:, data.df_mask], num_nodes=data.num_nodes, num_neg_samples=data.df_mask.sum()) z = model(data.x, data.train_pos_edge_index) logits = model.decode(z, data.train_pos_edge_index[:, data.df_mask]) label = torch.ones_like(logits, dtype=torch.float, device='cuda') loss = -F.binary_cross_entropy_with_logits(logits, label) # print('aaaaaaaaaaaaaa', data.df_mask.sum(), weight(model)) loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() end_time = time.time() epoch_time = end_time - start_time step_log = { 'Epoch': epoch, 'train_loss': loss.item(), 'train_time': epoch_time } wandb.log(step_log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in step_log.items()] tqdm.write(' | '.join(msg)) if (epoch + 1) % self.args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') valid_log['epoch'] = epoch train_log = { 'epoch': epoch, 'train_loss': loss.item(), 'train_time': epoch_time, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save ckpt = { 'model_state': {k: v.cpu() for k, v in model.state_dict().items()}, 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) def train_minibatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): best_metric = 0 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before data.edge_index = data.train_pos_edge_index data.node_id = torch.arange(data.x.shape[0]) loader = GraphSAINTRandomWalkSampler( data, batch_size=args.batch_size, walk_length=2, num_steps=args.num_steps, ) for epoch in trange(args.epochs, desc='Unlerning'): model.train() epoch_loss = 0 epoch_time = 0 for step, batch in enumerate(tqdm(loader, leave=False)): start_time = time.time() batch = batch.to(device) z = model(batch.x, batch.edge_index[:, batch.dr_mask]) # Positive and negative sample neg_edge_index = negative_sampling( edge_index=batch.edge_index[:, batch.df_mask], num_nodes=z.size(0)) logits = model.decode(z, batch.edge_index[:, batch.df_mask]) label = torch.ones_like(logits, dtype=torch.float, device=device) loss = -F.binary_cross_entropy_with_logits(logits, label) loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() end_time = time.time() epoch_loss += loss.item() epoch_time += end_time - start_time epoch_loss /= step epoch_time /= step if (epoch+1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': epoch_loss / step, 'train_time': epoch_time / step, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) torch.save(z, os.path.join(args.checkpoint_dir, 'node_embeddings.pt')) # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) class KGGradientAscentTrainer(KGTrainer): def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): model = model.to(device) start_time = time.time() best_metric = 0 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before loader = GraphSAINTRandomWalkSampler( data, batch_size=128, walk_length=args.walk_length, num_steps=args.num_steps, ) for epoch in trange(args.epochs, desc='Epoch'): model.train() epoch_loss = 0 for step, batch in enumerate(tqdm(loader, desc='Step', leave=False)): batch = batch.to(device) # Message passing edge_index = batch.edge_index[:, batch.dr_mask] edge_type = batch.edge_type[batch.dr_mask] z = model(batch.x, edge_index, edge_type) # Positive and negative sample decoding_edge_index = batch.edge_index[:, batch.df_mask] decoding_edge_type = batch.edge_type[batch.df_mask] decoding_mask = (decoding_edge_type < args.num_edge_type) # Only select directed edges for link prediction decoding_edge_index = decoding_edge_index[:, decoding_mask] decoding_edge_type = decoding_edge_type[decoding_mask] logits = model.decode(z, decoding_edge_index, decoding_edge_type) label = torch.ones_like(logits, dtype=torch.float, device=device) loss = -F.binary_cross_entropy_with_logits(logits, label) loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() log = { 'epoch': epoch, 'step': step, 'train_loss': loss.item(), } wandb.log(log) # msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] # tqdm.write(' | '.join(msg)) epoch_loss += loss.item() if (epoch + 1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': epoch_loss / step } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(train_log) self.trainer_log['log'].append(valid_log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) self.trainer_log['training_time'] = time.time() - start_time # Save models and node embeddings print('Saving final checkpoint') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) print(f'Training finished. Best checkpoint at epoch = {best_epoch:04d}, best valid loss = {best_metric:.4f}') self.trainer_log['best_epoch'] = best_epoch self.trainer_log['best_metric'] = best_metric self.trainer_log['training_time'] = np.mean([i['epoch_time'] for i in self.trainer_log['log'] if 'epoch_time' in i])
13,223
42.074919
128
py
GNNDelete
GNNDelete-main/framework/trainer/graph_eraser.py
import os import json import copy import math from tqdm import tqdm, trange import numpy as np import torch import torch.nn.functional as F from torch_geometric.utils import negative_sampling, subgraph from .base import Trainer from ..evaluation import * from ..utils import * class ConstrainedKmeans: '''This code is from https://github.com/MinChen00/Graph-Unlearning''' def __init__(self, args, data_feat, num_clusters, node_threshold, terminate_delta, max_iteration=20): self.args = args self.data_feat = data_feat self.num_clusters = num_clusters self.node_threshold = node_threshold self.terminate_delta = terminate_delta self.max_iteration = max_iteration def initialization(self): centroids = np.random.choice(np.arange(self.data_feat.shape[0]), self.num_clusters, replace=False) self.centroid = {} for i in range(self.num_clusters): self.centroid[i] = self.data_feat[centroids[i]] def clustering(self): centroid = copy.deepcopy(self.centroid) km_delta = [] # pbar = tqdm(total=self.max_iteration) # pbar.set_description('Clustering') for i in trange(self.max_iteration, desc='Graph partition'): # self.logger.info('iteration %s' % (i,)) self._node_reassignment() self._centroid_updating() # record the average change of centroids, if the change is smaller than a very small value, then terminate delta = self._centroid_delta(centroid, self.centroid) km_delta.append(delta) centroid = copy.deepcopy(self.centroid) if delta <= self.terminate_delta: break print("delta: %s" % delta) # pbar.close() return self.clusters, km_delta def _node_reassignment(self): self.clusters = {} for i in range(self.num_clusters): self.clusters[i] = np.zeros(0, dtype=np.uint64) distance = np.zeros([self.num_clusters, self.data_feat.shape[0]]) for i in range(self.num_clusters): distance[i] = np.sum(np.power((self.data_feat - self.centroid[i]), 2), axis=1) sort_indices = np.unravel_index(np.argsort(distance, axis=None), distance.shape) clusters = sort_indices[0] users = sort_indices[1] selected_nodes = np.zeros(0, dtype=np.int64) counter = 0 while len(selected_nodes) < self.data_feat.shape[0]: cluster = int(clusters[counter]) user = users[counter] if self.clusters[cluster].size < self.node_threshold: self.clusters[cluster] = np.append(self.clusters[cluster], np.array(int(user))) selected_nodes = np.append(selected_nodes, np.array(int(user))) # delete all the following pairs for the selected user user_indices = np.where(users == user)[0] a = np.arange(users.size) b = user_indices[user_indices > counter] remain_indices = a[np.where(np.logical_not(np.isin(a, b)))[0]] clusters = clusters[remain_indices] users = users[remain_indices] counter += 1 def _centroid_updating(self): for i in range(self.num_clusters): self.centroid[i] = np.mean(self.data_feat[self.clusters[i].astype(int)], axis=0) def _centroid_delta(self, centroid_pre, centroid_cur): delta = 0.0 for i in range(len(centroid_cur)): delta += np.sum(np.abs(centroid_cur[i] - centroid_pre[i])) return delta def generate_shard_data(self, data): shard_data = {} for shard in trange(self.args['num_shards'], desc='Generate shard data'): train_shard_indices = list(self.community_to_node[shard]) shard_indices = np.union1d(train_shard_indices, self.test_indices) x = data.x[shard_indices] y = data.y[shard_indices] edge_index = utils.filter_edge_index_1(data, shard_indices) data = Data(x=x, edge_index=torch.from_numpy(edge_index), y=y) data.train_mask = torch.from_numpy(np.isin(shard_indices, train_shard_indices)) data.test_mask = torch.from_numpy(np.isin(shard_indices, self.test_indices)) shard_data[shard] = data self.data_store.save_shard_data(self.shard_data) class OptimalAggregator: def __init__(self, run, target_model, data, args): self.args = args self.run = run self.target_model = target_model self.data = data self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') self.num_shards = args.num_clusters def generate_train_data(self): data_store = DataStore(self.args) train_indices, _ = data_store.load_train_test_split() # sample a set of nodes from train_indices if self.args["num_opt_samples"] == 1000: train_indices = np.random.choice(train_indices, size=1000, replace=False) elif self.args["num_opt_samples"] == 10000: train_indices = np.random.choice(train_indices, size=int(train_indices.shape[0] * 0.1), replace=False) elif self.args["num_opt_samples"] == 1: train_indices = np.random.choice(train_indices, size=int(train_indices.shape[0]), replace=False) train_indices = np.sort(train_indices) self.logger.info("Using %s samples for optimization" % (int(train_indices.shape[0]))) x = self.data.x[train_indices] y = self.data.y[train_indices] edge_index = utils.filter_edge_index(self.data.edge_index, train_indices) train_data = Data(x=x, edge_index=torch.from_numpy(edge_index), y=y) train_data.train_mask = torch.zeros(train_indices.shape[0], dtype=torch.bool) train_data.test_mask = torch.ones(train_indices.shape[0], dtype=torch.bool) self.true_labels = y self.posteriors = {} for shard in range(self.num_shards): self.target_model.data = train_data data_store.load_target_model(self.run, self.target_model, shard) self.posteriors[shard] = self.target_model.posterior().to(self.device) def optimization(self): weight_para = nn.Parameter(torch.full((self.num_shards,), fill_value=1.0 / self.num_shards), requires_grad=True) optimizer = optim.Adam([weight_para], lr=self.args['opt_lr']) scheduler = MultiStepLR(optimizer, milestones=[500, 1000], gamma=self.args['opt_lr']) train_dset = OptDataset(self.posteriors, self.true_labels) train_loader = DataLoader(train_dset, batch_size=32, shuffle=True, num_workers=0) min_loss = 1000.0 for epoch in range(self.args.epochs): loss_all = 0.0 for posteriors, labels in train_loader: labels = labels.to(self.device) optimizer.zero_grad() loss = self._loss_fn(posteriors, labels, weight_para) loss.backward() loss_all += loss optimizer.step() with torch.no_grad(): weight_para[:] = torch.clamp(weight_para, min=0.0) scheduler.step() if loss_all < min_loss: ret_weight_para = copy.deepcopy(weight_para) min_loss = loss_all self.logger.info('epoch: %s, loss: %s' % (epoch, loss_all)) return ret_weight_para / torch.sum(ret_weight_para) def _loss_fn(self, posteriors, labels, weight_para): aggregate_posteriors = torch.zeros_like(posteriors[0]) for shard in range(self.num_shards): aggregate_posteriors += weight_para[shard] * posteriors[shard] aggregate_posteriors = F.softmax(aggregate_posteriors, dim=1) loss_1 = F.cross_entropy(aggregate_posteriors, labels) loss_2 = torch.sqrt(torch.sum(weight_para ** 2)) return loss_1 + loss_2 class Aggregator: def __init__(self, run, target_model, data, shard_data, args): self.args = args self.run = run self.target_model = target_model self.data = data self.shard_data = shard_data self.num_shards = args.num_clusters def generate_posterior(self, suffix=""): self.true_label = self.shard_data[0].y[self.shard_data[0]['test_mask']].detach().cpu().numpy() self.posteriors = {} for shard in range(self.args.num_clusters): self.target_model.data = self.shard_data[shard] self.data_store.load_target_model(self.run, self.target_model, shard, suffix) self.posteriors[shard] = self.target_model.posterior() def _optimal_aggregator(self): optimal = OptimalAggregator(self.run, self.target_model, self.data, self.args) optimal.generate_train_data() weight_para = optimal.optimization() self.data_store.save_optimal_weight(weight_para, run=self.run) posterior = self.posteriors[0] * weight_para[0] for shard in range(1, self.num_shards): posterior += self.posteriors[shard] * weight_para[shard] return f1_score(self.true_label, posterior.argmax(axis=1).cpu().numpy(), average="micro") class GraphEraserTrainer(Trainer): def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): with torch.no_grad(): z = model(data.x, data.train_pos_edge_index[:, data.dr_mask]) # Retrain the model for c in model.children(): print('before', torch.norm(c.lin.weight), torch.norm(c.bias)) for c in model.children(): c.reset_parameters() for c in model.children(): print('after', torch.norm(c.lin.weight), torch.norm(c.bias)) model = model.cpu() num_nodes = data.num_nodes node_threshold = math.ceil( num_nodes / args.num_clusters + args.shard_size_delta * (num_nodes - num_nodes / args.num_clusters)) print(f'Number of nodes: {num_nodes}. Shard threshold: {node_threshold}') cluster = ConstrainedKmeans( args, z.cpu().numpy(), args.num_clusters, node_threshold, args.terminate_delta, args.kmeans_max_iters) cluster.initialization() community, km_deltas = cluster.clustering() # with open(os.path.join(args.checkpoint_dir, 'kmeans_delta.pkl'), 'wb') as f: # pickle.dump(km_deltas, f) community_to_node = {} for i in range(args.num_clusters): community_to_node[i] = np.array(community[i].astype(int)) models = {} test_result = [] for shard_id in trange(args.num_clusters, desc='Sharded retraining'): model_shard_id = copy.deepcopy(model).to('cuda') optimizer = torch.optim.Adam(model_shard_id.parameters(), lr=args.lr) subset_train, _ = subgraph( torch.tensor(community[shard_id], dtype=torch.long, device=device), data.train_pos_edge_index, num_nodes=data.num_nodes) self.train_model(model_shard_id, data, subset_train, optimizer, args, shard_id) with torch.no_grad(): z = model_shard_id(data.x, subset_train) logits = model_shard_id.decode(data.test_pos_edge_index, data.test_neg_edge_index) weight_para = nn.Parameter(torch.full((self.num_shards,), fill_value=1.0 / self.num_shards), requires_grad=True) optimizer = optim.Adam([weight_para], lr=self.args.lr) aggregator.generate_posterior() self.aggregate_f1_score = aggregator.aggregate() aggregate_time = time.time() - start_time self.logger.info("Partition cost %s seconds." % aggregate_time) self.logger.info("Final Test F1: %s" % (self.aggregate_f1_score,)) def train_model(self, model, data, subset_train, optimizer, args, shard_id): best_loss = 100000 for epoch in range(args.epochs): model.train() neg_edge_index = negative_sampling( edge_index=subset_train, num_nodes=data.num_nodes, num_neg_samples=subset_train.shape[1]) z = model(data.x, subset_train) logits = model.decode(z, subset_train, neg_edge_index) label = self.get_link_labels(subset_train, neg_edge_index) loss = F.binary_cross_entropy_with_logits(logits, label) loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() valid_loss, auc, aup, _, _, = self.eval_model(model, data, subset_train, 'val') log = { 'train_loss': loss.item(), 'valid_loss': valid_loss, 'valid_auc': auc, 'valid_aup': aup, } msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log[f'shard_{shard_id}'] = log torch.save(model.state_dict(), os.path.join(args.checkpoint_dir, f'model_{shard_id}.pt')) @torch.no_grad() def eval_model(self, model, data, subset_train, stage='val', pred_all=False): model.eval() pos_edge_index = data[f'{stage}_pos_edge_index'] neg_edge_index = data[f'{stage}_neg_edge_index'] z = model(data.x, subset_train) logits = model.decode(z, pos_edge_index, neg_edge_index).sigmoid() label = self.get_link_labels(pos_edge_index, neg_edge_index) loss = F.binary_cross_entropy_with_logits(logits, label).cpu().item() auc = roc_auc_score(label.cpu(), logits.cpu()) aup = average_precision_score(label.cpu(), logits.cpu()) if self.args.unlearning_model in ['original', 'retrain']: df_logit = float('nan') else: # df_logit = float('nan') df_logit = model.decode(z, subset_train).sigmoid().detach().cpu().item() if pred_all: logit_all_pair = (z @ z.t()).cpu() else: logit_all_pair = None log = { f'{stage}_loss': loss, f'{stage}_auc': auc, f'{stage}_aup': aup, f'{stage}_df_logit': df_logit, } wandb.log(log) msg = [f'{i}: {j:.4f}' if isinstance(j, (np.floating, float)) else f'{i}: {j:>4d}' for i, j in log.items()] tqdm.write(' | '.join(msg)) return loss, auc, aup, df_logit, logit_all_pair
14,714
38.24
120
py
GNNDelete
GNNDelete-main/framework/trainer/descent_to_delete.py
import os import time import wandb from tqdm import tqdm, trange import torch import torch.nn.functional as F from torch_geometric.utils import negative_sampling from .base import Trainer from ..evaluation import * from ..utils import * class DtdTrainer(Trainer): '''This code is adapte from https://github.com/ChrisWaites/descent-to-delete''' def compute_sigma(self, num_examples, iterations, lipshitz, smooth, strong, epsilon, delta): """Theorem 3.1 https://arxiv.org/pdf/2007.02923.pdf""" print('delta', delta) gamma = (smooth - strong) / (smooth + strong) numerator = 4 * np.sqrt(2) * lipshitz * np.power(gamma, iterations) denominator = (strong * num_examples * (1 - np.power(gamma, iterations))) * ((np.sqrt(np.log(1 / delta) + epsilon)) - np.sqrt(np.log(1 / delta))) # print('sigma', numerator, denominator, numerator / denominator) return numerator / denominator def publish(self, model, sigma): """Publishing function which adds Gaussian noise with scale sigma.""" with torch.no_grad(): for n, p in model.named_parameters(): p.copy_(p + torch.empty_like(p).normal_(0, sigma)) def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): start_time = time.time() best_valid_loss = 100000 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before for epoch in trange(args.epochs, desc='Unlerning'): model.train() # Positive and negative sample neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index[:, data.dr_mask], num_nodes=data.num_nodes, num_neg_samples=data.dr_mask.sum()) z = model(data.x, data.train_pos_edge_index[:, data.dr_mask]) logits = model.decode(z, data.train_pos_edge_index[:, data.dr_mask], neg_edge_index) label = get_link_labels(data.train_pos_edge_index[:, data.dr_mask], neg_edge_index) loss = F.binary_cross_entropy_with_logits(logits, label) loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() log = { 'Epoch': epoch, 'train_loss': loss.item(), } wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) valid_loss, auc, aup, df_logt, logit_all_pair = self.eval(model, data, 'val') self.trainer_log['log'].append({ 'dt_loss': valid_loss, 'dt_auc': auc, 'dt_aup': aup }) train_size = data.dr_mask.sum().cpu().item() sigma = self.compute_sigma( train_size, args.epochs, 1 + args.weight_decay, 4 - args.weight_decay, args.weight_decay, 5, 1 / train_size / train_size) self.publish(model, sigma) self.trainer_log['sigma'] = sigma self.trainer_log['training_time'] = time.time() - start_time # Save ckpt = { 'model_state': {k: v.cpu() for k, v in model.state_dict().items()}, 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt'))
4,135
38.390476
153
py
GNNDelete
GNNDelete-main/framework/trainer/approx_retrain.py
import os import wandb from tqdm import tqdm, trange import torch import torch.nn.functional as F from torch_geometric.utils import negative_sampling from torch.utils.data import DataLoader, TensorDataset from .base import Trainer from ..evaluation import * from ..utils import * DTYPE = np.float16 class ApproxTrainer(Trainer): '''This code is adapted from https://github.com/zleizzo/datadeletion''' def gram_schmidt(self, X): """ Uses numpy's qr factorization method to perform Gram-Schmidt. Args: X: (k x d matrix) X[i] = i-th vector Returns: U: (k x d matrix) U[i] = i-th orthonormal vector C: (k x k matrix) Coefficient matrix, C[i] = coeffs for X[i], X = CU """ (k, d) = X.shape if k <= d: q, r = np.linalg.qr(np.transpose(X)) else: q, r = np.linalg.qr(np.transpose(X), mode='complete') U = np.transpose(q) C = np.transpose(r) return U, C def LKO_pred(self, X, Y, ind, H=None, reg=1e-4): """ Computes the LKO model's prediction values on the left-out points. Args: X: (n x d matrix) Covariate matrix Y: (n x 1 vector) Response vector ind: (k x 1 list) List of indices to be removed H: (n x n matrix, optional) Hat matrix X (X^T X)^{-1} X^T Returns: LKO: (k x 1 vector) Retrained model's predictions on X[i], i in ind """ n = len(Y) k = len(ind) d = len(X[0, :]) if H is None: H = np.matmul(X, np.linalg.solve(np.matmul(X.T, X) + reg * np.eye(d), X.T)) LOO = np.zeros(k) for i in range(k): idx = ind[i] # This is the LOO residual y_i - \hat{y}^{LOO}_i LOO[i] = (Y[idx] - np.matmul(H[idx, :], Y)) / (1 - H[idx, idx]) # S = I - T from the paper S = np.eye(k) for i in range(k): for j in range(k): if j != i: idx_i = ind[i] idx_j = ind[j] S[i, j] = -H[idx_i, idx_j] / (1 - H[idx_i, idx_i]) LKO = np.linalg.solve(S, LOO) return Y[ind] - LKO def lin_res(self, X, Y, theta, ind, H=None, reg=1e-4): """ Approximate retraining via the projective residual update. Args: X: (n x d matrix) Covariate matrix Y: (n x 1 vector) Response vector theta: (d x 1 vector) Current value of parameters to be updated ind: (k x 1 list) List of indices to be removed H: (n x n matrix, optional) Hat matrix X (X^T X)^{-1} X^T Returns: updated: (d x 1 vector) Updated parameters """ d = len(X[0]) k = len(ind) # Step 1: Compute LKO predictions LKO = self.LKO_pred(X, Y, ind, H, reg) # Step 2: Eigendecompose B # 2.I U, C = self.gram_schmidt(X[ind, :]) # 2.II Cmatrix = np.matmul(C.T, C) eigenval, a = np.linalg.eigh(Cmatrix) V = np.matmul(a.T, U) # Step 3: Perform the update # 3.I grad = np.zeros_like(theta) # 2D grad for i in range(k): grad += (X[ind[i], :] * theta - LKO[i]) * X[ind[i], :] # 3.II step = np.zeros_like(theta) # 2D grad for i in range(k): factor = 1 / eigenval[i] if eigenval[i] > 1e-10 else 0 step += factor * V[i, :] * grad * V[i, :] # 3.III return step # update = theta - step # return update @torch.no_grad() def train(self, model, data, optimizer, args, logits_ori=None, attack_model=None): model.eval() best_loss = 100000 neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index[:, data.dr_mask], num_nodes=data.num_nodes, num_neg_samples=data.dr_mask.sum()) z = model(data.x, data.train_pos_edge_index[:, data.dr_mask]) edge_index_all = torch.cat([data.train_pos_edge_index[:, data.dr_mask], neg_edge_index], dim=1) X = z[edge_index_all[0]] * z[edge_index_all[1]] Y = self.get_link_labels(data.train_pos_edge_index[:, data.dr_mask], neg_edge_index) X = X.cpu() Y = Y.cpu() # According to the code, theta should be of (d, d). So only update the weights of the last layer theta = model.conv2.lin.weight.cpu().numpy() ind = [int(i) for i in self.args.df_idx.split(',')] # Not enough RAM for solving matrix inverse. So break into multiple batches update = [] loader = DataLoader(TensorDataset(X, Y), batch_size=4096, num_workers=8) for x, y in tqdm(loader, desc='Unlearning'): x = x.numpy() y = y.numpy() update_step = self.lin_res(x, y, theta.T, ind) update.append(torch.tensor(update_step)) update = torch.stack(update).mean(0) model.conv2.lin.weight = torch.nn.Parameter(model.conv2.lin.weight - update.t().cuda()) print(f'Update model weights from {torch.norm(torch.tensor(theta))} to {torch.norm(model.conv2.lin.weight)}') valid_loss, auc, aup, df_logt, logit_all_pair = self.eval(model, data, 'val') self.trainer_log['log'].append({ 'dt_loss': valid_loss, 'dt_auc': auc, 'dt_aup': aup }) # Save ckpt = { 'model_state': {k: v.cpu() for k, v in model.state_dict().items()}, 'node_emb': None, 'optimizer_state': None, } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt'))
5,736
33.14881
117
py
GNNDelete
GNNDelete-main/framework/trainer/gnndelete_embdis.py
import os import time import wandb from tqdm import tqdm, trange import torch import torch.nn as nn from torch_geometric.utils import negative_sampling, k_hop_subgraph from torch_geometric.loader import GraphSAINTRandomWalkSampler from .base import Trainer from ..evaluation import * from ..utils import * def BoundedKLD(logits, truth): return 1 - torch.exp(-F.kl_div(F.log_softmax(logits, -1), truth.softmax(-1), None, None, 'batchmean')) class GNNDeleteEmbeddingDistanceTrainer(Trainer): def train(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): if 'ogbl' in self.args.dataset: return self.train_minibatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) else: return self.train_fullbatch(model, data, optimizer, args, logits_ori, attack_model_all, attack_model_sub) def train_fullbatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): model = model.to('cuda') data = data.to('cuda') best_metric = 0 if 'kld' in args.unlearning_model: loss_fct = BoundedKLD else: loss_fct = nn.MSELoss() # neg_size = 10 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before # All node paris in S_Df without Df. For Local Causality ## S_Df all pair mask sdf_all_pair_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) idx = torch.combinations(torch.arange(data.num_nodes)[data.sdf_node_2hop_mask], with_replacement=True).t() sdf_all_pair_mask[idx[0], idx[1]] = True sdf_all_pair_mask[idx[1], idx[0]] = True # print(data.sdf_node_2hop_mask.sum()) # print(sdf_all_pair_mask.nonzero()) # print(data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]) assert sdf_all_pair_mask.sum().cpu() == data.sdf_node_2hop_mask.sum().cpu() * data.sdf_node_2hop_mask.sum().cpu() ## Remove Df itself sdf_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][0], data.train_pos_edge_index[:, data.df_mask][1]] = False sdf_all_pair_mask[data.train_pos_edge_index[:, data.df_mask][1], data.train_pos_edge_index[:, data.df_mask][0]] = False ## Lower triangular mask idx = torch.tril_indices(data.num_nodes, data.num_nodes, -1) lower_mask = torch.zeros(data.num_nodes, data.num_nodes, dtype=torch.bool) lower_mask[idx[0], idx[1]] = True ## The final mask is the intersection sdf_all_pair_without_df_mask = sdf_all_pair_mask & lower_mask # print('aaaaaaaaaaaa', data.sdf_node_2hop_mask.sum(), a, sdf_all_pair_mask.sum()) # print('aaaaaaaaaaaa', lower_mask.sum()) # print('aaaaaaaaaaaa', sdf_all_pair_without_df_mask.sum()) # print('aaaaaaaaaaaa', data.sdf_node_2hop_mask.sum()) # assert sdf_all_pair_without_df_mask.sum() == \ # data.sdf_node_2hop_mask.sum().cpu() * (data.sdf_node_2hop_mask.sum().cpu() - 1) // 2 - data.df_mask.sum().cpu() # Node representation for local causality with torch.no_grad(): z1_ori, z2_ori = model.get_original_embeddings(data.x, data.train_pos_edge_index[:, data.dtrain_mask], return_all_emb=True) total_time = 0 for epoch in trange(args.epochs, desc='Unlerning'): model.train() start_time = time.time() z1, z2 = model(data.x, data.train_pos_edge_index[:, data.sdf_mask], return_all_emb=True) print('current deletion weight', model.deletion1.deletion_weight.sum(), model.deletion2.deletion_weight.sum()) # Effectiveness and Randomness neg_size = data.df_mask.sum() neg_edge_index = negative_sampling( edge_index=data.train_pos_edge_index, num_nodes=data.num_nodes, num_neg_samples=neg_size) df_logits = model.decode(z2, data.train_pos_edge_index[:, data.df_mask], neg_edge_index) loss_e = loss_fct(df_logits[:neg_size], df_logits[neg_size:]) # df_logits = model.decode( # z, # data.train_pos_edge_index[:, data.df_mask].repeat(1, neg_size), # neg_edge_index).sigmoid() # loss_e = loss_fct(df_logits[:neg_size], df_logits[neg_size:]) # print('df_logits', df_logits) # raise # Local causality if sdf_all_pair_without_df_mask.sum() != 0: loss_l = loss_fct(z1_ori[data.sdf_node_1hop_mask], z1[data.sdf_node_1hop_mask]) + \ loss_fct(z2_ori[data.sdf_node_2hop_mask], z2[data.sdf_node_2hop_mask]) print('local proba', loss_l.item()) else: loss_l = torch.tensor(0) print('local proba', 0) alpha = 0.5 if 'ablation_random' in self.args.unlearning_model: loss_l = torch.tensor(0) loss = loss_e elif 'ablation_locality' in self.args.unlearning_model: loss_e = torch.tensor(0) loss = loss_l else: loss = alpha * loss_e + (1 - alpha) * loss_l loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() end_time = time.time() log = { 'epoch': epoch, 'train_loss': loss.item(), 'train_loss_l': loss_l.item(), 'train_loss_e': loss_e.item(), 'train_time': end_time - start_time, } # wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) if (epoch+1) % args.valid_freq == 0: valid_loss, dt_auc, dt_aup, df_auc, df_aup, df_logit, logit_all_pair, valid_log = self.eval(model, data, 'val') train_log = { 'epoch': epoch, 'train_loss': loss.item(), 'train_loss_l': loss_e.item(), 'train_loss_e': loss_l.item(), 'train_time': end_time - start_time, } for log in [train_log, valid_log]: wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) if dt_auc + df_auc > best_metric: best_metric = dt_auc + df_auc best_epoch = epoch print(f'Save best checkpoint at epoch {epoch:04d}. Valid loss = {valid_loss:.4f}') ckpt = { 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_final.pt')) # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt')) def train_minibatch(self, model, data, optimizer, args, logits_ori=None, attack_model_all=None, attack_model_sub=None): start_time = time.time() best_loss = 100000 if 'kld' in args.unlearning_model: loss_fct = BoundedKLD else: loss_fct = nn.MSELoss() # neg_size = 10 # MI Attack before unlearning if attack_model_all is not None: mi_logit_all_before, mi_sucrate_all_before = member_infer_attack(model, attack_model_all, data) self.trainer_log['mi_logit_all_before'] = mi_logit_all_before self.trainer_log['mi_sucrate_all_before'] = mi_sucrate_all_before if attack_model_sub is not None: mi_logit_sub_before, mi_sucrate_sub_before = member_infer_attack(model, attack_model_sub, data) self.trainer_log['mi_logit_sub_before'] = mi_logit_sub_before self.trainer_log['mi_sucrate_sub_before'] = mi_sucrate_sub_before z_ori = self.get_embedding(model, data, on_cpu=True) z_ori_two_hop = z_ori[data.sdf_node_2hop_mask] data.edge_index = data.train_pos_edge_index data.node_id = torch.arange(data.x.shape[0]) loader = GraphSAINTRandomWalkSampler( data, batch_size=args.batch_size, walk_length=2, num_steps=args.num_steps, ) for epoch in trange(args.epochs, desc='Unlerning'): model.train() print('current deletion weight', model.deletion1.deletion_weight.sum(), model.deletion2.deletion_weight.sum()) epoch_loss_e = 0 epoch_loss_l = 0 epoch_loss = 0 for step, batch in enumerate(tqdm(loader, leave=False)): # print('data', batch) # print('two hop nodes', batch.sdf_node_2hop_mask.sum()) batch = batch.to('cuda') train_pos_edge_index = batch.edge_index z = model(batch.x, train_pos_edge_index[:, batch.sdf_mask], batch.sdf_node_1hop_mask, batch.sdf_node_2hop_mask) z_two_hop = z[batch.sdf_node_2hop_mask] # Effectiveness and Randomness neg_size = batch.df_mask.sum() neg_edge_index = negative_sampling( edge_index=train_pos_edge_index, num_nodes=z.size(0), num_neg_samples=neg_size) df_logits = model.decode(z, train_pos_edge_index[:, batch.df_mask], neg_edge_index) loss_e = loss_fct(df_logits[:neg_size], df_logits[neg_size:]) # Local causality mask = torch.zeros(data.x.shape[0], dtype=torch.bool) mask[batch.node_id[batch.sdf_node_2hop_mask]] = True z_ori_subset = z_ori[mask].to('cuda') # Only take the lower triangular part num_nodes = z_ori_subset.shape[0] idx = torch.tril_indices(num_nodes, num_nodes, -1) local_lower_mask = torch.zeros(num_nodes, num_nodes, dtype=torch.bool) local_lower_mask[idx[0], idx[1]] = True logits_ori = (z_ori_subset @ z_ori_subset.t())[local_lower_mask].sigmoid() logits = (z_two_hop @ z_two_hop.t())[local_lower_mask].sigmoid() loss_l = loss_fct(logits, logits_ori) alpha = 0.5 if 'ablation_random' in self.args.unlearning_model: loss_l = torch.tensor(0) loss = loss_e elif 'ablation_locality' in self.args.unlearning_model: loss_e = torch.tensor(0) loss = loss_l else: loss = alpha * loss_e + (1 - alpha) * loss_l loss.backward() # torch.nn.utils.clip_grad_norm_(model.parameters(), 1) optimizer.step() optimizer.zero_grad() epoch_loss_e += loss_e.item() epoch_loss_l += loss_l.item() epoch_loss += loss.item() epoch_loss_e /= step epoch_loss_l /= step epoch_loss /= step if (epoch+1) % args.valid_freq == 0: valid_loss, auc, aup, df_logt, logit_all_pair = self.eval(model, data, 'val') log = { 'epoch': epoch, 'train_loss': epoch_loss, 'train_loss_e': epoch_loss_e, 'train_loss_l': epoch_loss_l, 'valid_dt_loss': valid_loss, 'valid_dt_auc': auc, 'valid_dt_aup': aup, } wandb.log(log) msg = [f'{i}: {j:>4d}' if isinstance(j, int) else f'{i}: {j:.4f}' for i, j in log.items()] tqdm.write(' | '.join(msg)) self.trainer_log['log'].append(log) self.trainer_log['training_time'] = time.time() - start_time # Save ckpt = { 'model_state': {k: v.to('cpu') for k, v in model.state_dict().items()}, 'optimizer_state': optimizer.state_dict(), } torch.save(ckpt, os.path.join(args.checkpoint_dir, 'model_best.pt'))
13,600
42.453674
135
py
GNNDelete
GNNDelete-main/framework/models/gin.py
import torch import torch.nn as nn import torch.nn.functional as F from torch_geometric.nn import GINConv class GIN(nn.Module): def __init__(self, args, **kwargs): super().__init__() self.conv1 = GINConv(nn.Linear(args.in_dim, args.hidden_dim)) self.conv2= GINConv(nn.Linear(args.hidden_dim, args.out_dim)) # self.transition = nn.Sequential( # nn.ReLU(), # # nn.Dropout(p=args.dropout) # ) # self.mlp1 = nn.Sequential( # nn.Linear(args.in_dim, args.hidden_dim), # nn.ReLU(), # ) # self.mlp2 = nn.Sequential( # nn.Linear(args.hidden_dim, args.out_dim), # nn.ReLU(), # ) def forward(self, x, edge_index, return_all_emb=False): x1 = self.conv1(x, edge_index) x = F.relu(x1) x2 = self.conv2(x, edge_index) if return_all_emb: return x1, x2 return x2 def decode(self, z, pos_edge_index, neg_edge_index=None): if neg_edge_index is not None: edge_index = torch.cat([pos_edge_index, neg_edge_index], dim=-1) logits = (z[edge_index[0]] * z[edge_index[1]]).sum(dim=-1) else: edge_index = pos_edge_index logits = (z[edge_index[0]] * z[edge_index[1]]).sum(dim=-1) return logits
1,373
28.869565
76
py
GNNDelete
GNNDelete-main/framework/models/rgat.py
import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from sklearn.metrics import roc_auc_score, average_precision_score from typing import Optional import torch import torch.nn.functional as F from torch import Tensor from torch.nn import Parameter, ReLU from torch_scatter import scatter_add from torch_sparse import SparseTensor from torch_geometric.nn.conv import MessagePassing from torch_geometric.nn.dense.linear import Linear from torch_geometric.nn.inits import glorot, ones, zeros from torch_geometric.typing import Adj, OptTensor, Size from torch_geometric.utils import softmax # Source: torch_geometric class RGATConv(MessagePassing): _alpha: OptTensor def __init__( self, in_channels: int, out_channels: int, num_relations: int, num_bases: Optional[int] = None, num_blocks: Optional[int] = None, mod: Optional[str] = None, attention_mechanism: str = "across-relation", attention_mode: str = "additive-self-attention", heads: int = 1, dim: int = 1, concat: bool = True, negative_slope: float = 0.2, dropout: float = 0.0, edge_dim: Optional[int] = None, bias: bool = True, **kwargs, ): kwargs.setdefault('aggr', 'add') super().__init__(node_dim=0, **kwargs) self.heads = heads self.negative_slope = negative_slope self.dropout = dropout self.mod = mod self.activation = ReLU() self.concat = concat self.attention_mode = attention_mode self.attention_mechanism = attention_mechanism self.dim = dim self.edge_dim = edge_dim self.in_channels = in_channels self.out_channels = out_channels self.num_relations = num_relations self.num_bases = num_bases self.num_blocks = num_blocks mod_types = ['additive', 'scaled', 'f-additive', 'f-scaled'] if (self.attention_mechanism != "within-relation" and self.attention_mechanism != "across-relation"): raise ValueError('attention mechanism must either be ' '"within-relation" or "across-relation"') if (self.attention_mode != "additive-self-attention" and self.attention_mode != "multiplicative-self-attention"): raise ValueError('attention mode must either be ' '"additive-self-attention" or ' '"multiplicative-self-attention"') if self.attention_mode == "additive-self-attention" and self.dim > 1: raise ValueError('"additive-self-attention" mode cannot be ' 'applied when value of d is greater than 1. ' 'Use "multiplicative-self-attention" instead.') if self.dropout > 0.0 and self.mod in mod_types: raise ValueError('mod must be None with dropout value greater ' 'than 0 in order to sample attention ' 'coefficients stochastically') if num_bases is not None and num_blocks is not None: raise ValueError('Can not apply both basis-decomposition and ' 'block-diagonal-decomposition at the same time.') # The learnable parameters to compute both attention logits and # attention coefficients: self.q = Parameter( torch.Tensor(self.heads * self.out_channels, self.heads * self.dim)) self.k = Parameter( torch.Tensor(self.heads * self.out_channels, self.heads * self.dim)) if bias and concat: self.bias = Parameter( torch.Tensor(self.heads * self.dim * self.out_channels)) elif bias and not concat: self.bias = Parameter(torch.Tensor(self.dim * self.out_channels)) else: self.register_parameter('bias', None) if edge_dim is not None: self.lin_edge = Linear(self.edge_dim, self.heads * self.out_channels, bias=False, weight_initializer='glorot') self.e = Parameter( torch.Tensor(self.heads * self.out_channels, self.heads * self.dim)) else: self.lin_edge = None self.register_parameter('e', None) if num_bases is not None: self.att = Parameter( torch.Tensor(self.num_relations, self.num_bases)) self.basis = Parameter( torch.Tensor(self.num_bases, self.in_channels, self.heads * self.out_channels)) elif num_blocks is not None: assert ( self.in_channels % self.num_blocks == 0 and (self.heads * self.out_channels) % self.num_blocks == 0), ( "both 'in_channels' and 'heads * out_channels' must be " "multiple of 'num_blocks' used") self.weight = Parameter( torch.Tensor(self.num_relations, self.num_blocks, self.in_channels // self.num_blocks, (self.heads * self.out_channels) // self.num_blocks)) else: self.weight = Parameter( torch.Tensor(self.num_relations, self.in_channels, self.heads * self.out_channels)) self.w = Parameter(torch.ones(self.out_channels)) self.l1 = Parameter(torch.Tensor(1, self.out_channels)) self.b1 = Parameter(torch.Tensor(1, self.out_channels)) self.l2 = Parameter(torch.Tensor(self.out_channels, self.out_channels)) self.b2 = Parameter(torch.Tensor(1, self.out_channels)) self._alpha = None self.reset_parameters() def reset_parameters(self): if self.num_bases is not None: glorot(self.basis) glorot(self.att) else: glorot(self.weight) glorot(self.q) glorot(self.k) zeros(self.bias) ones(self.l1) zeros(self.b1) torch.full(self.l2.size(), 1 / self.out_channels) zeros(self.b2) if self.lin_edge is not None: glorot(self.lin_edge) glorot(self.e) def forward(self, x: Tensor, edge_index: Adj, edge_type: OptTensor = None, edge_attr: OptTensor = None, size: Size = None, return_attention_weights=None): # propagate_type: (x: Tensor, edge_type: OptTensor, edge_attr: OptTensor) # noqa out = self.propagate(edge_index=edge_index, edge_type=edge_type, x=x, size=size, edge_attr=edge_attr) alpha = self._alpha assert alpha is not None self._alpha = None if isinstance(return_attention_weights, bool): if isinstance(edge_index, Tensor): return out, (edge_index, alpha) elif isinstance(edge_index, SparseTensor): return out, edge_index.set_value(alpha, layout='coo') else: return out def message(self, x_i: Tensor, x_j: Tensor, edge_type: Tensor, edge_attr: OptTensor, index: Tensor, ptr: OptTensor, size_i: Optional[int]) -> Tensor: if self.num_bases is not None: # Basis-decomposition ================= w = torch.matmul(self.att, self.basis.view(self.num_bases, -1)) w = w.view(self.num_relations, self.in_channels, self.heads * self.out_channels) if self.num_blocks is not None: # Block-diagonal-decomposition ======= if (x_i.dtype == torch.long and x_j.dtype == torch.long and self.num_blocks is not None): raise ValueError('Block-diagonal decomposition not supported ' 'for non-continuous input features.') w = self.weight x_i = x_i.view(-1, 1, w.size(1), w.size(2)) x_j = x_j.view(-1, 1, w.size(1), w.size(2)) w = torch.index_select(w, 0, edge_type) outi = torch.einsum('abcd,acde->ace', x_i, w) outi = outi.contiguous().view(-1, self.heads * self.out_channels) outj = torch.einsum('abcd,acde->ace', x_j, w) outj = outj.contiguous().view(-1, self.heads * self.out_channels) else: # No regularization/Basis-decomposition ======================== if self.num_bases is None: w = self.weight w = torch.index_select(w, 0, edge_type) outi = torch.bmm(x_i.unsqueeze(1), w).squeeze(-2) outj = torch.bmm(x_j.unsqueeze(1), w).squeeze(-2) qi = torch.matmul(outi, self.q) kj = torch.matmul(outj, self.k) alpha_edge, alpha = 0, torch.tensor([0]) if edge_attr is not None: if edge_attr.dim() == 1: edge_attr = edge_attr.view(-1, 1) assert self.lin_edge is not None, ( "Please set 'edge_dim = edge_attr.size(-1)' while calling the " "RGATConv layer") edge_attributes = self.lin_edge(edge_attr).view( -1, self.heads * self.out_channels) if edge_attributes.size(0) != edge_attr.size(0): edge_attributes = torch.index_select(edge_attributes, 0, edge_type) alpha_edge = torch.matmul(edge_attributes, self.e) if self.attention_mode == "additive-self-attention": if edge_attr is not None: alpha = torch.add(qi, kj) + alpha_edge else: alpha = torch.add(qi, kj) alpha = F.leaky_relu(alpha, self.negative_slope) elif self.attention_mode == "multiplicative-self-attention": if edge_attr is not None: alpha = (qi * kj) * alpha_edge else: alpha = qi * kj if self.attention_mechanism == "within-relation": across_out = torch.zeros_like(alpha) for r in range(self.num_relations): mask = edge_type == r across_out[mask] = softmax(alpha[mask], index[mask]) alpha = across_out elif self.attention_mechanism == "across-relation": alpha = softmax(alpha, index, ptr, size_i) self._alpha = alpha if self.mod == "additive": if self.attention_mode == "additive-self-attention": ones = torch.ones_like(alpha) h = (outj.view(-1, self.heads, self.out_channels) * ones.view(-1, self.heads, 1)) h = torch.mul(self.w, h) return (outj.view(-1, self.heads, self.out_channels) * alpha.view(-1, self.heads, 1) + h) elif self.attention_mode == "multiplicative-self-attention": ones = torch.ones_like(alpha) h = (outj.view(-1, self.heads, 1, self.out_channels) * ones.view(-1, self.heads, self.dim, 1)) h = torch.mul(self.w, h) return (outj.view(-1, self.heads, 1, self.out_channels) * alpha.view(-1, self.heads, self.dim, 1) + h) elif self.mod == "scaled": if self.attention_mode == "additive-self-attention": ones = alpha.new_ones(index.size()) degree = scatter_add(ones, index, dim_size=size_i)[index].unsqueeze(-1) degree = torch.matmul(degree, self.l1) + self.b1 degree = self.activation(degree) degree = torch.matmul(degree, self.l2) + self.b2 return torch.mul( outj.view(-1, self.heads, self.out_channels) * alpha.view(-1, self.heads, 1), degree.view(-1, 1, self.out_channels)) elif self.attention_mode == "multiplicative-self-attention": ones = alpha.new_ones(index.size()) degree = scatter_add(ones, index, dim_size=size_i)[index].unsqueeze(-1) degree = torch.matmul(degree, self.l1) + self.b1 degree = self.activation(degree) degree = torch.matmul(degree, self.l2) + self.b2 return torch.mul( outj.view(-1, self.heads, 1, self.out_channels) * alpha.view(-1, self.heads, self.dim, 1), degree.view(-1, 1, 1, self.out_channels)) elif self.mod == "f-additive": alpha = torch.where(alpha > 0, alpha + 1, alpha) elif self.mod == "f-scaled": ones = alpha.new_ones(index.size()) degree = scatter_add(ones, index, dim_size=size_i)[index].unsqueeze(-1) alpha = alpha * degree elif self.training and self.dropout > 0: alpha = F.dropout(alpha, p=self.dropout, training=True) else: alpha = alpha # original if self.attention_mode == "additive-self-attention": return alpha.view(-1, self.heads, 1) * outj.view( -1, self.heads, self.out_channels) else: return (alpha.view(-1, self.heads, self.dim, 1) * outj.view(-1, self.heads, 1, self.out_channels)) def update(self, aggr_out: Tensor) -> Tensor: if self.attention_mode == "additive-self-attention": if self.concat is True: aggr_out = aggr_out.view(-1, self.heads * self.out_channels) else: aggr_out = aggr_out.mean(dim=1) if self.bias is not None: aggr_out = aggr_out + self.bias return aggr_out else: if self.concat is True: aggr_out = aggr_out.view( -1, self.heads * self.dim * self.out_channels) else: aggr_out = aggr_out.mean(dim=1) aggr_out = aggr_out.view(-1, self.dim * self.out_channels) if self.bias is not None: aggr_out = aggr_out + self.bias return aggr_out def __repr__(self) -> str: return '{}({}, {}, heads={})'.format(self.__class__.__name__, self.in_channels, self.out_channels, self.heads) class RGAT(nn.Module): def __init__(self, args, num_nodes, num_edge_type, **kwargs): super().__init__() self.args = args self.num_edge_type = num_edge_type # Encoder: RGAT self.node_emb = nn.Embedding(num_nodes, args.in_dim) if num_edge_type > 20: self.conv1 = RGATConv(args.in_dim, args.hidden_dim, num_edge_type * 2, num_blocks=4) self.conv2 = RGATConv(args.hidden_dim, args.out_dim, num_edge_type * 2, num_blocks=4) else: self.conv1 = RGATConv(args.in_dim, args.hidden_dim, num_edge_type * 2) self.conv2 = RGATConv(args.hidden_dim, args.out_dim, num_edge_type * 2) self.relu = nn.ReLU() # Decoder: DistMult self.W = nn.Parameter(torch.Tensor(num_edge_type, args.out_dim)) nn.init.xavier_uniform_(self.W, gain=nn.init.calculate_gain('relu')) def forward(self, x, edge, edge_type, return_all_emb=False): x = self.node_emb(x) x1 = self.conv1(x, edge, edge_type) x = self.relu(x1) x2 = self.conv2(x, edge, edge_type) if return_all_emb: return x1, x2 return x2 def decode(self, z, edge_index, edge_type): h = z[edge_index[0]] t = z[edge_index[1]] r = self.W[edge_type] logits = torch.sum(h * r * t, dim=1) return logits
16,095
40.061224
97
py
GNNDelete
GNNDelete-main/framework/models/deletion.py
import torch import torch.nn as nn import torch.nn.functional as F import torch.nn.init as init from . import GCN, GAT, GIN, RGCN, RGAT class DeletionLayer(nn.Module): def __init__(self, dim, mask): super().__init__() self.dim = dim self.mask = mask self.deletion_weight = nn.Parameter(torch.ones(dim, dim) / 1000) # self.deletion_weight = nn.Parameter(torch.eye(dim, dim)) # init.xavier_uniform_(self.deletion_weight) def forward(self, x, mask=None): '''Only apply deletion operator to the local nodes identified by mask''' if mask is None: mask = self.mask if mask is not None: new_rep = x.clone() new_rep[mask] = torch.matmul(new_rep[mask], self.deletion_weight) return new_rep return x class DeletionLayerKG(nn.Module): def __init__(self, dim, mask): super().__init__() self.dim = dim self.mask = mask self.deletion_weight = nn.Parameter(torch.ones(dim, dim) / 1000) def forward(self, x, mask=None): '''Only apply deletion operator to the local nodes identified by mask''' if mask is None: mask = self.mask if mask is not None: new_rep = x.clone() new_rep[mask] = torch.matmul(new_rep[mask], self.deletion_weight) return new_rep return x class GCNDelete(GCN): def __init__(self, args, mask_1hop=None, mask_2hop=None, **kwargs): super().__init__(args) self.deletion1 = DeletionLayer(args.hidden_dim, mask_1hop) self.deletion2 = DeletionLayer(args.out_dim, mask_2hop) self.conv1.requires_grad = False self.conv2.requires_grad = False def forward(self, x, edge_index, mask_1hop=None, mask_2hop=None, return_all_emb=False): # with torch.no_grad(): x1 = self.conv1(x, edge_index) x1 = self.deletion1(x1, mask_1hop) x = F.relu(x1) x2 = self.conv2(x, edge_index) x2 = self.deletion2(x2, mask_2hop) if return_all_emb: return x1, x2 return x2 def get_original_embeddings(self, x, edge_index, return_all_emb=False): return super().forward(x, edge_index, return_all_emb) class GATDelete(GAT): def __init__(self, args, mask_1hop=None, mask_2hop=None, **kwargs): super().__init__(args) self.deletion1 = DeletionLayer(args.hidden_dim, mask_1hop) self.deletion2 = DeletionLayer(args.out_dim, mask_2hop) self.conv1.requires_grad = False self.conv2.requires_grad = False def forward(self, x, edge_index, mask_1hop=None, mask_2hop=None, return_all_emb=False): with torch.no_grad(): x1 = self.conv1(x, edge_index) x1 = self.deletion1(x1, mask_1hop) x = F.relu(x1) x2 = self.conv2(x, edge_index) x2 = self.deletion2(x2, mask_2hop) if return_all_emb: return x1, x2 return x2 def get_original_embeddings(self, x, edge_index, return_all_emb=False): return super().forward(x, edge_index, return_all_emb) class GINDelete(GIN): def __init__(self, args, mask_1hop=None, mask_2hop=None, **kwargs): super().__init__(args) self.deletion1 = DeletionLayer(args.hidden_dim, mask_1hop) self.deletion2 = DeletionLayer(args.out_dim, mask_2hop) self.conv1.requires_grad = False self.conv2.requires_grad = False def forward(self, x, edge_index, mask_1hop=None, mask_2hop=None, return_all_emb=False): with torch.no_grad(): x1 = self.conv1(x, edge_index) x1 = self.deletion1(x1, mask_1hop) x = F.relu(x1) x2 = self.conv2(x, edge_index) x2 = self.deletion2(x2, mask_2hop) if return_all_emb: return x1, x2 return x2 def get_original_embeddings(self, x, edge_index, return_all_emb=False): return super().forward(x, edge_index, return_all_emb) class RGCNDelete(RGCN): def __init__(self, args, num_nodes, num_edge_type, mask_1hop=None, mask_2hop=None, **kwargs): super().__init__(args, num_nodes, num_edge_type) self.deletion1 = DeletionLayer(args.hidden_dim, mask_1hop) self.deletion2 = DeletionLayer(args.out_dim, mask_2hop) self.node_emb.requires_grad = False self.conv1.requires_grad = False self.conv2.requires_grad = False def forward(self, x, edge_index, edge_type, mask_1hop=None, mask_2hop=None, return_all_emb=False): with torch.no_grad(): x = self.node_emb(x) x1 = self.conv1(x, edge_index, edge_type) x1 = self.deletion1(x1, mask_1hop) x = F.relu(x1) x2 = self.conv2(x, edge_index, edge_type) x2 = self.deletion2(x2, mask_2hop) if return_all_emb: return x1, x2 return x2 def get_original_embeddings(self, x, edge_index, edge_type, return_all_emb=False): return super().forward(x, edge_index, edge_type, return_all_emb) class RGATDelete(RGAT): def __init__(self, args, num_nodes, num_edge_type, mask_1hop=None, mask_2hop=None, **kwargs): super().__init__(args, num_nodes, num_edge_type) self.deletion1 = DeletionLayer(args.hidden_dim, mask_1hop) self.deletion2 = DeletionLayer(args.out_dim, mask_2hop) self.node_emb.requires_grad = False self.conv1.requires_grad = False self.conv2.requires_grad = False def forward(self, x, edge_index, edge_type, mask_1hop=None, mask_2hop=None, return_all_emb=False): with torch.no_grad(): x = self.node_emb(x) x1 = self.conv1(x, edge_index, edge_type) x1 = self.deletion1(x1, mask_1hop) x = F.relu(x1) x2 = self.conv2(x, edge_index, edge_type) x2 = self.deletion2(x2, mask_2hop) if return_all_emb: return x1, x2 return x2 def get_original_embeddings(self, x, edge_index, edge_type, return_all_emb=False): return super().forward(x, edge_index, edge_type, return_all_emb)
6,273
31.340206
102
py
GNNDelete
GNNDelete-main/framework/models/rgcn.py
import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from torch_geometric.nn import RGCNConv, FastRGCNConv from sklearn.metrics import roc_auc_score, average_precision_score class RGCN(nn.Module): def __init__(self, args, num_nodes, num_edge_type, **kwargs): super().__init__() self.args = args self.num_edge_type = num_edge_type # Encoder: RGCN self.node_emb = nn.Embedding(num_nodes, args.in_dim) if num_edge_type > 20: self.conv1 = RGCNConv(args.in_dim, args.hidden_dim, num_edge_type * 2, num_blocks=4) self.conv2 = RGCNConv(args.hidden_dim, args.out_dim, num_edge_type * 2, num_blocks=4) else: self.conv1 = RGCNConv(args.in_dim, args.hidden_dim, num_edge_type * 2) self.conv2 = RGCNConv(args.hidden_dim, args.out_dim, num_edge_type * 2) self.relu = nn.ReLU() # Decoder: DistMult self.W = nn.Parameter(torch.Tensor(num_edge_type, args.out_dim)) nn.init.xavier_uniform_(self.W, gain=nn.init.calculate_gain('relu')) def forward(self, x, edge, edge_type, return_all_emb=False): x = self.node_emb(x) x1 = self.conv1(x, edge, edge_type) x = self.relu(x1) x2 = self.conv2(x, edge, edge_type) if return_all_emb: return x1, x2 return x2 def decode(self, z, edge_index, edge_type): h = z[edge_index[0]] t = z[edge_index[1]] r = self.W[edge_type] logits = torch.sum(h * r * t, dim=1) return logits class RGCNDelete(RGCN): def __init__(self): pass
1,689
31.5
97
py
GNNDelete
GNNDelete-main/framework/models/gcn.py
import torch import torch.nn as nn import torch.nn.functional as F from torch_geometric.nn import GCNConv class GCN(nn.Module): def __init__(self, args, **kwargs): super().__init__() self.conv1 = GCNConv(args.in_dim, args.hidden_dim) self.conv2 = GCNConv(args.hidden_dim, args.out_dim) # self.dropout = nn.Dropout(args.dropout) def forward(self, x, edge_index, return_all_emb=False): x1 = self.conv1(x, edge_index) x = F.relu(x1) # x = self.dropout(x) x2 = self.conv2(x, edge_index) if return_all_emb: return x1, x2 return x2 def decode(self, z, pos_edge_index, neg_edge_index=None): if neg_edge_index is not None: edge_index = torch.cat([pos_edge_index, neg_edge_index], dim=-1) logits = (z[edge_index[0]] * z[edge_index[1]]).sum(dim=-1) else: edge_index = pos_edge_index logits = (z[edge_index[0]] * z[edge_index[1]]).sum(dim=-1) return logits
1,039
27.888889
76
py
GNNDelete
GNNDelete-main/framework/models/gat.py
import torch import torch.nn as nn import torch.nn.functional as F from torch_geometric.nn import GATConv class GAT(nn.Module): def __init__(self, args, **kwargs): super().__init__() self.conv1 = GATConv(args.in_dim, args.hidden_dim) self.conv2 = GATConv(args.hidden_dim, args.out_dim) # self.dropout = nn.Dropout(args.dropout) def forward(self, x, edge_index, return_all_emb=False): x1 = self.conv1(x, edge_index) x = F.relu(x1) # x = self.dropout(x) x2 = self.conv2(x, edge_index) if return_all_emb: return x1, x2 return x2 def decode(self, z, pos_edge_index, neg_edge_index=None): if neg_edge_index is not None: edge_index = torch.cat([pos_edge_index, neg_edge_index], dim=-1) logits = (z[edge_index[0]] * z[edge_index[1]]).sum(dim=-1) else: edge_index = pos_edge_index logits = (z[edge_index[0]] * z[edge_index[1]]).sum(dim=-1) return logits
1,039
27.888889
76
py
GNNDelete
GNNDelete-main/framework/models/graph_classification/gcn_delete.py
import torch import torch.nn as nn import torch.nn.functional as F from ogb.graphproppred.mol_encoder import AtomEncoder from torch_geometric.nn import GCNConv, MessagePassing, global_add_pool, global_mean_pool, global_max_pool, GlobalAttention, Set2Set from ogb.graphproppred.mol_encoder import AtomEncoder, BondEncoder from torch_geometric.utils import degree from torch_geometric.nn.inits import uniform from torch_scatter import scatter_mean from ..deletion import DeletionLayer def remove_edges(edge_index, edge_attr=None, ratio=0.025): row, col = edge_index mask = row < col row, col = row[mask], col[mask] if edge_attr is not None: edge_attr = edge_attr[mask] num_edges = len(row) num_remove = max(1, int(num_edges * ratio)) selected = torch.randperm(num_edges)[:num_edges - num_remove] row = row[selected] col = col[selected] edge_attr = edge_attr[selected] return torch.stack([row, col], dim=0), edge_attr ''' Source: OGB github https://github.com/snap-stanford/ogb/blob/master/examples/graphproppred/mol/main_pyg.py ''' ### GCN convolution along the graph structure class GCNConv(MessagePassing): def __init__(self, emb_dim): super().__init__(aggr='add') self.linear = nn.Linear(emb_dim, emb_dim) self.root_emb = nn.Embedding(1, emb_dim) self.bond_encoder = BondEncoder(emb_dim=emb_dim) def forward(self, x, edge_index, edge_attr): x = self.linear(x) edge_embedding = self.bond_encoder(edge_attr) row, col = edge_index #edge_weight = torch.ones((edge_index.size(1), ), device=edge_index.device) deg = degree(row, x.size(0), dtype=x.dtype) + 1 deg_inv_sqrt = deg.pow(-0.5) deg_inv_sqrt[deg_inv_sqrt == float('inf')] = 0 norm = deg_inv_sqrt[row] * deg_inv_sqrt[col] return self.propagate(edge_index, x=x, edge_attr=edge_embedding, norm=norm) + F.relu(x + self.root_emb.weight) * 1./deg.view(-1,1) def message(self, x_j, edge_attr, norm): return norm.view(-1, 1) * F.relu(x_j + edge_attr) def update(self, aggr_out): return aggr_out ### GNN to generate node embedding class GNN_node_delete(nn.Module): def __init__(self, num_layer, emb_dim, drop_ratio=0.5, JK="last", residual=False, mask_1hop=None, mask_2hop=None): super().__init__() self.num_layer = num_layer self.drop_ratio = drop_ratio self.JK = JK ### add residual connection or not self.residual = residual self.atom_encoder = AtomEncoder(emb_dim) ###List of GNNs self.deletes = nn.ModuleList([ DeletionLayer(emb_dim, None), DeletionLayer(emb_dim, None) ]) self.convs = nn.ModuleList() self.batch_norms = nn.ModuleList() for layer in range(num_layer): self.convs.append(GCNConv(emb_dim)) self.batch_norms.append(nn.BatchNorm1d(emb_dim)) def forward(self, batched_data): x, edge_index, edge_attr, batch = batched_data.x, batched_data.edge_index, batched_data.edge_attr, batched_data.batch edge_index, edge_attr = remove_edges(edge_index, edge_attr) ### computing input node embedding h_list = [self.atom_encoder(x)] for layer in range(self.num_layer): h = self.convs[layer](h_list[layer], edge_index, edge_attr) h = self.deletes[layer](h) h = self.batch_norms[layer](h) if layer == self.num_layer - 1: #remove relu for the last layer h = F.dropout(h, self.drop_ratio, training=self.training) else: h = F.dropout(F.relu(h), self.drop_ratio, training=self.training) if self.residual: h += h_list[layer] h_list.append(h) ### Different implementations of Jk-concat if self.JK == "last": node_representation = h_list[-1] elif self.JK == "sum": node_representation = 0 for layer in range(self.num_layer + 1): node_representation += h_list[layer] return node_representation class GNN(torch.nn.Module): def __init__(self, num_tasks, num_layer=2, emb_dim=300, virtual_node=True, residual=False, drop_ratio=0.5, JK="last", graph_pooling="mean"): super().__init__() self.num_layer = num_layer self.drop_ratio = drop_ratio self.JK = JK self.emb_dim = emb_dim self.num_tasks = num_tasks self.graph_pooling = graph_pooling ### GNN to generate node embeddings self.gnn_node = GNN_node_delete(num_layer, emb_dim, JK=JK, drop_ratio=drop_ratio, residual=residual) ### Pooling function to generate whole-graph embeddings if self.graph_pooling == "sum": self.pool = global_add_pool elif self.graph_pooling == "mean": self.pool = global_mean_pool elif self.graph_pooling == "max": self.pool = global_max_pool elif self.graph_pooling == "attention": self.pool = GlobalAttention(gate_nn=nn.Sequential(Linear(emb_dim, 2*emb_dim), nn.BatchNorm1d(2*emb_dim), nn.ReLU(), nn.Linear(2*emb_dim, 1))) elif self.graph_pooling == "set2set": self.pool = Set2Set(emb_dim, processing_steps = 2) else: raise ValueError("Invalid graph pooling type.") if graph_pooling == "set2set": self.graph_pred_linear = nn.Linear(2*self.emb_dim, self.num_tasks) else: self.graph_pred_linear = nn.Linear(self.emb_dim, self.num_tasks) def forward(self, batched_data): h_node = self.gnn_node(batched_data) h_graph = self.pool(h_node, batched_data.batch) return self.graph_pred_linear(h_graph)
5,857
34.719512
153
py
GNNDelete
GNNDelete-main/framework/models/graph_classification/gcn.py
import torch import torch.nn as nn import torch.nn.functional as F from ogb.graphproppred.mol_encoder import AtomEncoder from torch_geometric.nn import GCNConv, MessagePassing, global_add_pool, global_mean_pool, global_max_pool, GlobalAttention, Set2Set from ogb.graphproppred.mol_encoder import AtomEncoder, BondEncoder from torch_geometric.utils import degree from torch_geometric.nn.inits import uniform from torch_scatter import scatter_mean ''' Source: OGB github https://github.com/snap-stanford/ogb/blob/master/examples/graphproppred/mol/main_pyg.py ''' ### GCN convolution along the graph structure class GCNConv(MessagePassing): def __init__(self, emb_dim): super().__init__(aggr='add') self.linear = nn.Linear(emb_dim, emb_dim) self.root_emb = nn.Embedding(1, emb_dim) self.bond_encoder = BondEncoder(emb_dim=emb_dim) def forward(self, x, edge_index, edge_attr): x = self.linear(x) edge_embedding = self.bond_encoder(edge_attr) row, col = edge_index #edge_weight = torch.ones((edge_index.size(1), ), device=edge_index.device) deg = degree(row, x.size(0), dtype=x.dtype) + 1 deg_inv_sqrt = deg.pow(-0.5) deg_inv_sqrt[deg_inv_sqrt == float('inf')] = 0 norm = deg_inv_sqrt[row] * deg_inv_sqrt[col] return self.propagate(edge_index, x=x, edge_attr=edge_embedding, norm=norm) + F.relu(x + self.root_emb.weight) * 1./deg.view(-1,1) def message(self, x_j, edge_attr, norm): return norm.view(-1, 1) * F.relu(x_j + edge_attr) def update(self, aggr_out): return aggr_out ### GNN to generate node embedding class GNN_node(nn.Module): def __init__(self, num_layer, emb_dim, drop_ratio=0.5, JK="last", residual=False): super().__init__() self.num_layer = num_layer self.drop_ratio = drop_ratio self.JK = JK ### add residual connection or not self.residual = residual self.atom_encoder = AtomEncoder(emb_dim) ###List of GNNs self.convs = nn.ModuleList() self.batch_norms = nn.ModuleList() for layer in range(num_layer): self.convs.append(GCNConv(emb_dim)) self.batch_norms.append(nn.BatchNorm1d(emb_dim)) def forward(self, batched_data): x, edge_index, edge_attr, batch = batched_data.x, batched_data.edge_index, batched_data.edge_attr, batched_data.batch ### computing input node embedding h_list = [self.atom_encoder(x)] for layer in range(self.num_layer): h = self.convs[layer](h_list[layer], edge_index, edge_attr) h = self.batch_norms[layer](h) if layer == self.num_layer - 1: #remove relu for the last layer h = F.dropout(h, self.drop_ratio, training=self.training) else: h = F.dropout(F.relu(h), self.drop_ratio, training=self.training) if self.residual: h += h_list[layer] h_list.append(h) ### Different implementations of Jk-concat if self.JK == "last": node_representation = h_list[-1] elif self.JK == "sum": node_representation = 0 for layer in range(self.num_layer + 1): node_representation += h_list[layer] return node_representation class GNN(torch.nn.Module): def __init__(self, num_tasks, num_layer=2, emb_dim=300, virtual_node=True, residual=False, drop_ratio=0.5, JK="last", graph_pooling="mean"): super().__init__() self.num_layer = num_layer self.drop_ratio = drop_ratio self.JK = JK self.emb_dim = emb_dim self.num_tasks = num_tasks self.graph_pooling = graph_pooling ### GNN to generate node embeddings self.gnn_node = GNN_node(num_layer, emb_dim, JK=JK, drop_ratio=drop_ratio, residual=residual) ### Pooling function to generate whole-graph embeddings if self.graph_pooling == "sum": self.pool = global_add_pool elif self.graph_pooling == "mean": self.pool = global_mean_pool elif self.graph_pooling == "max": self.pool = global_max_pool elif self.graph_pooling == "attention": self.pool = GlobalAttention(gate_nn=nn.Sequential(Linear(emb_dim, 2*emb_dim), nn.BatchNorm1d(2*emb_dim), nn.ReLU(), nn.Linear(2*emb_dim, 1))) elif self.graph_pooling == "set2set": self.pool = Set2Set(emb_dim, processing_steps = 2) else: raise ValueError("Invalid graph pooling type.") if graph_pooling == "set2set": self.graph_pred_linear = nn.Linear(2*self.emb_dim, self.num_tasks) else: self.graph_pred_linear = nn.Linear(self.emb_dim, self.num_tasks) def forward(self, batched_data): h_node = self.gnn_node(batched_data) h_graph = self.pool(h_node, batched_data.batch) return self.graph_pred_linear(h_graph)
5,019
35.642336
153
py
DeeBERT
DeeBERT-master/setup.py
""" Simple check list from AllenNLP repo: https://github.com/allenai/allennlp/blob/master/setup.py To create the package for pypi. 1. Change the version in __init__.py, setup.py as well as docs/source/conf.py. 2. Commit these changes with the message: "Release: VERSION" 3. Add a tag in git to mark the release: "git tag VERSION -m'Adds tag VERSION for pypi' " Push the tag to git: git push --tags origin master 4. Build both the sources and the wheel. Do not change anything in setup.py between creating the wheel and the source distribution (obviously). For the wheel, run: "python setup.py bdist_wheel" in the top level directory. (this will build a wheel for the python version you use to build it - make sure you use python 3.x). For the sources, run: "python setup.py sdist" You should now have a /dist directory with both .whl and .tar.gz source versions. 5. Check that everything looks correct by uploading the package to the pypi test server: twine upload dist/* -r pypitest (pypi suggest using twine as other methods upload files via plaintext.) Check that you can install it in a virtualenv by running: pip install -i https://testpypi.python.org/pypi transformers 6. Upload the final version to actual pypi: twine upload dist/* -r pypi 7. Copy the release notes from RELEASE.md to the tag in github once everything is looking hunky-dory. """ from io import open from setuptools import find_packages, setup setup( name="transformers", version="2.1.1", author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors", author_email="thomas@huggingface.co", description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch", long_description=open("README.md", "r", encoding='utf-8').read(), long_description_content_type="text/markdown", keywords='NLP deep learning transformer pytorch tensorflow BERT GPT GPT-2 google openai CMU', license='Apache', url="https://github.com/huggingface/transformers", packages=find_packages(exclude=["*.tests", "*.tests.*", "tests.*", "tests"]), install_requires=['numpy', 'boto3', 'requests', 'tqdm', 'regex', 'sentencepiece', 'sacremoses'], entry_points={ 'console_scripts': [ "transformers=transformers.__main__:main", ] }, # python_requires='>=3.5.0', tests_require=['pytest'], classifiers=[ 'Intended Audience :: Science/Research', 'License :: OSI Approved :: Apache Software License', 'Programming Language :: Python :: 3', 'Topic :: Scientific/Engineering :: Artificial Intelligence', ], )
2,923
39.054795
183
py
DeeBERT
DeeBERT-master/hubconf.py
from transformers import ( AutoTokenizer, AutoConfig, AutoModel, AutoModelWithLMHead, AutoModelForSequenceClassification, AutoModelForQuestionAnswering ) from transformers.file_utils import add_start_docstrings dependencies = ['torch', 'tqdm', 'boto3', 'requests', 'regex', 'sentencepiece', 'sacremoses'] @add_start_docstrings(AutoConfig.__doc__) def config(*args, **kwargs): r""" # Using torch.hub ! import torch config = torch.hub.load('huggingface/transformers', 'config', 'bert-base-uncased') # Download configuration from S3 and cache. config = torch.hub.load('huggingface/transformers', 'config', './test/bert_saved_model/') # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')` config = torch.hub.load('huggingface/transformers', 'config', './test/bert_saved_model/my_configuration.json') config = torch.hub.load('huggingface/transformers', 'config', 'bert-base-uncased', output_attention=True, foo=False) assert config.output_attention == True config, unused_kwargs = torch.hub.load('huggingface/transformers', 'config', 'bert-base-uncased', output_attention=True, foo=False, return_unused_kwargs=True) assert config.output_attention == True assert unused_kwargs == {'foo': False} """ return AutoConfig.from_pretrained(*args, **kwargs) @add_start_docstrings(AutoTokenizer.__doc__) def tokenizer(*args, **kwargs): r""" # Using torch.hub ! import torch tokenizer = torch.hub.load('huggingface/transformers', 'tokenizer', 'bert-base-uncased') # Download vocabulary from S3 and cache. tokenizer = torch.hub.load('huggingface/transformers', 'tokenizer', './test/bert_saved_model/') # E.g. tokenizer was saved using `save_pretrained('./test/saved_model/')` """ return AutoTokenizer.from_pretrained(*args, **kwargs) @add_start_docstrings(AutoModel.__doc__) def model(*args, **kwargs): r""" # Using torch.hub ! import torch model = torch.hub.load('huggingface/transformers', 'model', 'bert-base-uncased') # Download model and configuration from S3 and cache. model = torch.hub.load('huggingface/transformers', 'model', './test/bert_model/') # E.g. model was saved using `save_pretrained('./test/saved_model/')` model = torch.hub.load('huggingface/transformers', 'model', 'bert-base-uncased', output_attention=True) # Update configuration during loading assert model.config.output_attention == True # Loading from a TF checkpoint file instead of a PyTorch model (slower) config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json') model = torch.hub.load('huggingface/transformers', 'model', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config) """ return AutoModel.from_pretrained(*args, **kwargs) @add_start_docstrings(AutoModelWithLMHead.__doc__) def modelWithLMHead(*args, **kwargs): r""" # Using torch.hub ! import torch model = torch.hub.load('huggingface/transformers', 'modelWithLMHead', 'bert-base-uncased') # Download model and configuration from S3 and cache. model = torch.hub.load('huggingface/transformers', 'modelWithLMHead', './test/bert_model/') # E.g. model was saved using `save_pretrained('./test/saved_model/')` model = torch.hub.load('huggingface/transformers', 'modelWithLMHead', 'bert-base-uncased', output_attention=True) # Update configuration during loading assert model.config.output_attention == True # Loading from a TF checkpoint file instead of a PyTorch model (slower) config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json') model = torch.hub.load('huggingface/transformers', 'modelWithLMHead', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config) """ return AutoModelWithLMHead.from_pretrained(*args, **kwargs) @add_start_docstrings(AutoModelForSequenceClassification.__doc__) def modelForSequenceClassification(*args, **kwargs): r""" # Using torch.hub ! import torch model = torch.hub.load('huggingface/transformers', 'modelForSequenceClassification', 'bert-base-uncased') # Download model and configuration from S3 and cache. model = torch.hub.load('huggingface/transformers', 'modelForSequenceClassification', './test/bert_model/') # E.g. model was saved using `save_pretrained('./test/saved_model/')` model = torch.hub.load('huggingface/transformers', 'modelForSequenceClassification', 'bert-base-uncased', output_attention=True) # Update configuration during loading assert model.config.output_attention == True # Loading from a TF checkpoint file instead of a PyTorch model (slower) config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json') model = torch.hub.load('huggingface/transformers', 'modelForSequenceClassification', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config) """ return AutoModelForSequenceClassification.from_pretrained(*args, **kwargs) @add_start_docstrings(AutoModelForQuestionAnswering.__doc__) def modelForQuestionAnswering(*args, **kwargs): r""" # Using torch.hub ! import torch model = torch.hub.load('huggingface/transformers', 'modelForQuestionAnswering', 'bert-base-uncased') # Download model and configuration from S3 and cache. model = torch.hub.load('huggingface/transformers', 'modelForQuestionAnswering', './test/bert_model/') # E.g. model was saved using `save_pretrained('./test/saved_model/')` model = torch.hub.load('huggingface/transformers', 'modelForQuestionAnswering', 'bert-base-uncased', output_attention=True) # Update configuration during loading assert model.config.output_attention == True # Loading from a TF checkpoint file instead of a PyTorch model (slower) config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json') model = torch.hub.load('huggingface/transformers', 'modelForQuestionAnswering', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config) """ return AutoModelForQuestionAnswering.from_pretrained(*args, **kwargs)
6,489
56.433628
189
py
DeeBERT
DeeBERT-master/examples/run_lm_finetuning.py
# coding=utf-8 # Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team. # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, BERT, RoBERTa). GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned using a masked language modeling (MLM) loss. """ from __future__ import absolute_import, division, print_function import argparse import glob import logging import os import pickle import random import re import shutil import numpy as np import torch from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler from torch.utils.data.distributed import DistributedSampler try: from torch.utils.tensorboard import SummaryWriter except: from tensorboardX import SummaryWriter from tqdm import tqdm, trange from transformers import (WEIGHTS_NAME, AdamW, get_linear_schedule_with_warmup, BertConfig, BertForMaskedLM, BertTokenizer, GPT2Config, GPT2LMHeadModel, GPT2Tokenizer, OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer, RobertaConfig, RobertaForMaskedLM, RobertaTokenizer, DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer) logger = logging.getLogger(__name__) MODEL_CLASSES = { 'gpt2': (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer), 'openai-gpt': (OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer), 'bert': (BertConfig, BertForMaskedLM, BertTokenizer), 'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer), 'distilbert': (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer) } class TextDataset(Dataset): def __init__(self, tokenizer, args, file_path='train', block_size=512): assert os.path.isfile(file_path) directory, filename = os.path.split(file_path) cached_features_file = os.path.join(directory, args.model_name_or_path + '_cached_lm_' + str(block_size) + '_' + filename) if os.path.exists(cached_features_file): logger.info("Loading features from cached file %s", cached_features_file) with open(cached_features_file, 'rb') as handle: self.examples = pickle.load(handle) else: logger.info("Creating features from dataset file at %s", directory) self.examples = [] with open(file_path, encoding="utf-8") as f: text = f.read() tokenized_text = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(text)) for i in range(0, len(tokenized_text)-block_size+1, block_size): # Truncate in block of block_size self.examples.append(tokenizer.build_inputs_with_special_tokens(tokenized_text[i:i+block_size])) # Note that we are loosing the last truncated example here for the sake of simplicity (no padding) # If your dataset is small, first you should loook for a bigger one :-) and second you # can change this behavior by adding (model specific) padding. logger.info("Saving features into cached file %s", cached_features_file) with open(cached_features_file, 'wb') as handle: pickle.dump(self.examples, handle, protocol=pickle.HIGHEST_PROTOCOL) def __len__(self): return len(self.examples) def __getitem__(self, item): return torch.tensor(self.examples[item]) def load_and_cache_examples(args, tokenizer, evaluate=False): dataset = TextDataset(tokenizer, args, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size) return dataset def set_seed(args): random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.n_gpu > 0: torch.cuda.manual_seed_all(args.seed) def _rotate_checkpoints(args, checkpoint_prefix, use_mtime=False): if not args.save_total_limit: return if args.save_total_limit <= 0: return # Check if we should delete older checkpoint(s) glob_checkpoints = glob.glob(os.path.join(args.output_dir, '{}-*'.format(checkpoint_prefix))) if len(glob_checkpoints) <= args.save_total_limit: return ordering_and_checkpoint_path = [] for path in glob_checkpoints: if use_mtime: ordering_and_checkpoint_path.append((os.path.getmtime(path), path)) else: regex_match = re.match('.*{}-([0-9]+)'.format(checkpoint_prefix), path) if regex_match and regex_match.groups(): ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path)) checkpoints_sorted = sorted(ordering_and_checkpoint_path) checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted] number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - args.save_total_limit) checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete] for checkpoint in checkpoints_to_be_deleted: logger.info("Deleting older checkpoint [{}] due to args.save_total_limit".format(checkpoint)) shutil.rmtree(checkpoint) def mask_tokens(inputs, tokenizer, args): """ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """ labels = inputs.clone() # We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa) probability_matrix = torch.full(labels.shape, args.mlm_probability) special_tokens_mask = [tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()] probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0) masked_indices = torch.bernoulli(probability_matrix).bool() labels[~masked_indices] = -1 # We only compute loss on masked tokens # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK]) indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices inputs[indices_replaced] = tokenizer.convert_tokens_to_ids(tokenizer.mask_token) # 10% of the time, we replace masked input tokens with random word indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced random_words = torch.randint(len(tokenizer), labels.shape, dtype=torch.long) inputs[indices_random] = random_words[indices_random] # The rest of the time (10% of the time) we keep the masked input tokens unchanged return inputs, labels def train(args, train_dataset, model, tokenizer): """ Train the model """ if args.local_rank in [-1, 0]: tb_writer = SummaryWriter() args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu) train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset) train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size) if args.max_steps > 0: t_total = args.max_steps args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1 else: t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs # Prepare optimizer and schedule (linear warmup and decay) no_decay = ['bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [ {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0} ] optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total) if args.fp16: try: from apex import amp except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) # multi-gpu training (should be after apex fp16 initialization) if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Distributed training (should be after apex fp16 initialization) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True) # Train! logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Num Epochs = %d", args.num_train_epochs) logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size) logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d", args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1)) logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps) logger.info(" Total optimization steps = %d", t_total) global_step = 0 tr_loss, logging_loss = 0.0, 0.0 model.zero_grad() train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]) set_seed(args) # Added here for reproducibility (even between python 2 and 3) for _ in train_iterator: epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0]) for step, batch in enumerate(epoch_iterator): inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch) inputs = inputs.to(args.device) labels = labels.to(args.device) model.train() outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels) loss = outputs[0] # model outputs are always tuple in transformers (see doc) if args.n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu parallel training if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() tr_loss += loss.item() if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) optimizer.step() scheduler.step() # Update learning rate schedule model.zero_grad() global_step += 1 if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0: # Log metrics if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well results = evaluate(args, model, tokenizer) for key, value in results.items(): tb_writer.add_scalar('eval_{}'.format(key), value, global_step) tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step) tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step) logging_loss = tr_loss if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0: checkpoint_prefix = 'checkpoint' # Save model checkpoint output_dir = os.path.join(args.output_dir, '{}-{}'.format(checkpoint_prefix, global_step)) if not os.path.exists(output_dir): os.makedirs(output_dir) model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save.save_pretrained(output_dir) torch.save(args, os.path.join(output_dir, 'training_args.bin')) logger.info("Saving model checkpoint to %s", output_dir) _rotate_checkpoints(args, checkpoint_prefix) if args.max_steps > 0 and global_step > args.max_steps: epoch_iterator.close() break if args.max_steps > 0 and global_step > args.max_steps: train_iterator.close() break if args.local_rank in [-1, 0]: tb_writer.close() return global_step, tr_loss / global_step def evaluate(args, model, tokenizer, prefix=""): # Loop to handle MNLI double evaluation (matched, mis-matched) eval_output_dir = args.output_dir eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True) if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]: os.makedirs(eval_output_dir) args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu) # Note that DistributedSampler samples randomly eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset) eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size) # multi-gpu evaluate if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Eval! logger.info("***** Running evaluation {} *****".format(prefix)) logger.info(" Num examples = %d", len(eval_dataset)) logger.info(" Batch size = %d", args.eval_batch_size) eval_loss = 0.0 nb_eval_steps = 0 model.eval() for batch in tqdm(eval_dataloader, desc="Evaluating"): inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch) inputs = inputs.to(args.device) labels = labels.to(args.device) with torch.no_grad(): outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels) lm_loss = outputs[0] eval_loss += lm_loss.mean().item() nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps perplexity = torch.exp(torch.tensor(eval_loss)) result = { "perplexity": perplexity } output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results {} *****".format(prefix)) for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) return result def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--train_data_file", default=None, type=str, required=True, help="The input training data file (a text file).") parser.add_argument("--output_dir", default=None, type=str, required=True, help="The output directory where the model predictions and checkpoints will be written.") ## Other parameters parser.add_argument("--eval_data_file", default=None, type=str, help="An optional input evaluation data file to evaluate the perplexity on (a text file).") parser.add_argument("--model_type", default="bert", type=str, help="The model architecture to be fine-tuned.") parser.add_argument("--model_name_or_path", default="bert-base-cased", type=str, help="The model checkpoint for weights initialization.") parser.add_argument("--mlm", action='store_true', help="Train with masked-language modeling loss instead of language modeling.") parser.add_argument("--mlm_probability", type=float, default=0.15, help="Ratio of tokens to mask for masked language modeling loss") parser.add_argument("--config_name", default="", type=str, help="Optional pretrained config name or path if not the same as model_name_or_path") parser.add_argument("--tokenizer_name", default="", type=str, help="Optional pretrained tokenizer name or path if not the same as model_name_or_path") parser.add_argument("--cache_dir", default="", type=str, help="Optional directory to store the pre-trained models downloaded from s3 (instread of the default one)") parser.add_argument("--block_size", default=-1, type=int, help="Optional input sequence length after tokenization." "The training dataset will be truncated in block of this size for training." "Default to the model max input length for single sentence inputs (take into account special tokens).") parser.add_argument("--do_train", action='store_true', help="Whether to run training.") parser.add_argument("--do_eval", action='store_true', help="Whether to run eval on the dev set.") parser.add_argument("--evaluate_during_training", action='store_true', help="Run evaluation during training at each logging step.") parser.add_argument("--do_lower_case", action='store_true', help="Set this flag if you are using an uncased model.") parser.add_argument("--per_gpu_train_batch_size", default=4, type=int, help="Batch size per GPU/CPU for training.") parser.add_argument("--per_gpu_eval_batch_size", default=4, type=int, help="Batch size per GPU/CPU for evaluation.") parser.add_argument('--gradient_accumulation_steps', type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.") parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight deay if we apply some.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") parser.add_argument("--num_train_epochs", default=1.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--max_steps", default=-1, type=int, help="If > 0: set total number of training steps to perform. Override num_train_epochs.") parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.") parser.add_argument('--logging_steps', type=int, default=50, help="Log every X updates steps.") parser.add_argument('--save_steps', type=int, default=50, help="Save checkpoint every X updates steps.") parser.add_argument('--save_total_limit', type=int, default=None, help='Limit the total amount of checkpoints, delete the older checkpoints in the output_dir, does not delete by default') parser.add_argument("--eval_all_checkpoints", action='store_true', help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number") parser.add_argument("--no_cuda", action='store_true', help="Avoid using CUDA when available") parser.add_argument('--overwrite_output_dir', action='store_true', help="Overwrite the content of the output directory") parser.add_argument('--overwrite_cache', action='store_true', help="Overwrite the cached training and evaluation sets") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument('--fp16', action='store_true', help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit") parser.add_argument('--fp16_opt_level', type=str, default='O1', help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." "See details at https://nvidia.github.io/apex/amp.html") parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.") parser.add_argument('--server_port', type=str, default='', help="For distant debugging.") args = parser.parse_args() if args.model_type in ["bert", "roberta", "distilbert"] and not args.mlm: raise ValueError("BERT and RoBERTa do not have LM heads but masked LM heads. They must be run using the --mlm " "flag (masked language modeling).") if args.eval_data_file is None and args.do_eval: raise ValueError("Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file " "or remove the --do_eval argument.") if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir: raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir)) # Setup distant debugging if needed if args.server_ip and args.server_port: # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script import ptvsd print("Waiting for debugger attach") ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True) ptvsd.wait_for_attach() # Setup CUDA, GPU & distributed training if args.local_rank == -1 or args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") args.n_gpu = torch.cuda.device_count() else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) torch.distributed.init_process_group(backend='nccl') args.n_gpu = 1 args.device = device # Setup logging logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt = '%m/%d/%Y %H:%M:%S', level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN) logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16) # Set seed set_seed(args) # Load pretrained model and tokenizer if args.local_rank not in [-1, 0]: torch.distributed.barrier() # Barrier to make sure only the first process in distributed training download model & vocab config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type] config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, cache_dir=args.cache_dir if args.cache_dir else None) tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case, cache_dir=args.cache_dir if args.cache_dir else None) if args.block_size <= 0: args.block_size = tokenizer.max_len_single_sentence # Our input block size will be the max possible for the model args.block_size = min(args.block_size, tokenizer.max_len_single_sentence) model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config, cache_dir=args.cache_dir if args.cache_dir else None) model.to(args.device) if args.local_rank == 0: torch.distributed.barrier() # End of barrier to make sure only the first process in distributed training download model & vocab logger.info("Training/evaluation parameters %s", args) # Training if args.do_train: if args.local_rank not in [-1, 0]: torch.distributed.barrier() # Barrier to make sure only the first process in distributed training process the dataset, and the others will use the cache train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False) if args.local_rank == 0: torch.distributed.barrier() global_step, tr_loss = train(args, train_dataset, model, tokenizer) logger.info(" global_step = %s, average loss = %s", global_step, tr_loss) # Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained() if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0): # Create output directory if needed if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]: os.makedirs(args.output_dir) logger.info("Saving model checkpoint to %s", args.output_dir) # Save a trained model, configuration and tokenizer using `save_pretrained()`. # They can then be reloaded using `from_pretrained()` model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir) # Good practice: save your training arguments together with the trained model torch.save(args, os.path.join(args.output_dir, 'training_args.bin')) # Load a trained model and vocabulary that you have fine-tuned model = model_class.from_pretrained(args.output_dir) tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case) model.to(args.device) # Evaluation results = {} if args.do_eval and args.local_rank in [-1, 0]: checkpoints = [args.output_dir] if args.eval_all_checkpoints: checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True))) logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging logger.info("Evaluate the following checkpoints: %s", checkpoints) for checkpoint in checkpoints: global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else "" prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else "" model = model_class.from_pretrained(checkpoint) model.to(args.device) result = evaluate(args, model, tokenizer, prefix=prefix) result = dict((k + '_{}'.format(global_step), v) for k, v in result.items()) results.update(result) return results if __name__ == "__main__": main()
28,845
50.881295
165
py
DeeBERT
DeeBERT-master/examples/run_squad.py
# coding=utf-8 # Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team. # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet).""" from __future__ import absolute_import, division, print_function import argparse import logging import os import random import glob import timeit import numpy as np import torch from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset) from torch.utils.data.distributed import DistributedSampler try: from torch.utils.tensorboard import SummaryWriter except: from tensorboardX import SummaryWriter from tqdm import tqdm, trange from transformers import (WEIGHTS_NAME, BertConfig, BertForQuestionAnswering, BertTokenizer, XLMConfig, XLMForQuestionAnswering, XLMTokenizer, XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer, DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer) from transformers import AdamW, get_linear_schedule_with_warmup from utils_squad import (read_squad_examples, convert_examples_to_features, RawResult, write_predictions, RawResultExtended, write_predictions_extended) # The follwing import is the official SQuAD evaluation script (2.0). # You can remove it from the dependencies if you are using this script outside of the library # We've added it here for automated tests (see examples/test_examples.py file) from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad logger = logging.getLogger(__name__) ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \ for conf in (BertConfig, XLNetConfig, XLMConfig)), ()) MODEL_CLASSES = { 'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer), 'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer), 'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer), 'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer) } def set_seed(args): random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.n_gpu > 0: torch.cuda.manual_seed_all(args.seed) def to_list(tensor): return tensor.detach().cpu().tolist() def train(args, train_dataset, model, tokenizer): """ Train the model """ if args.local_rank in [-1, 0]: tb_writer = SummaryWriter() args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu) train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset) train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size) if args.max_steps > 0: t_total = args.max_steps args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1 else: t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs # Prepare optimizer and schedule (linear warmup and decay) no_decay = ['bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [ {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0} ] optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total) if args.fp16: try: from apex import amp except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) # multi-gpu training (should be after apex fp16 initialization) if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Distributed training (should be after apex fp16 initialization) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True) # Train! logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Num Epochs = %d", args.num_train_epochs) logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size) logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d", args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1)) logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps) logger.info(" Total optimization steps = %d", t_total) global_step = 0 tr_loss, logging_loss = 0.0, 0.0 model.zero_grad() train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]) set_seed(args) # Added here for reproductibility (even between python 2 and 3) for _ in train_iterator: epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0]) for step, batch in enumerate(epoch_iterator): model.train() batch = tuple(t.to(args.device) for t in batch) inputs = {'input_ids': batch[0], 'attention_mask': batch[1], 'start_positions': batch[3], 'end_positions': batch[4]} if args.model_type != 'distilbert': inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2] if args.model_type in ['xlnet', 'xlm']: inputs.update({'cls_index': batch[5], 'p_mask': batch[6]}) outputs = model(**inputs) loss = outputs[0] # model outputs are always tuple in transformers (see doc) if args.n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() tr_loss += loss.item() if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) optimizer.step() scheduler.step() # Update learning rate schedule model.zero_grad() global_step += 1 if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0: # Log metrics if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well results = evaluate(args, model, tokenizer) for key, value in results.items(): tb_writer.add_scalar('eval_{}'.format(key), value, global_step) tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step) tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step) logging_loss = tr_loss if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0: # Save model checkpoint output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step)) if not os.path.exists(output_dir): os.makedirs(output_dir) model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save.save_pretrained(output_dir) torch.save(args, os.path.join(output_dir, 'training_args.bin')) logger.info("Saving model checkpoint to %s", output_dir) if args.max_steps > 0 and global_step > args.max_steps: epoch_iterator.close() break if args.max_steps > 0 and global_step > args.max_steps: train_iterator.close() break if args.local_rank in [-1, 0]: tb_writer.close() return global_step, tr_loss / global_step def evaluate(args, model, tokenizer, prefix=""): dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True) if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]: os.makedirs(args.output_dir) args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu) # Note that DistributedSampler samples randomly eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset) eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size) # multi-gpu evaluate if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Eval! logger.info("***** Running evaluation {} *****".format(prefix)) logger.info(" Num examples = %d", len(dataset)) logger.info(" Batch size = %d", args.eval_batch_size) all_results = [] start_time = timeit.default_timer() for batch in tqdm(eval_dataloader, desc="Evaluating"): model.eval() batch = tuple(t.to(args.device) for t in batch) with torch.no_grad(): inputs = {'input_ids': batch[0], 'attention_mask': batch[1] } if args.model_type != 'distilbert': inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2] # XLM don't use segment_ids example_indices = batch[3] if args.model_type in ['xlnet', 'xlm']: inputs.update({'cls_index': batch[4], 'p_mask': batch[5]}) outputs = model(**inputs) for i, example_index in enumerate(example_indices): eval_feature = features[example_index.item()] unique_id = int(eval_feature.unique_id) if args.model_type in ['xlnet', 'xlm']: # XLNet uses a more complex post-processing procedure result = RawResultExtended(unique_id = unique_id, start_top_log_probs = to_list(outputs[0][i]), start_top_index = to_list(outputs[1][i]), end_top_log_probs = to_list(outputs[2][i]), end_top_index = to_list(outputs[3][i]), cls_logits = to_list(outputs[4][i])) else: result = RawResult(unique_id = unique_id, start_logits = to_list(outputs[0][i]), end_logits = to_list(outputs[1][i])) all_results.append(result) evalTime = timeit.default_timer() - start_time logger.info(" Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset)) # Compute predictions output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix)) output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix)) if args.version_2_with_negative: output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix)) else: output_null_log_odds_file = None if args.model_type in ['xlnet', 'xlm']: # XLNet uses a more complex post-processing procedure write_predictions_extended(examples, features, all_results, args.n_best_size, args.max_answer_length, output_prediction_file, output_nbest_file, output_null_log_odds_file, args.predict_file, model.config.start_n_top, model.config.end_n_top, args.version_2_with_negative, tokenizer, args.verbose_logging) else: write_predictions(examples, features, all_results, args.n_best_size, args.max_answer_length, args.do_lower_case, output_prediction_file, output_nbest_file, output_null_log_odds_file, args.verbose_logging, args.version_2_with_negative, args.null_score_diff_threshold) # Evaluate with the official SQuAD script evaluate_options = EVAL_OPTS(data_file=args.predict_file, pred_file=output_prediction_file, na_prob_file=output_null_log_odds_file) results = evaluate_on_squad(evaluate_options) return results def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False): if args.local_rank not in [-1, 0] and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache # Load data features from cache or dataset file input_file = args.predict_file if evaluate else args.train_file cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format( 'dev' if evaluate else 'train', list(filter(None, args.model_name_or_path.split('/'))).pop(), str(args.max_seq_length))) if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples: logger.info("Loading features from cached file %s", cached_features_file) features = torch.load(cached_features_file) else: logger.info("Creating features from dataset file at %s", input_file) examples = read_squad_examples(input_file=input_file, is_training=not evaluate, version_2_with_negative=args.version_2_with_negative) features = convert_examples_to_features(examples=examples, tokenizer=tokenizer, max_seq_length=args.max_seq_length, doc_stride=args.doc_stride, max_query_length=args.max_query_length, is_training=not evaluate, cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0, pad_token_segment_id=3 if args.model_type in ['xlnet'] else 0, cls_token_at_end=True if args.model_type in ['xlnet'] else False, sequence_a_is_doc=True if args.model_type in ['xlnet'] else False) if args.local_rank in [-1, 0]: logger.info("Saving features into cached file %s", cached_features_file) torch.save(features, cached_features_file) if args.local_rank == 0 and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache # Convert to Tensors and build dataset all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long) all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long) all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float) if evaluate: all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long) dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index, all_cls_index, all_p_mask) else: all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long) all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long) dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_start_positions, all_end_positions, all_cls_index, all_p_mask) if output_examples: return dataset, examples, features return dataset def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--train_file", default=None, type=str, required=True, help="SQuAD json for training. E.g., train-v1.1.json") parser.add_argument("--predict_file", default=None, type=str, required=True, help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json") parser.add_argument("--model_type", default=None, type=str, required=True, help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys())) parser.add_argument("--model_name_or_path", default=None, type=str, required=True, help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS)) parser.add_argument("--output_dir", default=None, type=str, required=True, help="The output directory where the model checkpoints and predictions will be written.") ## Other parameters parser.add_argument("--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name") parser.add_argument("--tokenizer_name", default="", type=str, help="Pretrained tokenizer name or path if not the same as model_name") parser.add_argument("--cache_dir", default="", type=str, help="Where do you want to store the pre-trained models downloaded from s3") parser.add_argument('--version_2_with_negative', action='store_true', help='If true, the SQuAD examples contain some that do not have an answer.') parser.add_argument('--null_score_diff_threshold', type=float, default=0.0, help="If null_score - best_non_null is greater than the threshold predict null.") parser.add_argument("--max_seq_length", default=384, type=int, help="The maximum total input sequence length after WordPiece tokenization. Sequences " "longer than this will be truncated, and sequences shorter than this will be padded.") parser.add_argument("--doc_stride", default=128, type=int, help="When splitting up a long document into chunks, how much stride to take between chunks.") parser.add_argument("--max_query_length", default=64, type=int, help="The maximum number of tokens for the question. Questions longer than this will " "be truncated to this length.") parser.add_argument("--do_train", action='store_true', help="Whether to run training.") parser.add_argument("--do_eval", action='store_true', help="Whether to run eval on the dev set.") parser.add_argument("--evaluate_during_training", action='store_true', help="Rul evaluation during training at each logging step.") parser.add_argument("--do_lower_case", action='store_true', help="Set this flag if you are using an uncased model.") parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.") parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation.") parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument('--gradient_accumulation_steps', type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.") parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--max_steps", default=-1, type=int, help="If > 0: set total number of training steps to perform. Override num_train_epochs.") parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.") parser.add_argument("--n_best_size", default=20, type=int, help="The total number of n-best predictions to generate in the nbest_predictions.json output file.") parser.add_argument("--max_answer_length", default=30, type=int, help="The maximum length of an answer that can be generated. This is needed because the start " "and end predictions are not conditioned on one another.") parser.add_argument("--verbose_logging", action='store_true', help="If true, all of the warnings related to data processing will be printed. " "A number of warnings are expected for a normal SQuAD evaluation.") parser.add_argument('--logging_steps', type=int, default=50, help="Log every X updates steps.") parser.add_argument('--save_steps', type=int, default=50, help="Save checkpoint every X updates steps.") parser.add_argument("--eval_all_checkpoints", action='store_true', help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number") parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available") parser.add_argument('--overwrite_output_dir', action='store_true', help="Overwrite the content of the output directory") parser.add_argument('--overwrite_cache', action='store_true', help="Overwrite the cached training and evaluation sets") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument('--fp16', action='store_true', help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit") parser.add_argument('--fp16_opt_level', type=str, default='O1', help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." "See details at https://nvidia.github.io/apex/amp.html") parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.") parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.") args = parser.parse_args() if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir: raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir)) # Setup distant debugging if needed if args.server_ip and args.server_port: # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script import ptvsd print("Waiting for debugger attach") ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True) ptvsd.wait_for_attach() # Setup CUDA, GPU & distributed training if args.local_rank == -1 or args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") args.n_gpu = torch.cuda.device_count() else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) torch.distributed.init_process_group(backend='nccl') args.n_gpu = 1 args.device = device # Setup logging logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt = '%m/%d/%Y %H:%M:%S', level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN) logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16) # Set seed set_seed(args) # Load pretrained model and tokenizer if args.local_rank not in [-1, 0]: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab args.model_type = args.model_type.lower() config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type] config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, cache_dir=args.cache_dir if args.cache_dir else None) tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case, cache_dir=args.cache_dir if args.cache_dir else None) model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config, cache_dir=args.cache_dir if args.cache_dir else None) if args.local_rank == 0: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab model.to(args.device) logger.info("Training/evaluation parameters %s", args) # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set. # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will # remove the need for this code, but it is still valid. if args.fp16: try: import apex apex.amp.register_half_function(torch, 'einsum') except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") # Training if args.do_train: train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False) global_step, tr_loss = train(args, train_dataset, model, tokenizer) logger.info(" global_step = %s, average loss = %s", global_step, tr_loss) # Save the trained model and the tokenizer if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0): # Create output directory if needed if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]: os.makedirs(args.output_dir) logger.info("Saving model checkpoint to %s", args.output_dir) # Save a trained model, configuration and tokenizer using `save_pretrained()`. # They can then be reloaded using `from_pretrained()` model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir) # Good practice: save your training arguments together with the trained model torch.save(args, os.path.join(args.output_dir, 'training_args.bin')) # Load a trained model and vocabulary that you have fine-tuned model = model_class.from_pretrained(args.output_dir) tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case) model.to(args.device) # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory results = {} if args.do_eval and args.local_rank in [-1, 0]: checkpoints = [args.output_dir] if args.eval_all_checkpoints: checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True))) logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce model loading logs logger.info("Evaluate the following checkpoints: %s", checkpoints) for checkpoint in checkpoints: # Reload the model global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else "" model = model_class.from_pretrained(checkpoint) model.to(args.device) # Evaluate result = evaluate(args, model, tokenizer, prefix=global_step) result = dict((k + ('_{}'.format(global_step) if global_step else ''), v) for k, v in result.items()) results.update(result) logger.info("Results: {}".format(results)) return results if __name__ == "__main__": main()
31,570
54.001742
151
py
DeeBERT
DeeBERT-master/examples/run_highway_glue.py
# coding=utf-8 # Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team. # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Finetuning the library models for sequence classification on GLUE (Bert, XLM, XLNet, RoBERTa).""" from __future__ import absolute_import, division, print_function import argparse import glob import logging import os import random import time import numpy as np import torch from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset) from torch.utils.data.distributed import DistributedSampler try: from torch.utils.tensorboard import SummaryWriter except: from tensorboardX import SummaryWriter from tqdm import tqdm, trange from transformers import (WEIGHTS_NAME, BertConfig, BertTokenizer, RobertaConfig, RobertaTokenizer, XLMConfig, XLMForSequenceClassification, XLMTokenizer, XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer, DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer) from transformers.modeling_highway_bert import BertForSequenceClassification from transformers.modeling_highway_roberta import RobertaForSequenceClassification from transformers import AdamW, get_linear_schedule_with_warmup from transformers import glue_compute_metrics as compute_metrics from transformers import glue_output_modes as output_modes from transformers import glue_processors as processors from transformers import glue_convert_examples_to_features as convert_examples_to_features logger = logging.getLogger(__name__) ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig, DistilBertConfig)), ()) MODEL_CLASSES = { 'bert': (BertConfig, BertForSequenceClassification, BertTokenizer), 'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer), 'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer), 'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer), 'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer) } def set_seed(args): random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.n_gpu > 0: torch.cuda.manual_seed_all(args.seed) def get_wanted_result(result): if "spearmanr" in result: print_result = result["spearmanr"] elif "f1" in result: print_result = result["f1"] elif "mcc" in result: print_result = result["mcc"] elif "acc" in result: print_result = result["acc"] else: print(result) exit(1) return print_result def train(args, train_dataset, model, tokenizer, train_highway=False): """ Train the model """ if args.local_rank in [-1, 0]: tb_writer = SummaryWriter() args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu) train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset) train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size) if args.max_steps > 0: t_total = args.max_steps args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1 else: t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs # Prepare optimizer and schedule (linear warmup and decay) no_decay = ['bias', 'LayerNorm.weight'] if train_highway: optimizer_grouped_parameters = [ {'params': [p for n, p in model.named_parameters() if ("highway" in n) and (not any(nd in n for nd in no_decay))], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if ("highway" in n) and (any(nd in n for nd in no_decay))], 'weight_decay': 0.0} ] else: optimizer_grouped_parameters = [ {'params': [p for n, p in model.named_parameters() if ("highway" not in n) and (not any(nd in n for nd in no_decay))], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if ("highway" not in n) and (any(nd in n for nd in no_decay))], 'weight_decay': 0.0} ] optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total) if args.fp16: try: from apex import amp except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) # multi-gpu training (should be after apex fp16 initialization) if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Distributed training (should be after apex fp16 initialization) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True) # Train! logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Num Epochs = %d", args.num_train_epochs) logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size) logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d", args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1)) logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps) logger.info(" Total optimization steps = %d", t_total) global_step = 0 tr_loss, logging_loss = 0.0, 0.0 model.zero_grad() train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]) set_seed(args) # Added here for reproductibility (even between python 2 and 3) for _ in train_iterator: epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0]) for step, batch in enumerate(epoch_iterator): model.train() batch = tuple(t.to(args.device) for t in batch) inputs = {'input_ids': batch[0], 'attention_mask': batch[1], 'labels': batch[3]} if args.model_type != 'distilbert': inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None # XLM, DistilBERT and RoBERTa don't use segment_ids inputs['train_highway'] = train_highway outputs = model(**inputs) loss = outputs[0] # model outputs are always tuple in transformers (see doc) if args.n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu parallel training if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() tr_loss += loss.item() if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) optimizer.step() scheduler.step() # Update learning rate schedule model.zero_grad() global_step += 1 if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0: # Log metrics if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well results = evaluate(args, model, tokenizer) for key, value in results.items(): tb_writer.add_scalar('eval_{}'.format(key), value, global_step) tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step) tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step) logging_loss = tr_loss if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0: # Save model checkpoint output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step)) if not os.path.exists(output_dir): os.makedirs(output_dir) model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save.save_pretrained(output_dir) torch.save(args, os.path.join(output_dir, 'training_args.bin')) logger.info("Saving model checkpoint to %s", output_dir) if args.max_steps > 0 and global_step > args.max_steps: epoch_iterator.close() break if args.max_steps > 0 and global_step > args.max_steps: train_iterator.close() break if args.local_rank in [-1, 0]: tb_writer.close() return global_step, tr_loss / global_step def evaluate(args, model, tokenizer, prefix="", output_layer=-1, eval_highway=False): # Loop to handle MNLI double evaluation (matched, mis-matched) eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,) eval_outputs_dirs = (args.output_dir, args.output_dir + '-MM') if args.task_name == "mnli" else (args.output_dir,) results = {} for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs): eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True) if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]: os.makedirs(eval_output_dir) args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu) # Note that DistributedSampler samples randomly eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset) eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size) # multi-gpu eval if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Eval! logger.info("***** Running evaluation {} *****".format(prefix)) logger.info(" Num examples = %d", len(eval_dataset)) logger.info(" Batch size = %d", args.eval_batch_size) eval_loss = 0.0 nb_eval_steps = 0 preds = None out_label_ids = None exit_layer_counter = {(i+1):0 for i in range(model.num_layers)} st = time.time() for batch in tqdm(eval_dataloader, desc="Evaluating"): model.eval() batch = tuple(t.to(args.device) for t in batch) with torch.no_grad(): inputs = {'input_ids': batch[0], 'attention_mask': batch[1], 'labels': batch[3]} if args.model_type != 'distilbert': inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None # XLM, DistilBERT and RoBERTa don't use segment_ids if output_layer >= 0: inputs['output_layer'] = output_layer outputs = model(**inputs) if eval_highway: exit_layer_counter[outputs[-1]] += 1 tmp_eval_loss, logits = outputs[:2] eval_loss += tmp_eval_loss.mean().item() nb_eval_steps += 1 if preds is None: preds = logits.detach().cpu().numpy() out_label_ids = inputs['labels'].detach().cpu().numpy() else: preds = np.append(preds, logits.detach().cpu().numpy(), axis=0) out_label_ids = np.append(out_label_ids, inputs['labels'].detach().cpu().numpy(), axis=0) eval_time = time.time() - st print("Eval time:", eval_time) eval_loss = eval_loss / nb_eval_steps if args.output_mode == "classification": preds = np.argmax(preds, axis=1) elif args.output_mode == "regression": preds = np.squeeze(preds) result = compute_metrics(eval_task, preds, out_label_ids) results.update(result) if eval_highway: print("Exit layer counter", exit_layer_counter) actual_cost = sum([l*c for l, c in exit_layer_counter.items()]) full_cost = len(eval_dataloader) * model.num_layers print("Expected saving", actual_cost/full_cost) if args.early_exit_entropy>=0: save_fname = args.plot_data_dir + '/' +\ args.model_name_or_path[2:] +\ "/entropy_{}.npy".format(args.early_exit_entropy) if not os.path.exists(os.path.dirname(save_fname)): os.makedirs(os.path.dirname(save_fname)) print_result = get_wanted_result(result) np.save(save_fname, np.array([exit_layer_counter, eval_time, actual_cost/full_cost, print_result])) output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results {} *****".format(prefix)) for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) return results def load_and_cache_examples(args, task, tokenizer, evaluate=False): if args.local_rank not in [-1, 0] and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache processor = processors[task]() output_mode = output_modes[task] # Load data features from cache or dataset file cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}'.format( 'dev' if evaluate else 'train', list(filter(None, args.model_name_or_path.split('/'))).pop(), str(args.max_seq_length), str(task))) if os.path.exists(cached_features_file) and not args.overwrite_cache: logger.info("Loading features from cached file %s", cached_features_file) features = torch.load(cached_features_file) else: logger.info("Creating features from dataset file at %s", args.data_dir) label_list = processor.get_labels() if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta']: # HACK(label indices are swapped in RoBERTa pretrained model) label_list[1], label_list[2] = label_list[2], label_list[1] examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir) features = convert_examples_to_features(examples, tokenizer, label_list=label_list, max_length=args.max_seq_length, output_mode=output_mode, pad_on_left=bool(args.model_type in ['xlnet']), # pad on the left for xlnet pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0], pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0, ) if args.local_rank in [-1, 0]: logger.info("Saving features into cached file %s", cached_features_file) torch.save(features, cached_features_file) if args.local_rank == 0 and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache # Convert to Tensors and build dataset all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long) all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long) all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long) if output_mode == "classification": all_labels = torch.tensor([f.label for f in features], dtype=torch.long) elif output_mode == "regression": all_labels = torch.tensor([f.label for f in features], dtype=torch.float) dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels) return dataset def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--data_dir", default=None, type=str, required=True, help="The input data dir. Should contain the .tsv files (or other data files) for the task.") parser.add_argument("--model_type", default=None, type=str, required=True, help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys())) parser.add_argument("--model_name_or_path", default=None, type=str, required=True, help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS)) parser.add_argument("--task_name", default=None, type=str, required=True, help="The name of the task to train selected in the list: " + ", ".join(processors.keys())) parser.add_argument("--output_dir", default=None, type=str, required=True, help="The output directory where the model predictions and checkpoints will be written.") parser.add_argument("--plot_data_dir", default="./plotting/", type=str, required=False, help="The directory to store data for plotting figures.") ## Other parameters parser.add_argument("--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name") parser.add_argument("--tokenizer_name", default="", type=str, help="Pretrained tokenizer name or path if not the same as model_name") parser.add_argument("--cache_dir", default="", type=str, help="Where do you want to store the pre-trained models downloaded from s3") parser.add_argument("--max_seq_length", default=128, type=int, help="The maximum total input sequence length after tokenization. Sequences longer " "than this will be truncated, sequences shorter will be padded.") parser.add_argument("--do_train", action='store_true', help="Whether to run training.") parser.add_argument("--do_eval", action='store_true', help="Whether to run eval on the dev set.") parser.add_argument("--evaluate_during_training", action='store_true', help="Rul evaluation during training at each logging step.") parser.add_argument("--do_lower_case", action='store_true', help="Set this flag if you are using an uncased model.") parser.add_argument("--eval_each_highway", action='store_true', help="Set this flag to evaluate each highway.") parser.add_argument("--eval_after_first_stage", action='store_true', help="Set this flag to evaluate after training only bert (not highway).") parser.add_argument("--eval_highway", action='store_true', help="Set this flag if it's evaluating highway models") parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.") parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation.") parser.add_argument('--gradient_accumulation_steps', type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.") parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight deay if we apply some.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--max_steps", default=-1, type=int, help="If > 0: set total number of training steps to perform. Override num_train_epochs.") parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.") parser.add_argument("--early_exit_entropy", default=-1, type=float, help = "Entropy threshold for early exit.") parser.add_argument('--logging_steps', type=int, default=50, help="Log every X updates steps.") parser.add_argument('--save_steps', type=int, default=50, help="Save checkpoint every X updates steps.") parser.add_argument("--eval_all_checkpoints", action='store_true', help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number") parser.add_argument("--no_cuda", action='store_true', help="Avoid using CUDA when available") parser.add_argument('--overwrite_output_dir', action='store_true', help="Overwrite the content of the output directory") parser.add_argument('--overwrite_cache', action='store_true', help="Overwrite the cached training and evaluation sets") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument('--fp16', action='store_true', help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit") parser.add_argument('--fp16_opt_level', type=str, default='O1', help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." "See details at https://nvidia.github.io/apex/amp.html") parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.") parser.add_argument('--server_port', type=str, default='', help="For distant debugging.") args = parser.parse_args() if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir: raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir)) # Setup distant debugging if needed if args.server_ip and args.server_port: # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script import ptvsd print("Waiting for debugger attach") ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True) ptvsd.wait_for_attach() # Setup CUDA, GPU & distributed training if args.local_rank == -1 or args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") args.n_gpu = torch.cuda.device_count() else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) torch.distributed.init_process_group(backend='nccl') args.n_gpu = 1 args.device = device # Setup logging logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt = '%m/%d/%Y %H:%M:%S', level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN) logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16) # Set seed set_seed(args) # Prepare GLUE task args.task_name = args.task_name.lower() if args.task_name not in processors: raise ValueError("Task not found: %s" % (args.task_name)) processor = processors[args.task_name]() args.output_mode = output_modes[args.task_name] label_list = processor.get_labels() num_labels = len(label_list) # Load pretrained model and tokenizer if args.local_rank not in [-1, 0]: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab args.model_type = args.model_type.lower() config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type] config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name, cache_dir=args.cache_dir if args.cache_dir else None) tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case, cache_dir=args.cache_dir if args.cache_dir else None) model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config, cache_dir=args.cache_dir if args.cache_dir else None) if args.model_type == "bert": model.bert.encoder.set_early_exit_entropy(args.early_exit_entropy) model.bert.init_highway_pooler() else: model.roberta.encoder.set_early_exit_entropy(args.early_exit_entropy) model.roberta.init_highway_pooler() if args.local_rank == 0: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab model.to(args.device) logger.info("Training/evaluation parameters %s", args) # Training if args.do_train: train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False) global_step, tr_loss = train(args, train_dataset, model, tokenizer) logger.info(" global_step = %s, average loss = %s", global_step, tr_loss) if args.eval_after_first_stage: result = evaluate(args, model, tokenizer, prefix="") print_result = get_wanted_result(result) train(args, train_dataset, model, tokenizer, train_highway=True) # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained() if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0): # Create output directory if needed if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]: os.makedirs(args.output_dir) logger.info("Saving model checkpoint to %s", args.output_dir) # Save a trained model, configuration and tokenizer using `save_pretrained()`. # They can then be reloaded using `from_pretrained()` model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir) # Good practice: save your training arguments together with the trained model torch.save(args, os.path.join(args.output_dir, 'training_args.bin')) # Load a trained model and vocabulary that you have fine-tuned model = model_class.from_pretrained(args.output_dir) tokenizer = tokenizer_class.from_pretrained(args.output_dir) model.to(args.device) # Evaluation results = {} if args.do_eval and args.local_rank in [-1, 0]: tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case) checkpoints = [args.output_dir] if args.eval_all_checkpoints: checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True))) logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging logger.info("Evaluate the following checkpoints: %s", checkpoints) for checkpoint in checkpoints: global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else "" prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else "" model = model_class.from_pretrained(checkpoint) if args.model_type=="bert": model.bert.encoder.set_early_exit_entropy(args.early_exit_entropy) else: model.roberta.encoder.set_early_exit_entropy(args.early_exit_entropy) model.to(args.device) result = evaluate(args, model, tokenizer, prefix=prefix, eval_highway=args.eval_highway) print_result = get_wanted_result(result) print("Result: {}".format(print_result)) if args.eval_each_highway: last_layer_results = print_result each_layer_results = [] for i in range(model.num_layers): logger.info("\n") _result = evaluate(args, model, tokenizer, prefix=prefix, output_layer=i, eval_highway=args.eval_highway) if i+1 < model.num_layers: each_layer_results.append(get_wanted_result(_result)) each_layer_results.append(last_layer_results) save_fname = args.plot_data_dir + '/' + args.model_name_or_path[2:] + "/each_layer.npy" if not os.path.exists(os.path.dirname(save_fname)): os.makedirs(os.path.dirname(save_fname)) np.save(save_fname, np.array(each_layer_results)) result = dict((k + '_{}'.format(global_step), v) for k, v in result.items()) results.update(result) return results if __name__ == "__main__": main()
33,078
51.423138
158
py
DeeBERT
DeeBERT-master/examples/run_glue.py
# coding=utf-8 # Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team. # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Finetuning the library models for sequence classification on GLUE (Bert, XLM, XLNet, RoBERTa).""" from __future__ import absolute_import, division, print_function import argparse import glob import logging import os import random import numpy as np import torch from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset) from torch.utils.data.distributed import DistributedSampler try: from torch.utils.tensorboard import SummaryWriter except: from tensorboardX import SummaryWriter from tqdm import tqdm, trange from transformers import (WEIGHTS_NAME, BertConfig, BertForSequenceClassification, BertTokenizer, RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer, XLMConfig, XLMForSequenceClassification, XLMTokenizer, XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer, DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer) # from transformers.modeling_highway_roberta import RobertaForSequenceClassification from transformers import AdamW, get_linear_schedule_with_warmup from transformers import glue_compute_metrics as compute_metrics from transformers import glue_output_modes as output_modes from transformers import glue_processors as processors from transformers import glue_convert_examples_to_features as convert_examples_to_features logger = logging.getLogger(__name__) ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig, DistilBertConfig)), ()) MODEL_CLASSES = { 'bert': (BertConfig, BertForSequenceClassification, BertTokenizer), 'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer), 'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer), 'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer), 'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer) } def set_seed(args): random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.n_gpu > 0: torch.cuda.manual_seed_all(args.seed) def train(args, train_dataset, model, tokenizer): """ Train the model """ if args.local_rank in [-1, 0]: tb_writer = SummaryWriter() args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu) train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset) train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size) if args.max_steps > 0: t_total = args.max_steps args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1 else: t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs # Prepare optimizer and schedule (linear warmup and decay) no_decay = ['bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [ {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0} ] optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total) if args.fp16: try: from apex import amp except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) # multi-gpu training (should be after apex fp16 initialization) if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Distributed training (should be after apex fp16 initialization) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True) # Train! logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Num Epochs = %d", args.num_train_epochs) logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size) logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d", args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1)) logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps) logger.info(" Total optimization steps = %d", t_total) global_step = 0 tr_loss, logging_loss = 0.0, 0.0 model.zero_grad() train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]) set_seed(args) # Added here for reproductibility (even between python 2 and 3) for _ in train_iterator: epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0]) for step, batch in enumerate(epoch_iterator): model.train() batch = tuple(t.to(args.device) for t in batch) inputs = {'input_ids': batch[0], 'attention_mask': batch[1], 'labels': batch[3]} if args.model_type != 'distilbert': inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None # XLM, DistilBERT and RoBERTa don't use segment_ids outputs = model(**inputs) loss = outputs[0] # model outputs are always tuple in transformers (see doc) if args.n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu parallel training if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() tr_loss += loss.item() if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) optimizer.step() scheduler.step() # Update learning rate schedule model.zero_grad() global_step += 1 if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0: # Log metrics if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well results = evaluate(args, model, tokenizer) for key, value in results.items(): tb_writer.add_scalar('eval_{}'.format(key), value, global_step) tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step) tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step) logging_loss = tr_loss if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0: # Save model checkpoint output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step)) if not os.path.exists(output_dir): os.makedirs(output_dir) model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save.save_pretrained(output_dir) torch.save(args, os.path.join(output_dir, 'training_args.bin')) logger.info("Saving model checkpoint to %s", output_dir) if args.max_steps > 0 and global_step > args.max_steps: epoch_iterator.close() break if args.max_steps > 0 and global_step > args.max_steps: train_iterator.close() break if args.local_rank in [-1, 0]: tb_writer.close() return global_step, tr_loss / global_step def evaluate(args, model, tokenizer, prefix=""): # Loop to handle MNLI double evaluation (matched, mis-matched) eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,) eval_outputs_dirs = (args.output_dir, args.output_dir + '-MM') if args.task_name == "mnli" else (args.output_dir,) results = {} for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs): eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True) if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]: os.makedirs(eval_output_dir) args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu) # Note that DistributedSampler samples randomly eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset) eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size) # multi-gpu eval if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Eval! logger.info("***** Running evaluation {} *****".format(prefix)) logger.info(" Num examples = %d", len(eval_dataset)) logger.info(" Batch size = %d", args.eval_batch_size) eval_loss = 0.0 nb_eval_steps = 0 preds = None out_label_ids = None for batch in tqdm(eval_dataloader, desc="Evaluating"): model.eval() batch = tuple(t.to(args.device) for t in batch) with torch.no_grad(): inputs = {'input_ids': batch[0], 'attention_mask': batch[1], 'labels': batch[3]} if args.model_type != 'distilbert': inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None # XLM, DistilBERT and RoBERTa don't use segment_ids outputs = model(**inputs) tmp_eval_loss, logits = outputs[:2] eval_loss += tmp_eval_loss.mean().item() nb_eval_steps += 1 if preds is None: preds = logits.detach().cpu().numpy() out_label_ids = inputs['labels'].detach().cpu().numpy() else: preds = np.append(preds, logits.detach().cpu().numpy(), axis=0) out_label_ids = np.append(out_label_ids, inputs['labels'].detach().cpu().numpy(), axis=0) eval_loss = eval_loss / nb_eval_steps if args.output_mode == "classification": preds = np.argmax(preds, axis=1) elif args.output_mode == "regression": preds = np.squeeze(preds) result = compute_metrics(eval_task, preds, out_label_ids) results.update(result) output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results {} *****".format(prefix)) for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) return results def load_and_cache_examples(args, task, tokenizer, evaluate=False): if args.local_rank not in [-1, 0] and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache processor = processors[task]() output_mode = output_modes[task] # Load data features from cache or dataset file cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}'.format( 'dev' if evaluate else 'train', list(filter(None, args.model_name_or_path.split('/'))).pop(), str(args.max_seq_length), str(task))) if os.path.exists(cached_features_file) and not args.overwrite_cache: logger.info("Loading features from cached file %s", cached_features_file) features = torch.load(cached_features_file) else: logger.info("Creating features from dataset file at %s", args.data_dir) label_list = processor.get_labels() if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta']: # HACK(label indices are swapped in RoBERTa pretrained model) label_list[1], label_list[2] = label_list[2], label_list[1] examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir) features = convert_examples_to_features(examples, tokenizer, label_list=label_list, max_length=args.max_seq_length, output_mode=output_mode, pad_on_left=bool(args.model_type in ['xlnet']), # pad on the left for xlnet pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0], pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0, ) if args.local_rank in [-1, 0]: logger.info("Saving features into cached file %s", cached_features_file) torch.save(features, cached_features_file) if args.local_rank == 0 and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache # Convert to Tensors and build dataset all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long) all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long) all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long) if output_mode == "classification": all_labels = torch.tensor([f.label for f in features], dtype=torch.long) elif output_mode == "regression": all_labels = torch.tensor([f.label for f in features], dtype=torch.float) dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels) return dataset def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--data_dir", default=None, type=str, required=True, help="The input data dir. Should contain the .tsv files (or other data files) for the task.") parser.add_argument("--model_type", default=None, type=str, required=True, help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys())) parser.add_argument("--model_name_or_path", default=None, type=str, required=True, help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS)) parser.add_argument("--task_name", default=None, type=str, required=True, help="The name of the task to train selected in the list: " + ", ".join(processors.keys())) parser.add_argument("--output_dir", default=None, type=str, required=True, help="The output directory where the model predictions and checkpoints will be written.") ## Other parameters parser.add_argument("--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name") parser.add_argument("--tokenizer_name", default="", type=str, help="Pretrained tokenizer name or path if not the same as model_name") parser.add_argument("--cache_dir", default="", type=str, help="Where do you want to store the pre-trained models downloaded from s3") parser.add_argument("--max_seq_length", default=128, type=int, help="The maximum total input sequence length after tokenization. Sequences longer " "than this will be truncated, sequences shorter will be padded.") parser.add_argument("--do_train", action='store_true', help="Whether to run training.") parser.add_argument("--do_eval", action='store_true', help="Whether to run eval on the dev set.") parser.add_argument("--evaluate_during_training", action='store_true', help="Rul evaluation during training at each logging step.") parser.add_argument("--do_lower_case", action='store_true', help="Set this flag if you are using an uncased model.") parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.") parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation.") parser.add_argument('--gradient_accumulation_steps', type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.") parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight deay if we apply some.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--max_steps", default=-1, type=int, help="If > 0: set total number of training steps to perform. Override num_train_epochs.") parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.") parser.add_argument('--logging_steps', type=int, default=50, help="Log every X updates steps.") parser.add_argument('--save_steps', type=int, default=50, help="Save checkpoint every X updates steps.") parser.add_argument("--eval_all_checkpoints", action='store_true', help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number") parser.add_argument("--no_cuda", action='store_true', help="Avoid using CUDA when available") parser.add_argument('--overwrite_output_dir', action='store_true', help="Overwrite the content of the output directory") parser.add_argument('--overwrite_cache', action='store_true', help="Overwrite the cached training and evaluation sets") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument('--fp16', action='store_true', help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit") parser.add_argument('--fp16_opt_level', type=str, default='O1', help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." "See details at https://nvidia.github.io/apex/amp.html") parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.") parser.add_argument('--server_port', type=str, default='', help="For distant debugging.") args = parser.parse_args() if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir: raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir)) # Setup distant debugging if needed if args.server_ip and args.server_port: # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script import ptvsd print("Waiting for debugger attach") ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True) ptvsd.wait_for_attach() # Setup CUDA, GPU & distributed training if args.local_rank == -1 or args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") args.n_gpu = torch.cuda.device_count() else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) torch.distributed.init_process_group(backend='nccl') args.n_gpu = 1 args.device = device # Setup logging logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt = '%m/%d/%Y %H:%M:%S', level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN) logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16) # Set seed set_seed(args) # Prepare GLUE task args.task_name = args.task_name.lower() if args.task_name not in processors: raise ValueError("Task not found: %s" % (args.task_name)) processor = processors[args.task_name]() args.output_mode = output_modes[args.task_name] label_list = processor.get_labels() num_labels = len(label_list) # Load pretrained model and tokenizer if args.local_rank not in [-1, 0]: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab args.model_type = args.model_type.lower() config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type] config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name, cache_dir=args.cache_dir if args.cache_dir else None) tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case, cache_dir=args.cache_dir if args.cache_dir else None) model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config, cache_dir=args.cache_dir if args.cache_dir else None) if args.local_rank == 0: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab model.to(args.device) logger.info("Training/evaluation parameters %s", args) # Training if args.do_train: train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False) global_step, tr_loss = train(args, train_dataset, model, tokenizer) logger.info(" global_step = %s, average loss = %s", global_step, tr_loss) # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained() if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0): # Create output directory if needed if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]: os.makedirs(args.output_dir) logger.info("Saving model checkpoint to %s", args.output_dir) # Save a trained model, configuration and tokenizer using `save_pretrained()`. # They can then be reloaded using `from_pretrained()` model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir) # Good practice: save your training arguments together with the trained model torch.save(args, os.path.join(args.output_dir, 'training_args.bin')) # Load a trained model and vocabulary that you have fine-tuned model = model_class.from_pretrained(args.output_dir) tokenizer = tokenizer_class.from_pretrained(args.output_dir) model.to(args.device) # Evaluation results = {} if args.do_eval and args.local_rank in [-1, 0]: tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case) checkpoints = [args.output_dir] if args.eval_all_checkpoints: checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True))) logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging logger.info("Evaluate the following checkpoints: %s", checkpoints) for checkpoint in checkpoints: global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else "" prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else "" model = model_class.from_pretrained(checkpoint) model.to(args.device) result = evaluate(args, model, tokenizer, prefix=prefix) result = dict((k + '_{}'.format(global_step), v) for k, v in result.items()) results.update(result) return results if __name__ == "__main__": main()
28,140
52.398482
158
py
DeeBERT
DeeBERT-master/examples/benchmarks.py
# coding=utf-8 # Copyright 2018 The HuggingFace Inc. team. # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Benchmarking the library on inference and training """ # If checking the tensors placement # tf.debugging.set_log_device_placement(True) from typing import List import timeit from transformers import is_tf_available, is_torch_available from time import time import argparse import csv if is_tf_available(): import tensorflow as tf from transformers import TFAutoModel if is_torch_available(): import torch from transformers import AutoModel from transformers import AutoConfig, AutoTokenizer input_text = """Bent over their instruments, three hundred Fertilizers were plunged, as the Director of Hatcheries and Conditioning entered the room, in the scarcely breathing silence, the absent-minded, soliloquizing hum or whistle, of absorbed concentration. A troop of newly arrived students, very young, pink and callow, followed nervously, rather abjectly, at the Director's heels. Each of them carried a notebook, in which, whenever the great man spoke, he desperately scribbled. Straight from the horse's mouth. It was a rare privilege. The D. H. C. for Central London always made a point of personally conducting his new students round the various departments. "Just to give you a general idea," he would explain to them. For of course some sort of general idea they must have, if they were to do their work intelligently-though as little of one, if they were to be good and happy members of society, as possible. For particulars, as every one knows, make for virtue and happiness; generalities are intellectu- ally necessary evils. Not philosophers but fret-sawyers and stamp col- lectors compose the backbone of society. "To-morrow," he would add, smiling at them with a slightly menacing geniality, "you'll be settling down to serious work. You won't have time for generalities. Meanwhile ..." Meanwhile, it was a privilege. Straight from the horse's mouth into the notebook. The boys scribbled like mad. Tall and rather thin but upright, the Director advanced into the room. He had a long chin and big rather prominent teeth, just covered, when he was not talking, by his full, floridly curved lips. Old, young? Thirty? Fifty? Fifty-five? It was hard to say. And anyhow the question didn't arise; in this year of stability, A. F. 632, it didn't occur to you to ask it. "I shall begin at the beginning," said the D.H.C. and the more zealous students recorded his intention in their notebooks: Begin at the begin- ning. "These," he waved his hand, "are the incubators." And opening an insulated door he showed them racks upon racks of numbered test- tubes. "The week's supply of ova. Kept," he explained, "at blood heat; whereas the male gametes," and here he opened another door, "they have to be kept at thirty-five instead of thirty-seven. Full blood heat sterilizes." Rams wrapped in theremogene beget no lambs. Still leaning against the incubators he gave them, while the pencils scurried illegibly across the pages, a brief description of the modern fertilizing process; spoke first, of course, of its surgical introduc- tion-"the operation undergone voluntarily for the good of Society, not to mention the fact that it carries a bonus amounting to six months' salary"; continued with some account of the technique for preserving the excised ovary alive and actively developing; passed on to a consid- eration of optimum temperature, salinity, viscosity; referred to the liq- uor in which the detached and ripened eggs were kept; and, leading his charges to the work tables, actually showed them how this liquor was drawn off from the test-tubes; how it was let out drop by drop onto the specially warmed slides of the microscopes; how the eggs which it contained were inspected for abnormalities, counted and transferred to a porous receptacle; how (and he now took them to watch the operation) this receptacle was immersed in a warm bouillon containing free-swimming spermatozoa-at a minimum concentration of one hundred thousand per cubic centimetre, he insisted; and how, after ten minutes, the container was lifted out of the liquor and its contents re-examined; how, if any of the eggs remained unfertilized, it was again immersed, and, if necessary, yet again; how the fertilized ova went back to the incubators; where the Alphas and Betas re- mained until definitely bottled; while the Gammas, Deltas and Epsilons were brought out again, after only thirty-six hours, to undergo Bo- kanovsky's Process. "Bokanovsky's Process," repeated the Director, and the students un- derlined the words in their little notebooks. One egg, one embryo, one adult-normality. But a bokanovskified egg will bud, will proliferate, will divide. From eight to ninety-six buds, and every bud will grow into a perfectly formed embryo, and every embryo into a full-sized adult. Making ninety-six human beings grow where only one grew before. Progress. "Essentially," the D.H.C. concluded, "bokanovskification consists of a series of arrests of development. We check the normal growth and, paradoxically enough, the egg responds by budding." Responds by budding. The pencils were busy. He pointed. On a very slowly moving band a rack-full of test-tubes was entering a large metal box, another, rack-full was emerging. Machinery faintly purred. It took eight minutes for the tubes to go through, he told them. Eight minutes of hard X-rays being about as much as an egg can stand. A few died; of the rest, the least susceptible divided into two; most put out four buds; some eight; all were returned to the incubators, where the buds began to develop; then, after two days, were suddenly chilled, chilled and checked. Two, four, eight, the buds in their turn budded; and having budded were dosed almost to death with alcohol; consequently burgeoned again and having budded-bud out of bud out of bud-were thereafter-further arrest being generally fatal-left to develop in peace. By which time the original egg was in a fair way to becoming anything from eight to ninety-six embryos- a prodigious improvement, you will agree, on nature. Identical twins-but not in piddling twos and threes as in the old viviparous days, when an egg would sometimes accidentally divide; actually by dozens, by scores at a time. "Scores," the Director repeated and flung out his arms, as though he were distributing largesse. "Scores." But one of the students was fool enough to ask where the advantage lay. "My good boy!" The Director wheeled sharply round on him. "Can't you see? Can't you see?" He raised a hand; his expression was solemn. "Bokanovsky's Process is one of the major instruments of social stabil- ity!" Major instruments of social stability. Standard men and women; in uniform batches. The whole of a small factory staffed with the products of a single bokanovskified egg. "Ninety-six identical twins working ninety-six identical machines!" The voice was almost tremulous with enthusiasm. "You really know where you are. For the first time in history." He quoted the planetary motto. "Community, Identity, Stability." Grand words. "If we could bo- kanovskify indefinitely the whole problem would be solved." Solved by standard Gammas, unvarying Deltas, uniform Epsilons. Mil- lions of identical twins. The principle of mass production at last applied to biology. "But, alas," the Director shook his head, "we can't bokanovskify indefi- nitely." Ninety-six seemed to be the limit; seventy-two a good average. From the same ovary and with gametes of the same male to manufacture as many batches of identical twins as possible-that was the best (sadly a second best) that they could do. And even that was difficult. "For in nature it takes thirty years for two hundred eggs to reach ma- turity. But our business is to stabilize the population at this moment, here and now. Dribbling out twins over a quarter of a century-what would be the use of that?" Obviously, no use at all. But Podsnap's Technique had immensely ac- celerated the process of ripening. They could make sure of at least a hundred and fifty mature eggs within two years. Fertilize and bo- kanovskify-in other words, multiply by seventy-two-and you get an average of nearly eleven thousand brothers and sisters in a hundred and fifty batches of identical twins, all within two years of the same age. "And in exceptional cases we can make one ovary yield us over fifteen thousand adult individuals." Beckoning to a fair-haired, ruddy young man who happened to be passing at the moment. "Mr. Foster," he called. The ruddy young man approached. "Can you tell us the record for a single ovary, Mr. Foster?" "Sixteen thousand and twelve in this Centre," Mr. Foster replied with- out hesitation. He spoke very quickly, had a vivacious blue eye, and took an evident pleasure in quoting figures. "Sixteen thousand and twelve; in one hundred and eighty-nine batches of identicals. But of course they've done much better," he rattled on, "in some of the tropi- cal Centres. Singapore has often produced over sixteen thousand five hundred; and Mombasa has actually touched the seventeen thousand mark. But then they have unfair advantages. You should see the way a negro ovary responds to pituitary! It's quite astonishing, when you're used to working with European material. Still," he added, with a laugh (but the light of combat was in his eyes and the lift of his chin was challenging), "still, we mean to beat them if we can. I'm working on a wonderful Delta-Minus ovary at this moment. Only just eighteen months old. Over twelve thousand seven hundred children already, ei- ther decanted or in embryo. And still going strong. We'll beat them yet." "That's the spirit I like!" cried the Director, and clapped Mr. Foster on the shoulder. "Come along with us, and give these boys the benefit of your expert knowledge." Mr. Foster smiled modestly. "With pleasure." They went. In the Bottling Room all was harmonious bustle and ordered activity. Flaps of fresh sow's peritoneum ready cut to the proper size came shooting up in little lifts from the Organ Store in the sub-basement. Whizz and then, click! the lift-hatches hew open; the bottle-liner had only to reach out a hand, take the flap, insert, smooth-down, and be- fore the lined bottle had had time to travel out of reach along the end- less band, whizz, click! another flap of peritoneum had shot up from the depths, ready to be slipped into yet another bottle, the next of that slow interminable procession on the band. Next to the Liners stood the Matriculators. The procession advanced; one by one the eggs were transferred from their test-tubes to the larger containers; deftly the peritoneal lining was slit, the morula dropped into place, the saline solution poured in ... and already the bottle had passed, and it was the turn of the labellers. Heredity, date of fertilization, membership of Bokanovsky Group-details were trans- ferred from test-tube to bottle. No longer anonymous, but named, identified, the procession marched slowly on; on through an opening in the wall, slowly on into the Social Predestination Room. "Eighty-eight cubic metres of card-index," said Mr. Foster with relish, as they entered.""" def create_setup_and_compute(model_names: List[str], gpu: bool = True, tensorflow: bool = False, average_over: int = 3, torchscript: bool = False, xla: bool = False, amp: bool = False, fp16: bool = False, save_to_csv: bool = False, csv_filename: str = f"results_{round(time())}.csv"): if xla: tf.config.optimizer.set_jit(True) if amp: tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True}) if tensorflow: dictionary = {model_name: {} for model_name in model_names} results = _compute_tensorflow(model_names, dictionary, average_over, amp) else: device = 'cuda' if (gpu and torch.cuda.is_available()) else 'cpu' dictionary = {model_name: {} for model_name in model_names} results = _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16) print("=========== RESULTS ===========") for model_name in model_names: print("\t" + f"======= MODEL CHECKPOINT: {model_name} =======") for batch_size in results[model_name]["bs"]: print("\t\t" + f"===== BATCH SIZE: {batch_size} =====") for slice_size in results[model_name]["ss"]: result = results[model_name]['results'][batch_size][slice_size] if isinstance(result, str): print(f"\t\t{model_name}/{batch_size}/{slice_size}: " f"{result}") else: print(f"\t\t{model_name}/{batch_size}/{slice_size}: " f"{(round(1000 * result) / 1000)}" f"s") if save_to_csv: with open(csv_filename, mode='w') as csv_file: fieldnames = ['model', '1x8', '1x64', '1x128', '1x256', '1x512', '1x1024', '2x8', '2x64', '2x128', '2x256', '2x512', '2x1024', '4x8', '4x64', '4x128', '4x256', '4x512', '4x1024', '8x8', '8x64', '8x128', '8x256', '8x512', '8x1024', ] writer = csv.DictWriter(csv_file, fieldnames=fieldnames) writer.writeheader() for model_name in model_names: model_results = { f'{bs}x{ss}': results[model_name]['results'][bs][ss] for bs in results[model_name]["results"] for ss in results[model_name]['results'][bs] } writer.writerow({'model': model_name, **model_results}) def _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16): for c, model_name in enumerate(model_names): print(f"{c + 1} / {len(model_names)}") config = AutoConfig.from_pretrained(model_name, torchscript=torchscript) model = AutoModel.from_pretrained(model_name, config=config) tokenizer = AutoTokenizer.from_pretrained(model_name) tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False) max_input_size = tokenizer.max_model_input_sizes[model_name] batch_sizes = [1, 2, 4, 8] slice_sizes = [8, 64, 128, 256, 512, 1024] dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}} dictionary[model_name]["results"] = {i: {} for i in batch_sizes} for batch_size in batch_sizes: if fp16: model.half() model.to(device) model.eval() for slice_size in slice_sizes: if max_input_size is not None and slice_size > max_input_size: dictionary[model_name]["results"][batch_size][slice_size] = "N/A" else: sequence = torch.tensor(tokenized_sequence[:slice_size], device=device).repeat(batch_size, 1) try: if torchscript: print("Tracing model with sequence size", sequence.shape) inference = torch.jit.trace(model, sequence) inference(sequence) else: inference = model inference(sequence) print("Going through model with sequence of shape", sequence.shape) runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3) average_time = sum(runtimes)/float(len(runtimes)) / 3.0 dictionary[model_name]["results"][batch_size][slice_size] = average_time except RuntimeError as e: print("Doesn't fit on GPU.", e) torch.cuda.empty_cache() dictionary[model_name]["results"][batch_size][slice_size] = "N/A" return dictionary def _compute_tensorflow(model_names, dictionary, average_over, amp): for c, model_name in enumerate(model_names): print(f"{c + 1} / {len(model_names)}") config = AutoConfig.from_pretrained(model_name) model = TFAutoModel.from_pretrained(model_name, config=config) tokenizer = AutoTokenizer.from_pretrained(model_name) tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False) max_input_size = tokenizer.max_model_input_sizes[model_name] batch_sizes = [1, 2, 4, 8] slice_sizes = [8, 64, 128, 256, 512, 1024] dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}} dictionary[model_name]["results"] = {i: {} for i in batch_sizes} print("Using model", model) @tf.function def inference(inputs): return model(inputs) for batch_size in batch_sizes: for slice_size in slice_sizes: if max_input_size is not None and slice_size > max_input_size: dictionary[model_name]["results"][batch_size][slice_size] = "N/A" else: sequence = tf.stack([tf.squeeze(tf.constant(tokenized_sequence[:slice_size])[None, :])] * batch_size) try: print("Going through model with sequence of shape", sequence.shape) # To make sure that the model is traced + that the tensors are on the appropriate device inference(sequence) runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3) average_time = sum(runtimes)/float(len(runtimes)) / 3.0 dictionary[model_name]["results"][batch_size][slice_size] = average_time except tf.errors.ResourceExhaustedError as e: print("Doesn't fit on GPU.", e) torch.cuda.empty_cache() dictionary[model_name]["results"][batch_size][slice_size] = "N/A" return dictionary def main(): parser = argparse.ArgumentParser() parser.add_argument("--models", required=False, type=str, default='all', help="Model checkpoints to be provided " "to the AutoModel classes. Leave " "blank to benchmark the base version " "of all available model " "architectures.") parser.add_argument("--torch", required=False, action="store_true", help="Benchmark the Pytorch version of the " "models") parser.add_argument("--torch_cuda", required=False, action="store_true", help="Pytorch only: run on available " "cuda devices") parser.add_argument("--torchscript", required=False, action="store_true", help="Pytorch only: trace the models " "using torchscript") parser.add_argument("--tensorflow", required=False, action="store_true", help="Benchmark the TensorFlow version " "of the models. Will run on GPU if " "the correct dependencies are " "installed") parser.add_argument("--xla", required=False, action="store_true", help="TensorFlow only: use XLA acceleration.") parser.add_argument("--amp", required=False, action="store_true", help="TensorFlow only: use automatic mixed precision acceleration.") parser.add_argument("--fp16", required=False, action="store_true", help="PyTorch only: use FP16 to accelerate inference.") parser.add_argument("--keras_predict", required=False, action="store_true", help="Whether to use model.predict " "instead of model() to do a " "forward pass.") parser.add_argument("--save_to_csv", required=False, action="store_true", help="Save to a CSV file.") parser.add_argument("--csv_filename", required=False, default=None, help="CSV filename used if saving results to csv.") parser.add_argument("--average_over", required=False, default=30, type=int, help="Times an experiment will be run.") args = parser.parse_args() if args.models == 'all': args.models = [ "gpt2", "bert-base-cased", "xlnet-base-cased", "xlm-mlm-en-2048", "transfo-xl-wt103", "openai-gpt", "distilbert-base-uncased", "distilgpt2", "roberta-base", "ctrl" ] else: args.models = args.models.split() print("Running with arguments", args) if args.torch: if is_torch_available(): create_setup_and_compute( model_names=args.models, tensorflow=False, gpu=args.torch_cuda, torchscript=args.torchscript, fp16=args.fp16, save_to_csv=args.save_to_csv, csv_filename=args.csv_filename, average_over=args.average_over ) else: raise ImportError("Trying to run a PyTorch benchmark but PyTorch was not found in the environment.") if args.tensorflow: if is_tf_available(): create_setup_and_compute( model_names=args.models, tensorflow=True, xla=args.xla, amp=args.amp, save_to_csv=args.save_to_csv, csv_filename=args.csv_filename, average_over=args.average_over ) else: raise ImportError("Trying to run a TensorFlow benchmark but TensorFlow was not found in the environment.") if __name__ == '__main__': main()
23,631
48.439331
138
py
DeeBERT
DeeBERT-master/examples/run_summarization_finetuning.py
# coding=utf-8 # Copyright 2019 The HuggingFace Inc. team. # Copyright (c) 2019 The HuggingFace Inc. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Finetuning seq2seq models for sequence generation.""" import argparse import functools import logging import os import random import sys import numpy as np from tqdm import tqdm, trange import torch from torch.optim import Adam from torch.utils.data import DataLoader, RandomSampler, SequentialSampler from transformers import ( AutoTokenizer, BertForMaskedLM, BertConfig, PreTrainedEncoderDecoder, Model2Model, ) from utils_summarization import ( CNNDailyMailDataset, encode_for_summarization, fit_to_block_size, build_lm_labels, build_mask, compute_token_type_ids, ) logger = logging.getLogger(__name__) logging.basicConfig(stream=sys.stdout, level=logging.INFO) def set_seed(args): random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) # ------------ # Load dataset # ------------ def load_and_cache_examples(args, tokenizer): dataset = CNNDailyMailDataset(tokenizer, data_dir=args.data_dir) return dataset def collate(data, tokenizer, block_size): """ List of tuple as an input. """ # remove the files with empty an story/summary, encode and fit to block data = filter(lambda x: not (len(x[0]) == 0 or len(x[1]) == 0), data) data = [ encode_for_summarization(story, summary, tokenizer) for story, summary in data ] data = [ ( fit_to_block_size(story, block_size, tokenizer.pad_token_id), fit_to_block_size(summary, block_size, tokenizer.pad_token_id), ) for story, summary in data ] stories = torch.tensor([story for story, summary in data]) summaries = torch.tensor([summary for story, summary in data]) encoder_token_type_ids = compute_token_type_ids(stories, tokenizer.cls_token_id) encoder_mask = build_mask(stories, tokenizer.pad_token_id) decoder_mask = build_mask(summaries, tokenizer.pad_token_id) lm_labels = build_lm_labels(summaries, tokenizer.pad_token_id) return ( stories, summaries, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels, ) # ---------- # Optimizers # ---------- class BertSumOptimizer(object): """ Specific optimizer for BertSum. As described in [1], the authors fine-tune BertSum for abstractive summarization using two Adam Optimizers with different warm-up steps and learning rate. They also use a custom learning rate scheduler. [1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders." arXiv preprint arXiv:1908.08345 (2019). """ def __init__(self, model, lr, warmup_steps, beta_1=0.99, beta_2=0.999, eps=1e-8): self.encoder = model.encoder self.decoder = model.decoder self.lr = lr self.warmup_steps = warmup_steps self.optimizers = { "encoder": Adam( model.encoder.parameters(), lr=lr["encoder"], betas=(beta_1, beta_2), eps=eps, ), "decoder": Adam( model.decoder.parameters(), lr=lr["decoder"], betas=(beta_1, beta_2), eps=eps, ), } self._step = 0 def _update_rate(self, stack): return self.lr[stack] * min( self._step ** (-0.5), self._step * self.warmup_steps[stack] ** (-0.5) ) def zero_grad(self): self.optimizer_decoder.zero_grad() self.optimizer_encoder.zero_grad() def step(self): self._step += 1 for stack, optimizer in self.optimizers.items(): new_rate = self._update_rate(stack) for param_group in optimizer.param_groups: param_group["lr"] = new_rate optimizer.step() # ------------ # Train # ------------ def train(args, model, tokenizer): """ Fine-tune the pretrained model on the corpus. """ set_seed(args) # Load the data args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu) train_dataset = load_and_cache_examples(args, tokenizer) train_sampler = RandomSampler(train_dataset) model_collate_fn = functools.partial(collate, tokenizer=tokenizer, block_size=512) train_dataloader = DataLoader( train_dataset, sampler=train_sampler, batch_size=args.train_batch_size, collate_fn=model_collate_fn, ) # Training schedule if args.max_steps > 0: t_total = args.max_steps args.num_train_epochs = t_total // ( len(train_dataloader) // args.gradient_accumulation_steps + 1 ) else: t_total = ( len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs ) # Prepare the optimizer lr = {"encoder": 0.002, "decoder": 0.2} warmup_steps = {"encoder": 20000, "decoder": 10000} optimizer = BertSumOptimizer(model, lr, warmup_steps) # Train logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Num Epochs = %d", args.num_train_epochs) logger.info( " Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size ) logger.info( " Total train batch size (w. parallel, distributed & accumulation) = %d", args.train_batch_size * args.gradient_accumulation_steps # * (torch.distributed.get_world_size() if args.local_rank != -1 else 1), ) logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps) logger.info(" Total optimization steps = %d", t_total) model.zero_grad() train_iterator = trange(args.num_train_epochs, desc="Epoch", disable=True) global_step = 0 tr_loss = 0.0 for _ in train_iterator: epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=True) for step, batch in enumerate(epoch_iterator): source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch source = source.to(args.device) target = target.to(args.device) encoder_token_type_ids = encoder_token_type_ids.to(args.device) encoder_mask = encoder_mask.to(args.device) decoder_mask = decoder_mask.to(args.device) lm_labels = lm_labels.to(args.device) model.train() outputs = model( source, target, encoder_token_type_ids=encoder_token_type_ids, encoder_attention_mask=encoder_mask, decoder_attention_mask=decoder_mask, decoder_lm_labels=lm_labels, ) loss = outputs[0] print(loss) if args.gradient_accumulation_steps > 1: loss /= args.gradient_accumulation_steps loss.backward() tr_loss += loss.item() if (step + 1) % args.gradient_accumulation_steps == 0: torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) optimizer.step() model.zero_grad() global_step += 1 if args.max_steps > 0 and global_step > args.max_steps: epoch_iterator.close() break if args.max_steps > 0 and global_step > args.max_steps: train_iterator.close() break return global_step, tr_loss / global_step # ------------ # Train # ------------ def evaluate(args, model, tokenizer, prefix=""): set_seed(args) args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu) eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True) eval_sampler = SequentialSampler(eval_dataset) eval_dataloader = DataLoader( eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size ) # multi-gpu evaluate if args.n_gpu > 1: model = torch.nn.DataParallel(model) logger.info("***** Running evaluation {} *****".format(prefix)) logger.info(" Num examples = %d", len(eval_dataset)) logger.info(" Batch size = %d", args.eval_batch_size) eval_loss = 0.0 nb_eval_steps = 0 model.eval() for batch in tqdm(eval_dataloader, desc="Evaluating"): source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch source = source.to(args.device) target = target.to(args.device) encoder_token_type_ids = encoder_token_type_ids.to(args.device) encoder_mask = encoder_mask.to(args.device) decoder_mask = decoder_mask.to(args.device) lm_labels = lm_labels.to(args.device) with torch.no_grad(): outputs = model( source, target, encoder_token_type_ids=encoder_token_type_ids, encoder_attention_mask=encoder_mask, decoder_attention_mask=decoder_mask, decoder_lm_labels=lm_labels, ) lm_loss = outputs[0] eval_loss += lm_loss.mean().item() nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps perplexity = torch.exp(torch.tensor(eval_loss)) result = {"perplexity": perplexity} # Save the evaluation's results output_eval_file = os.path.join(args.output_dir, "eval_results.txt") if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) with open(output_eval_file, "w") as writer: logger.info("***** Eval results {} *****".format(prefix)) for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) return result def main(): parser = argparse.ArgumentParser() # Required parameters parser.add_argument( "--data_dir", default=None, type=str, required=True, help="The input training data file (a text file).", ) parser.add_argument( "--output_dir", default=None, type=str, required=True, help="The output directory where the model predictions and checkpoints will be written.", ) # Optional parameters parser.add_argument( "--gradient_accumulation_steps", type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.", ) parser.add_argument( "--do_evaluate", type=bool, default=False, help="Run model evaluation on out-of-sample data.", ) parser.add_argument("--do_train", type=bool, default=False, help="Run training.") parser.add_argument( "--do_overwrite_output_dir", type=bool, default=False, help="Whether to overwrite the output dir.", ) parser.add_argument( "--model_name_or_path", default="bert-base-cased", type=str, help="The model checkpoint to initialize the encoder and decoder's weights with.", ) parser.add_argument( "--model_type", default="bert", type=str, help="The decoder architecture to be fine-tuned.", ) parser.add_argument( "--max_grad_norm", default=1.0, type=float, help="Max gradient norm." ) parser.add_argument( "--max_steps", default=-1, type=int, help="If > 0: set total number of training steps to perform. Override num_train_epochs.", ) parser.add_argument( "--to_cpu", default=False, type=bool, help="Whether to force training on CPU." ) parser.add_argument( "--num_train_epochs", default=10, type=int, help="Total number of training epochs to perform.", ) parser.add_argument( "--per_gpu_train_batch_size", default=4, type=int, help="Batch size per GPU/CPU for training.", ) parser.add_argument("--seed", default=42, type=int) args = parser.parse_args() if ( os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.do_overwrite_output_dir ): raise ValueError( "Output directory ({}) already exists and is not empty. Use --do_overwrite_output_dir to overwrite.".format( args.output_dir ) ) # Set up training device if args.to_cpu or not torch.cuda.is_available(): args.device = torch.device("cpu") args.n_gpu = 0 else: args.device = torch.device("cuda") args.n_gpu = torch.cuda.device_count() # Load pretrained model and tokenizer. The decoder's weights are randomly initialized. tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path) config = BertConfig.from_pretrained(args.model_name_or_path) decoder_model = BertForMaskedLM(config) model = Model2Model.from_pretrained( args.model_name_or_path, decoder_model=decoder_model ) # Setup logging logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger.warning( "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", 0, args.device, args.n_gpu, False, False, ) logger.info("Training/evaluation parameters %s", args) # Train the model model.to(args.device) if args.do_train: global_step, tr_loss = train(args, model, tokenizer) logger.info(" global_step = %s, average loss = %s", global_step, tr_loss) if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) logger.info("Saving model checkpoint to %s", args.output_dir) # Save a trained model, configuration and tokenizer using `save_pretrained()`. # They can then be reloaded using `from_pretrained()` model_to_save = ( model.module if hasattr(model, "module") else model ) # Take care of distributed/parallel training model_to_save.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir) torch.save(args, os.path.join(args.output_dir, "training_arguments.bin")) # Evaluate the model results = {} if args.do_evaluate: checkpoints = [] logger.info("Evaluate the following checkpoints: %s", checkpoints) for checkpoint in checkpoints: encoder_checkpoint = os.path.join(checkpoint, "encoder") decoder_checkpoint = os.path.join(checkpoint, "decoder") model = PreTrainedEncoderDecoder.from_pretrained( encoder_checkpoint, decoder_checkpoint ) model.to(args.device) results = "placeholder" return results if __name__ == "__main__": main()
15,727
30.902637
120
py
DeeBERT
DeeBERT-master/examples/run_bertology.py
#!/usr/bin/env python3 # Copyright 2018 CMU and The HuggingFace Inc. team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Bertology: this script shows how you can explore the internals of the models in the library to: - compute the entropy of the head attentions - compute the importance of each head - prune (remove) the low importance head. Some parts of this script are adapted from the code of Michel et al. (http://arxiv.org/abs/1905.10650) which is available at https://github.com/pmichel31415/are-16-heads-really-better-than-1 """ import os import argparse import logging from datetime import timedelta, datetime from tqdm import tqdm import numpy as np import torch from torch.utils.data import DataLoader, SequentialSampler, TensorDataset, Subset from torch.utils.data.distributed import DistributedSampler from torch.nn import CrossEntropyLoss, MSELoss from transformers import (WEIGHTS_NAME, BertConfig, BertForSequenceClassification, BertTokenizer, XLMConfig, XLMForSequenceClassification, XLMTokenizer, XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer) from run_glue import set_seed, load_and_cache_examples, ALL_MODELS, MODEL_CLASSES from transformers import glue_compute_metrics as compute_metrics from transformers import glue_output_modes as output_modes from transformers import glue_processors as processors logger = logging.getLogger(__name__) def entropy(p): """ Compute the entropy of a probability distribution """ plogp = p * torch.log(p) plogp[p == 0] = 0 return -plogp.sum(dim=-1) def print_2d_tensor(tensor): """ Print a 2D tensor """ logger.info("lv, h >\t" + "\t".join(f"{x + 1}" for x in range(len(tensor)))) for row in range(len(tensor)): if tensor.dtype != torch.long: logger.info(f"layer {row + 1}:\t" + "\t".join(f"{x:.5f}" for x in tensor[row].cpu().data)) else: logger.info(f"layer {row + 1}:\t" + "\t".join(f"{x:d}" for x in tensor[row].cpu().data)) def compute_heads_importance(args, model, eval_dataloader, compute_entropy=True, compute_importance=True, head_mask=None): """ This method shows how to compute: - head attention entropy - head importance scores according to http://arxiv.org/abs/1905.10650 """ # Prepare our tensors n_layers, n_heads = model.bert.config.num_hidden_layers, model.bert.config.num_attention_heads head_importance = torch.zeros(n_layers, n_heads).to(args.device) attn_entropy = torch.zeros(n_layers, n_heads).to(args.device) if head_mask is None: head_mask = torch.ones(n_layers, n_heads).to(args.device) head_mask.requires_grad_(requires_grad=True) preds = None labels = None tot_tokens = 0.0 for step, batch in enumerate(tqdm(eval_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])): batch = tuple(t.to(args.device) for t in batch) input_ids, input_mask, segment_ids, label_ids = batch # Do a forward pass (not with torch.no_grad() since we need gradients for importance score - see below) outputs = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids, head_mask=head_mask) loss, logits, all_attentions = outputs[0], outputs[1], outputs[-1] # Loss and logits are the first, attention the last loss.backward() # Backpropagate to populate the gradients in the head mask if compute_entropy: for layer, attn in enumerate(all_attentions): masked_entropy = entropy(attn.detach()) * input_mask.float().unsqueeze(1) attn_entropy[layer] += masked_entropy.sum(-1).sum(0).detach() if compute_importance: head_importance += head_mask.grad.abs().detach() # Also store our logits/labels if we want to compute metrics afterwards if preds is None: preds = logits.detach().cpu().numpy() labels = label_ids.detach().cpu().numpy() else: preds = np.append(preds, logits.detach().cpu().numpy(), axis=0) labels = np.append(labels, label_ids.detach().cpu().numpy(), axis=0) tot_tokens += input_mask.float().detach().sum().data # Normalize attn_entropy /= tot_tokens head_importance /= tot_tokens # Layerwise importance normalization if not args.dont_normalize_importance_by_layer: exponent = 2 norm_by_layer = torch.pow(torch.pow(head_importance, exponent).sum(-1), 1/exponent) head_importance /= norm_by_layer.unsqueeze(-1) + 1e-20 if not args.dont_normalize_global_importance: head_importance = (head_importance - head_importance.min()) / (head_importance.max() - head_importance.min()) # Print/save matrices np.save(os.path.join(args.output_dir, 'attn_entropy.npy'), attn_entropy.detach().cpu().numpy()) np.save(os.path.join(args.output_dir, 'head_importance.npy'), head_importance.detach().cpu().numpy()) logger.info("Attention entropies") print_2d_tensor(attn_entropy) logger.info("Head importance scores") print_2d_tensor(head_importance) logger.info("Head ranked by importance scores") head_ranks = torch.zeros(head_importance.numel(), dtype=torch.long, device=args.device) head_ranks[head_importance.view(-1).sort(descending=True)[1]] = torch.arange(head_importance.numel(), device=args.device) head_ranks = head_ranks.view_as(head_importance) print_2d_tensor(head_ranks) return attn_entropy, head_importance, preds, labels def mask_heads(args, model, eval_dataloader): """ This method shows how to mask head (set some heads to zero), to test the effect on the network, based on the head importance scores, as described in Michel et al. (http://arxiv.org/abs/1905.10650) """ _, head_importance, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False) preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds) original_score = compute_metrics(args.task_name, preds, labels)[args.metric_name] logger.info("Pruning: original score: %f, threshold: %f", original_score, original_score * args.masking_threshold) new_head_mask = torch.ones_like(head_importance) num_to_mask = max(1, int(new_head_mask.numel() * args.masking_amount)) current_score = original_score while current_score >= original_score * args.masking_threshold: head_mask = new_head_mask.clone() # save current head mask # heads from least important to most - keep only not-masked heads head_importance[head_mask == 0.0] = float('Inf') current_heads_to_mask = head_importance.view(-1).sort()[1] if len(current_heads_to_mask) <= num_to_mask: break # mask heads current_heads_to_mask = current_heads_to_mask[:num_to_mask] logger.info("Heads to mask: %s", str(current_heads_to_mask.tolist())) new_head_mask = new_head_mask.view(-1) new_head_mask[current_heads_to_mask] = 0.0 new_head_mask = new_head_mask.view_as(head_mask) print_2d_tensor(new_head_mask) # Compute metric and head importance again _, head_importance, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False, head_mask=new_head_mask) preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds) current_score = compute_metrics(args.task_name, preds, labels)[args.metric_name] logger.info("Masking: current score: %f, remaning heads %d (%.1f percents)", current_score, new_head_mask.sum(), new_head_mask.sum()/new_head_mask.numel() * 100) logger.info("Final head mask") print_2d_tensor(head_mask) np.save(os.path.join(args.output_dir, 'head_mask.npy'), head_mask.detach().cpu().numpy()) return head_mask def prune_heads(args, model, eval_dataloader, head_mask): """ This method shows how to prune head (remove heads weights) based on the head importance scores as described in Michel et al. (http://arxiv.org/abs/1905.10650) """ # Try pruning and test time speedup # Pruning is like masking but we actually remove the masked weights before_time = datetime.now() _, _, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False, compute_importance=False, head_mask=head_mask) preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds) score_masking = compute_metrics(args.task_name, preds, labels)[args.metric_name] original_time = datetime.now() - before_time original_num_params = sum(p.numel() for p in model.parameters()) heads_to_prune = dict((layer, (1 - head_mask[layer].long()).nonzero().tolist()) for layer in range(len(head_mask))) assert sum(len(h) for h in heads_to_prune.values()) == (1 - head_mask.long()).sum().item() model.prune_heads(heads_to_prune) pruned_num_params = sum(p.numel() for p in model.parameters()) before_time = datetime.now() _, _, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False, compute_importance=False, head_mask=None) preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds) score_pruning = compute_metrics(args.task_name, preds, labels)[args.metric_name] new_time = datetime.now() - before_time logger.info("Pruning: original num of params: %.2e, after pruning %.2e (%.1f percents)", original_num_params, pruned_num_params, pruned_num_params/original_num_params * 100) logger.info("Pruning: score with masking: %f score with pruning: %f", score_masking, score_pruning) logger.info("Pruning: speed ratio (new timing / original timing): %f percents", original_time/new_time * 100) def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--data_dir", default=None, type=str, required=True, help="The input data dir. Should contain the .tsv files (or other data files) for the task.") parser.add_argument("--model_name_or_path", default=None, type=str, required=True, help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join( ALL_MODELS)) parser.add_argument("--task_name", default=None, type=str, required=True, help="The name of the task to train selected in the list: " + ", ".join(processors.keys())) parser.add_argument("--output_dir", default=None, type=str, required=True, help="The output directory where the model predictions and checkpoints will be written.") ## Other parameters parser.add_argument("--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name_or_path") parser.add_argument("--tokenizer_name", default="", type=str, help="Pretrained tokenizer name or path if not the same as model_name_or_path") parser.add_argument("--cache_dir", default="", type=str, help="Where do you want to store the pre-trained models downloaded from s3") parser.add_argument("--data_subset", type=int, default=-1, help="If > 0: limit the data to a subset of data_subset instances.") parser.add_argument("--overwrite_output_dir", action='store_true', help="Whether to overwrite data in output directory") parser.add_argument('--overwrite_cache', action='store_true', help="Overwrite the cached training and evaluation sets") parser.add_argument("--dont_normalize_importance_by_layer", action='store_true', help="Don't normalize importance score by layers") parser.add_argument("--dont_normalize_global_importance", action='store_true', help="Don't normalize all importance scores between 0 and 1") parser.add_argument("--try_masking", action='store_true', help="Whether to try to mask head until a threshold of accuracy.") parser.add_argument("--masking_threshold", default=0.9, type=float, help="masking threshold in term of metrics (stop masking when metric < threshold * original metric value).") parser.add_argument("--masking_amount", default=0.1, type=float, help="Amount to heads to masking at each masking step.") parser.add_argument("--metric_name", default="acc", type=str, help="Metric to use for head masking.") parser.add_argument("--max_seq_length", default=128, type=int, help="The maximum total input sequence length after WordPiece tokenization. \n" "Sequences longer than this will be truncated, sequences shorter padded.") parser.add_argument("--batch_size", default=1, type=int, help="Batch size.") parser.add_argument("--seed", type=int, default=42) parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available") parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.") parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.") args = parser.parse_args() if args.server_ip and args.server_port: # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script import ptvsd print("Waiting for debugger attach") ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True) ptvsd.wait_for_attach() # Setup devices and distributed training if args.local_rank == -1 or args.no_cuda: args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") args.n_gpu = torch.cuda.device_count() else: torch.cuda.set_device(args.local_rank) args.device = torch.device("cuda", args.local_rank) args.n_gpu = 1 torch.distributed.init_process_group(backend='nccl') # Initializes the distributed backend # Setup logging logging.basicConfig(level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN) logger.info("device: {} n_gpu: {}, distributed: {}".format(args.device, args.n_gpu, bool(args.local_rank != -1))) # Set seeds set_seed(args) # Prepare GLUE task args.task_name = args.task_name.lower() if args.task_name not in processors: raise ValueError("Task not found: %s" % (args.task_name)) processor = processors[args.task_name]() args.output_mode = output_modes[args.task_name] label_list = processor.get_labels() num_labels = len(label_list) # Load pretrained model and tokenizer if args.local_rank not in [-1, 0]: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab args.model_type = "" for key in MODEL_CLASSES: if key in args.model_name_or_path.lower(): args.model_type = key # take the first match in model types break config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type] config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name, output_attentions=True, cache_dir=args.cache_dir if args.cache_dir else None) tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, cache_dir=args.cache_dir if args.cache_dir else None) model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config, cache_dir=args.cache_dir if args.cache_dir else None) if args.local_rank == 0: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab # Distributed and parallel training model.to(args.device) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True) elif args.n_gpu > 1: model = torch.nn.DataParallel(model) # Print/save training arguments torch.save(args, os.path.join(args.output_dir, 'run_args.bin')) logger.info("Training/evaluation parameters %s", args) # Prepare dataset for the GLUE task eval_data = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=True) if args.data_subset > 0: eval_data = Subset(eval_data, list(range(min(args.data_subset, len(eval_data))))) eval_sampler = SequentialSampler(eval_data) if args.local_rank == -1 else DistributedSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.batch_size) # Compute head entropy and importance score compute_heads_importance(args, model, eval_dataloader) # Try head masking (set heads to zero until the score goes under a threshole) # and head pruning (remove masked heads and see the effect on the network) if args.try_masking and args.masking_threshold > 0.0 and args.masking_threshold < 1.0: head_mask = mask_heads(args, model, eval_dataloader) prune_heads(args, model, eval_dataloader, head_mask) if __name__ == '__main__': main()
18,901
51.798883
177
py
DeeBERT
DeeBERT-master/examples/utils_summarization_test.py
# coding=utf-8 # Copyright 2019 HuggingFace Inc. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import unittest import numpy as np import torch from utils_summarization import ( compute_token_type_ids, fit_to_block_size, build_mask, build_lm_labels, process_story, ) class SummarizationDataProcessingTest(unittest.TestCase): def setUp(self): self.block_size = 10 def test_fit_to_block_sequence_too_small(self): """ Pad the sequence with 0 if the sequence is smaller than the block size.""" sequence = [1, 2, 3, 4] expected_output = [1, 2, 3, 4, 0, 0, 0, 0, 0, 0] self.assertEqual( fit_to_block_size(sequence, self.block_size, 0), expected_output ) def test_fit_to_block_sequence_fit_exactly(self): """ Do nothing if the sequence is the right size. """ sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] self.assertEqual( fit_to_block_size(sequence, self.block_size, 0), expected_output ) def test_fit_to_block_sequence_too_big(self): """ Truncate the sequence if it is too long. """ sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13] expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] self.assertEqual( fit_to_block_size(sequence, self.block_size, 0), expected_output ) def test_process_story_no_highlights(self): """ Processing a story with no highlights returns an empty list for the summary. """ raw_story = """It was the year of Our Lord one thousand seven hundred and seventy-five.\n\nSpiritual revelations were conceded to England at that favoured period, as at this.""" _, summary_lines = process_story(raw_story) self.assertEqual(summary_lines, []) def test_process_empty_story(self): """ An empty story returns an empty collection of lines. """ raw_story = "" story_lines, summary_lines = process_story(raw_story) self.assertEqual(story_lines, []) self.assertEqual(summary_lines, []) def test_process_story_with_missing_period(self): raw_story = ( "It was the year of Our Lord one thousand seven hundred and " "seventy-five\n\nSpiritual revelations were conceded to England " "at that favoured period, as at this.\n@highlight\n\nIt was the best of times" ) story_lines, summary_lines = process_story(raw_story) expected_story_lines = [ "It was the year of Our Lord one thousand seven hundred and seventy-five.", "Spiritual revelations were conceded to England at that favoured period, as at this.", ] self.assertEqual(expected_story_lines, story_lines) expected_summary_lines = ["It was the best of times."] self.assertEqual(expected_summary_lines, summary_lines) def test_build_lm_labels_no_padding(self): sequence = torch.tensor([1, 2, 3, 4]) expected = sequence np.testing.assert_array_equal( build_lm_labels(sequence, 0).numpy(), expected.numpy() ) def test_build_lm_labels(self): sequence = torch.tensor([1, 2, 3, 4, 0, 0, 0]) expected = torch.tensor([1, 2, 3, 4, -1, -1, -1]) np.testing.assert_array_equal( build_lm_labels(sequence, 0).numpy(), expected.numpy() ) def test_build_mask_no_padding(self): sequence = torch.tensor([1, 2, 3, 4]) expected = torch.tensor([1, 1, 1, 1]) np.testing.assert_array_equal(build_mask(sequence, 0).numpy(), expected.numpy()) def test_build_mask(self): sequence = torch.tensor([1, 2, 3, 4, 23, 23, 23]) expected = torch.tensor([1, 1, 1, 1, 0, 0, 0]) np.testing.assert_array_equal( build_mask(sequence, 23).numpy(), expected.numpy() ) def test_build_mask_with_padding_equal_to_one(self): sequence = torch.tensor([8, 2, 3, 4, 1, 1, 1]) expected = torch.tensor([1, 1, 1, 1, 0, 0, 0]) np.testing.assert_array_equal(build_mask(sequence, 1).numpy(), expected.numpy()) def test_compute_token_type_ids(self): separator = 101 batch = torch.tensor( [[1, 2, 3, 4, 5, 6], [1, 2, 3, 101, 5, 6], [1, 101, 3, 4, 101, 6]] ) expected = torch.tensor( [[0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1], [0, 1, 1, 1, 0, 0]] ) result = compute_token_type_ids(batch, separator) np.testing.assert_array_equal(result, expected) if __name__ == "__main__": unittest.main()
5,178
36.80292
98
py
DeeBERT
DeeBERT-master/examples/run_generation.py
#!/usr/bin/env python3 # coding=utf-8 # Copyright 2018 Google AI, Google Brain and Carnegie Mellon University Authors and the HuggingFace Inc. team. # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Conditional text generation with the auto-regressive models of the library (GPT/GPT-2/CTRL/Transformer-XL/XLNet) """ from __future__ import absolute_import, division, print_function, unicode_literals import argparse import logging from tqdm import trange import torch import torch.nn.functional as F import numpy as np from transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig from transformers import GPT2LMHeadModel, GPT2Tokenizer from transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer from transformers import XLNetLMHeadModel, XLNetTokenizer from transformers import TransfoXLLMHeadModel, TransfoXLTokenizer from transformers import CTRLLMHeadModel, CTRLTokenizer from transformers import XLMWithLMHeadModel, XLMTokenizer logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt = '%m/%d/%Y %H:%M:%S', level = logging.INFO) logger = logging.getLogger(__name__) MAX_LENGTH = int(10000) # Hardcoded max length to avoid infinite loop ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig)), ()) MODEL_CLASSES = { 'gpt2': (GPT2LMHeadModel, GPT2Tokenizer), 'ctrl': (CTRLLMHeadModel, CTRLTokenizer), 'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer), 'xlnet': (XLNetLMHeadModel, XLNetTokenizer), 'transfo-xl': (TransfoXLLMHeadModel, TransfoXLTokenizer), 'xlm': (XLMWithLMHeadModel, XLMTokenizer), } # Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia # in https://github.com/rusiaaman/XLNet-gen#methodology # and https://medium.com/@amanrusia/xlnet-speaks-comparison-to-gpt-2-ea1a4e9ba39e PADDING_TEXT = """ In 1991, the remains of Russian Tsar Nicholas II and his family (except for Alexei and Maria) are discovered. The voice of Nicholas's young son, Tsarevich Alexei Nikolaevich, narrates the remainder of the story. 1883 Western Siberia, a young Grigori Rasputin is asked by his father and a group of men to perform magic. Rasputin has a vision and denounces one of the men as a horse thief. Although his father initially slaps him for making such an accusation, Rasputin watches as the man is chased outside and beaten. Twenty years later, Rasputin sees a vision of the Virgin Mary, prompting him to become a priest. Rasputin quickly becomes famous, with people, even a bishop, begging for his blessing. <eod> </s> <eos>""" def set_seed(args): np.random.seed(args.seed) torch.manual_seed(args.seed) if args.n_gpu > 0: torch.cuda.manual_seed_all(args.seed) def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')): """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering Args: logits: logits distribution shape (batch size x vocabulary size) top_k > 0: keep only top k tokens with highest probability (top-k filtering). top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering). Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751) From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317 """ top_k = min(top_k, logits.size(-1)) # Safety check if top_k > 0: # Remove all tokens with a probability less than the last token of the top-k indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None] logits[indices_to_remove] = filter_value if top_p > 0.0: sorted_logits, sorted_indices = torch.sort(logits, descending=True) cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1) # Remove tokens with cumulative probability above the threshold sorted_indices_to_remove = cumulative_probs > top_p # Shift the indices to the right to keep also the first token above the threshold sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone() sorted_indices_to_remove[..., 0] = 0 # scatter sorted tensors to original indexing indices_to_remove = sorted_indices_to_remove.scatter(dim=1, index=sorted_indices, src=sorted_indices_to_remove) logits[indices_to_remove] = filter_value return logits def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0, is_xlnet=False, is_xlm_mlm=False, xlm_mask_token=None, xlm_lang=None, device='cpu'): context = torch.tensor(context, dtype=torch.long, device=device) context = context.unsqueeze(0).repeat(num_samples, 1) generated = context with torch.no_grad(): for _ in trange(length): inputs = {'input_ids': generated} if is_xlnet: # XLNet is a direct (predict same token, not next token) and bi-directional model by default # => need one additional dummy token in the input (will be masked), attention mask and target mapping (see model docstring) input_ids = torch.cat((generated, torch.zeros((1, 1), dtype=torch.long, device=device)), dim=1) perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float, device=device) perm_mask[:, :, -1] = 1.0 # Previous tokens don't see last token target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float, device=device) target_mapping[0, 0, -1] = 1.0 # predict last token inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping} if is_xlm_mlm and xlm_mask_token: # XLM MLM models are direct models (predict same token, not next token) # => need one additional dummy token in the input (will be masked and guessed) input_ids = torch.cat((generated, torch.full((1, 1), xlm_mask_token, dtype=torch.long, device=device)), dim=1) inputs = {'input_ids': input_ids} if xlm_lang is not None: inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1) outputs = model(**inputs) # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states) next_token_logits = outputs[0][:, -1, :] / (temperature if temperature > 0 else 1.) # repetition penalty from CTRL (https://arxiv.org/abs/1909.05858) for i in range(num_samples): for _ in set(generated[i].tolist()): next_token_logits[i, _] /= repetition_penalty filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p) if temperature == 0: # greedy sampling: next_token = torch.argmax(filtered_logits, dim=-1).unsqueeze(-1) else: next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1) generated = torch.cat((generated, next_token), dim=1) return generated def main(): parser = argparse.ArgumentParser() parser.add_argument("--model_type", default=None, type=str, required=True, help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys())) parser.add_argument("--model_name_or_path", default=None, type=str, required=True, help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS)) parser.add_argument("--prompt", type=str, default="") parser.add_argument("--padding_text", type=str, default="") parser.add_argument("--xlm_lang", type=str, default="", help="Optional language when used with the XLM model.") parser.add_argument("--length", type=int, default=20) parser.add_argument("--num_samples", type=int, default=1) parser.add_argument("--temperature", type=float, default=1.0, help="temperature of 0 implies greedy sampling") parser.add_argument("--repetition_penalty", type=float, default=1.0, help="primarily useful for CTRL model; in that case, use 1.2") parser.add_argument("--top_k", type=int, default=0) parser.add_argument("--top_p", type=float, default=0.9) parser.add_argument("--no_cuda", action='store_true', help="Avoid using CUDA when available") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument('--stop_token', type=str, default=None, help="Token at which text generation is stopped") args = parser.parse_args() args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") args.n_gpu = torch.cuda.device_count() set_seed(args) args.model_type = args.model_type.lower() model_class, tokenizer_class = MODEL_CLASSES[args.model_type] tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path) model = model_class.from_pretrained(args.model_name_or_path) model.to(args.device) model.eval() if args.length < 0 and model.config.max_position_embeddings > 0: args.length = model.config.max_position_embeddings elif 0 < model.config.max_position_embeddings < args.length: args.length = model.config.max_position_embeddings # No generation bigger than model size elif args.length < 0: args.length = MAX_LENGTH # avoid infinite loop logger.info(args) if args.model_type in ["ctrl"]: if args.temperature > 0.7: logger.info('CTRL typically works better with lower temperatures (and lower top_k).') while True: xlm_lang = None # XLM Language usage detailed in the issues #1414 if args.model_type in ["xlm"] and hasattr(tokenizer, 'lang2id') and hasattr(model.config, 'use_lang_emb') \ and model.config.use_lang_emb: if args.xlm_lang: language = args.xlm_lang else: language = None while language not in tokenizer.lang2id.keys(): language = input("Using XLM. Select language in " + str(list(tokenizer.lang2id.keys())) + " >>> ") xlm_lang = tokenizer.lang2id[language] # XLM masked-language modeling (MLM) models need masked token (see details in sample_sequence) is_xlm_mlm = args.model_type in ["xlm"] and 'mlm' in args.model_name_or_path if is_xlm_mlm: xlm_mask_token = tokenizer.mask_token_id else: xlm_mask_token = None raw_text = args.prompt if args.prompt else input("Model prompt >>> ") if args.model_type in ["transfo-xl", "xlnet"]: # Models with memory likes to have a long prompt for short inputs. raw_text = (args.padding_text if args.padding_text else PADDING_TEXT) + raw_text context_tokens = tokenizer.encode(raw_text, add_special_tokens=False) if args.model_type == "ctrl": if not any(context_tokens[0] == x for x in tokenizer.control_codes.values()): logger.info("WARNING! You are not starting your generation from a control code so you won't get good results") out = sample_sequence( model=model, context=context_tokens, num_samples=args.num_samples, length=args.length, temperature=args.temperature, top_k=args.top_k, top_p=args.top_p, repetition_penalty=args.repetition_penalty, is_xlnet=bool(args.model_type == "xlnet"), is_xlm_mlm=is_xlm_mlm, xlm_mask_token=xlm_mask_token, xlm_lang=xlm_lang, device=args.device, ) out = out[:, len(context_tokens):].tolist() for o in out: text = tokenizer.decode(o, clean_up_tokenization_spaces=True) text = text[: text.find(args.stop_token) if args.stop_token else None] print(text) if args.prompt: break return text if __name__ == '__main__': main()
13,112
49.241379
167
py
DeeBERT
DeeBERT-master/examples/run_ner.py
# coding=utf-8 # Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team. # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Fine-tuning the library models for named entity recognition on CoNLL-2003 (Bert or Roberta). """ from __future__ import absolute_import, division, print_function import argparse import glob import logging import os import random import numpy as np import torch from seqeval.metrics import precision_score, recall_score, f1_score from tensorboardX import SummaryWriter from torch.nn import CrossEntropyLoss from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset from torch.utils.data.distributed import DistributedSampler from tqdm import tqdm, trange from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file from transformers import AdamW, get_linear_schedule_with_warmup from transformers import WEIGHTS_NAME, BertConfig, BertForTokenClassification, BertTokenizer from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer from transformers import DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer from transformers import CamembertConfig, CamembertForTokenClassification, CamembertTokenizer logger = logging.getLogger(__name__) ALL_MODELS = sum( (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)), ()) MODEL_CLASSES = { "bert": (BertConfig, BertForTokenClassification, BertTokenizer), "roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer), "distilbert": (DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer), "camembert": (CamembertConfig, CamembertForTokenClassification, CamembertTokenizer), } def set_seed(args): random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.n_gpu > 0: torch.cuda.manual_seed_all(args.seed) def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id): """ Train the model """ if args.local_rank in [-1, 0]: tb_writer = SummaryWriter() args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu) train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset) train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size) if args.max_steps > 0: t_total = args.max_steps args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1 else: t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs # Prepare optimizer and schedule (linear warmup and decay) no_decay = ["bias", "LayerNorm.weight"] optimizer_grouped_parameters = [ {"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], "weight_decay": args.weight_decay}, {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0} ] optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total) if args.fp16: try: from apex import amp except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) # multi-gpu training (should be after apex fp16 initialization) if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Distributed training (should be after apex fp16 initialization) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True) # Train! logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Num Epochs = %d", args.num_train_epochs) logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size) logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d", args.train_batch_size * args.gradient_accumulation_steps * ( torch.distributed.get_world_size() if args.local_rank != -1 else 1)) logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps) logger.info(" Total optimization steps = %d", t_total) global_step = 0 tr_loss, logging_loss = 0.0, 0.0 model.zero_grad() train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]) set_seed(args) # Added here for reproductibility (even between python 2 and 3) for _ in train_iterator: epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0]) for step, batch in enumerate(epoch_iterator): model.train() batch = tuple(t.to(args.device) for t in batch) inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]} if args.model_type != "distilbert": inputs["token_type_ids"]: batch[2] if args.model_type in ["bert", "xlnet"] else None # XLM and RoBERTa don"t use segment_ids outputs = model(**inputs) loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc) if args.n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu parallel training if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() tr_loss += loss.item() if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) scheduler.step() # Update learning rate schedule optimizer.step() model.zero_grad() global_step += 1 if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0: # Log metrics if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well results, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev") for key, value in results.items(): tb_writer.add_scalar("eval_{}".format(key), value, global_step) tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step) tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step) logging_loss = tr_loss if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0: # Save model checkpoint output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step)) if not os.path.exists(output_dir): os.makedirs(output_dir) model_to_save = model.module if hasattr(model, "module") else model # Take care of distributed/parallel training model_to_save.save_pretrained(output_dir) torch.save(args, os.path.join(output_dir, "training_args.bin")) logger.info("Saving model checkpoint to %s", output_dir) if args.max_steps > 0 and global_step > args.max_steps: epoch_iterator.close() break if args.max_steps > 0 and global_step > args.max_steps: train_iterator.close() break if args.local_rank in [-1, 0]: tb_writer.close() return global_step, tr_loss / global_step def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""): eval_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode=mode) args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu) # Note that DistributedSampler samples randomly eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset) eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size) # multi-gpu evaluate if args.n_gpu > 1: model = torch.nn.DataParallel(model) # Eval! logger.info("***** Running evaluation %s *****", prefix) logger.info(" Num examples = %d", len(eval_dataset)) logger.info(" Batch size = %d", args.eval_batch_size) eval_loss = 0.0 nb_eval_steps = 0 preds = None out_label_ids = None model.eval() for batch in tqdm(eval_dataloader, desc="Evaluating"): batch = tuple(t.to(args.device) for t in batch) with torch.no_grad(): inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]} if args.model_type != "distilbert": inputs["token_type_ids"]: batch[2] if args.model_type in ["bert", "xlnet"] else None # XLM and RoBERTa don"t use segment_ids outputs = model(**inputs) tmp_eval_loss, logits = outputs[:2] if args.n_gpu > 1: tmp_eval_loss = tmp_eval_loss.mean() # mean() to average on multi-gpu parallel evaluating eval_loss += tmp_eval_loss.item() nb_eval_steps += 1 if preds is None: preds = logits.detach().cpu().numpy() out_label_ids = inputs["labels"].detach().cpu().numpy() else: preds = np.append(preds, logits.detach().cpu().numpy(), axis=0) out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0) eval_loss = eval_loss / nb_eval_steps preds = np.argmax(preds, axis=2) label_map = {i: label for i, label in enumerate(labels)} out_label_list = [[] for _ in range(out_label_ids.shape[0])] preds_list = [[] for _ in range(out_label_ids.shape[0])] for i in range(out_label_ids.shape[0]): for j in range(out_label_ids.shape[1]): if out_label_ids[i, j] != pad_token_label_id: out_label_list[i].append(label_map[out_label_ids[i][j]]) preds_list[i].append(label_map[preds[i][j]]) results = { "loss": eval_loss, "precision": precision_score(out_label_list, preds_list), "recall": recall_score(out_label_list, preds_list), "f1": f1_score(out_label_list, preds_list) } logger.info("***** Eval results %s *****", prefix) for key in sorted(results.keys()): logger.info(" %s = %s", key, str(results[key])) return results, preds_list def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode): if args.local_rank not in [-1, 0] and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache # Load data features from cache or dataset file cached_features_file = os.path.join(args.data_dir, "cached_{}_{}_{}".format(mode, list(filter(None, args.model_name_or_path.split("/"))).pop(), str(args.max_seq_length))) if os.path.exists(cached_features_file) and not args.overwrite_cache: logger.info("Loading features from cached file %s", cached_features_file) features = torch.load(cached_features_file) else: logger.info("Creating features from dataset file at %s", args.data_dir) examples = read_examples_from_file(args.data_dir, mode) features = convert_examples_to_features(examples, labels, args.max_seq_length, tokenizer, cls_token_at_end=bool(args.model_type in ["xlnet"]), # xlnet has a cls token at the end cls_token=tokenizer.cls_token, cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0, sep_token=tokenizer.sep_token, sep_token_extra=bool(args.model_type in ["roberta"]), # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805 pad_on_left=bool(args.model_type in ["xlnet"]), # pad on the left for xlnet pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0], pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0, pad_token_label_id=pad_token_label_id ) if args.local_rank in [-1, 0]: logger.info("Saving features into cached file %s", cached_features_file) torch.save(features, cached_features_file) if args.local_rank == 0 and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache # Convert to Tensors and build dataset all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long) all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long) dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) return dataset def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--data_dir", default=None, type=str, required=True, help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.") parser.add_argument("--model_type", default=None, type=str, required=True, help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys())) parser.add_argument("--model_name_or_path", default=None, type=str, required=True, help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS)) parser.add_argument("--output_dir", default=None, type=str, required=True, help="The output directory where the model predictions and checkpoints will be written.") ## Other parameters parser.add_argument("--labels", default="", type=str, help="Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.") parser.add_argument("--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name") parser.add_argument("--tokenizer_name", default="", type=str, help="Pretrained tokenizer name or path if not the same as model_name") parser.add_argument("--cache_dir", default="", type=str, help="Where do you want to store the pre-trained models downloaded from s3") parser.add_argument("--max_seq_length", default=128, type=int, help="The maximum total input sequence length after tokenization. Sequences longer " "than this will be truncated, sequences shorter will be padded.") parser.add_argument("--do_train", action="store_true", help="Whether to run training.") parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.") parser.add_argument("--do_predict", action="store_true", help="Whether to run predictions on the test set.") parser.add_argument("--evaluate_during_training", action="store_true", help="Whether to run evaluation during training at each logging step.") parser.add_argument("--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model.") parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.") parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation.") parser.add_argument("--gradient_accumulation_steps", type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.") parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--max_steps", default=-1, type=int, help="If > 0: set total number of training steps to perform. Override num_train_epochs.") parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.") parser.add_argument("--logging_steps", type=int, default=50, help="Log every X updates steps.") parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.") parser.add_argument("--eval_all_checkpoints", action="store_true", help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number") parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available") parser.add_argument("--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory") parser.add_argument("--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets") parser.add_argument("--seed", type=int, default=42, help="random seed for initialization") parser.add_argument("--fp16", action="store_true", help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit") parser.add_argument("--fp16_opt_level", type=str, default="O1", help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." "See details at https://nvidia.github.io/apex/amp.html") parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.") parser.add_argument("--server_port", type=str, default="", help="For distant debugging.") args = parser.parse_args() if os.path.exists(args.output_dir) and os.listdir( args.output_dir) and args.do_train and not args.overwrite_output_dir: raise ValueError( "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format( args.output_dir)) # Setup distant debugging if needed if args.server_ip and args.server_port: # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script import ptvsd print("Waiting for debugger attach") ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True) ptvsd.wait_for_attach() # Setup CUDA, GPU & distributed training if args.local_rank == -1 or args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") args.n_gpu = torch.cuda.device_count() else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) torch.distributed.init_process_group(backend="nccl") args.n_gpu = 1 args.device = device # Setup logging logging.basicConfig(format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN) logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16) # Set seed set_seed(args) # Prepare CONLL-2003 task labels = get_labels(args.labels) num_labels = len(labels) # Use cross entropy ignore index as padding label id so that only real label ids contribute to the loss later pad_token_label_id = CrossEntropyLoss().ignore_index # Load pretrained model and tokenizer if args.local_rank not in [-1, 0]: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab args.model_type = args.model_type.lower() config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type] config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, cache_dir=args.cache_dir if args.cache_dir else None) tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case, cache_dir=args.cache_dir if args.cache_dir else None) model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path), config=config, cache_dir=args.cache_dir if args.cache_dir else None) if args.local_rank == 0: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab model.to(args.device) logger.info("Training/evaluation parameters %s", args) # Training if args.do_train: train_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode="train") global_step, tr_loss = train(args, train_dataset, model, tokenizer, labels, pad_token_label_id) logger.info(" global_step = %s, average loss = %s", global_step, tr_loss) # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained() if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0): # Create output directory if needed if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]: os.makedirs(args.output_dir) logger.info("Saving model checkpoint to %s", args.output_dir) # Save a trained model, configuration and tokenizer using `save_pretrained()`. # They can then be reloaded using `from_pretrained()` model_to_save = model.module if hasattr(model, "module") else model # Take care of distributed/parallel training model_to_save.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir) # Good practice: save your training arguments together with the trained model torch.save(args, os.path.join(args.output_dir, "training_args.bin")) # Evaluation results = {} if args.do_eval and args.local_rank in [-1, 0]: tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case) checkpoints = [args.output_dir] if args.eval_all_checkpoints: checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))) logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging logger.info("Evaluate the following checkpoints: %s", checkpoints) for checkpoint in checkpoints: global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else "" model = model_class.from_pretrained(checkpoint) model.to(args.device) result, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev", prefix=global_step) if global_step: result = {"{}_{}".format(global_step, k): v for k, v in result.items()} results.update(result) output_eval_file = os.path.join(args.output_dir, "eval_results.txt") with open(output_eval_file, "w") as writer: for key in sorted(results.keys()): writer.write("{} = {}\n".format(key, str(results[key]))) if args.do_predict and args.local_rank in [-1, 0]: tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case) model = model_class.from_pretrained(args.output_dir) model.to(args.device) result, predictions = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="test") # Save results output_test_results_file = os.path.join(args.output_dir, "test_results.txt") with open(output_test_results_file, "w") as writer: for key in sorted(result.keys()): writer.write("{} = {}\n".format(key, str(result[key]))) # Save predictions output_test_predictions_file = os.path.join(args.output_dir, "test_predictions.txt") with open(output_test_predictions_file, "w") as writer: with open(os.path.join(args.data_dir, "test.txt"), "r") as f: example_id = 0 for line in f: if line.startswith("-DOCSTART-") or line == "" or line == "\n": writer.write(line) if not predictions[example_id]: example_id += 1 elif predictions[example_id]: output_line = line.split()[0] + " " + predictions[example_id].pop(0) + "\n" writer.write(output_line) else: logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0]) return results if __name__ == "__main__": main()
28,786
53.009381
184
py
DeeBERT
DeeBERT-master/examples/utils_summarization.py
from collections import deque import os import torch from torch.utils.data import Dataset # ------------ # Data loading # ------------ class CNNDailyMailDataset(Dataset): """ Abstracts the dataset used to train seq2seq models. CNN/Daily News: The CNN/Daily News raw datasets are downloaded from [1]. The stories are stored in different files; the summary appears at the end of the story as sentences that are prefixed by the special `@highlight` line. To process the data, untar both datasets in the same folder, and pass the path to this folder as the "data_dir argument. The formatting code was inspired by [2]. [1] https://cs.nyu.edu/~kcho/ [2] https://github.com/abisee/cnn-dailymail/ """ def __init__(self, tokenizer, prefix="train", data_dir=""): assert os.path.isdir(data_dir) self.tokenizer = tokenizer # We initialize the class by listing all the files that contain # stories and summaries. Files are not read in memory given # the size of the corpus. self.stories_path = [] datasets = ("cnn", "dailymail") for dataset in datasets: path_to_stories = os.path.join(data_dir, dataset, "stories") story_filenames_list = os.listdir(path_to_stories) for story_filename in story_filenames_list: path_to_story = os.path.join(path_to_stories, story_filename) if not os.path.isfile(path_to_story): continue self.stories_path.append(path_to_story) def __len__(self): return len(self.stories_path) def __getitem__(self, idx): story_path = self.stories_path[idx] with open(story_path, encoding="utf-8") as source: raw_story = source.read() story_lines, summary_lines = process_story(raw_story) return story_lines, summary_lines def process_story(raw_story): """ Extract the story and summary from a story file. Attributes: raw_story (str): content of the story file as an utf-8 encoded string. Raises: IndexError: If the stoy is empty or contains no highlights. """ nonempty_lines = list( filter(lambda x: len(x) != 0, [line.strip() for line in raw_story.split("\n")]) ) # for some unknown reason some lines miss a period, add it nonempty_lines = [_add_missing_period(line) for line in nonempty_lines] # gather article lines story_lines = [] lines = deque(nonempty_lines) while True: try: element = lines.popleft() if element.startswith("@highlight"): break story_lines.append(element) except IndexError: # if "@highlight" is absent from the file we pop # all elements until there is None. return story_lines, [] # gather summary lines summary_lines = list(filter(lambda t: not t.startswith("@highlight"), lines)) return story_lines, summary_lines def _add_missing_period(line): END_TOKENS = [".", "!", "?", "...", "'", "`", '"', u"\u2019", u"\u2019", ")"] if line.startswith("@highlight"): return line if line[-1] in END_TOKENS: return line return line + "." # -------------------------- # Encoding and preprocessing # -------------------------- def fit_to_block_size(sequence, block_size, pad_token): """ Adapt the source and target sequences' lengths to the block size. If the sequence is shorter than the block size we pad it with -1 ids which correspond to padding tokens. """ if len(sequence) > block_size: return sequence[:block_size] else: sequence.extend([pad_token] * (block_size - len(sequence))) return sequence def build_lm_labels(sequence, pad_token): """ Padding token, encoded as 0, are represented by the value -1 so they are not taken into account in the loss computation. """ padded = sequence.clone() padded[padded == pad_token] = -1 return padded def build_mask(sequence, pad_token): """ Builds the mask. The attention mechanism will only attend to positions with value 1. """ mask = torch.ones_like(sequence) idx_pad_tokens = sequence == pad_token mask[idx_pad_tokens] = 0 return mask def encode_for_summarization(story_lines, summary_lines, tokenizer): """ Encode the story and summary lines, and join them as specified in [1] by using `[SEP] [CLS]` tokens to separate sentences. """ story_lines_token_ids = [ tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line)) for line in story_lines ] summary_lines_token_ids = [ tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line)) for line in summary_lines ] story_token_ids = [ token for sentence in story_lines_token_ids for token in sentence ] summary_token_ids = [ token for sentence in summary_lines_token_ids for token in sentence ] return story_token_ids, summary_token_ids def compute_token_type_ids(batch, separator_token_id): """ Segment embeddings as described in [1] The values {0,1} were found in the repository [2]. Attributes: batch: torch.Tensor, size [batch_size, block_size] Batch of input. separator_token_id: int The value of the token that separates the segments. [1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders." arXiv preprint arXiv:1908.08345 (2019). [2] https://github.com/nlpyang/PreSumm (/src/prepro/data_builder.py, commit fac1217) """ batch_embeddings = [] for sequence in batch: sentence_num = 0 embeddings = [] for s in sequence: if s == separator_token_id: sentence_num += 1 embeddings.append(sentence_num % 2) batch_embeddings.append(embeddings) return torch.tensor(batch_embeddings)
6,022
31.556757
88
py