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
RPMG
RPMG-main/ModelNet_Img/pylibs/pytorch_util/libtrain/__init__.py
from .init_torch import list_models, rm_models, copy_weights, cfg, init_weights_by_filling, count_parameters_all, \ count_parameters_trainable try: from hooks import Forward_Hook_Handlers, Backward_Hook_Handlers, fw_hook_percentile except: pass
258
31.375
115
py
RPMG
RPMG-main/ModelNet_Img/pylibs/pytorch_util/libtrain/init_torch.py
import os, sys from basic.common import env, Open, add_path # rdict import numpy as np import math import torch import torch.nn as nn import torchvision import torch.utils.model_zoo as model_zoo # `pip install easydict` if you don't have it from easydict import EasyDict as edict # Pixel mean values (BGR order) as a (1, 1, 3) array # These are the values originally used for training VGG16 # __C.PIXEL_MEANS = np.array([[[102.9801, 115.9465, 122.7717]]]) _cfg = edict(caffemodel=edict(), # To be define later torchmodel=edict(), # To be define later PIXEL_MEANS=np.array([[[102.9801, 115.9465, 122.7717]]]), # PIXEL_MEANS_Imagenet = np.array([[[104.006987932, 116.668767617, 122.678914341]]]), ) cfg = _cfg this_dir = os.path.realpath(os.path.dirname(__file__)) base_dir = os.path.realpath(this_dir + '/../pretrained_model.cache') # default models and pretrained weights. cfg.caffemodel.alexnet = edict( proto=os.path.join(base_dir + '/bvlc_alexnet/deploy.prototxt'), model=os.path.join(base_dir + '/bvlc_alexnet/bvlc_alexnet.caffemodel'), pkl=os.path.join(base_dir + '/bvlc_alexnet/bvlc_alexnet.pkl'), input_size=(227, 227), ) cfg.caffemodel.caffenet = edict( proto=os.path.join(base_dir + '/bvlc_reference_caffenet/deploy.prototxt'), model=os.path.join(base_dir + '/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel'), pkl=os.path.join(base_dir + '/bvlc_reference_caffenet/bvlc_reference_caffenet.pkl'), input_size=(227, 227), ) cfg.caffemodel.vgg16 = edict( proto=os.path.join(base_dir + '/vgg_net/VGG_ILSVRC_16_layers_deploy.prototxt'), model=os.path.join(base_dir + '/vgg_net/VGG_ILSVRC_16_layers.caffemodel'), pkl=os.path.join(base_dir + '/vgg_net/VGG_ILSVRC_16_layers.pkl'), input_size=(224, 224), ) cfg.caffemodel.vgg19 = edict( proto=os.path.join(base_dir + '/vgg_net/VGG_ILSVRC_19_layers_deploy.prototxt'), model=os.path.join(base_dir + '/vgg_net/VGG_ILSVRC_19_layers.caffemodel'), pkl=os.path.join(base_dir + '/vgg_net/VGG_ILSVRC_19_layers.pkl'), input_size=(224, 224), ) cfg.caffemodel.GoogLeNet = edict( proto=os.path.join(base_dir + '/bvlc_googlenet/deploy.prototxt'), model=os.path.join(base_dir + '/bvlc_googlenet/bvlc_googlenet.caffemodel'), input_size=(224, 224), ) cfg.caffemodel.vggm = edict( proto=os.path.join(base_dir + '/vgg_net/VGG_CNN_M_deploy.prototxt'), model=os.path.join(base_dir + '/vgg_net/VGG_CNN_M.caffemodel'), pkl=os.path.join(base_dir + '/vgg_net/VGG_CNN_M.pkl'), input_size=(224, 224), ) cfg.caffemodel.resnet50 = edict( model=os.path.join(base_dir + '/resnet50-caffe.pth'), ) cfg.caffemodel.resnet101 = edict( model=os.path.join(base_dir + '/resnet101-caffe.pth'), ) cfg.caffemodel.resnet152 = edict( model=os.path.join(base_dir + '/resnet152-caffe.pth'), ) # cfg.caffemodel.vgg16 = edict( # model = os.path.join( base_dir + '/vgg16_caffe.pth'), # ) # --------------------------------------------------------- [torchmodel] cfg.torchmodel.alexnet = edict( # module = torchvision.models.alexnet, model_url='https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth', model=os.path.join(env.Home, '.torch/models/alexnet-owt-4df8aa71.pth'), input_size=(224, 224), ) cfg.torchmodel.inception_v3_google = edict( # module = torchvision.models.alexnet, model=os.path.join(env.Home, '.torch/models/inception_v3_google-1a9a5a14.pth'), input_size=(224, 224), ) cfg.torchmodel.resnet101 = edict( # module = torchvision.models.alexnet, model=os.path.join(env.Home, '.torch/models/resnet101-5d3b4d8f.pth'), # input_size = (224,224), ) cfg.torchmodel.vgg16 = edict( model_url='https://download.pytorch.org/models/vgg16-397923af.pth', # model = os.path.join( env.Home, '.torch/models/resnet101-5d3b4d8f.pth'), # input_size = (224,224), ) cfg.torchmodel.vgg19 = edict( model_url='https://download.pytorch.org/models/vgg19-dcbb9e9d.pth', # model = os.path.join( env.Home, '.torch/models/resnet101-5d3b4d8f.pth'), # input_size = (224,224), ) cfg.torchmodel.vgg16_bn = edict( model_url='https://download.pytorch.org/models/vgg16_bn-6c64b313.pth', # model = os.path.join( env.Home, '.torch/models/resnet101-5d3b4d8f.pth'), # input_size = (224,224), ) cfg.torchmodel.vgg19_bn = edict( model_url='https://download.pytorch.org/models/vgg19_bn-c79401a0.pth', # model = os.path.join( env.Home, '.torch/models/resnet101-5d3b4d8f.pth'), # input_size = (224,224), ) def list_models(snapshots_dir, marker='iter'): import re # collect ".solverstate" and ".caffemodel" files. _files = os.listdir(snapshots_dir) reg_valid_name_str = '([()\[\]a-zA-Z0-9\s\.,_+-]+)' # train_iter_20000.pth.tar reg_model = r"^%s_%s_(\d+).pth.tar$" % (reg_valid_name_str, marker) # reg_state = r"^%s_iter_(\d+).solverstate$" % reg_valid_name_str net_names_model = set() iter_num_model = [] model_files = [x for x in _files if x.endswith('.pth.tar')] for name in model_files: match = re.search(reg_model, name) if match: # match is not None gs = match.groups() else: raise Exception( 'No model matches in snapshots_dir: %s\n\tExist model list: %s' % (snapshots_dir, model_files)) assert len(gs) == 2, '[Exception] in matching .pth.tar file: "%s"' % name net_name, iter_num = gs[0], int(gs[1]) net_names_model.add(net_name) iter_num_model.append(iter_num) assert len(net_names_model) == 1, "None or Multiple net models in this dir: %s " % len(net_names_model) assert len(set(iter_num_model)) == len(iter_num_model) iter_num_model.sort() existed_nums = iter_num_model net_name = list(net_names_model)[0] return existed_nums, net_name def rm_models(snapshots_dir, type='KEEP_LATEST', marker='iter'): saved_nums, net_name = list_models(snapshots_dir, marker=marker) # ('snapshots') assert len(saved_nums) > 0, "No models available" if type == 'KEEP_LATEST': latest_model_name = '%s_%s_%s.pth.tar' % (net_name, marker, saved_nums[-1]) # '%s_iter_%s.pth.tar' latest_model_path = os.path.join(snapshots_dir, latest_model_name) assert os.path.exists(latest_model_path), latest_model_path if len(saved_nums) == 1: # only has latest model. return latest_model_name else: for it in saved_nums[:-1]: model_prefix = os.path.join(snapshots_dir, '%s_%s_%s') % (net_name, marker, it) model_path = model_prefix + '.pth.tar' # '.caffemodel' assert os.path.exists(model_path) os.system('rm -f %s' % (model_path)) return latest_model_name else: raise NotImplementedError def get_weights_from_caffesnapeshot(proto_file, model_file): import caffe from collections import OrderedDict import cPickle as pickle caffe.set_mode_cpu() net = caffe.Net(proto_file, model_file, caffe.TEST) model_dict = OrderedDict() for layer_name, param in net.params.iteritems(): # Most param blob has w and b, but for PReLU there's only w. learnable_weight = [] # {} if len(param) == 2: # learnable_weight['w'] = param[0].data.copy() # learnable_weight['b'] = param[1].data.copy() learnable_weight.append(param[0].data.copy()) learnable_weight.append(param[1].data.copy()) elif len(param) == 1: # learnable_weight['w'] = param[0].data.copy() learnable_weight.append(param[0].data.copy()) else: raise NotImplementedError model_dict[layer_name] = learnable_weight return model_dict def _copy_weights_from_caffemodel(own_state, pretrained_type='alexnet', ignore_missing_dst=False, src2dsts=dict(conv1='conv1', conv2='conv2', conv3='conv3', conv4='conv4', conv5='conv5', fc6='fc6', fc7='fc7')): """ src2dsts = dict(conv1='conv1', conv2='conv2', conv3='conv3', conv4='conv4', conv5='conv5') Or in list: src2dsts = dict(conv1=['conv1'], conv2=['conv2'], conv3=['conv3'], conv4=['conv4'], conv5=['conv5']) """ print("-----------------------") print("[Info] Copy from %s " % pretrained_type) print("-----------------------") if isinstance(pretrained_type, tuple): proto_file, model_file = pretrained_type pretrained_weights = get_weights_from_caffesnapeshot(proto_file, model_file) else: import pickle print('Loading: %s' % cfg.caffemodel[pretrained_type].pkl) pretrained_weights = pickle.load(open(cfg.caffemodel[pretrained_type].pkl, 'rb'), encoding='bytes') # 'bytes' ) print(pretrained_weights.keys()) # print (list(pretrained_weights.keys())[0].decode()) not_copied = list(own_state.keys()) src_list = sorted(src2dsts.keys()) for src in src_list: # src2dsts.iteritems(): dsts = src2dsts[src] if not isinstance(dsts, list): dsts = [dsts] w, b = pretrained_weights[src.encode('utf-8')] w = torch.from_numpy(w) # cast as pytorch tensor b = torch.from_numpy(b) # cast as pytorch tensor # one src can be copied to multiple dsts for dst in dsts: if ignore_missing_dst and dst not in own_state.keys(): # net.params.keys(): print('%-20s --> %-20s [ignored] Missing dst.' % (src, dst)) continue print('%-20s --> %-20s' % (src, dst)) dst_w_name = '%s.weight' % dst dst_b_name = '%s.bias' % dst assert dst_w_name in own_state.keys(), "[Error] %s not in %s" % (dst_w_name, own_state.keys()) assert dst_b_name in own_state.keys(), "[Error] %s not in %s" % (dst_b_name, own_state.keys()) # -- Copy w assert own_state[dst_w_name].shape == w.shape, '[%s] w: dest. %s != src. %s' % ( dst_w_name, own_state[dst_w_name].shape, w.shape) own_state[dst_w_name].copy_(w) not_copied.remove(dst_w_name) # -- Copy b assert own_state[dst_b_name].shape == b.shape, '[%s] w: dest. %s != src. %s' % ( dst_b_name, own_state[dst_b_name].shape, b.shape) own_state[dst_b_name].copy_(b) not_copied.remove(dst_b_name) for name in not_copied: if name.endswith('.weight'): own_state[name].normal_(mean=0.0, std=0.005) # guassian for fc std=0.005, for conv std=0.01 print('%-20s --> %-20s' % ('[filler] gaussian005', name)) elif name.endswith('.bias'): own_state[name].fill_(0) print('%-20s --> %-20s' % ('[filler] 0', name)) else: print("Unknow parameter type: ", name) raise NotImplementedError print("-----------------------") def _copy_weights_from_torchmodel(own_state, pretrained_type='alexnet', strict=True, src2dsts=None): # wrapper for load_state_dict from torch.nn.parameter import Parameter print("-----------------------") print("[Info] Copy from %s " % pretrained_type) print("-----------------------") src_state = model_zoo.load_url(cfg.torchmodel[pretrained_type].model_url) not_copied = list(own_state.keys()) # print(type(not_copied)) if src2dsts is not None: for src, dsts in src2dsts.items(): if not isinstance(dsts, list): dsts = [dsts] # one src can be copied to multiple dsts for dst in dsts: if dst in own_state.keys(): own_state[dst].copy_(src_state[src]) not_copied.remove(dst) print('%-20s --> %-20s' % (src, dst)) else: dst_w_name, src_w_name = '%s.weight' % dst, '%s.weight' % src dst_b_name, src_b_name = '%s.bias' % dst, '%s.bias' % src if ( dst_w_name not in own_state.keys() or dst_b_name not in own_state.keys()) and not strict: # net.params.keys(): print('%-20s --> %-20s [ignored] Missing dst.' % (src, dst)) continue print('%-20s --> %-20s' % (src, dst)) # -- Copy w assert own_state[dst_w_name].shape == src_state[src_w_name].shape, '[%s] w: dest. %s != src. %s' % ( dst_w_name, own_state[dst_w_name].shape, src_state[src_w_name].shape) own_state[dst_w_name].copy_(src_state[src_w_name]) not_copied.remove(dst_w_name) # -- Copy b assert own_state[dst_b_name].shape == src_state[src_b_name].shape, '[%s] w: dest. %s != src. %s' % ( dst_b_name, own_state[dst_b_name].shape, src_state[src_b_name].shape) own_state[dst_b_name].copy_(src_state[src_b_name]) not_copied.remove(dst_b_name) else: for name, param in src_state.items(): if name in own_state: # find in own parameter if isinstance(param, Parameter): # backwards compatibility for serialized parameters param = param.data try: print('%-30s --> %-30s' % (name, name)) own_state[name].copy_(param) not_copied.remove(name) except Exception: raise RuntimeError('While copying the parameter named {}, ' 'whose dimensions in the model are {} and ' 'whose dimensions in the checkpoint are {}.' .format(name, own_state[name].size(), param.size())) elif strict: raise KeyError('unexpected key "{}" in state_dict' .format(name)) else: print('%-30s --> %-30s [ignored] Missing dst.' % (name, name)) for name in not_copied: if name.endswith('.weight'): # -# torch.nn.init.xavier_normal(own_state[name]) # -# print '%-30s --> %-30s' % ('[filler] xavier_normal', name) own_state[name].normal_(mean=0.0, std=0.005) # guassian for fc std=0.005, for conv std=0.01 print('%-20s --> %-20s' % ('[filler] gaussian005', name)) elif name.endswith('.bias'): own_state[name].fill_(0) print('%-30s --> %-30s' % ('[filler] 0', name)) elif name.endswith('.running_mean') or name.endswith('.running_var') or name.endswith('num_batches_tracked'): print('*************************** pass', name) else: print("Unknow parameter type: ", name) raise NotImplementedError if strict: missing = set(own_state.keys()) - set(src_state.keys()) if len(missing) > 0: raise KeyError('missing keys in state_dict: "{}"'.format(missing)) print("-----------------------") # Copy pretrained weights to net. # By specify scrs, dsts layer names, this can prevent problem of # "same layer name has different shape" in traditional weight copy. def copy_weights(own_state, pretrained_type, **kwargs): ''' Usage example: copy_weights(own_state, 'torchmodel.alexnet', strict=False) copy_weights(own_state, 'caffemodel.alexnet', ignore_missing_dst=True, src2dsts={}) copy_weights(own_state, '') ''' if isinstance(pretrained_type, str): if pretrained_type.startswith('torchmodel.'): pretrained_type = pretrained_type[11:] # remove header copy_func = _copy_weights_from_torchmodel elif pretrained_type.startswith('caffemodel.'): pretrained_type = pretrained_type[11:] # remove header copy_func = _copy_weights_from_caffemodel else: print("Unkonw pretrained_type: ", pretrained_type) raise NotImplementedError elif isinstance(pretrained_type, tuple): copy_func = _copy_weights_from_caffemodel else: print("Unkonw type(pretrained_type): ", type(pretrained_type)) raise NotImplementedError copy_func(own_state, pretrained_type, **kwargs) def init_weights_by_filling(nn_module_or_seq, gaussian_std=0.01, kaiming_normal=True, silent=False): """ Note: gaussian_std is fully connected layer (nn.Linear) only. For nn.Conv2d: If kaiming_normal is enable, nn.Conv2d is initialized by kaiming_normal. Otherwise, initialized based on kernel size. """ if not silent: print('[init_weights_by_filling] gaussian_std=%s kaiming_normal=%s \n %s' % ( gaussian_std, kaiming_normal, nn_module_or_seq)) for name, m in nn_module_or_seq.named_modules(): if isinstance(m, nn.Conv2d): if kaiming_normal: nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') # 'relu' else: n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): if m.weight is not None: m.weight.data.fill_(1) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.Linear): m.weight.data.normal_(mean=0, std=gaussian_std) # std=0.01) # ('[filler] gaussian005', name) m.bias.data.zero_() # torch.nn.init.xavier_normal(own_state[name]) return nn_module_or_seq def count_parameters_all(model): pp = 0 for p in list(model.parameters()): nn = 1 for s in list(p.size()): nn = nn * s pp += nn return pp def count_parameters_trainable(model): return sum(p.numel() for p in model.parameters() if p.requires_grad)
18,165
43.415648
139
py
RPMG
RPMG-main/poselstm-pytorch/train.py
import time from options.train_options import TrainOptions from data.data_loader import CreateDataLoader from models.models import create_model from util.visualizer import Visualizer opt = TrainOptions().parse() ## SEEDING import torch import numpy import random torch.manual_seed(opt.seed) numpy.random.seed(opt.seed) random.seed(opt.seed) # torch.backends.cudnn.enabled = False torch.backends.cudnn.deterministic = True ## SEEDING data_loader = CreateDataLoader(opt) dataset = data_loader.load_data() dataset_size = len(data_loader) print('#training images = %d' % dataset_size) model = create_model(opt) visualizer = Visualizer(opt) total_steps = 0 for epoch in range(opt.epoch_count, opt.niter + opt.niter_decay + 1): epoch_start_time = time.time() epoch_iter = 0 for i, data in enumerate(dataset): iter_start_time = time.time() visualizer.reset() total_steps += opt.batchSize epoch_iter += opt.batchSize ratio = min(epoch // (opt.niter // 10), 9) / 9 tau = 1/20 + ratio * (1/4 - 1/20) model.set_input(data, tau) model.optimize_parameters() # if total_steps % opt.display_freq == 0: # save_result = total_steps % opt.update_html_freq == 0 # visualizer.display_current_results(model.get_current_visuals(), epoch, save_result) if total_steps % opt.print_freq == 0: errors = model.get_current_errors() t = (time.time() - iter_start_time) / opt.batchSize visualizer.print_current_errors(epoch, epoch_iter, errors, t) if opt.display_id > 0: visualizer.plot_current_errors(epoch, float(epoch_iter)/dataset_size, opt, errors) # if total_steps % opt.save_latest_freq == 0: # print('saving the latest model (epoch %d, total_steps %d)' % # (epoch, total_steps)) # model.save('latest') if epoch % opt.save_epoch_freq == 0: print('saving the model at the end of epoch %d, iters %d' % (epoch, total_steps)) model.save('latest') model.save(epoch) print('End of epoch %d / %d \t Time Taken: %d sec' % (epoch, opt.niter + opt.niter_decay, time.time() - epoch_start_time)) model.update_learning_rate()
2,288
33.681818
98
py
RPMG
RPMG-main/poselstm-pytorch/options/base_options.py
import argparse import os from util import util import torch class BaseOptions(): def __init__(self): self.parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter) self.initialized = False def initialize(self): self.parser.add_argument('--dataroot', required=True, help='path to images (should have subfolders trainA, trainB, valA, valB, etc)') self.parser.add_argument('--batchSize', type=int, default=75, help='input batch size') self.parser.add_argument('--loadSize', type=int, default=256, help='scale images to this size') self.parser.add_argument('--fineSize', type=int, default=224, help='then crop to this size') self.parser.add_argument('--input_nc', type=int, default=3, help='# of input image channels') self.parser.add_argument('--output_nc', type=int, default=7, help='# of output image channels') self.parser.add_argument('--lstm_hidden_size', type=int, default=256, help='hidden size of the LSTM layer in PoseLSTM') self.parser.add_argument('--gpu_ids', type=str, default='0', help='gpu ids: e.g. 0 0,1,2, 0,2. use -1 for CPU') self.parser.add_argument('--name', type=str, default='experiment_name', help='name of the experiment. It decides where to store samples and models') self.parser.add_argument('--dataset_mode', type=str, default='unaligned_posenet', help='chooses how datasets are loaded. [unaligned | aligned | single]') self.parser.add_argument('--model', type=str, default='posenet', help='chooses which model to use. [posenet | poselstm]') self.parser.add_argument('--nThreads', default=8, type=int, help='# threads for loading data') self.parser.add_argument('--checkpoints_dir', type=str, default='./checkpoints', help='models are saved here') self.parser.add_argument('--serial_batches', action='store_true', help='if true, takes images in order to make batches, otherwise takes them randomly') self.parser.add_argument('--display_winsize', type=int, default=224, help='display window size') self.parser.add_argument('--display_id', type=int, default=0, help='window id of the web display') self.parser.add_argument('--display_port', type=int, default=8097, help='visdom port of the web display') self.parser.add_argument('--no_dropout', action='store_true', help='no dropout for the generator') self.parser.add_argument('--max_dataset_size', type=int, default=float("inf"), help='Maximum number of samples allowed per dataset. If the dataset directory contains more than max_dataset_size, only a subset is loaded.') self.parser.add_argument('--resize_or_crop', type=str, default='scale_width_and_crop', help='scaling and cropping of images at load time [resize_and_crop|crop|scale_width|scale_width_and_crop]') self.parser.add_argument('--no_flip', action='store_true', default=True, help='if specified, do not flip the images for data augmentation') self.parser.add_argument('--seed', type=int, default=0, help='initial random seed for deterministic results') self.parser.add_argument('--beta', type=float, default=500, help='beta factor used in posenet.') self.parser.add_argument('--mode', type=str, default='9D_inf', choices=['9D_SVD', '9D_inf', '9D_RPMG', '6D_GM', '4D_Axis', '6D_RPMG', '4D_norm', '3D_Euler', '4D_RPMG', '10D', '10D_RPMG']) self.initialized = True def parse(self): if not self.initialized: self.initialize() self.opt = self.parser.parse_args() self.opt.isTrain = self.isTrain # train or test str_ids = self.opt.gpu_ids.split(',') self.opt.gpu_ids = [] for str_id in str_ids: id = int(str_id) if id >= 0: self.opt.gpu_ids.append(id) # set gpu ids if len(self.opt.gpu_ids) > 0: torch.cuda.set_device(self.opt.gpu_ids[0]) args = vars(self.opt) print('------------ Options -------------') for k, v in sorted(args.items()): print('%s: %s' % (str(k), str(v))) print('-------------- End ----------------') # save to the disk expr_dir = os.path.join(self.opt.checkpoints_dir, self.opt.name) util.mkdirs(expr_dir) file_name = os.path.join(expr_dir, 'opt_'+self.opt.phase+'.txt') with open(file_name, 'wt') as opt_file: opt_file.write('------------ Options -------------\n') for k, v in sorted(args.items()): opt_file.write('%s: %s\n' % (str(k), str(v))) opt_file.write('-------------- End ----------------\n') return self.opt
4,881
63.236842
228
py
RPMG
RPMG-main/poselstm-pytorch/models/base_model.py
import os import torch class BaseModel(): def name(self): return 'BaseModel' def initialize(self, opt): self.opt = opt self.gpu_ids = opt.gpu_ids self.isTrain = opt.isTrain self.Tensor = torch.cuda.FloatTensor if self.gpu_ids else torch.Tensor self.save_dir = os.path.join(opt.checkpoints_dir, opt.name) def set_input(self, input): self.input = input def forward(self): pass # used in test time, no backprop def test(self): pass def get_image_paths(self): pass def optimize_parameters(self): pass def get_current_visuals(self): return self.input def get_current_errors(self): return {} def save(self, label): pass # helper saving function that can be used by subclasses def save_network(self, network, network_label, epoch_label, gpu_ids): save_filename = '%s_net_%s.pth' % (epoch_label, network_label) save_path = os.path.join(self.save_dir, save_filename) torch.save(network.cpu().state_dict(), save_path) if len(gpu_ids) and torch.cuda.is_available(): network.cuda(gpu_ids[0]) # helper loading function that can be used by subclasses def load_network(self, network, network_label, epoch_label): save_filename = '%s_net_%s.pth' % (epoch_label, network_label) save_path = os.path.join(self.save_dir, save_filename) network.load_state_dict(torch.load(save_path)) # update learning rate (called once every epoch) def update_learning_rate(self): for scheduler in self.schedulers: scheduler.step() lr = self.optimizers[0].param_groups[0]['lr'] print('learning rate = %.7f' % lr)
1,774
28.098361
78
py
RPMG
RPMG-main/poselstm-pytorch/models/networks.py
import torch import torch.nn as nn from torch.nn import init from torch.nn import functional as F import functools from torch.autograd import Variable from torch.optim import lr_scheduler import numpy as np ############################################################################### # Functions ############################################################################### def weight_init_googlenet(key, module, weights=None): if key == "LSTM": for name, param in module.named_parameters(): if 'bias' in name: init.constant_(param, 0.0) elif 'weight' in name: init.xavier_normal_(param) elif weights is None: init.constant_(module.bias.data, 0.0) if key == "XYZ": init.normal_(module.weight.data, 0.0, 0.5) elif key == "LSTM": init.xavier_normal_(module.weight.data) else: init.normal_(module.weight.data, 0.0, 0.01) else: # print(key, weights[(key+"_1").encode()].shape, module.bias.size()) module.bias.data[...] = torch.from_numpy(weights[(key+"_1").encode()]) module.weight.data[...] = torch.from_numpy(weights[(key+"_0").encode()]) return module def get_scheduler(optimizer, opt): if opt.lr_policy == 'lambda': def lambda_rule(epoch): lr_l = 1.0 - max(0, epoch + 1 + opt.epoch_count - opt.niter) / float(opt.niter_decay + 1) return lr_l scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda_rule) elif opt.lr_policy == 'step': scheduler = lr_scheduler.StepLR(optimizer, step_size=opt.lr_decay_iters, gamma=0.1) elif opt.lr_policy == 'plateau': scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.2, threshold=0.01, patience=5) else: return NotImplementedError('learning rate policy [%s] is not implemented', opt.lr_policy) return scheduler def define_network(mode, input_nc, lstm_hidden_size, model, init_from=None, isTest=False, gpu_ids=[]): netG = None use_gpu = len(gpu_ids) > 0 if use_gpu: assert(torch.cuda.is_available()) if model == 'posenet': netG = PoseNet(input_nc, weights=init_from, isTest=isTest, gpu_ids=gpu_ids) elif model == 'poselstm': netG = PoseLSTM(mode, input_nc, lstm_hidden_size, weights=init_from, isTest=isTest, gpu_ids=gpu_ids) else: raise NotImplementedError('Model name [%s] is not recognized' % model) if len(gpu_ids) > 0: netG.cuda(gpu_ids[0]) return netG ############################################################################## # Classes ############################################################################## # defines the regression heads for googlenet class RegressionHead(nn.Module): def __init__(self, outdim, lossID, weights=None, lstm_hidden_size=None): super(RegressionHead, self).__init__() self.has_lstm = lstm_hidden_size != None dropout_rate = 0.5 if lossID == "loss3" else 0.7 nc_loss = {"loss1": 512, "loss2": 528} nc_cls = [1024, 2048] if lstm_hidden_size is None else [lstm_hidden_size*4, lstm_hidden_size*4] self.dropout = nn.Dropout(p=dropout_rate) if lossID != "loss3": self.projection = nn.Sequential(*[nn.AvgPool2d(kernel_size=5, stride=3), weight_init_googlenet(lossID+"/conv", nn.Conv2d(nc_loss[lossID], 128, kernel_size=1), weights), nn.ReLU(inplace=True)]) self.cls_fc_pose = nn.Sequential(*[weight_init_googlenet(lossID+"/fc", nn.Linear(2048, 1024), weights), nn.ReLU(inplace=True)]) self.cls_fc_xy = weight_init_googlenet("XYZ", nn.Linear(nc_cls[0], 3)) self.cls_fc_wpqr = weight_init_googlenet("WPQR", nn.Linear(nc_cls[0], outdim)) if lstm_hidden_size is not None: self.lstm_pose_lr = weight_init_googlenet("LSTM", nn.LSTM(input_size=32, hidden_size=lstm_hidden_size, bidirectional=True, batch_first=True)) self.lstm_pose_ud = weight_init_googlenet("LSTM", nn.LSTM(input_size=32, hidden_size=lstm_hidden_size, bidirectional=True, batch_first=True)) else: self.projection = nn.AvgPool2d(kernel_size=7, stride=1) self.cls_fc_pose = nn.Sequential(*[weight_init_googlenet("pose", nn.Linear(1024, 2048)), nn.ReLU(inplace=True)]) self.cls_fc_xy = weight_init_googlenet("XYZ", nn.Linear(nc_cls[1], 3)) self.cls_fc_wpqr = weight_init_googlenet("WPQR", nn.Linear(nc_cls[1], outdim)) if lstm_hidden_size is not None: self.lstm_pose_lr = weight_init_googlenet("LSTM", nn.LSTM(input_size=64, hidden_size=lstm_hidden_size, bidirectional=True, batch_first=True)) self.lstm_pose_ud = weight_init_googlenet("LSTM", nn.LSTM(input_size=32, hidden_size=lstm_hidden_size, bidirectional=True, batch_first=True)) def forward(self, input): output = self.projection(input) output = self.cls_fc_pose(output.view(output.size(0), -1)) if self.has_lstm: output = output.view(output.size(0),32, -1) _, (hidden_state_lr, _) = self.lstm_pose_lr(output.permute(0,1,2)) _, (hidden_state_ud, _) = self.lstm_pose_ud(output.permute(0,2,1)) output = torch.cat((hidden_state_lr[0,:,:], hidden_state_lr[1,:,:], hidden_state_ud[0,:,:], hidden_state_ud[1,:,:]), 1) output = self.dropout(output) output_xy = self.cls_fc_xy(output) output_wpqr = self.cls_fc_wpqr(output) return [output_xy, output_wpqr] # define inception block for GoogleNet class InceptionBlock(nn.Module): def __init__(self, incp, input_nc, x1_nc, x3_reduce_nc, x3_nc, x5_reduce_nc, x5_nc, proj_nc, weights=None, gpu_ids=[]): super(InceptionBlock, self).__init__() self.gpu_ids = gpu_ids # first self.branch_x1 = nn.Sequential(*[ weight_init_googlenet("inception_"+incp+"/1x1", nn.Conv2d(input_nc, x1_nc, kernel_size=1), weights), nn.ReLU(inplace=True)]) self.branch_x3 = nn.Sequential(*[ weight_init_googlenet("inception_"+incp+"/3x3_reduce", nn.Conv2d(input_nc, x3_reduce_nc, kernel_size=1), weights), nn.ReLU(inplace=True), weight_init_googlenet("inception_"+incp+"/3x3", nn.Conv2d(x3_reduce_nc, x3_nc, kernel_size=3, padding=1), weights), nn.ReLU(inplace=True)]) self.branch_x5 = nn.Sequential(*[ weight_init_googlenet("inception_"+incp+"/5x5_reduce", nn.Conv2d(input_nc, x5_reduce_nc, kernel_size=1), weights), nn.ReLU(inplace=True), weight_init_googlenet("inception_"+incp+"/5x5", nn.Conv2d(x5_reduce_nc, x5_nc, kernel_size=5, padding=2), weights), nn.ReLU(inplace=True)]) self.branch_proj = nn.Sequential(*[ nn.MaxPool2d(kernel_size=3, stride=1, padding=1), weight_init_googlenet("inception_"+incp+"/pool_proj", nn.Conv2d(input_nc, proj_nc, kernel_size=1), weights), nn.ReLU(inplace=True)]) if incp in ["3b", "4e"]: self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) else: self.pool = None def forward(self, input): outputs = [self.branch_x1(input), self.branch_x3(input), self.branch_x5(input), self.branch_proj(input)] # print([[o.size()] for o in outputs]) output = torch.cat(outputs, 1) if self.pool is not None: return self.pool(output) return output class PoseNet(nn.Module): def __init__(self, input_nc, weights=None, isTest=False, gpu_ids=[]): super(PoseNet, self).__init__() self.gpu_ids = gpu_ids self.isTest = isTest self.before_inception = nn.Sequential(*[ weight_init_googlenet("conv1/7x7_s2", nn.Conv2d(input_nc, 64, kernel_size=7, stride=2, padding=3), weights), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=3, stride=2, padding=1), nn.LocalResponseNorm(size=5, alpha=0.0001, beta=0.75, k=1), weight_init_googlenet("conv2/3x3_reduce", nn.Conv2d(64, 64, kernel_size=1), weights), nn.ReLU(inplace=True), weight_init_googlenet("conv2/3x3", nn.Conv2d(64, 192, kernel_size=3, padding=1), weights), nn.ReLU(inplace=True), nn.LocalResponseNorm(size=5, alpha=0.0001, beta=0.75, k=1), nn.MaxPool2d(kernel_size=3, stride=2, padding=1) ]) self.inception_3a = InceptionBlock("3a", 192, 64, 96, 128, 16, 32, 32, weights, gpu_ids) self.inception_3b = InceptionBlock("3b", 256, 128, 128, 192, 32, 96, 64, weights, gpu_ids) self.inception_4a = InceptionBlock("4a", 480, 192, 96, 208, 16, 48, 64, weights, gpu_ids) self.inception_4b = InceptionBlock("4b", 512, 160, 112, 224, 24, 64, 64, weights, gpu_ids) self.inception_4c = InceptionBlock("4c", 512, 128, 128, 256, 24, 64, 64, weights, gpu_ids) self.inception_4d = InceptionBlock("4d", 512, 112, 144, 288, 32, 64, 64, weights, gpu_ids) self.inception_4e = InceptionBlock("4e", 528, 256, 160, 320, 32, 128, 128, weights, gpu_ids) self.inception_5a = InceptionBlock("5a", 832, 256, 160, 320, 32, 128, 128, weights, gpu_ids) self.inception_5b = InceptionBlock("5b", 832, 384, 192, 384, 48, 128, 128, weights, gpu_ids) # self.cls1_fc = RegressionHead(lossID="loss1", weights=weights) # self.cls2_fc = RegressionHead(lossID="loss2", weights=weights) # self.cls3_fc = RegressionHead(lossID="loss3", weights=weights) # self.model = nn.Sequential(*[self.inception_3a, self.inception_3b, # self.inception_4a, self.inception_4b, # self.inception_4c, self.inception_4d, # self.inception_4e, self.inception_5a, # self.inception_5b, self.cls1_fc, # self.cls2_fc, self.cls3_fc # ]) # if self.isTest: # self.model.eval() # ensure Dropout is deactivated during test def forward(self, input): output_bf = self.before_inception(input) output_3a = self.inception_3a(output_bf) output_3b = self.inception_3b(output_3a) output_4a = self.inception_4a(output_3b) output_4b = self.inception_4b(output_4a) output_4c = self.inception_4c(output_4b) output_4d = self.inception_4d(output_4c) output_4e = self.inception_4e(output_4d) output_5a = self.inception_5a(output_4e) output_5b = self.inception_5b(output_5a) if not self.isTest: return self.cls1_fc(output_4a) + self.cls2_fc(output_4d) + self.cls3_fc(output_5b) return self.cls3_fc(output_5b) class PoseLSTM(PoseNet): def __init__(self, mode, input_nc, lstm_hidden_size, weights=None, isTest=False, gpu_ids=[]): super(PoseLSTM, self).__init__(input_nc, weights, isTest, gpu_ids) outdim = int(mode.split('D')[0]) self.cls1_fc = RegressionHead(outdim,lossID="loss1", weights=weights, lstm_hidden_size=lstm_hidden_size) self.cls2_fc = RegressionHead(outdim,lossID="loss2", weights=weights, lstm_hidden_size=lstm_hidden_size) self.cls3_fc = RegressionHead(outdim,lossID="loss3", weights=weights, lstm_hidden_size=lstm_hidden_size) self.model = nn.Sequential(*[self.inception_3a, self.inception_3b, self.inception_4a, self.inception_4b, self.inception_4c, self.inception_4d, self.inception_4e, self.inception_5a, self.inception_5b, self.cls1_fc, self.cls2_fc, self.cls3_fc ]) if self.isTest: self.model.eval() # ensure Dropout is deactivated during test
12,387
51.05042
157
py
RPMG
RPMG-main/poselstm-pytorch/models/poselstm_model.py
from tracemalloc import get_traced_memory from builtins import NotImplementedError import numpy as np import torch import torch.nn.functional as F import os from collections import OrderedDict from torch.autograd import Variable import util.util as util from util.image_pool import ImagePool from .base_model import BaseModel from . import networks import pickle import numpy import sys BASEPATH = os.path.dirname(__file__) sys.path.append(os.path.join(BASEPATH, '..', '..', 'utils')) import tools import rpmg class PoseLSTModel(BaseModel): def name(self): return 'PoseLSTModel' def initialize(self, opt): BaseModel.initialize(self, opt) self.isTrain = opt.isTrain # define tensors self.input_A = self.Tensor(opt.batchSize, opt.input_nc, opt.fineSize, opt.fineSize) self.input_B = self.Tensor(opt.batchSize, opt.output_nc) # load/define networks googlenet_weights = None if self.isTrain and opt.init_weights != '': googlenet_file = open(opt.init_weights, "rb") googlenet_weights = pickle.load(googlenet_file, encoding="bytes") googlenet_file.close() print('initializing the weights from '+ opt.init_weights) self.mean_image = np.load(os.path.join(opt.dataroot , 'mean_image.npy')) self.netG = networks.define_network(opt.mode, opt.input_nc, opt.lstm_hidden_size, opt.model, init_from=googlenet_weights, isTest=not self.isTrain, gpu_ids = self.gpu_ids) if not self.isTrain or opt.continue_train: self.load_network(self.netG, 'G', opt.which_epoch) if self.isTrain: self.old_lr = opt.lr # define loss functions self.sum_criterion = torch.nn.MSELoss(reduction='sum') self.mean_criterion = torch.nn.MSELoss(reduction='mean') # initialize optimizers self.schedulers = [] self.optimizers = [] self.optimizer_G = torch.optim.Adam(self.netG.parameters(), lr=opt.lr, eps=1, weight_decay=0.0625, betas=(self.opt.adambeta1, self.opt.adambeta2)) self.optimizers.append(self.optimizer_G) # for optimizer in self.optimizers: # self.schedulers.append(networks.get_scheduler(optimizer, opt)) print('---------- Networks initialized -------------') # networks.print_network(self.netG) # print('-----------------------------------------------') def set_input(self, input, tau): input_A = input['A'] input_B = input['B'] self.image_paths = input['A_paths'] self.input_A.resize_(input_A.size()).copy_(input_A) self.input_B.resize_(input_B.size()).copy_(input_B) self.tau = tau self.gt_r = tools.compute_rotation_matrix_from_quaternion(self.input_B[:, 3:]) def forward(self): self.pred_B = self.netG(self.input_A) l = len(self.pred_B) // 2 loss_weights = [0.3, 0.3, 1] for i in range(l): out_nd = self.pred_B[2*i+1] if 'RPMG' in self.opt.mode: out_rmat = rpmg.simple_RPMG.apply(out_nd, self.tau, 0.01, self.opt.beta * loss_weights[i]) else: if(self.opt.mode == "4D_norm"): out_rmat = tools.compute_rotation_matrix_from_quaternion(out_nd) #b*3*3 elif(self.opt.mode=="6D_GM"): out_rmat = tools.compute_rotation_matrix_from_ortho6d(out_nd) #b*3*3 elif(self.opt.mode=="9D_SVD"): out_rmat = tools.symmetric_orthogonalization(out_nd) # b*3*3 elif (self.opt.mode == "10D"): out_rmat = tools.compute_rotation_matrix_from_10d(out_nd) # b*3*3 elif (self.opt.mode == "3D_Euler"): out_rmat = tools.compute_rotation_matrix_from_euler(out_nd) # b*3*3 elif (self.opt.mode == "4D_Axis"): out_rmat = tools.compute_rotation_matrix_from_axisAngle(out_nd) # b*3*3 else: raise NotImplementedError self.pred_B[2*i+1] = out_rmat # no backprop gradients def test(self): self.forward() # get image paths def get_image_paths(self): return self.image_paths def backward(self): self.loss_G = 0 # self.loss_pos = 0 # self.loss_ori = 0 loss_weights = [0.3, 0.3, 1] for l, w in enumerate(loss_weights): mse_pos = self.mean_criterion(self.pred_B[2*l], self.input_B[:, 0:3]) if 'RPMG' in self.opt.mode: mse_ori = self.sum_criterion(self.pred_B[2*l+1], self.gt_r) self.loss_G += mse_pos * w + mse_ori else: mse_ori = self.mean_criterion(self.pred_B[2*l+1], self.gt_r) self.loss_G += (mse_pos + mse_ori*self.opt.beta) * w # mse_ori = self.sum_criterion(self.pred_B[2*l+1], self.gt_r) # self.loss_pos += mse_pos.item() * w # self.loss_ori += mse_ori.item() * w * self.opt.beta self.loss_G.backward() def optimize_parameters(self): self.forward() self.optimizer_G.zero_grad() self.backward() self.optimizer_G.step() def get_current_errors(self): pos_err = torch.dist(self.pred_B[0], self.input_B[:, 0:3]) ori_err = tools.compute_geodesic_distance_from_two_matrices(self.gt_r, self.pred_B[1]) ori_err = ori_err.mean() * 180 / np.pi if self.opt.isTrain: return OrderedDict([('pos_err', pos_err), ('ori_err', ori_err), ]) else: return [pos_err.item(), ori_err.item()] def get_current_pose(self): return numpy.concatenate((self.pred_B[0].data[0].cpu().numpy(), self.pred_B[1].data[0].cpu().numpy())) def get_current_visuals(self): input_A = util.tensor2im(self.input_A.data) # pred_B = util.tensor2im(self.pred_B.data) # input_B = util.tensor2im(self.input_B.data) return OrderedDict([('input_A', input_A)]) def save(self, label): self.save_network(self.netG, 'G', label, self.gpu_ids)
6,529
39.308642
106
py
RPMG
RPMG-main/poselstm-pytorch/util/image_pool.py
import random import numpy as np import torch from torch.autograd import Variable class ImagePool(): def __init__(self, pool_size): self.pool_size = pool_size if self.pool_size > 0: self.num_imgs = 0 self.images = [] def query(self, images): if self.pool_size == 0: return Variable(images) return_images = [] for image in images: image = torch.unsqueeze(image, 0) if self.num_imgs < self.pool_size: self.num_imgs = self.num_imgs + 1 self.images.append(image) return_images.append(image) else: p = random.uniform(0, 1) if p > 0.5: random_id = random.randint(0, self.pool_size-1) tmp = self.images[random_id].clone() self.images[random_id] = image return_images.append(tmp) else: return_images.append(image) return_images = Variable(torch.cat(return_images, 0)) return return_images
1,116
30.914286
67
py
RPMG
RPMG-main/poselstm-pytorch/util/util.py
from __future__ import print_function import torch import numpy as np from PIL import Image import inspect, re import numpy as np import os import collections # Converts a Tensor into a Numpy array # |imtype|: the desired type of the converted numpy array def tensor2im(image_tensor, imtype=np.uint8): image_numpy = image_tensor[0].cpu().float().numpy() if image_numpy.shape[0] == 1: image_numpy = np.tile(image_numpy, (3, 1, 1)) image_numpy = (np.transpose(image_numpy, (1, 2, 0)) + 1) / 2.0 * 255.0 return image_numpy.astype(imtype) def diagnose_network(net, name='network'): mean = 0.0 count = 0 for param in net.parameters(): if param.grad is not None: mean += torch.mean(torch.abs(param.grad.data)) count += 1 if count > 0: mean = mean / count print(name) print(mean) def save_image(image_numpy, image_path): image_pil = Image.fromarray(image_numpy) image_pil.save(image_path) def info(object, spacing=10, collapse=1): """Print methods and doc strings. Takes module, class, list, dictionary, or string.""" methodList = [e for e in dir(object) if isinstance(getattr(object, e), collections.Callable)] processFunc = collapse and (lambda s: " ".join(s.split())) or (lambda s: s) print( "\n".join(["%s %s" % (method.ljust(spacing), processFunc(str(getattr(object, method).__doc__))) for method in methodList]) ) def varname(p): for line in inspect.getframeinfo(inspect.currentframe().f_back)[3]: m = re.search(r'\bvarname\s*\(\s*([A-Za-z_][A-Za-z0-9_]*)\s*\)', line) if m: return m.group(1) def print_numpy(x, val=True, shp=False): x = x.astype(np.float64) if shp: print('shape,', x.shape) if val: x = x.flatten() print('mean = %3.3f, min = %3.3f, max = %3.3f, median = %3.3f, std=%3.3f' % ( np.mean(x), np.min(x), np.max(x), np.median(x), np.std(x))) def mkdirs(paths): if isinstance(paths, list) and not isinstance(paths, str): for path in paths: mkdir(path) else: mkdir(paths) def mkdir(path): if not os.path.exists(path): os.makedirs(path)
2,265
29.621622
97
py
RPMG
RPMG-main/poselstm-pytorch/data/custom_dataset_data_loader.py
import torch.utils.data from data.base_data_loader import BaseDataLoader def CreateDataset(opt): dataset = None if opt.dataset_mode == 'unaligned_posenet': from data.unaligned_posenet_dataset import UnalignedPoseNetDataset dataset = UnalignedPoseNetDataset() else: raise ValueError("Dataset [%s] not recognized." % opt.dataset_mode) print("dataset [%s] was created" % (dataset.name())) dataset.initialize(opt) return dataset class CustomDatasetDataLoader(BaseDataLoader): def name(self): return 'CustomDatasetDataLoader' def initialize(self, opt): BaseDataLoader.initialize(self, opt) self.dataset = CreateDataset(opt) def init_fn(worker_id): torch.manual_seed(opt.seed) self.dataloader = torch.utils.data.DataLoader( self.dataset, batch_size=opt.batchSize, shuffle=not opt.serial_batches, num_workers=int(opt.nThreads), worker_init_fn=init_fn) def load_data(self): return self def __len__(self): return min(len(self.dataset), self.opt.max_dataset_size) def __iter__(self): for i, data in enumerate(self.dataloader): if i >= self.opt.max_dataset_size: break yield data
1,324
27.191489
75
py
RPMG
RPMG-main/poselstm-pytorch/data/unaligned_posenet_dataset.py
import os.path import torchvision.transforms as transforms from data.base_dataset import BaseDataset, get_posenet_transform from data.image_folder import make_dataset from PIL import Image import PIL import random import numpy class UnalignedPoseNetDataset(BaseDataset): def initialize(self, opt): self.opt = opt self.root = opt.dataroot split_file = os.path.join(self.root , 'dataset_'+opt.phase+'.txt') self.A_paths = numpy.loadtxt(split_file, dtype=str, delimiter=' ', skiprows=3, usecols=(0)) self.A_paths = [os.path.join(self.root, path) for path in self.A_paths] self.A_poses = numpy.loadtxt(split_file, dtype=float, delimiter=' ', skiprows=3, usecols=(1,2,3,4,5,6,7)) self.mean_image = numpy.load(os.path.join(self.root , 'mean_image.npy')) if opt.model == "poselstm": self.mean_image = None print("mean image subtraction is deactivated") self.A_size = len(self.A_paths) self.transform = get_posenet_transform(opt, self.mean_image) def __getitem__(self, index): A_path = self.A_paths[index % self.A_size] index_A = index % self.A_size # print('(A, B) = (%d, %d)' % (index_A, index_B)) A_img = Image.open(A_path).convert('RGB') A_pose = self.A_poses[index % self.A_size] A = self.transform(A_img) return {'A': A, 'B': A_pose, 'A_paths': A_path} def __len__(self): return self.A_size def name(self): return 'UnalignedPoseNetDataset'
1,554
34.340909
113
py
RPMG
RPMG-main/poselstm-pytorch/data/base_dataset.py
import torch.utils.data as data from PIL import Image import torchvision.transforms as transforms import numpy import torch class BaseDataset(data.Dataset): def __init__(self): super(BaseDataset, self).__init__() def name(self): return 'BaseDataset' def initialize(self, opt): pass def get_transform(opt): transform_list = [] if opt.resize_or_crop == 'resize_and_crop': osize = [opt.loadSize, opt.loadSize] transform_list.append(transforms.Scale(osize, Image.BICUBIC)) transform_list.append(transforms.RandomCrop(opt.fineSize)) elif opt.resize_or_crop == 'crop': transform_list.append(transforms.RandomCrop(opt.fineSize)) elif opt.resize_or_crop == 'scale_width': transform_list.append(transforms.Lambda( lambda img: __scale_width(img, opt.fineSize))) elif opt.resize_or_crop == 'scale_width_and_crop': transform_list.append(transforms.Resize(opt.loadSize, Image.BICUBIC)) # transform_list.append(transforms.Lambda( # lambda img: __scale_width(img, opt.loadSize))) if opt.isTrain: transform_list.append(transforms.RandomCrop(opt.fineSize)) else: transform_list.append(transforms.CenterCrop(opt.fineSize)) if opt.isTrain and not opt.no_flip: transform_list.append(transforms.RandomHorizontalFlip()) transform_list += [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))] return transforms.Compose(transform_list) def get_posenet_transform(opt, mean_image): transform_list = [] transform_list.append(transforms.Resize(opt.loadSize, Image.BICUBIC)) transform_list.append(transforms.Lambda( lambda img: __subtract_mean(img, mean_image))) transform_list.append(transforms.Lambda( lambda img: __crop_image(img, opt.fineSize, opt.isTrain))) transform_list.append(transforms.Lambda( lambda img: __to_tensor(img))) return transforms.Compose(transform_list) def __scale_width(img, target_width): ow, oh = img.size if (ow == target_width): return img w = target_width h = int(target_width * oh / ow) return img.resize((w, h), Image.BICUBIC) def __subtract_mean(img, mean_image): if mean_image is None: return numpy.array(img).astype('float') return numpy.array(img).astype('float') - mean_image.astype('float') def __crop_image(img, size, isTrain): h, w = img.shape[0:2] # w, h = img.size if isTrain: if w == size and h == size: return img x = numpy.random.randint(0, w - size) y = numpy.random.randint(0, h - size) else: x = int(round((w - size) / 2.)) y = int(round((h - size) / 2.)) return img[y:y+size, x:x+size, :] # return img.crop((x, y, x + size, y + size)) def __to_tensor(img): return torch.from_numpy(img.transpose((2, 0, 1)))
2,999
34.294118
77
py
RPMG
RPMG-main/poselstm-pytorch/data/image_folder.py
############################################################################### # Code from # https://github.com/pytorch/vision/blob/master/torchvision/datasets/folder.py # Modified the original code so that it also loads images from the current # directory as well as the subdirectories ############################################################################### import torch.utils.data as data from PIL import Image import os import os.path IMG_EXTENSIONS = [ '.jpg', '.JPG', '.jpeg', '.JPEG', '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', ] def is_image_file(filename): return any(filename.endswith(extension) for extension in IMG_EXTENSIONS) def make_dataset(dir): images = [] assert os.path.isdir(dir), '%s is not a valid directory' % dir for root, _, fnames in sorted(os.walk(dir)): for fname in fnames: if is_image_file(fname): path = os.path.join(root, fname) images.append(path) return sorted(images) def default_loader(path): return Image.open(path).convert('RGB') class ImageFolder(data.Dataset): def __init__(self, root, transform=None, return_paths=False, loader=default_loader): imgs = make_dataset(root) if len(imgs) == 0: raise(RuntimeError("Found 0 images in: " + root + "\n" "Supported image extensions are: " + ",".join(IMG_EXTENSIONS))) self.root = root self.imgs = imgs self.transform = transform self.return_paths = return_paths self.loader = loader def __getitem__(self, index): path = self.imgs[index] img = self.loader(path) if self.transform is not None: img = self.transform(img) if self.return_paths: return img, path else: return img def __len__(self): return len(self.imgs)
1,954
27.333333
79
py
RPMG
RPMG-main/Pascal3D_Img/S3.3D_Rotation/dataset.py
import numpy as np import torch from torch.utils.data import Dataset, DataLoader import torchvision.transforms as tfs from PIL import Image import scipy.io import os from enum import IntEnum from pytorch3d import transforms as trans # format # data['record']: { # 'filename': string, filename # 'folder': string # 'source': struct for database etc # 'imgname': string, path # 'size': dimensions # 'height' # 'width' # 'depth' # 'segmented': 0/1 ? # 'imgsize': dimensions [h,w,c] # 'database': source database # 'objects': list of objects # 'class': string, class # 'view': string, Frontal/Rear/Left/Right # 'bbox': bounding box # 'bndbox': bounding box as map (might only exist sometimes) # 'orglabel': string includes a bunch of stuff (might only exist sometimes) # 'truncated': 0/1 # 'occluded': 0/1 # 'difficult': 0/1 # 'anchors': list of anchors and their coordinates in image, contents depend on class, map between strings and anchors # 'location': [], or [x,y] position in image # 'status': something # 'viewpoint': essentially angle # 'azimuth_coarse': # 'azimuth': # 'elevation_coarse': # 'elevation': # 'distance': # 'px': center # 'py': center # 'theta': # 'error': # 'interval_azimuth': # 'interval_elevation': # 'num_anchor': # 'viewport': # 'cad_index': related to which cad was used # 'polygon': empty list # 'point': empty list # 'part': empty list # 'hasparts': 0/1 # 'actions': list # 'hasactions': 0/1 # 'mask': 0/1 # sometimes additional dimensions of size 1 is inserted probably due to matlab def get_mat_element(data): while isinstance(data, np.ndarray): if len(data) == 0: raise PascalParseError("Encountered Empty List") if len(data) > 1: x = data[0] for y in data: if y != x: print(data[0]) print(data[1]) raise (Exception("blah" + str(data))) data = data[0] return data def get_mat_list(data): data_old = data while len(data) == 1: data_old = data data = data[0] if isinstance(data, np.void): return data_old return data def pascal3d_get_bbox(data): names = data.dtype.names if 'bbox' in names: bbox = data['bbox'] bbox = get_mat_list(bbox) return list(map(float, bbox)) elif 'bndbox' in names: raise Exception("NOT IMPLEMENTED") raise PascalParseError("could not parse bounding box") class PascalParseError(Exception): def __init__(self, string): super().__init__(string) class PascalClasses(IntEnum): AEROPLANE = 1 BICYCLE = 2 BOAT = 3 BOTTLE = 4 BUS = 5 CAR = 6 CHAIR = 7 DININGTABLE = 8 MOTORBIKE = 9 SOFA = 10 TRAIN = 11 TVMONITOR = 12 def __str__(self): return self.name.lower() pascal_3d_str_enum_map = {} for v in PascalClasses: pascal_3d_str_enum_map[str(v)] = v failed_parse_strings = set() def pascal3d_get_class(data): class_str = get_mat_element(data['class']) try: return pascal_3d_str_enum_map[class_str.lower()] except KeyError: failed_parse_strings.add(class_str.lower()) # print("unknown class: " + class_str) raise PascalParseError("could not parse class") def pascal3d_idx_to_str(idx): return str(PascalClasses(idx)) def parse_single_angle(viewpoint, angle_name): names = viewpoint.dtype.names if angle_name in names: try: angle = get_mat_element(viewpoint[angle_name]) return float(angle) except PascalParseError: pass angle_name_coarse = angle_name + "_coarse" if angle_name_coarse in names: angle = get_mat_element(viewpoint[angle_name_coarse]) return float(angle) raise PascalParseError("No angle found") def pascal3d_get_angle(data): viewpoint = get_mat_element(data['viewpoint']) azimuth = parse_single_angle(viewpoint, 'azimuth') elevation = parse_single_angle(viewpoint, 'elevation') theta = parse_single_angle(viewpoint, 'theta') if azimuth == 0 and elevation == 0 and theta == 0: raise PascalParseError("Angle probably not entered") return [azimuth, elevation, theta] # note in degree def pascal3d_get_point(data): viewpoint = get_mat_element(data['viewpoint']) px = float(get_mat_element(viewpoint['px'])) py = float(get_mat_element(viewpoint['py'])) return [px, py] def pascal3d_get_distance(data): viewpoint = get_mat_element(data['viewpoint']) return float(get_mat_element(viewpoint['distance'])) DICT_BOUNDING_BOX = 'bounding_box' DICT_CLASS = 'class' DICT_ANGLE = 'angle' DICT_OCCLUDED = 'occluded' DICT_TRUNCATED = 'truncated' DICT_DIFFICULT = 'difficult' DICT_POINT = 'px' DICT_OBJECT_LIST = 'obj_list' DICT_OBJECT_INSTANCE = 'obj_instance' DICT_FILENAME = 'filename' DICT_DISTANCE = 'distance' DICT_CAMERA = 'camera' DICT_CAD_INDEX = 'cad_index' def get_pascal_camera_params(mat_data): viewpoint = get_mat_element(mat_data['viewpoint']) try: focal = get_mat_element(viewpoint['focal']) except PascalParseError: print("default_focal") focal = 1 if focal != 1: print("focal {}".format(focal)) try: viewport = get_mat_element(viewpoint['viewport']) except PascalParseError: print("default_viewpoer") viewport = 3000 if viewport != 3000: print("viewport {}".format(viewport)) return float(focal), float(viewport) def mat_data_to_dict_data(mat_data, folder): record = get_mat_element(mat_data['record']) ret = {} objects = [] mat_objects = get_mat_list(record['objects']) for obj in mat_objects: ret_obj = {} try: ret_obj[DICT_BOUNDING_BOX] = pascal3d_get_bbox(obj) ret_obj[DICT_CLASS] = pascal3d_get_class(obj).value ret_obj[DICT_ANGLE] = pascal3d_get_angle(obj) ret_obj[DICT_OCCLUDED] = bool(get_mat_element(obj['occluded'])) ret_obj[DICT_TRUNCATED] = bool(get_mat_element(obj['truncated'])) ret_obj[DICT_POINT] = pascal3d_get_point(obj) ret_obj[DICT_DIFFICULT] = bool(get_mat_element(obj['difficult'])) ret_obj[DICT_DISTANCE] = pascal3d_get_distance(obj) ret_obj[DICT_CAMERA] = get_pascal_camera_params(obj) ret_obj[DICT_CAD_INDEX] = int(get_mat_element(obj['cad_index'])) objects.append(ret_obj) except PascalParseError as e: pass ret[DICT_OBJECT_LIST] = objects ret[DICT_FILENAME] = os.path.join(folder, get_mat_element(record['filename'])) return ret def process_annotated_image( im, left, top, right, bottom, azimuth, elevation, theta, augment, reverse_theta, crop): # perturb bbox randomly # inputs are matlab (start at 1) if augment: max_shift = 7 left = left - 1 + np.random.randint(-max_shift, max_shift+1) top = top - 1 + np.random.randint(-max_shift, max_shift+1) right = right - 1 + np.random.randint(-max_shift, max_shift+1) bottom = bottom - 1 + np.random.randint(-max_shift, max_shift+1) else: left = left - 1 top = top - 1 right = right - 1 bottom = bottom - 1 width, height = im.size left = min(max(left,0), width) top = min(max(top, 0), height) right = min(max(right, 0), width) bottom = min(max(bottom, 0), height) # Resizing can change aspect ratio, so we could adjust the ground truth # rotation accordingly (leaving as-is since initial results didn't change) if crop: im = im.crop((left, top, right, bottom)) im = im.resize([224,224]) # Inputs are in degrees, convert to rad. az = azimuth* np.pi / 180.0 el = elevation * np.pi / 180.0 th = theta * np.pi / 180.0 # Reversing theta for RenderForCNN data since that theta was set from filename # which has negative theta (see github.com/ShapeNet/RenderForCNN). if reverse_theta: th = -th if augment: # Flip rand = np.random.uniform(0,1) if rand < 0.5: az = -az th = -th im = im.transpose(Image.FLIP_LEFT_RIGHT) im = tfs.ColorJitter(brightness=0.1, contrast=0.5, hue=0.2, saturation=0.5)(im) #im_crop = tf.clip_by_value(im_crop, 0.0, 1.0) # R = R_z(th) * R_x(el−pi/2) * R_z(−az) R1 = trans.euler_angles_to_matrix(torch.Tensor([-az, 0, 0]).unsqueeze(0), 'ZYX') R2 = trans.euler_angles_to_matrix(torch.Tensor([th, 0, el - np.pi / 2.0]).unsqueeze(0), 'ZYX') R = torch.bmm(R2, R1) return tfs.ToTensor()(im), R class PascalParseError(Exception): def __init__(self, string): super().__init__(string) def get_mat_element(data): while isinstance(data, np.ndarray): if len(data) == 0: raise PascalParseError("Encountered Empty List") if len(data) > 1: x = data[0] for y in data: if y != x: print(data[0]) print(data[1]) raise(Exception("blah" + str(data))) data = data[0] return data def create_imagenet_anno(data_folder, save_folder, category, validation_split_size=0.3): ImageNet_anno_folder = os.path.join(data_folder, 'Annotations', category + '_imagenet') ImageNet_img_folder = os.path.join(data_folder, 'Images', category + '_imagenet') imagenet_split_train_path = os.path.join(data_folder, 'Image_sets', category + '_imagenet_train.txt') imagenet_split_test_path = os.path.join(data_folder, 'Image_sets', category + '_imagenet_val.txt') # train and val train_val_split = [] with open(imagenet_split_train_path,'r') as f: while True: l = f.readline() if len(l) == 0: break while l[-1] in ('\n', '\r'): l = l[:-1] if len(l) == 0: continue train_val_split.append(l) train_val_split = sorted(train_val_split) val_idx = (np.arange(len(train_val_split) * validation_split_size) / validation_split_size).astype(np.int) val_split = [train_val_split[i] for i in val_idx] train_split = sorted(list(set(train_val_split) - set(val_split))) # test test_split = [] with open(imagenet_split_test_path, 'r') as f: while True: l = f.readline() if len(l) == 0: break while l[-1] in ('\n', '\r'): l = l[:-1] if len(l) == 0: continue test_split.append(l) split = [] split.append(train_split) split.append(val_split) split.append(test_split) save_path = [] save_path.append(os.path.join(save_folder, category+'_imagenet_train')) save_path.append(os.path.join(save_folder, category + '_imagenet_val')) save_path.append(os.path.join(save_folder, category + '_imagenet_test')) name_lst = ['train', 'val', 'test'] #create new annotation of ImageNet for i in range(len(name_lst)): if not os.path.isdir(save_path[i]): os.mkdir(save_path[i]) for instance in split[i]: annopath = os.path.join(ImageNet_anno_folder, instance+'.mat') anno = scipy.io.loadmat(annopath) dict = mat_data_to_dict_data(anno, ImageNet_img_folder) for num, obj in enumerate(dict['obj_list']): obj['imgpath'] = dict['filename'] if obj['occluded'] or obj['truncated'] or obj['difficult']: continue save_file = os.path.join(save_path[i], instance+'_'+str(num)+'.npy') np.save(save_file, obj) print('Total imagenet %s obj number for %s: %d' %(category, name_lst[i], len(os.listdir(save_path[i])))) def create_pascal_anno(data_folder, save_folder, category): pascal_anno_folder = os.path.join(data_folder, 'Annotations', category + '_pascal') pascal_img_folder = os.path.join(data_folder, 'Images', category + '_pascal') # train train_split = os.listdir(pascal_anno_folder) save_path = os.path.join(save_folder, category+'_pascal_train') if not os.path.isdir(save_path): os.mkdir(save_path) for instance in train_split: annopath = os.path.join(pascal_anno_folder, instance) anno = scipy.io.loadmat(annopath) dict = mat_data_to_dict_data(anno, pascal_img_folder) for num, obj in enumerate(dict['obj_list']): obj['imgpath'] = dict['filename'] if obj['occluded'] or obj['truncated'] or obj['difficult']: continue save_file = os.path.join(save_path, instance[:-4] + '_' + str(num) + '.npy') np.save(save_file, obj) print('Total pascal %s obj number: %d' %(category, len(os.listdir(save_path)))) class Pascal3dDataset(torch.utils.data.Dataset): def __init__(self, anno_folder, augment=False, voc_train_addition_folder=None): self.anno_paths = [os.path.join(anno_folder, i) for i in os.listdir(anno_folder)] if voc_train_addition_folder != None: self.anno_paths.extend(os.listdir(voc_train_addition_folder)) self.augment = augment self.size = len(self.anno_paths) print('Load Pascal Dataset, Length:', self.size) def __getitem__(self, idx): annopath = self.anno_paths[idx] anno = np.load(annopath, allow_pickle=True).item() imgpath = anno['imgpath'] img = Image.open(imgpath) left, top, right, bottom = anno['bounding_box'] azimuth, elevation, theta = anno['angle'] img, R = process_annotated_image(img, left, top, right, bottom, azimuth, elevation, theta, augment=self.augment, reverse_theta=False, crop=True) return img, R def __len__(self): return self.size class SyntheticDataset(torch.utils.data.Dataset): def __init__(self, syn_folder, category_id): img_folder = os.path.join(syn_folder, category_id) instance_lst = os.listdir(img_folder) self.paths = [] for i in instance_lst: path_i = os.path.join(img_folder, i) lst_i = os.listdir(path_i) for j in lst_i: if '.png' in j: self.paths.append(os.path.join(path_i, j)) self.size = len(self.paths) print('Load Synthetic Dataset, Length:', self.size) def __getitem__(self, idx): fpath = self.paths[idx] img = Image.open(fpath) anno = fpath.split('_') azimuth = int(anno[-4][1:]) elevation = int(anno[-3][1:]) theta = int(anno[-2][1:]) img, R = process_annotated_image(img, 0, 0, 0, 0, azimuth, elevation, theta, augment=False, reverse_theta=True, crop=False) return img, R def __len__(self): return self.size name2id = { 'aeroplane':'02691156', 'bicycle':'02834778', 'boat':'02858304', 'bottle':'02876657', 'bus':'02924116', 'car':'02958343', 'chair':'03001627', 'diningtable':'04379243', 'motorbike':'03790512', 'sofa': '04256520', 'train':'04468005', 'tvmonitor':'03211117', } def get_dataloader(mode, config): pascal3d_path = config.pascal3d_path syn_path = config.syn_path save_anno_path = os.path.join(config.pascal3d_path, 'my_anno') category = config.category if config.create_anno and mode == 'train': if not os.path.isdir(save_anno_path): os.mkdir(save_anno_path) create_imagenet_anno(pascal3d_path, save_anno_path, category) create_pascal_anno(pascal3d_path, save_anno_path, category) if mode == 'train': anno_path_train = os.path.join(save_anno_path, category + '_imagenet_train') if config.voc_train: anno_path_train_voc = os.path.join(save_anno_path, category + '_pascal_train') pascal3d_train_dataset = Pascal3dDataset(anno_path_train, augment=True,voc_train_addition_folder=anno_path_train_voc) else: pascal3d_train_dataset = Pascal3dDataset(anno_path_train, augment=True) train_loader1 = DataLoader(pascal3d_train_dataset, batch_size=int(config.batch_size/2), shuffle=True, num_workers=config.num_workers) syn_train_dataset = SyntheticDataset(syn_path, name2id[category]) train_loader2 = DataLoader(syn_train_dataset, batch_size=int(config.batch_size/2), shuffle=True, num_workers=config.num_workers) return train_loader1, train_loader2 elif mode == 'test': anno_path_test = os.path.join(save_anno_path, category + '_imagenet_test') pascal3d_test_dataset = Pascal3dDataset(anno_path_test) test_loader = DataLoader(pascal3d_test_dataset, batch_size=config.batch_size, shuffle=False, num_workers=config.num_workers) return test_loader if __name__ == '__main__': import argparse from os.path import join arg_parser = argparse.ArgumentParser() arg_parser.add_argument("--data_root", type=str, default='../dataset') args = arg_parser.parse_args() pascal3d_path = join(args.data_root, 'PASCAL3D+_release1.1') syn_path = join(args.data_root, 'syn_images_cropped_bkg_overlaid') save_anno_path = os.path.join(pascal3d_path, 'my_anno') category_lst = ['aeroplane','sofa', 'bicycle','boat','bottle','bus', 'car', 'chair', 'diningtable', 'motorbike', 'train', 'tvmonitor'] if not os.path.isdir(save_anno_path): os.mkdir(save_anno_path) #create annotation first!!! for category in category_lst: create_imagenet_anno(pascal3d_path, save_anno_path, category) # # No need to do below line # create_pascal_anno(pascal3d_path, save_anno_path, category) ''' anno_path_train = os.path.join(save_anno_path, category+'_imagenet_train') anno_path_val = os.path.join(save_anno_path, category+'_imagenet_val') anno_path_test = os.path.join(save_anno_path, category+'_imagenet_test') pascal3d_train_dataset = Pascal3dDataset(anno_path_train, augment=True) pascal3d_test_dataset = Pascal3dDataset(anno_path_test) pascal3d_val_dataset = Pascal3dDataset(anno_path_val) syn_train_dataset = SyntheticDataset(syn_path, name2id[category]) train_loader1 = DataLoader(pascal3d_train_dataset, batch_size=16, shuffle=True, num_workers=4) train_loader2 = DataLoader(syn_train_dataset, batch_size=16, shuffle=True, num_workers=4) #training for real_data, syn_data in zip(train_loader1,train_loader2): real_img, real_gt = real_data syn_img, syn_gt = syn_data train_data = torch.cat((real_img,syn_img),0) #[32,3,224,224] train_gt = torch.cat((real_gt,syn_gt),0) #[32,1,3,3] print(train_data.shape, train_gt.shape) '''
19,144
35.328273
152
py
RPMG
RPMG-main/Pascal3D_Img/S3.3D_Rotation/networks.py
import torch from torch import nn from torch import Tensor from typing import Callable, Any, Optional, List def get_network(config): return MobileNetV2(config.num_classes) def set_requires_grad(nets, requires_grad=False): """Set requies_grad=Fasle for all the networks to avoid unnecessary computations Parameters: nets (network list) -- a list of networks requires_grad (bool) -- whether the networks require gradients or not """ if not isinstance(nets, list): nets = [nets] for net in nets: if net is not None: for param in net.parameters(): param.requires_grad = requires_grad def _make_divisible(v: float, divisor: int, min_value: Optional[int] = None) -> int: """ This function is taken from the original tf repo. It ensures that all layers have a channel number that is divisible by 8 It can be seen here: https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py """ if min_value is None: min_value = divisor new_v = max(min_value, int(v + divisor / 2) // divisor * divisor) # Make sure that round down does not go down by more than 10%. if new_v < 0.9 * v: new_v += divisor return new_v class ConvBNActivation(nn.Sequential): def __init__( self, in_planes: int, out_planes: int, kernel_size: int = 3, stride: int = 1, groups: int = 1, norm_layer: Optional[Callable[..., nn.Module]] = None, activation_layer: Optional[Callable[..., nn.Module]] = None, dilation: int = 1, ) -> None: padding = (kernel_size - 1) // 2 * dilation if norm_layer is None: norm_layer = nn.BatchNorm2d if activation_layer is None: activation_layer = nn.ReLU6 super(ConvBNReLU, self).__init__( nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, dilation=dilation, groups=groups, bias=False), norm_layer(out_planes), activation_layer(inplace=True) ) self.out_channels = out_planes # necessary for backwards compatibility ConvBNReLU = ConvBNActivation class InvertedResidual(nn.Module): def __init__( self, inp: int, oup: int, stride: int, expand_ratio: int, norm_layer: Optional[Callable[..., nn.Module]] = None ) -> None: super(InvertedResidual, self).__init__() self.stride = stride assert stride in [1, 2] if norm_layer is None: norm_layer = nn.BatchNorm2d hidden_dim = int(round(inp * expand_ratio)) self.use_res_connect = self.stride == 1 and inp == oup layers: List[nn.Module] = [] if expand_ratio != 1: # pw layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1, norm_layer=norm_layer)) layers.extend([ # dw ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim, norm_layer=norm_layer), # pw-linear nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False), norm_layer(oup), ]) self.conv = nn.Sequential(*layers) self.out_channels = oup self._is_cn = stride > 1 def forward(self, x: Tensor) -> Tensor: if self.use_res_connect: return x + self.conv(x) else: return self.conv(x) class MobileNetV2(nn.Module): def __init__( self, num_classes: int = 1000, width_mult: float = 1.0, inverted_residual_setting: Optional[List[List[int]]] = None, round_nearest: int = 8, block: Optional[Callable[..., nn.Module]] = None, norm_layer: Optional[Callable[..., nn.Module]] = None ) -> None: """ MobileNet V2 main class Args: num_classes (int): Number of classes width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount inverted_residual_setting: Network structure round_nearest (int): Round the number of channels in each layer to be a multiple of this number Set to 1 to turn off rounding block: Module specifying inverted residual building block for mobilenet norm_layer: Module specifying the normalization layer to use """ super(MobileNetV2, self).__init__() if block is None: block = InvertedResidual if norm_layer is None: norm_layer = nn.BatchNorm2d input_channel = 32 last_channel = 1280 if inverted_residual_setting is None: inverted_residual_setting = [ # t, c, n, s [1, 16, 1, 1], [6, 24, 2, 2], [6, 32, 3, 2], [6, 64, 4, 2], [6, 96, 3, 1], [6, 160, 3, 2], [6, 320, 1, 1], ] # only check the first element, assuming user knows t,c,n,s are required if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4: raise ValueError("inverted_residual_setting should be non-empty " "or a 4-element list, got {}".format(inverted_residual_setting)) # building first layer input_channel = _make_divisible(input_channel * width_mult, round_nearest) self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest) features: List[nn.Module] = [ConvBNReLU(3, input_channel, stride=2, norm_layer=norm_layer)] # building inverted residual blocks for t, c, n, s in inverted_residual_setting: output_channel = _make_divisible(c * width_mult, round_nearest) for i in range(n): stride = s if i == 0 else 1 features.append(block(input_channel, output_channel, stride, expand_ratio=t, norm_layer=norm_layer)) input_channel = output_channel # building last several layers features.append(ConvBNReLU(input_channel, self.last_channel, kernel_size=1, norm_layer=norm_layer)) # make it nn.Sequential self.features = nn.Sequential(*features) # building classifier self.classifier = nn.Sequential( nn.Dropout(0.2), nn.Linear(self.last_channel, 256), nn.BatchNorm1d(256), nn.ReLU6(inplace=True), nn.Linear(256, 64), nn.BatchNorm1d(64), nn.ReLU6(inplace=True), nn.Linear(64, num_classes), ) # weight initialization for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out') if m.bias is not None: nn.init.zeros_(m.bias) elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): nn.init.ones_(m.weight) nn.init.zeros_(m.bias) elif isinstance(m, nn.Linear): nn.init.normal_(m.weight, 0, 0.01) nn.init.zeros_(m.bias) def _forward_impl(self, x: Tensor) -> Tensor: # This exists since TorchScript doesn't support inheritance, so the superclass method # (this one) needs to have a name other than `forward` that can be accessed in a subclass x = self.features(x) # Cannot use "squeeze" as batch-size can be 1 => must use reshape with x.shape[0] x = nn.functional.adaptive_avg_pool2d(x, (1, 1)).reshape(x.shape[0], -1) x = self.classifier(x) return x def forward(self, x: Tensor) -> Tensor: return self._forward_impl(x) if __name__ == '__main__': net = MobileNetV2(6) input = torch.randn(4, 3, 227, 227) print(net(input).shape)
8,009
35.244344
116
py
RPMG
RPMG-main/Pascal3D_Img/S3.3D_Rotation/trainval_workdir.py
""" @Author : Shuai Liao """ import torch import torch.optim import os, sys, time from time import gmtime, strftime import numpy as np from math import pi from easydict import EasyDict as edict from collections import OrderedDict as odict # from basic.common import Open, env, add_path, RefObj as rdict, argv2dict, is_py3 this_dir = os.path.dirname(os.path.realpath(__file__)) # from pytorch_util.libtrain.yaml_netconf import parse_yaml, import_module_v2 from pytorch_util.libtrain.tools import get_stripped_DataParallel_state_dict, patch_saved_DataParallel_state_dict from txt_table_v1 import TxtTable # from tensorboardX import SummaryWriter from tqdm import tqdm # =========== Parsing from working path ========= pwd = os.getcwd() # Assume: $base_dir/S3.3D_Rotation/{MtdFamily}/{MtdType} MtdFamily, MtdType = pwd.split(os.sep)[-2:] # ================================================ # ------- args from convenient run yaml --------- # For the purpose that no need to specific each run (without argparse) convenient_run_argv_yaml = \ ''' MtdFamily : {MtdFamily} # e.g. regQuatNet MtdType : {MtdType} # e.g. reg_Direct, reg_Sexp, reg_Sflat net_module : {net_module} # same as MtdType here, namely import from 'MtdType'.py net_arch : {net_arch} # e.g. alexnet, vgg16 base_dir : {base_dir} # e.g. path/to/S3.3D_Rotation LIB_DIR : {base_dir}/lib train_view : 100V # 20V # work_dir : './snapshots/{net_arch}' nr_epoch : 150 test_step_epoch : 10 this_dir : {this_dir} base_lr : 0.001 '''.format(net_arch='alexnet', base_dir=this_dir, this_dir=pwd, # MtdFamily=MtdFamily, MtdType=MtdType, net_module=MtdFamily, ) run_args = parse_yaml(convenient_run_argv_yaml) # ------- arg from argparse ----------------- import argparse parser = argparse.ArgumentParser(description='PyTorch Training') parser.add_argument('conf_yml_file', default='', type=str, metavar='PATH', help='path to conf_yml_file (default: none)') # parser.add_argument('work_dir' , default='', type=str, metavar='PATH',) # help='path to work_dir (default: none)') parser.add_argument('gpu_ids', default='', type=str, metavar='PATH', help='e.g. 0 or 0,1,2,3') parser.add_argument('--resume', action="store_true", default=False, help='to resume by the last checkpoint') parser.add_argument('--pretrain', default=None, type=str, metavar='PATH', help='path to pretrained checkpoint (default: none)') parser.add_argument('--optimizer', default='SGD', type=str, help='SGD or Adam') parser.add_argument('--test_only', action="store_true", default=False, help='only do test once.') _pargs, _rest = parser.parse_known_args() # parser.parse_args() # parse the rest undefined args with "--key=value" form. _cmd_args = argv2dict(_rest) _cmd_args.update(vars(_pargs)) # run_args.update(_cmd_args) from string import Template template_str = open(os.path.join(this_dir, 'conf_template.yml')).read() template = Template(template_str) print(run_args) conf_yml_str = template.substitute(run_args) # -- parse module_yml_file opt = parse_yaml(conf_yml_str) # opt.update(run_args) # from ordered_easydict import Ordered_EasyDict as oedict opt = oedict(opt) # Use opt for reference all configurations. # ------ Import modules ---------- [(_dataset_module, _dataset_kwargs), netcfg] = import_module_v2(opt.IMPORT_dataset) # pred2angle [(_net_module, _net_kwargs)] = import_module_v2(opt.IMPORT_makenet) # [_net_type] [(eval_cates, _), (compute_geo_dists, _)] = import_module_v2(opt.IMPORT_eval.GTbox) net_arch = opt.net_arch # or _net_kwargs.net_arch _cfg = netcfg[net_arch] # [opt.net_arch] np.random.seed(_cfg.RNG_SEED) torch.manual_seed(_cfg.RNG_SEED) if opt.use_gpu: torch.cuda.manual_seed(_cfg.RNG_SEED) # ---------------------------------------------------------------------------------------------------[dataset] dataset_test = _dataset_module(collection='test', sampling=0.2, **_dataset_kwargs) # dataset_train = _dataset_module(collection='train', **_dataset_kwargs) # # change the sampling of dataset: e.g. sampling: {imagenet:1.0, synthetic:1.0} # 'ModelNet10/SO3_100V.white_BG_golden_FG' # From default.run.conf.yml.sh if opt.cates is None: opt.cates = dataset_train.cates cates = opt.cates if 'batch_size' in opt: batch_size = opt.batch_size else: batch_size = _cfg.TRAIN.BATCH_SIZE nr_GPUs = len(opt.gpu_ids) assert nr_GPUs >= 1, opt.gpu_ids if nr_GPUs > 1: print('--------------------- Use multiple-GPU %s -------------------------' % opt.gpu_ids) print(' batch_size = %s' % batch_size) # (batch_size*nr_GPUs) print(' num_workers = %s' % (opt.num_workers * nr_GPUs)) # train_loader = torch.utils.data.DataLoader(dataset_train, batch_size=batch_size, shuffle=True, # batch_size*nr_GPUs num_workers=opt.num_workers * nr_GPUs, pin_memory=opt.pin_memory, sampler=None) # ---------------------------------------------------------------------------------------------------[model] print('[makenet] nr_cate: ', len(cates)) model = _net_module(nr_cate=len(cates), **_net_kwargs) # len(_cfg.cates)) if 'fix_conv1_conv2' in opt.keys() and opt.fix_conv1_conv2: model.fix_conv1_conv2() # watch_targets = model.targets # ---------------------------------------------------------------------------------------------------[optimizer] params = [] for name, param in model.named_parameters(): print('----(*) ', name) if param.requires_grad: params.append(param) print('[Optimizer] %s' % opt.optimizer) if opt.optimizer == 'Adam': optimizer = torch.optim.Adam(params, lr=opt.base_lr) # , weight_decay=opt.weight_decay) elif opt.optimizer == 'SGD': optimizer = torch.optim.SGD(params, opt.base_lr, # model.parameters(), opt.base_lr, momentum=opt.momentum, weight_decay=opt.weight_decay) else: raise NotImplementedError work_dir = opt.work_dir work_dir += '.%s' % opt.train_view _short_work_dir = os.path.abspath(work_dir)[len(os.path.abspath(opt.base_dir)) + 1:] # global state variables. start_it = 0 start_epoch = 0 from pytorch_util.libtrain import rm_models, list_models from pytorch_util.libtrain.reducer import reducer, reducer_group # Log file. script_name, _ = os.path.splitext(os.path.basename(__file__)) log_filename = '%s/%s.log' % (work_dir, script_name) if os.path.exists(log_filename): # backup previous content. pre_log_content = open(log_filename).read() logf = Open(log_filename, 'w') def logprint(s): print("\r%s " % s) logf.write(s + "\n") # -- Resume or use pretrained (Note not imagenet pretrain.) assert not (opt.resume and opt.pretrain is not None), 'Only resume or pretrain can exist.' if opt.resume: iter_nums, net_name = list_models(work_dir) # ('snapshots') assert len(iter_nums) > 0, "No models available" latest_model_name = os.path.join(work_dir, '%s_iter_%s.pth.tar' % (net_name, iter_nums[-1])) print('\n\nResuming from: %s \n\n' % latest_model_name) if os.path.isfile(latest_model_name): print("=> loading checkpoint '{}'".format(latest_model_name)) checkpoint = torch.load(latest_model_name) start_it, start_epoch = checkpoint['it_and_epoch'] # mainly for control lr: (it, epoch) model.load_state_dict(checkpoint['state_dict']) optimizer.load_state_dict(checkpoint['optimizer']) # fix for pytorch 4.0.x [https://github.com/jwyang/faster-rcnn.pytorch/issues/222] for state in optimizer.state.values(): for k, v in state.items(): if isinstance(v, torch.Tensor): state[k] = v.cuda() print("=> loaded checkpoint '{}' (it_and_epoch {})" .format(latest_model_name, checkpoint['it_and_epoch'])) else: print("=> no checkpoint found at '{}'".format(latest_model_name)) # unnecessary line elif opt.pretrain is not None: print('\n\nUsing pretrained: %s \n\n' % opt.pretrain) if os.path.isfile(opt.pretrain): print("=> loading checkpoint '{}'".format(opt.pretrain)) checkpoint = torch.load(opt.pretrain) model.load_state_dict(checkpoint['state_dict']) optimizer.load_state_dict(checkpoint['optimizer']) print("=> loaded checkpoint '{}' (it_and_epoch {})" .format(opt.pretrain, checkpoint['it_and_epoch'])) else: print("=> no checkpoint found at '{}'".format(opt.pretrain)) # Check if use multi-gpus. # Note should be after any "model.load_state_dict()" call! if opt.use_gpu: # model.cuda() if nr_GPUs > 1: # multi-GPUs opt['mGPUs']: # see: https://github.com/pytorch/examples/blob/master/imagenet/main.py if net_arch.startswith('alexnet'): model.trunk.Convs = torch.nn.DataParallel(model.trunk.Convs) elif net_arch.startswith('vgg'): model.trunk.features = torch.nn.DataParallel(model.trunk.features) else: model.trunk = torch.nn.DataParallel(model.trunk) model.cuda() if not os.path.exists(work_dir): print("[Make new dir] ", work_dir) os.makedirs(work_dir) disp_interval = 10 if ('disp_interval' not in opt) else opt['disp_interval'] """ (nr_iter * batch_size)/nr_train = nr_epoch where nr_train=28647/29786 (40000*200)/29786. = 268.6 (40000* 50)/29786. = 67.1 (40000*_cfg.TRAIN.BATCH_SIZE)/29786. / 2 /10 """ def adjust_learning_rate(optimizer, epoch): """Sets the learning rate to the initial LR decayed by 10 every _N_ epochs""" _N_ = int((40000 * batch_size) / 29786. / 2 / 10) * 10 # '2' mean at most decay 2 times. lr = opt.base_lr * (0.1 ** (epoch // _N_)) # 300)) for param_group in optimizer.param_groups: param_group['lr'] = lr return lr def adjust_learning_rate_by_iter(optimizer, cur_iter, max_iter): """Sets the learning rate to the initial LR decayed by 10 every _N_ epochs""" _N_ = max_iter // 3 # add just learning rate 3 times. lr = opt.base_lr * (0.1 ** (max(cur_iter, 0) // _N_)) # 300)) for param_group in optimizer.param_groups: param_group['lr'] = lr return lr def adjust_learning_rate_by_epoch(optimizer, cur_epoch, max_epoch): """Sets the learning rate to the initial LR decayed by 10 every _N_ epochs""" _N_ = max_epoch // 3 # add just learning rate 3 times. lr = opt.base_lr * (0.1 ** (max(cur_epoch, 0) // _N_)) # 300)) for param_group in optimizer.param_groups: param_group['lr'] = lr return lr def save_checkpoint(state, filename='checkpoint.pth.tar'): torch.save(state, filename) def test(dataset_test, work_dir, test_model=None, marker='epoch'): out_rslt_path = work_dir + '/temp.out_rslt_path.txt' out_eval_path = work_dir + '/temp.out_eval_path.txt' if test_model is None: test_model = model # ---- Load trained weights here.------ assert os.path.exists(work_dir) # iter_nums, net_name = list_models(work_dir, marker=marker) saved_iter_num = iter_nums[-1] pretrained_model = work_dir + '/%s_%s_%s.pth.tar' % ( net_name, marker, saved_iter_num) # select maxmun iter number. print('[pretrained_model] ', pretrained_model) checkpoint = torch.load(pretrained_model) # load weights here. _state_dict = patch_saved_DataParallel_state_dict(checkpoint['state_dict']) test_model.load_state_dict(_state_dict) # switch to train mode test_model.eval() gLoss_redu = reducer_group(*watch_targets) gPred_redu = reducer_group(*['quat']) pre_time = time.time() it = -1 epoch = -1 # keys = dataset_test.keys test_loader = torch.utils.data.DataLoader(dataset_test, batch_size=_cfg.TEST.BATCH_SIZE * nr_GPUs, shuffle=False, num_workers=opt.num_workers * nr_GPUs, pin_memory=opt.pin_memory, sampler=None) with torch.no_grad(): pbar = tqdm(test_loader) for _i_, sample_batched in enumerate(pbar): pbar.set_description("[work_dir] %s " % _short_work_dir) it += 1 # Note: Tensor.cuda() Returns a copy of this object in CUDA memory. label = torch.autograd.Variable(sample_batched['label'].cuda(non_blocking=True)) data = torch.autograd.Variable(sample_batched['data'].cuda(non_blocking=True)) # formulate GT dict _gt_targets = test_model.gt_targets if hasattr(test_model, 'gt_targets') else test_model.targets GT = edict() for tgt in _gt_targets: GT[tgt] = torch.autograd.Variable(sample_batched[tgt].cuda(non_blocking=True)) # compute Pred output Prob = test_model(data, label) # compute Loss for each target and formulate Loss dictionary. Loss = test_model.compute_loss(Prob, GT) total_loss = 0 for tgt in watch_targets: total_loss += Loss[tgt] # predict target angles value Pred = test_model.compute_pred(Prob) gLoss_redu.collect(Loss) # pass in dict of all loss (loss_a, loss_e, loss_t). gPred_redu.collect(Pred, squeeze=False) # print loss info cur_time = time.time() time_consume = cur_time - pre_time pre_time = cur_time print('\r %s [test-iter] %5d / %5d ---------[time_consume] %.2f' % ( strftime("%Y-%m-%d %H:%M:%S", gmtime()), it, len(test_loader), time_consume)) for trgt in watch_targets: _loss = Loss[trgt].data.cpu().numpy().copy() print(' %-15s loss=%.3f' % (trgt, _loss)) if np.isnan(_loss): print("[Warning] Weights explode! Stop training ... ") exit(-1) # pbar.set_description("[work_dir] %s " % os.path.abspath(work_dir)[len(os.path.abspath(opt.base_dir))+1:]) # print ("\r[work_dir] %s \r" % os.path.abspath(work_dir)[len(os.path.abspath(opt.base_dir))+1:],end='') sys.stdout.flush() pred_quats = gPred_redu.reduce()['quat'] # -- Write result to file (Format: # {obj_id} {a} {e} {t} ) txtTbl = TxtTable('{obj_id:<20s} {a:>6.4f} {b:>6.4f} {c:>6.4f} {d:>6.4f}') rslt_lines = [txtTbl.getHeader()] for _k, _quat in zip(keys, pred_quats): _a, _b, _c, _d = _quat rslt_line = txtTbl.format(_k, _a, _b, _c, _d) rslt_lines.append(rslt_line) rslt_lines = '\n'.join(rslt_lines) Open(out_rslt_path, 'w').write(rslt_lines) # print('[out_rslt_path]', out_rslt_path) # -- Do evaluation ('MedError', 'Acc@theta') from numpy_db import npy_table rc_tbl = npy_table(dataset_test.recs) # summary_str = eval_cates(out_rslt_path, rc_tbl, cates=opt.cates, theta_levels_str='pi/6 pi/12 pi/24') # ['aeroplane','boat','car']) Open(out_eval_path, 'w').write(summary_str) print(summary_str) reca = TxtTable().load_as_recarr(out_eval_path, fields=['MedError', 'Acc@pi/6', 'Acc@pi/12', 'Acc@pi/24']) return reca[-1] def train(nr_disp=5000): os.system('rm -rf %s ' % (work_dir + "/logs")) logger = SummaryWriter(work_dir + "/logs") nr_epoch = opt.nr_epoch nr_iter = opt.nr_epoch * (len(dataset_train) / batch_size) # 130800 # based on iter disp_interval = int(nr_iter / nr_disp) pre_time = time.time() it = start_it - 1 # -1 epoch = start_epoch - 1 # -1 # while epoch < nr_epoch: epoch += 1 # Do test first if epoch % opt.test_step_epoch == 0: mederr, acc6, acc12, acc24 = test(dataset_test, work_dir, model) logger.add_scalars('acc/test', {'MedError': mederr, 'Acc@pi/6': acc6, 'Acc@pi/12': acc12, 'Acc@pi/24': acc24}, epoch + 1) # switch to train mode model.train() lr = adjust_learning_rate_by_epoch(optimizer, epoch, nr_epoch) # opt.base_lr pbar = tqdm(train_loader) for _i_, sample_batched in enumerate(pbar): pbar.set_description("[work_dir] %s B=%s " % (_short_work_dir, batch_size)) rec_inds = sample_batched['idx'].numpy() # it += 1 label = torch.autograd.Variable(sample_batched['label'].cuda(non_blocking=True)) data = torch.autograd.Variable(sample_batched['data'].cuda(non_blocking=True)) # formulate GT dict _gt_targets = model.gt_targets if hasattr(model, 'gt_targets') else model.targets GT = edict() for tgt in _gt_targets: GT[tgt] = torch.autograd.Variable(sample_batched[tgt].cuda(non_blocking=True)) # compute Pred output Prob = model(data, label) # compute Loss for each target and formulate Loss dictionary. Loss = model.compute_loss(Prob, GT) total_loss = 0 for tgt in watch_targets: total_loss += Loss[tgt] # * loss_weight # compute gradient and do SGD step optimizer.zero_grad() # Clears the gradients of all optimized Variable s. total_loss.backward() optimizer.step() logger.add_scalars('loss_iter', Loss, it + 1) # logger.add_scalar('grad_norm/fc7', fc7_gradNorm, it+1) # print loss info if it % disp_interval == 0: # or (it+1)==len(dataset_train)/batch_size: cur_time = time.time() time_consume = cur_time - pre_time pre_time = cur_time logprint( '%s [epoch] %3d/%3d [iter] %5d -----------------------------------[time_consume] %.2f lr=%.8f' % (strftime("%Y-%m-%d %H:%M:%S", gmtime()), epoch + 1, nr_epoch, it + 1, time_consume, lr)) for tgt in watch_targets: _loss = Loss[tgt].data.cpu().numpy().copy() logprint(' %-15s loss=%.3f' % (tgt, _loss)) if np.isnan(_loss): print("[Warning] Weights explode! Stop training ... ") exit(-1) # Compute Acc@theta recs = dataset_train.recs[rec_inds] Pred = model.compute_pred(Prob) geo_dists = compute_geo_dists(Pred['quat'], recs.so3.quaternion) MedError = np.median(geo_dists) / np.pi * 180. theta_levels = odict(zip(['pi/6', 'pi/12', 'pi/24'], [np.pi / 6, np.pi / 12, np.pi / 24])) # # {'pi/6':np.pi/6, 'pi/12':np.pi/12, 'pi/24':np.pi/24}) Acc_at_ts = odict([(tname, sum(geo_dists < tvalue) / float(len(geo_dists))) for tname, tvalue in theta_levels.items()]) logger.add_scalars('acc/train', Acc_at_ts, it + 1) acc_str = ' '.join(['[%s] %3.1f%%' % (k, Acc_at_ts[k] * 100) for k, v in theta_levels.items()]) logprint(' Acc@{ %s } ' % acc_str) # pbar.set_description("[work_dir] %s B=%s \r" % (os.path.abspath(work_dir)[len(os.path.abspath(opt.base_dir))+1:], batch_size)) # print ("\r[work_dir] %s B=%s \r" % (os.path.abspath(work_dir)[len(os.path.abspath(opt.base_dir))+1:], batch_size), end='') sys.stdout.flush() # # if (epoch + 1) % opt.snapshot_step_epoch == 0: save_checkpoint({ 'it_and_epoch': (it, epoch), 'state_dict': model.state_dict(), 'optimizer': optimizer.state_dict(), }, filename=os.path.join(work_dir, 'train_epoch_%s.pth.tar' % (epoch + 1))) logger.close() if __name__ == '__main__': if opt.test_only: test(dataset_test, work_dir) else: train() rm_models(work_dir, marker='epoch')
20,434
39.951904
146
py
RPMG
RPMG-main/Pascal3D_Img/S3.3D_Rotation/agent.py
from utils import TrainClock, KSchedule import os import torch import torch.optim as optim import torch.nn as nn from torch.utils.tensorboard import SummaryWriter import sys BASEPATH = os.path.dirname(__file__) sys.path.append(os.path.join(BASEPATH, '..', '..', 'utils')) import tools import rpmg from networks import get_network def get_agent(config): return MyAgent(config) class MyAgent(object): """Base trainer that provides common training behavior. All customized trainer should be subclass of this class. """ def __init__(self, config): self.config = config self.clock = TrainClock() self.k_schedule = KSchedule(config.k_init, config.k_safe, config.max_iters) self.net = get_network(config).cuda() self.optimizer = optim.Adam(self.net.parameters(), config.lr) self.criterion = torch.nn.MSELoss(reduction='sum') self.writer = SummaryWriter(log_dir=self.config.log_dir) rpmg.logger_init(self.writer) def adjust_learning_rate_by_epoch(self, optimizer, cur_epoch, max_epoch): """Sets the learning rate to the initial LR decayed by 10 every _N_ epochs""" _N_ = max_epoch // 3 # add just learning rate 3 times. lr = self.config.lr * (0.1 ** (max(cur_epoch, 0) // _N_)) # 300)) for param_group in optimizer.param_groups: param_group['lr'] = lr return lr def save_ckpt(self, name=None): """save checkpoint during training for future restore""" if name is None: save_path = os.path.join(self.config.model_dir, "ckpt_epoch{}.pth".format(self.clock.epoch)) print("Saving checkpoint epoch {}...".format(self.clock.epoch)) else: save_path = os.path.join(self.config.model_dir, "{}.pth".format(name)) if isinstance(self.net, nn.DataParallel): model_state_dict = self.net.module.cpu().state_dict() else: model_state_dict = self.net.cpu().state_dict() torch.save({ 'clock': self.clock.make_checkpoint(), 'model_state_dict': model_state_dict, 'optimizer_state_dict': self.optimizer.state_dict(), }, save_path) self.net.cuda() def load_ckpt(self, name=None): """load checkpoint from saved checkpoint""" name = name if name == 'latest' else "ckpt_epoch{}".format(name) load_path = os.path.join(self.config.model_dir, "{}.pth".format(name)) if not os.path.exists(load_path): raise ValueError("Checkpoint {} not exists.".format(load_path)) checkpoint = torch.load(load_path) print("Loading checkpoint from {} ...".format(load_path)) if isinstance(self.net, nn.DataParallel): self.net.module.load_state_dict(checkpoint['model_state_dict']) else: self.net.load_state_dict(checkpoint['model_state_dict']) self.optimizer.load_state_dict(checkpoint['optimizer_state_dict']) self.clock.restore_checkpoint(checkpoint['clock']) def forward(self, img,gt): img = img.cuda() gt = gt.cuda() # (b, 3, 3) pred = self.net(img) # (b, 9) if 'RPMG' in self.config.mode: k = self.k_schedule.get_k(self.clock.iteration) pred_orth = rpmg.RPMG.apply(pred, k, 0.01, gt, self.clock.iteration) loss = self.criterion(pred_orth, gt) elif '9D' in self.config.mode: pred_orth = tools.symmetric_orthogonalization(pred) if self.config.mode == '9D_SVD': loss = self.criterion(pred_orth, gt) elif self.config.mode == '9D_inf': loss = self.criterion(pred, gt.flatten(1)) else: raise NotImplementedError elif '6D' in self.config.mode: pred_orth = tools.compute_rotation_matrix_from_ortho6d(pred) if self.config.mode == '6D_GM': loss = self.criterion(pred_orth, gt) elif self.config.mode == '6D_inf': gt_6d = torch.cat(gt[:, :, 0], gt[:, :, 1], 1) loss = self.criterion(pred, gt_6d) else: raise NotImplementedError elif '4D' in self.config.mode: pred_orth = tools.compute_rotation_matrix_from_quaternion(pred) if self.config.mode == '4D_norm': loss = self.criterion(pred_orth, gt) elif self.config.mode == '4D_inf': gt_q = tools.compute_quaternions_from_rotation_matrices(gt) # (b, 4) loss = self.criterion(pred, gt_q) elif self.config.mode == '4D_Axis': pred_orth = tools.compute_rotation_matrix_from_axisAngle(pred) loss = self.criterion(pred_orth, gt) else: raise NotImplementedError elif self.config.mode == '3D_Euler': pred_orth = tools.compute_rotation_matrix_from_euler(pred) loss = self.criterion(pred_orth, gt) elif self.config.mode == '10D': pred_orth = tools.compute_rotation_matrix_from_10d(pred) loss = self.criterion(pred_orth, gt) else: raise NotImplementedError err_deg = torch.rad2deg(tools.compute_geodesic_distance_from_two_matrices(pred_orth, gt)) # batch return pred, loss, err_deg def train_func(self, real_data,syn_data): """one step of training""" self.net.train() img = torch.cat((real_data[0],syn_data[0]),0) gt = torch.cat((real_data[1],syn_data[1]),0).squeeze(1) pred, loss, err_deg = self.forward(img, gt) self.optimizer.zero_grad() loss.backward() self.optimizer.step() return pred, loss, err_deg def val_func(self, data): """one step of validation""" self.net.eval() with torch.no_grad(): pred, loss, err_deg = self.forward(data[0],data[1].squeeze(1)) return pred, loss, err_deg if __name__ == '__main__': max_epoch = 1000 _N_ = max_epoch // 3 # add just learning rate 3 times. for cur_epoch in range(1000): lr = 1e-3 * (0.1 ** (max(cur_epoch, 0) // _N_)) # 300)) if cur_epoch % 10 == 0: print(f'epoch {cur_epoch}: {lr}')
6,286
38.540881
106
py
RPMG
RPMG-main/Pascal3D_Img/S3.3D_Rotation/regQuatNet/regQuatNet.py
# coding: utf8 """ @Author : Shuai Liao """ import torch.nn as nn import torch.utils.model_zoo as model_zoo from torch.autograd import Variable import torch import torch.nn.functional as F from basic.common import rdict import numpy as np from easydict import EasyDict as edict from collections import OrderedDict as odict from itertools import product from pytorch_util.netutil.common_v2.trunk_alexnet_bvlc import AlexNet_Trunk from pytorch_util.netutil.common_v2.trunk_vgg import VGG16_Trunk from pytorch_util.netutil.common_v2.trunk_resnet import ResNet101_Trunk, ResNet50_Trunk net_arch2Trunk = dict( alexnet=AlexNet_Trunk, vgg16=VGG16_Trunk, resnet101=ResNet101_Trunk, resnet50=ResNet50_Trunk ) from pytorch_util.libtrain import copy_weights, init_weights_by_filling from pytorch_util.torch_v4_feature import LocalResponseNorm # * from pytorch_util.torch_3rd_layers import Maskout from pytorch_util.torch_3rd_funcs import norm2unit, exp_Normalization from basic.common import env, add_path from lib.helper import * loss_balance = 4. def cls_pred(output, topk=(1,)): maxk = max(topk) batch_size = output.size(0) _, pred = output.topk(maxk, 1, True, True) return pred def reg2d_pred2tgt(pr_sin, pr_cos): theta = torch.atan2(pr_sin, pr_cos) return theta class _BaseReg_Net(nn.Module): # @staticmethod def head_seq(in_size, reg_n_D, nr_cate=12, nr_fc8=334, init_weights=True): # in_size=4096 seq = nn.Sequential( nn.Linear(in_size, nr_fc8), # Fc8 nn.ReLU(inplace=True), # nn.Dropout(), nn.Linear(nr_fc8, nr_cate * reg_n_D), # Prob ) if init_weights: init_weights_by_filling(seq, gaussian_std=0.005, kaiming_normal=True) # fill weight with gaussian filler return seq """BVLC alexnet architecture (Note: slightly different from pytorch implementation.)""" def __init__(self, nr_cate=12, net_arch='alexnet', init_weights=True): # AlexNet_Trunk super(_BaseReg_Net, self).__init__() _Trunk = net_arch2Trunk[net_arch] self.trunk = _Trunk(init_weights=init_weights) self.nr_cate = nr_cate self.top_size = 4096 if not self.trunk.net_arch.startswith('resnet') else 2048 def forword(self, x, label): raise NotImplementedError # ---------------------------------------------------------------------[reg_Direct] class reg_Direct_Net(_BaseReg_Net): # No L2 norm at all, """ No any L2 normalization to guarantee prediction is on n-sphere, smooth l1 loss is used. """ """BVLC alexnet architecture (Note: slightly different from pytorch implementation.)""" def __init__(self, nr_cate=12, net_arch='alexnet', init_weights=True): _BaseReg_Net.__init__(self, nr_cate=nr_cate, net_arch=net_arch, init_weights=init_weights) self.nr_cate = nr_cate self.reg_n_D = 4 # -- Head architecture # Note: for quaternion, there's only one regression head (instead of 3 Euler angles (a,e,t)). # Thus, nr_fc8=996 (see design.py) self.head_quat = self.head_seq(self.top_size, self.reg_n_D, nr_cate=nr_cate, nr_fc8=996, init_weights=init_weights) # for maskout specific category self.maskout = Maskout(nr_cate=nr_cate) # loss module self.loss_handler = Smooth_L1_Loss_Handler() self.targets = ['quat'] def forward(self, x, label): """label shape (batchsize, ) """ x = self.trunk(x) # Forward Conv and Fc6,Fc7 # batchsize = x.size(0) # .split(1, dim=1) # Note: quat(a,b,c,d) is on a 4d sphere and (x^2+y^2=1) x_quat = self.maskout(self.head_quat(x).view(batchsize, self.nr_cate, self.reg_n_D), label) # -- Normalize coordinate to a unit # x_quat = norm2unit(x_quat) #, nr_cate=self.nr_cate) Prob = edict(quat=x_quat) return Prob def compute_loss(self, Prob, GT): Loss = self.loss_handler.compute_loss(self.targets, Prob, GT) return Loss @staticmethod def compute_pred(Prob): x_quat = Prob['quat'] # -- Normalize coordinate to a unit x_quat = norm2unit(x_quat) # Note: here we do l2 normalization. Just to make predicted quaternion a unit norm. # batchsize = x_quat.size(0) # Get cpu data. batch_data = x_quat.data.cpu().numpy().copy() assert batch_data.shape == (batchsize, 4), batch_data.shape Pred = edict(quat=batch_data) return Pred # --------------------------------------------------------------------- # ---------------------------------------------------------------------[reg_Sflat] class reg_Sflat_Net(_BaseReg_Net): """ L2 normalization activation, with cosine proximity loss. """ """BVLC alexnet architecture (Note: slightly different from pytorch implementation.)""" def __init__(self, nr_cate=12, net_arch='alexnet', init_weights=True): _BaseReg_Net.__init__(self, nr_cate=nr_cate, net_arch=net_arch, init_weights=init_weights) self.nr_cate = nr_cate self.reg_n_D = 4 # -- Head architecture # Note: for quaternion, there's only one regression head (instead of 3 (for a,e,t)). # Thus, nr_fc8=996 (see design.py) self.head_quat = self.head_seq(self.top_size, self.reg_n_D, nr_cate=nr_cate, nr_fc8=996, init_weights=init_weights) # for maskout a,e,t self.maskout = Maskout(nr_cate=nr_cate) # loss module self.loss_handler = Cos_Proximity_Loss_Handler() self.targets = ['quat'] def forward(self, x, label): """label shape (batchsize, ) """ x = self.trunk(x) # batchsize = x.size(0) # Note: quat(a,b,c,d) is on a 4d sphere and (x^2+y^2=1) x_quat = self.maskout(self.head_quat(x).view(batchsize, self.nr_cate, self.reg_n_D), label) # -- Normalize coordinate to a unit x_quat = norm2unit(x_quat) # , nr_cate=self.nr_cate) Prob = edict(quat=x_quat) return Prob def compute_loss(self, Prob, GT): Loss = self.loss_handler.compute_loss(self.targets, Prob, GT) return Loss @staticmethod def compute_pred(Prob): x_quat = Prob['quat'] batchsize = x_quat.size(0) # Get cpu data. batch_data = x_quat.data.cpu().numpy().copy() assert batch_data.shape == (batchsize, 4), batch_data.shape Pred = edict(quat=batch_data) return Pred # --------------------------------------------------------------------- # ---------------------------------------------------------------------[reg_Sexp] class reg_Sexp_Net(_BaseReg_Net): """ Spherical exponential activation + Sign classification, with cosine proximity loss """ """BVLC alexnet architecture (Note: slightly different from pytorch implementation.)""" def __init__(self, nr_cate=12, net_arch='alexnet', init_weights=True): _BaseReg_Net.__init__(self, nr_cate=nr_cate, net_arch=net_arch, init_weights=init_weights) self.nr_cate = nr_cate self.reg_n_D = 4 # Note: for a quaternion q=(a,b,c,d), we always ensure a>0, that this cos(theta/2)>0 --> theta in [0,pi] # Thus only b,c,d need sign prediction. dim_need_sign = 3 _signs = list(product(*([(-1, 1)] * dim_need_sign))) # [(-1, -1, -1), (-1, -1, 1), ..., (1, 1, 1)], with len=8 self.signs = [(1,) + x for x in _signs] # [(1, -1, -1, -1), (1, -1, -1, 1), ..., (1, 1, 1, 1)], with len=8 self.signs2label = odict(zip(self.signs, range(len(self.signs)))) self.label2signs = Variable(torch.FloatTensor(self.signs)).cuda() # make it as a Variable # -- Head architecture # Note: for quaternion, there's only one regression head (instead of 3 (for a,e,t)). # Thus, nr_fc8=996 (see design.py) self.head_sqrdprob_quat = self.head_seq(self.top_size, self.reg_n_D, nr_cate=nr_cate, nr_fc8=996, init_weights=init_weights) # each of 3 quaternion complex component can be + or -, that totally 2**3 possible sign categories. self.head_signcate_quat = self.head_seq(self.top_size, len(self.signs), nr_cate=nr_cate, nr_fc8=996, init_weights=init_weights) # for abs branch self.maskout = Maskout(nr_cate=nr_cate) self.softmax = nn.Softmax(dim=1).cuda() # for sgc branch self.maskout_sgc = Maskout(nr_cate=nr_cate) # make a new layer to maskout sign classification only. # loss module self.loss_handler_abs_quat = Cos_Proximity_Loss_Handler() # Neg_Dot_Loss_Handler() # Cos_Proximity_Loss_Handler() # self.loss_handler_sgc_quat = Cross_Entropy_Loss_Handler() self.targets = ['abs_quat', 'sgc_quat'] self.gt_targets = ['quat'] def forward(self, x, label): """label shape (batchsize, ) """ x = self.trunk(x) # Forward Conv and Fc6,Fc7 # batchsize = x.size(0) # Note: squared probability x_sqr_quat = self.maskout(self.head_sqrdprob_quat(x).view(batchsize, self.nr_cate, self.reg_n_D), label) # ========>>>>> Maskout output (B,4) hook gradient. # -- Exp and Normalize coordinate to a unit x_sqr_quat = self.softmax(x_sqr_quat) # , nr_cate=self.nr_cate) # sign category head (totally 2^4=16 category) x_sgc_quat = self.maskout_sgc(self.head_signcate_quat(x).view(batchsize, self.nr_cate, len(self.signs)), label) Prob = edict(abs_quat=torch.sqrt(x_sqr_quat), sgc_quat=x_sgc_quat) return Prob def compute_loss(self, Prob, GT): # First get sign label from GT # == Formulate absolute value of quaternion GT_abs_quat = torch.abs(GT.quat) # == Formulate signs label of quaternion GT_sign_quat = torch.sign(GT.quat) GT_sign_quat[GT_sign_quat == 0] = 1 # make sign of '0' as 1 signs_tuples = [tuple(x) for x in GT_sign_quat.data.cpu().numpy().astype(np.int32).tolist()] for signs_tuple in signs_tuples: # q and -q gives the same rotation. assert signs_tuple[ 0] > 0, "Need GT to be all positive on first dim of quaternion: %s" % GT # assert all quaternion first dim is positive. # signs label GT_sgc_quat = Variable(torch.LongTensor([self.signs2label[signs_tuple] for signs_tuple in signs_tuples])) if GT.quat.is_cuda: GT_sgc_quat = GT_sgc_quat.cuda() # here just because compute_loss need a same key from Prob and GT, # so we just give a fake name to GT.sqr_quat as '_GT.logsqr_quat'. _GT = edict(abs_quat=GT_abs_quat, sgc_quat=GT_sgc_quat) Loss_abs_quat = self.loss_handler_abs_quat.compute_loss(['abs_quat'], Prob, _GT) Loss_sgc_quat = self.loss_handler_sgc_quat.compute_loss(['sgc_quat'], Prob, _GT) # To add loss weights here. Loss = edict(abs_quat=Loss_abs_quat['abs_quat'] * 10, # / 5. sgc_quat=Loss_sgc_quat['sgc_quat'], ) return Loss def compute_pred(self, Prob): x_abs_quat = Prob['abs_quat'] # torch.sqrt(torch.exp(Prob['logsqr_quat'])) x_sgc_quat = Prob['sgc_quat'] batchsize = x_abs_quat.size(0) # sign_ind = cls_pred(x_sgc_quat, topk=(1,)).data.view(-1, ) item_inds = torch.from_numpy(np.arange(batchsize)).cuda() _label_shape = self.label2signs.size() x_sign_quat = self.label2signs.expand(batchsize, *_label_shape)[item_inds, sign_ind] x_quat = x_abs_quat * x_sign_quat # Get cpu data. batch_quat = x_quat.data.cpu().numpy().copy() batchsize = x_quat.size(0) assert batch_quat.shape == (batchsize, 4), batch_quat.shape # Pred = edict(quat=batch_quat) return Pred # --------------------------------------------------------------------- # ---------------------------------------------------------------------------------- if __name__ == '__main__': model = reg2D_Net().copy_weights() # import numpy as np dummy_batch_data = np.zeros((2, 3, 227, 227), dtype=np.float32) dummy_batch_label = np.zeros((2, 1), dtype=np.int64) dummy_batch_data = torch.autograd.Variable(torch.from_numpy(dummy_batch_data)) dummy_batch_label = torch.autograd.Variable(torch.from_numpy(dummy_batch_label)) Pred = model(dummy_batch_data, dummy_batch_label) # print (Prob.a)
12,708
38.715625
143
py
RPMG
RPMG-main/Pascal3D_Img/S3.3D_Rotation/lib/helper.py
# coding: utf8 """ @Author : Shuai Liao """ import torch import torch.nn as nn from torch.autograd import Variable import numpy as np from easydict import EasyDict as edict from collections import OrderedDict as odict from itertools import product # from collections import OrderedDict class Cross_Entropy_Loss_Handler: def __init__(self): self.cross_entropy_loss = nn.CrossEntropyLoss().cuda() # interface function def compute_loss(self, tgts, Pred, GT): """ tgts: list of target names e.g. tgts=['a', 'e', 't'] GT : dict of ground truth for each target Pred: dict of prediction for each target """ Loss = edict() for tgt in tgts: Loss[tgt] = self.cross_entropy_loss(Pred[tgt], GT[tgt]) return Loss class Neg_Dot_Loss_Handler: def __init_(self): pass def compute_loss(self, tgts, Pred, GT): Loss = edict() for tgt in tgts: """ Bug fixed on 22 Aug 2018 torch.dot can only be applied to 1-dim tensor Don't know why there's no error. """ # Loss[tgt] = torch.mean( -torch.dot(GT[tgt],Pred[tgt]) ) # In fact here only does -GT[tgt]*Pred[tgt] Loss[tgt] = torch.mean(-torch.sum(GT[tgt] * Pred[tgt], dim=1)) return Loss class Cos_Proximity_Loss_Handler: def __init__(self): # self.period_l1_loss = PeriodL1Loss(period=period).cuda() self.cos_sim = nn.CosineSimilarity().cuda() # interface function def compute_loss(self, tgts, Pred, GT): """ tgts: list of target names. In this case has to be tgts=['quat'] GT : dict of ground truth for each target Pred: dict of prediction for each target """ # assert tgts==['quat'], tgts Loss = edict() for tgt in tgts: Loss[tgt] = torch.mean(1 - self.cos_sim(Pred[tgt], GT[tgt])) # use 1-cos(theta) to make loss as positive. return Loss class Smooth_L1_Loss_Handler: def __init__(self): self.smooth_l1_loss = nn.SmoothL1Loss().cuda() def compute_loss(self, tgts, Pred, GT): """ tgts: list of target names e.g. tgts=['a', 'e', 't'] GT : dict of ground truth for each target Pred: dict of prediction for each target """ Loss = edict() for tgt in tgts: Loss[tgt] = self.smooth_l1_loss(Pred[tgt], GT[tgt]) # [warning] pred first, gt second return Loss
2,518
30.4875
118
py
RPMG
RPMG-main/Pascal3D_Img/S3.3D_Rotation/lib/datasets/Dataset_Base.py
""" @Author : Shuai Liao """ import os, sys import numpy as np from math import ceil, floor, pi import torch from torch.utils.data import Dataset, DataLoader from collections import OrderedDict as odict import cv2 path = os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))) sys.path.append(os.path.join(path, 'pylibs')) from basic.common import add_path, env, rdict, cv2_wait, cv2_putText, is_py3 if is_py3: import pickle else: import cPickle as pickle from lmdb_util import ImageData_lmdb from numpy_db import npy_table, npy_db, dtype_summary, reorder_dtype this_dir = os.path.dirname(os.path.abspath(__file__)) base_dir = this_dir + '/../../../dataset' # where the dataset directory is. assert os.path.exists(base_dir) cate10 = ['bathtub', 'bed', 'chair', 'desk', 'dresser', 'monitor', 'night_stand', 'sofa', 'table', 'toilet'] cate40 = ['airplane', 'bathtub', 'bed', 'bench', 'bookshelf', 'bottle', 'bowl', 'car', 'chair', 'cone', 'cup', 'curtain', 'desk', 'door', 'dresser', 'flower_pot', 'glass_box', 'guitar', 'keyboard', 'lamp', 'laptop', 'mantel', 'monitor', 'night_stand', 'person', 'piano', 'plant', 'radio', 'range_hood', 'sink', 'sofa', 'stairs', 'stool', 'table', 'tent', 'toilet', 'tv_stand', 'vase', 'wardrobe', 'xbox'] ## Net configurations that are independent of task netcfg = rdict( # configuration for alexnet alexnet=rdict(TRAIN=rdict(BATCH_SIZE=200), TEST=rdict(BATCH_SIZE=200), INPUT_SHAPE=(227, 227), # resize_shape PIXEL_MEANS=np.array([[[102.9801, 115.9465, 122.7717]]]), RNG_SEED=3, ), # ignore_label=-1, # configuration for vgg vgg16=rdict(TRAIN=rdict(BATCH_SIZE=40), # 64 20 TEST=rdict(BATCH_SIZE=20), INPUT_SHAPE=(224, 224), PIXEL_MEANS=np.array([[[102.9801, 115.9465, 122.7717]]]), RNG_SEED=3, ), # configuration for resnet50 resnet50=rdict(TRAIN=rdict(BATCH_SIZE=100), # 128 TEST=rdict(BATCH_SIZE=64), INPUT_SHAPE=(224, 224), PIXEL_MEANS=np.array([[[102.9801, 115.9465, 122.7717]]]), RNG_SEED=3, ), # configuration for resnet101 resnet101=rdict(TRAIN=rdict(BATCH_SIZE=64), TEST=rdict(BATCH_SIZE=20), INPUT_SHAPE=(224, 224), PIXEL_MEANS=np.array([[[102.9801, 115.9465, 122.7717]]]), RNG_SEED=3, ), # configuration for resnet152 resnet152=rdict(TRAIN=rdict(BATCH_SIZE=32), TEST=rdict(BATCH_SIZE=10), INPUT_SHAPE=(224, 224), PIXEL_MEANS=np.array([[[102.9801, 115.9465, 122.7717]]]), RNG_SEED=3, ), ) def get_anno(db_path, category_target): # target='' # TO Move to generation of data db. viewID2quat = pickle.load(open(os.path.join(db_path, 'viewID2quat.pkl'), 'rb'), encoding='latin1') viewID2euler = pickle.load(open(os.path.join(db_path, 'viewID2euler.pkl'), 'rb'), encoding='latin1') keys = np.array(list(viewID2quat.keys())) idx_target = [] add_path(this_dir) from db_type import img_view_anno rcs = np.zeros((len(viewID2quat.keys()),), dtype=img_view_anno).view(np.recarray) for i, (key, quat) in enumerate(viewID2quat.items()): rc = rcs[i] rc.img_id = key # bathtub_0107.v001 cad_id, viewId = key.split('.') category = cad_id[:cad_id.rfind('_')] if category == category_target: idx_target.append(i) rc.category = category rc.cad_id = cad_id rc.so3.quaternion = quat if quat[0] > 0 else -quat # q and -q give the same rotation matrix. # Make sure all q[0]>0, that is rotation angle in [0,pi] rc.so3.euler = viewID2euler[key] rcs_cate = rcs[idx_target] keys_cate = keys[idx_target] return keys_cate, rcs_cate class Dataset_Base(Dataset): collection2dbname = \ dict(train='train_100V.Rawjpg.lmdb', # 'train_20V.Rawjpg.lmdb' test='test_20V.Rawjpg.lmdb', ) def __init__(self, category, collection='train', net_arch='alexnet', sampling=1.0): self.net_arch = net_arch self.cfg = netcfg[net_arch] self.collection = collection self.cates = cate10 # self.cate2ind = odict(zip(self.cates, range(len(self.cates)))) # get im_db self.db_path = os.path.join(base_dir, 'ModelNet10-SO3', self.collection2dbname[collection]) assert self.db_path is not None, '%s is not exist.' % (self.db_path) self.datadb = ImageData_lmdb(self.db_path) # Get anno self.keys, self.recs = get_anno(self.db_path, category) assert sampling > 0 and sampling <= 1.0, sampling if sampling < 1.0: print('Sampling dataset: %s' % sampling) _inds = np.arange(len(self.keys)) sample_inds = np.random.choice(_inds, size=int(len(_inds) * sampling), replace=False) sample_inds.sort() self.keys, self.recs = [self.keys[x] for x in sample_inds], self.recs[sample_inds] self.key2ind = dict(zip(self.keys, range(len(self.keys)))) # self.resize_shape = rsz_shape self.mean_pxl = np.array([102.9801, 115.9465, 122.7717], np.float32) # def _image2data(self, img, data_normal_type='caffe'): if self.net_arch == 'alexnet': img = np.pad(img, [(0, 3), (0, 3), (0, 0)], mode='edge') # (0,0,3,3) # caffe-style if data_normal_type == 'caffe': # Subtract mean pixel data = (img - self.mean_pxl).astype(np.float32) / 255. # Transpose data = data.transpose((2, 0, 1)) # H,W,C -> C,H,W elif data_normal_type == 'pytorch': # -# img = cv2.cvtColor( img, cv2.COLOR_GRAY2RGB ) # -# if self.transform is not None: # -# img = self.transform(img) # return (3,224,224) raise NotImplementedError else: raise NotImplementedError return data def _get_image(self, rc): img_id = rc.img_id img = self._image2data(self.datadb[img_id]) return img # if not flip else cv2.flip( img, 1 ) def __len__(self): return len(self.recs) def __getitem__(self, idx): rcobj = self.recs[idx] cate = rcobj.category obj_id = rcobj.obj_id image_id = rcobj.src_img.image_id """ To implement construction of sample dictionary. To get image data: call 'self.roiloader(rcobj)' """ print('This is an interface method, and you need to implement it in inherited class.') raise NotImplementedError def get_recs(self, query_keys): inds = [self.key2ind[k] for k in query_keys] return self.recs[inds] # interface method def _vis_minibatch(self, sample_batched): """Visualize a mini-batch for debugging.""" for i, (idx, label, quat, data) in enumerate(zip(sample_batched['idx'], # note: these are tensors sample_batched['label'], sample_batched['quat'], sample_batched['data'])): rc = self.recs[idx] # print idx im = data.numpy().transpose((1, 2, 0)).copy() im += self.cfg.PIXEL_MEANS im = im.astype(np.uint8) # xmin, ymax a, b, c, d = quat cv2_putText(im, (0, 20), rc.category, bgcolor=(255, 255, 255)) text = '%.1f %.1f %.1f %.1f' % (a, b, c, d) cv2_putText(im, (0, 40), text, bgcolor=(255, 255, 255)) cv2.imshow('im', im) cv2_wait() # pass class Dataset_Example(Dataset_Base): def __getitem__(self, idx): rc = self.recs[idx] cate = rc.category # img_id = rc.img_id quat = rc.so3.quaternion # sample = dict(idx=idx, label=self.cate2ind[cate], quat=quat, data=self._get_image(rc)) return sample if __name__ == '__main__': def test_dataloader(): dataset = Dataset_Example(collection='test', sampling=0.2) # dataloader = DataLoader(dataset, batch_size=4, shuffle=False, num_workers=1) for i_batch, sample_batched in enumerate(dataloader): dataset._vis_minibatch(sample_batched) test_dataloader()
8,723
38.654545
108
py
RPMG
RPMG-main/Pascal3D_Img/S3.3D_Rotation/lib/datasets/dataset_regQuatNet.py
""" @Author : Shuai Liao """ import numpy as np from Dataset_Base import Dataset_Base, netcfg import torch from torch.utils.data import Dataset, DataLoader import cv2 from basic.common import add_path, env, rdict, cv2_wait, cv2_putText # =============================================================== def pred2angle(a, e, t): return (a * 180. / np.pi) % 360, e * 180. / np.pi, t * 180. / np.pi def pred2angle_shift45(a, e, t): # shift 45 degree back return (a * 180. / np.pi - 45) % 360, e * 180. / np.pi, t * 180. / np.pi class Dataset_regQuatNet(Dataset_Base): def __init__(self, *args, **kwargs): Dataset_Base.__init__(self, *args, **kwargs) def __getitem__(self, idx): rc = self.recs[idx] cate = rc.category quat = rc.so3.quaternion sample = dict(idx=idx, label=self.cate2ind[cate], quat=quat, data=self._get_image(rc)) return sample # build class alias Dataset_reg_Direct = Dataset_regQuatNet Dataset_reg_Sexp = Dataset_regQuatNet Dataset_reg_Sflat = Dataset_regQuatNet if __name__ == '__main__': np.random.seed(3) def test_dataloader(collection='test', sampling=0.2): dataset = Dataset_reg_Sexp(collection=collection, sampling=sampling) print(len(dataset.keys)) anno_path = dataset.db_path sampling_file = anno_path + '/%s_sampling%.2f.txt' % (collection, sampling) with open(sampling_file, 'w') as f: f.write('\n'.join(dataset.keys)) print('sampling_file:', sampling_file) # dataloader = DataLoader(dataset, batch_size=50, shuffle=False, num_workers=1, sampler=None) for i_batch, sample_batched in enumerate(dataloader): print(sample_batched) dataset._vis_minibatch(sample_batched) test_dataloader()
1,914
27.58209
83
py
BVFSM
BVFSM-main/BVFSM.py
import matplotlib import matplotlib.pyplot as plt import torch import hypergrad as hg import numpy as np from sklearn.model_selection import train_test_split import torch.nn.functional as F import copy import time import csv import math import os import psutil as psutil import argparse import utils import function def train(args, lf=function.lf, uF=function.uF): args.y_size = args.y_size a = args.a b = args.b x0 = args.x0 y0 = args.y0 d = 1 seed = 1 np.random.seed(seed) z_loop = args.z_loop y_loop = args.y_loop x_loop = args.x_loop z_L2 = args.z_L2_reg y_L2 = args.y_L2_reg y_ln = args.y_ln_reg z_lr = args.z_lr y_lr = args.y_lr x_lr = args.x_lr TK = 1 z0 = y0 dc = args.decay dcr = 1.1 total_time = 0 total_hyper_time = 0 log_path = "result_{}.csv".format(time.strftime("%Y_%m_%d_%H_%M_%S")) with open(log_path, 'a', encoding='utf-8', newline='') as f: csv_writer = csv.writer(f) csv_writer.writerow( ['z_loop{}-y_loop{}-z_L2{}-y_L2{}-y_ln{}-z_lr{}-y_lr{}-x_lr{}'.format(z_loop, y_loop, z_L2, y_L2, y_ln, z_lr, y_lr, x_lr), 'd', 'x_itr', 'test loss', 'h_norm', 'step_time', 'total_time', 'x', 'y']) w = (float(y0) * torch.ones(args.y_size)).cuda().requires_grad_(True) h = (float(x0) * torch.ones(args.x_size)).cuda().requires_grad_(True) w_z = (float(z0) * torch.ones(args.y_size)).cuda().requires_grad_(True) w_opt = torch.optim.SGD([w], lr=y_lr) h_opt = torch.optim.Adam([h], lr=x_lr) w_z_opt = torch.optim.SGD([w_z], lr=z_lr) for x_itr in range(x_loop * TK): h_opt.zero_grad() step_start_time = time.time() lr_decay_rate = 1 / (1 ** (math.floor(x_itr / TK))) if dc == 'log': reg_decay_rate = 1 / (math.log(dcr * math.floor((x_itr + 1) / TK))) elif dc == 'power1': reg_decay_rate = 1 / (dcr ** math.floor((x_itr + 1) / TK)) elif dc == 'power2': reg_decay_rate = 1 / (math.floor((x_itr + 1) / TK) ** dcr) elif dc == 'linear': reg_decay_rate = 1 / (math.floor((x_itr + 1) / TK) * dcr) else: assert 1 w_opt.param_groups[0]['lr'] = w_opt.defaults['lr'] * lr_decay_rate h_opt.param_groups[0]['lr'] = h_opt.defaults['lr'] * lr_decay_rate w_z_opt.param_groups[0]['lr'] = w_z_opt.defaults['lr'] * lr_decay_rate for z_itr in range(z_loop): w_z_opt.zero_grad() loss_z = lf(h, w_z) + z_L2 * reg_decay_rate * torch.norm( w_z) ** 2 loss_z.backward() w_z_opt.step() for y_itr in range(y_loop): w_opt.zero_grad() loss_w_f = lf(h, w) loss_z = lf(h, w_z) + z_L2 * reg_decay_rate * torch.norm( w_z) ** 2 loss_w_L2 = y_L2 * reg_decay_rate * torch.norm(w) ** 2 loss_w_ln = y_ln * reg_decay_rate * torch.log( loss_w_f.detach() + args.y_size + 1e-8 + loss_z.detach() - loss_w_f) loss_w_ = uF(h, w) loss_w = loss_w_ + loss_w_L2 - loss_w_ln loss_w.backward() w_opt.step() low_time = time.time() - step_start_time hyper_time = time.time() h_opt.zero_grad() loss_w_f = lf(h, w) loss_z = lf(h, w_z) + z_L2 * reg_decay_rate * torch.norm( w_z) ** 2 loss_h_ = uF(h, w) loss_h_L2 = y_L2 * reg_decay_rate * torch.norm(w) ** 2 loss_h_ln = y_ln * reg_decay_rate * torch.log(loss_w_f.detach() + args.y_size + 1e-8 + loss_z - loss_w_f) loss_h = loss_h_ + loss_h_L2 - loss_h_ln loss_h.backward() h_opt.step() step_time = time.time() - step_start_time total_time += step_time total_hyper_time += (time.time() - hyper_time) if x_itr % TK == 0: with torch.no_grad(): loss_test = uF(h, w) hnp = h.detach().cpu().numpy() wnp = w.detach().cpu().numpy() print( 'd={:d},x_itr={:d},test loss={:.4f}, h_norm={:.4f},step_time={:.4f},total_time={:.4f},x={:.4f}'.format( d, x_itr, loss_test.data, h.norm() / d, step_time, total_time, hnp[0])) with open(log_path, 'a', encoding='utf-8', newline='') as f: csv_writer = csv.writer(f) csv_writer.writerow( [d, x_itr, loss_test.data, h.norm() / d, step_time, total_time, hnp[0]] + [ws for ws in utils.np_to_list(wnp)])
4,871
34.562044
123
py
BVFSM
BVFSM-main/utils.py
import matplotlib import matplotlib.pyplot as plt import torch import hypergrad as hg import numpy as np from sklearn.model_selection import train_test_split import torch.nn.functional as F import copy import time import csv import math import os import psutil as psutil import argparse def np_to_list(arr): this_type_str = type(arr) if this_type_str is np.ndarray: arr = arr.tolist() elif this_type_str in [np.int, np.int32, np.int64]: arr = [int(arr), ] else: arr = arr return arr def show_memory_info(hint): pid = os.getpid() p = psutil.Process(pid) info = p.memory_full_info() memory = info.uss / 1024. / 1024 print(f"{hint} memory used: {memory} MB ") def loss_L2(parameters): loss = 0 for w in parameters: loss += torch.norm(w, 2) ** 2 return loss def penalty(x): return torch.log(x) def tpenalty(x, d): if x > d: return torch.log(x)
952
17.326923
55
py
BVFSM
BVFSM-main/BVFSM_constraint.py
#!/usr/bin/env python # coding: utf-8 # In[1]: # get_ipython().run_line_magic('matplotlib', 'inline') import matplotlib import matplotlib.pyplot as plt import torch import hypergrad as hg import numpy as np from sklearn.model_selection import train_test_split import torch.nn.functional as F import copy import time import csv import math import os import psutil as psutil def np_to_list(arr): this_type_str = type(arr) if this_type_str is np.ndarray: arr = arr.tolist() elif this_type_str in [np.int, np.int32, np.int64]: arr = [int(arr),] else: arr = arr return arr def show_memory_info(hint): # 获取当前进程的进程号 pid = os.getpid() # psutil 是一个获取系统信息的库 p = psutil.Process(pid) info = p.memory_full_info() memory = info.uss/1024./1024 print(f"{hint} memory used: {memory} MB ") cuda = False double_precision = False default_tensor_str = 'torch.cuda' if cuda else 'torch' default_tensor_str += '.DoubleTensor' if double_precision else '.FloatTensor' torch.set_default_tensor_type(default_tensor_str) def frnp(x): t = torch.from_numpy(x).cuda() if cuda else torch.from_numpy(x) return t if double_precision else t.float() def tonp(x, cuda=cuda): return x.detach().cpu().numpy() if cuda else x.detach().numpy() def loss_L2(parameters): loss = 0 for w in parameters: loss += torch.norm(w, 2) ** 2 return loss def penalty(x): return torch.log(x) def tpenalty(x,d): if x>d: return torch.log(x) initial=[0,2,4,6,8,10] # In[3]: for y_ln in [0.01]: for x0,y0 in zip(initial,initial): for con1 in range(1): z_loop = 50 y_loop=25 tSize=2 a=2 b=2 # x0=2 # y0=2 # con1=0 d=1 iu=2 # synthetic data generation seed = 1 n = 100 val_perc = 0.5 np.random.seed(seed) # z_loop = 50 # y_loop = 25 x_loop = 500 z_L2 = 0.01 y_L2 = 0.01 # y_ln = 0.01 z_lr = 0.01 y_lr = 0.01 x_lr = 0.01 TK=1 # x0=3. # y0=3. # z0=3. # a=0. # b=0. z0=y0 decay=['log','power1','power2','poly','linear'] dc=decay[iu] dcr=1.1 conlist=['log','quad','None'] con=conlist[con1] # In[4]: total_time = 0 total_hyper_time = 0 log_path = "x1_constraint{}_yln{}_yzloop{}_{}_tSize{}_dc{}{}_TK{}_xyz{}{}{}_ab{}{}._{}.csv".format(con,y_ln,y_loop,z_loop, tSize,dc,dcr,TK,x0,y0,z0,a,b, time.strftime("%Y_%m_%d_%H_%M_%S")) with open(log_path, 'a', encoding='utf-8', newline='') as f: csv_writer = csv.writer(f) csv_writer.writerow( ['z_loop{}-y_loop{}-z_L2{}-y_L2{}-y_ln{}-z_lr{}-y_lr{}-x_lr{}'.format(z_loop, y_loop, z_L2, y_L2, y_ln, z_lr, y_lr, x_lr), 'd','x_itr','test loss', 'h_norm','step_time','total_time','x','gx','y']) # problem definition use_gpu = False # tSize=2000 Fmin=0 xmin=[] xmind=0 for i in range(tSize): Fmin=Fmin+(-np.pi/4/(i+1)-a)**2+(-np.pi/4/(i+1)-b)**2 xmin.append(-np.pi/4/(i+1)) xmind=xmind+(-np.pi/4/(i+1))**2 w = (float(0)*torch.ones(tSize)).cuda().requires_grad_(True) h =(float(0)*torch.ones(1)).cuda().requires_grad_(True) w_z = (float(0)*torch.ones(tSize)).cuda().requires_grad_(True) w_opt = torch.optim.Adam([w], lr=y_lr) h_opt = torch.optim.Adam([h], lr=x_lr) w_z_opt = torch.optim.SGD([w_z], lr=z_lr) C = (float(1) * torch.ones(tSize)).cuda().requires_grad_(False) def lf(x,y): out = 0 for i in range(tSize): out = out + torch.sin((x + y[i] - C[i])) return out def uF(x,y): return torch.norm(x - a) ** 2 + torch.norm(y - a ) ** 2 for x_itr in range(x_loop*TK): print('-'*50) h_opt.zero_grad() step_start_time = time.time() yhis=[] lr_decay_rate = 1 / (1 ** (math.floor(x_itr / TK))) if dc=='log': reg_decay_rate = 1 / (math.log(dcr*math.floor((x_itr+1) / TK))) elif dc=='power1': reg_decay_rate = 1 / (dcr ** math.floor((x_itr + 1) / TK)) elif dc=='power2': reg_decay_rate = 1 / (math.floor((x_itr + 1) / TK)**dcr) elif dc == 'linear': reg_decay_rate = 1 / (math.floor((x_itr + 1) / TK) * dcr) else: assert 1 w_opt.param_groups[0]['lr'] = w_opt.defaults['lr'] * lr_decay_rate h_opt.param_groups[0]['lr'] = h_opt.defaults['lr'] * lr_decay_rate w_z_opt.param_groups[0]['lr'] = w_z_opt.defaults['lr'] * lr_decay_rate loss_z_l = 0 for z_itr in range(z_loop): w_z_opt.zero_grad() loss_z = lf(h,w_z) + z_L2 * reg_decay_rate * torch.norm( w_z) ** 2 loss_z.backward() w_z_opt.step() if x_itr==466: print('y={}'.format(w_z[0].item())) loss_y_l = 0 wl = w for y_itr in range(y_loop): w_opt.zero_grad() loss_w_f = lf(h,w) loss_z = lf(h,w_z) + z_L2 * reg_decay_rate * torch.norm( w_z) ** 2 loss_w_L2 = y_L2 * reg_decay_rate * torch.norm(w) ** 2 loss_w_ln = y_ln * reg_decay_rate * torch.log(loss_w_f.detach()+tSize+1e-7+loss_z.detach() - loss_w_f) loss_w_ =uF(h,w) if con=='log': ls=(h.detach()+w)*(h.detach()+w-1) # print(-(h.detach()+w)*(h.detach()+w-1)+ls.detach()+1e-1) loss_w = loss_w_ + loss_w_L2 - loss_w_ln- y_ln/ reg_decay_rate *torch.sum(torch.log(-(h.detach()+w)*(h.detach()+w-1)+ls.detach()+2e-1)) elif con=='quad': loss_w = loss_w_ + loss_w_L2 - loss_w_ln + y_ln / reg_decay_rate * torch.sum( torch.relu((h.detach() + w) * (h.detach() + w - 1)) ** 2) else: loss_w = loss_w_ + loss_w_L2 - loss_w_ln loss_w.backward() w_opt.step() wl = w low_time = time.time() - step_start_time hyper_time = time.time() h_opt.zero_grad() loss_w_f = lf(h,w) loss_z = lf(h,w_z) + z_L2 * reg_decay_rate * torch.norm( w_z) ** 2 loss_h_ = uF(h,w) loss_h_L2 = y_L2 * reg_decay_rate * torch.norm(w) ** 2 loss_h_ln = y_ln * reg_decay_rate * torch.log(loss_w_f.detach() +tSize+1e-8+ loss_z - loss_w_f) if con=='log': ls = (h.detach() + w) * (h.detach() + w - 1) loss_h = loss_h_ + loss_h_L2 - loss_h_ln- y_ln/ reg_decay_rate *torch.sum(torch.log(-(h+w)*(h+w-1)+ls.detach()+2e-1))#+ y_ln/ reg_decay_rate *torch.relu((h)*(h-1))**2 elif con=='quad': loss_h = loss_h_ + loss_h_L2 - loss_h_ln+ y_ln/ reg_decay_rate *torch.sum(torch.relu((h+w)*(h+w-1)))**2#+ y_ln/ reg_decay_rate *torch.relu((h)*(h-1))**2 else: loss_h = loss_h_ + loss_h_L2 - loss_h_ln grad_h=torch.autograd.grad(loss_h,[h],retain_graph=True) grad_h_ = torch.autograd.grad(loss_h_, [h], retain_graph=True, allow_unused=True) grad_h_L2 = torch.autograd.grad(loss_h_L2, [h], retain_graph=True, allow_unused=True) grad_h_ln = torch.autograd.grad(loss_h_ln, [h], retain_graph=True, allow_unused=True) loss_h.backward() h_opt.step() step_time = time.time() - step_start_time total_time += step_time total_hyper_time += (time.time() - hyper_time) if x_itr % TK == 0: with torch.no_grad(): loss_test = uF(h,w) for g in grad_h: gnp=g.detach().cpu().numpy() hnp=h.detach().cpu().numpy() wnp=w.detach().cpu().numpy() print( 'd={:d},x_itr={:d},test loss={:.4f}, h_norm={:.4f},step_time={:.4f},total_time={:.4f},x={:.4f}'.format( d, x_itr, loss_test.data, h.norm() / d, step_time, total_time,hnp[0])) with open(log_path, 'a', encoding='utf-8', newline='') as f: csv_writer = csv.writer(f) csv_writer.writerow( [d, x_itr, loss_test.data, h.norm() / d, step_time, total_time,hnp[0],gnp[0]]+[ws for ws in np_to_list(wnp)])
9,895
36.06367
200
py
BVFSM
BVFSM-main/demo.py
import matplotlib import matplotlib.pyplot as plt import torch import hypergrad as hg import numpy as np from sklearn.model_selection import train_test_split import torch.nn.functional as F import copy import time import csv import math import os import psutil as psutil import argparse import utils import BVFSM parser = argparse.ArgumentParser() parser.add_argument('--x_size', type=int, default=1) parser.add_argument('--y_size', type=int, default=2) parser.add_argument('--z_loop', type=int, default=50) parser.add_argument('--y_loop', type=int, default=25) parser.add_argument('--x_loop', type=int, default=500) parser.add_argument('--z_lr', type=float, default=0.01) parser.add_argument('--y_lr', type=float, default=0.01) parser.add_argument('--x_lr', type=float, default=0.01) parser.add_argument('--z_L2_reg', type=float, default=0.01) parser.add_argument('--y_L2_reg', type=float, default=0.01) parser.add_argument('--y_ln_reg', type=float, default=0.001) parser.add_argument('--x0', type=float, default=0.) parser.add_argument('--y0', type=float, default=0.) parser.add_argument('--a', type=float, default=2.) parser.add_argument('--b', type=float, default=2.) parser.add_argument('--c', type=float, default=2.) parser.add_argument('--decay', type=str, default='log', help='log, power1, power2, poly, linear') args = parser.parse_args() a = args.a b = args.b C = (float(args.c) * torch.ones(args.y_size)).cuda().requires_grad_(False) def lf(x, y): out = 0 for i in range(args.y_size): out = out + torch.sin((x + y[i] - C[i])) return out def uF(x, y): return torch.norm(x - a) ** 2 + torch.norm(y - a - C) ** 2 BVFSM.train(args,lf,uF)
1,684
26.177419
97
py
BVFSM
BVFSM-main/BVFSM_PBO.py
import matplotlib import matplotlib.pyplot as plt import torch import hypergrad as hg import numpy as np from sklearn.model_selection import train_test_split import torch.nn.functional as F import copy import time import csv import math import os import psutil as psutil def show_memory_info(hint): pid = os.getpid() p = psutil.Process(pid) info = p.memory_full_info() memory = info.uss/1024./1024 print(f"{hint} memory used: {memory} MB ") cuda = False double_precision = False default_tensor_str = 'torch.cuda' if cuda else 'torch' default_tensor_str += '.DoubleTensor' if double_precision else '.FloatTensor' torch.set_default_tensor_type(default_tensor_str) def frnp(x): t = torch.from_numpy(x).cuda() if cuda else torch.from_numpy(x) return t if double_precision else t.float() def tonp(x, cuda=cuda): return x.detach().cpu().numpy() if cuda else x.detach().numpy() def loss_L2(parameters): loss = 0 for w in parameters: loss += torch.norm(w, 2) ** 2 return loss def penalty(x): return torch.log(x) def tpenalty(x,d): if x>d: return torch.log(x) save_log=True initial=[0] for iu in [2]: for a,b in zip([2],[2]): for x0,y0 in zip([1],[7]): d=1 tSize=2 # synthetic data generation seed = 1 n = 100 val_perc = 0.5 np.random.seed(seed) z_loop = 50 y_loop = 25 x_loop = 500 z_L2 = 0.001 y_L2 = 0.001 y_ln = 0.7 z_ln = 0.1 z_lr = 0.01 y_lr = 0.01 x_lr = 0.01 TK=1 z0=y0 decay=['log','power1','power2','poly','linear'] dc=decay[iu] dcr=1.001 # In[4]: total_time = 0 total_hyper_time = 0 if save_log: log_path = "PBO\\x1_nonconvex_dc{}{}_TK{}_xyz{}{}{}_ab{}{}._{}.csv".format( dc,dcr,TK,x0,y0,z0,a,b, time.strftime("%Y_%m_%d_%H_%M_%S")) with open(log_path, 'a', encoding='utf-8', newline='') as f: csv_writer = csv.writer(f) csv_writer.writerow( ['z_loop{}-y_loop{}-z_L2{}-y_L2{}-y_ln{}-z_lr{}-y_lr{}-x_lr{}'.format(z_loop, y_loop, z_L2, y_L2, y_ln, z_lr, y_lr, x_lr), 'd','x_itr','test loss', 'h_norm','step_time','total_time','x','y','x+y']) use_gpu = False Fmin=0 xmin=[] xmind=0 for i in range(tSize): Fmin=Fmin+(-np.pi/4/(i+1)-a)**2+(-np.pi/4/(i+1)-b)**2 xmin.append(-np.pi/4/(i+1)) xmind=xmind+(-np.pi/4/(i+1))**2 w = (float(y0)*torch.ones(tSize)).cuda().requires_grad_(True) h =(float(x0)*torch.ones(1)).cuda().requires_grad_(True) w_z = (float(y0)*torch.ones(tSize)).cuda().requires_grad_(True) w_opt = torch.optim.SGD([w], lr=y_lr) h_opt = torch.optim.Adam([h], lr=x_lr) w_z_opt = torch.optim.SGD([w_z], lr=z_lr) C = (float(2) * torch.ones(tSize)).cuda().requires_grad_(False) def lf(x,y): out = 0 for i in range(tSize): out = out + torch.sin((x + y[i] - C[i])) # print(out) return out def uF(x,y): return torch.norm(x - a) ** 2 - torch.norm(y - a - C) ** 2 print(uF(float(-0.42)*torch.ones(1).cuda(),float(7.14)*torch.ones(tSize).cuda())) print(lf(float(-0.42)*torch.ones(1).cuda(),float(7.14)*torch.ones(tSize).cuda())) # z_loss_min=1e-6 for x_itr in range(x_loop*TK): h_opt.zero_grad() step_start_time = time.time() yhis=[] lr_decay_rate = 1 / (1 ** (math.floor(x_itr / TK))) if dc=='log': reg_decay_rate = 1 / (math.log(dcr*math.floor((x_itr+1) / TK))) elif dc=='power1': reg_decay_rate = 1 / (dcr ** math.floor((x_itr + 1) / TK)) elif dc=='power2': reg_decay_rate = 1 / (math.floor((x_itr + 1) / TK)**dcr) elif dc == 'linear': reg_decay_rate = 1 / (math.floor((x_itr + 1) / TK) * dcr) else: assert 1 w_opt.param_groups[0]['lr'] = w_opt.defaults['lr'] * lr_decay_rate h_opt.param_groups[0]['lr'] = h_opt.defaults['lr'] * lr_decay_rate w_z_opt.param_groups[0]['lr'] = w_z_opt.defaults['lr'] * lr_decay_rate loss_z_l = 0 for z_itr in range(z_loop): w_z_opt.zero_grad() loss_z = lf(h,w_z) + z_L2 * reg_decay_rate * torch.norm( w_z) ** 2 loss_z.backward() # print(loss_z_l-loss_z.item()) w_z_opt.step() loss_y_l = 0 wl = w # print(w_z.item()) for y_itr in range(y_loop): w_opt.zero_grad() loss_w_f = lf(h,w) loss_z = lf(h,w_z) + z_L2 * reg_decay_rate * torch.norm( w_z) ** 2 loss_w_L2 = y_L2 * reg_decay_rate * torch.norm(w) ** 2 loss_w_ln = y_ln * torch.log(loss_w_f.detach()+tSize+1e-8+loss_z.detach() - loss_w_f) loss_w_ =uF(h,w) loss_w = -loss_w_ + loss_w_L2 - loss_w_ln grad_w_ = torch.autograd.grad(loss_w_, [w], retain_graph=True, allow_unused=True) grad_w_L2 = torch.autograd.grad(loss_w_L2, [w], retain_graph=True, allow_unused=True) grad_w_ln = torch.autograd.grad(loss_w_ln, [w], retain_graph=True, allow_unused=True) loss_w.backward() torch.nn.utils.clip_grad_norm_([w],10) w_opt.step() wl = w yhis.append(w[0].item()) low_time = time.time() - step_start_time hyper_time = time.time() h_opt.zero_grad() loss_w_f = lf(h,w) loss_z = lf(h,w_z) + z_L2 * reg_decay_rate * torch.norm( w_z) ** 2 loss_h_ = uF(h,w) loss_h_L2 = y_L2 * reg_decay_rate * torch.norm(w) ** 2 loss_h_ln = y_ln * torch.log(loss_w_f.detach()+tSize+1e-8+ loss_z - loss_w_f) loss_h = loss_h_ + loss_h_ln grad_h=torch.autograd.grad(loss_h,[h],retain_graph=True) grad_h_ = torch.autograd.grad(loss_h_, [h], retain_graph=True, allow_unused=True) grad_h_L2 = torch.autograd.grad(loss_h_L2, [h], retain_graph=True, allow_unused=True) grad_h_ln = torch.autograd.grad(loss_h_ln, [h], retain_graph=True, allow_unused=True) loss_h.backward() torch.nn.utils.clip_grad_norm_([h], 100) # print(loss_L2([grad_h])) h_opt.step() step_time = time.time() - step_start_time total_time += step_time total_hyper_time += (time.time() - hyper_time) if x_itr % TK == 0: with torch.no_grad(): loss_test = uF(h,w) loss_train = lf(h,w) print( 'd={:d},x_itr={:d},F={:.2f},x={:.2f},y={:.2f},x+y={:.2f},gw_={:.2f}, gw_L2={:.2f}, gw_ln=={:.2f},gh={:.2f}, gh_={:.2f}, gh_ln=={:.2f}'.format( d, x_itr, loss_test.data, h.item(),w[0].item(),h.item()+w[0].item(),grad_w_[0][0].item(),grad_w_L2[0][0].item(),grad_w_ln[0][0].item(),grad_h[0].item(),grad_h_[0].item(),grad_h_ln[0].item())) if save_log: with open(log_path, 'a', encoding='utf-8', newline='') as f: csv_writer = csv.writer(f) csv_writer.writerow( [d, x_itr, loss_test.data, h.norm() / d, step_time, total_time,h.item(),w[0].item(),h.item()+w[0].item()]+[yh for yh in yhis])
8,740
36.515021
197
py
deep-speaker
deep-speaker-master/setup.py
import os import platform from setuptools import setup tensorflow = 'tensorflow' if platform.system() == 'Darwin' and platform.processor() == 'arm': tensorflow = 'tensorflow-macos' # https://github.com/grpc/grpc/issues/25082 os.environ['GRPC_PYTHON_BUILD_SYSTEM_OPENSSL'] = '1' os.environ['GRPC_PYTHON_BUILD_SYSTEM_ZLIB'] = '1' install_requires = [ 'numpy', tensorflow, 'natsort', 'librosa', 'dill', 'python_speech_features', 'tqdm', 'click', 'pandas', 'matplotlib', 'keras', 'numba' ] setup( name='deep-speaker', version='1.0', description='Deep Speaker', author='Philippe Remy', license='MIT', long_description_content_type='text/markdown', long_description=open('README.md').read(), packages=['deep_speaker'], install_requires=install_requires )
853
20.897436
67
py
deep-speaker
deep-speaker-master/cli.py
#!/usr/bin/env python # -*- coding: utf-8 -*- import logging import os import click from deep_speaker.audio import Audio from deep_speaker.batcher import KerasFormatConverter from deep_speaker.constants import SAMPLE_RATE, NUM_FRAMES from deep_speaker.test import test from deep_speaker.train import start_training from deep_speaker.utils import ClickType as Ct, ensures_dir from deep_speaker.utils import init_pandas logger = logging.getLogger(__name__) VERSION = '3.0a' @click.group() def cli(): logging.basicConfig(format='%(asctime)12s - %(levelname)s - %(message)s', level=logging.INFO) init_pandas() @cli.command('version', short_help='Prints the version.') def version(): print(f'Version is {VERSION}.') @cli.command('build-mfcc-cache', short_help='Build audio cache.') @click.option('--working_dir', required=True, type=Ct.output_dir()) @click.option('--audio_dir', default=None) @click.option('--sample_rate', default=SAMPLE_RATE, show_default=True, type=int) def build_audio_cache(working_dir, audio_dir, sample_rate): ensures_dir(working_dir) if audio_dir is None: audio_dir = os.path.join(working_dir, 'LibriSpeech') Audio(cache_dir=working_dir, audio_dir=audio_dir, sample_rate=sample_rate) @cli.command('build-keras-inputs', short_help='Build inputs to Keras.') @click.option('--working_dir', required=True, type=Ct.input_dir()) @click.option('--counts_per_speaker', default='600,100', show_default=True, type=str) # train,test def build_keras_inputs(working_dir, counts_per_speaker): # counts_per_speaker: If you specify --counts_per_speaker 600,100, that means for each speaker, # you're going to generate 600 samples for training and 100 for testing. One sample is 160 frames # by default (~roughly 1.6 seconds). counts_per_speaker = [int(b) for b in counts_per_speaker.split(',')] kc = KerasFormatConverter(working_dir) kc.generate(max_length=NUM_FRAMES, counts_per_speaker=counts_per_speaker) kc.persist_to_disk() @cli.command('test-model', short_help='Test a Keras model.') @click.option('--working_dir', required=True, type=Ct.input_dir()) @click.option('--checkpoint_file', required=True, type=Ct.input_file()) def test_model(working_dir, checkpoint_file=None): # export CUDA_VISIBLE_DEVICES=0; python cli.py test-model # --working_dir /home/philippe/ds-test/triplet-training/ # --checkpoint_file ../ds-test/checkpoints-softmax/ResCNN_checkpoint_102.h5 # f-measure = 0.789, true positive rate = 0.733, accuracy = 0.996, equal error rate = 0.043 # export CUDA_VISIBLE_DEVICES=0; python cli.py test-model # --working_dir /home/philippe/ds-test/triplet-training/ # --checkpoint_file ../ds-test/checkpoints-triplets/ResCNN_checkpoint_175.h5 # f-measure = 0.849, true positive rate = 0.798, accuracy = 0.997, equal error rate = 0.025 test(working_dir, checkpoint_file) @cli.command('train-model', short_help='Train a Keras model.') @click.option('--working_dir', required=True, type=Ct.input_dir()) @click.option('--pre_training_phase/--no_pre_training_phase', default=False, show_default=True) def train_model(working_dir, pre_training_phase): # PRE TRAINING # commit a5030dd7a1b53cd11d5ab7832fa2d43f2093a464 # Merge: a11d13e b30e64e # Author: Philippe Remy <premy.enseirb@gmail.com> # Date: Fri Apr 10 10:37:59 2020 +0900 # LibriSpeech train-clean-data360 (600, 100). 0.985 on test set (enough for pre-training). # TRIPLET TRAINING # [...] # Epoch 175/1000 # 2000/2000 [==============================] - 919s 459ms/step - loss: 0.0077 - val_loss: 0.0058 # Epoch 176/1000 # 2000/2000 [==============================] - 917s 458ms/step - loss: 0.0075 - val_loss: 0.0059 # Epoch 177/1000 # 2000/2000 [==============================] - 927s 464ms/step - loss: 0.0075 - val_loss: 0.0059 # Epoch 178/1000 # 2000/2000 [==============================] - 948s 474ms/step - loss: 0.0073 - val_loss: 0.0058 start_training(working_dir, pre_training_phase) if __name__ == '__main__': cli()
4,089
39.9
101
py
deep-speaker
deep-speaker-master/deep_speaker/batcher.py
import json import logging import os from collections import deque, Counter from random import choice from time import time import dill import numpy as np from tqdm import tqdm from deep_speaker.audio import pad_mfcc, Audio from deep_speaker.constants import NUM_FRAMES, NUM_FBANKS from deep_speaker.conv_models import DeepSpeakerModel from deep_speaker.utils import ensures_dir, load_pickle, load_npy, train_test_sp_to_utt logger = logging.getLogger(__name__) def extract_speaker(utt_file): return utt_file.split('/')[-1].split('_')[0] def sample_from_mfcc(mfcc, max_length): if mfcc.shape[0] >= max_length: r = choice(range(0, len(mfcc) - max_length + 1)) s = mfcc[r:r + max_length] else: s = pad_mfcc(mfcc, max_length) return np.expand_dims(s, axis=-1) def sample_from_mfcc_file(utterance_file, max_length): mfcc = np.load(utterance_file) return sample_from_mfcc(mfcc, max_length) class KerasFormatConverter: def __init__(self, working_dir, load_test_only=False): self.working_dir = working_dir self.output_dir = os.path.join(self.working_dir, 'keras-inputs') ensures_dir(self.output_dir) self.categorical_speakers = load_pickle(os.path.join(self.output_dir, 'categorical_speakers.pkl')) if not load_test_only: self.kx_train = load_npy(os.path.join(self.output_dir, 'kx_train.npy')) self.ky_train = load_npy(os.path.join(self.output_dir, 'ky_train.npy')) self.kx_test = load_npy(os.path.join(self.output_dir, 'kx_test.npy')) self.ky_test = load_npy(os.path.join(self.output_dir, 'ky_test.npy')) self.audio = Audio(cache_dir=self.working_dir, audio_dir=None) if self.categorical_speakers is None: self.categorical_speakers = SparseCategoricalSpeakers(self.audio.speaker_ids) def persist_to_disk(self): with open(os.path.join(self.output_dir, 'categorical_speakers.pkl'), 'wb') as w: dill.dump(self.categorical_speakers, w) np.save(os.path.join(self.output_dir, 'kx_train.npy'), self.kx_train) np.save(os.path.join(self.output_dir, 'kx_test.npy'), self.kx_test) np.save(os.path.join(self.output_dir, 'ky_train.npy'), self.ky_train) np.save(os.path.join(self.output_dir, 'ky_test.npy'), self.ky_test) def generate_per_phase(self, max_length=NUM_FRAMES, num_per_speaker=3000, is_test=False): # train OR test. num_speakers = len(self.audio.speaker_ids) sp_to_utt = train_test_sp_to_utt(self.audio, is_test) # 64 fbanks 1 channel(s). # float32 kx = np.zeros((num_speakers * num_per_speaker, max_length, NUM_FBANKS, 1), dtype=np.float32) ky = np.zeros((num_speakers * num_per_speaker, 1), dtype=np.float32) desc = f'Converting to Keras format [{"test" if is_test else "train"}]' for i, speaker_id in enumerate(tqdm(self.audio.speaker_ids, desc=desc)): utterances_files = sp_to_utt[speaker_id] for j, utterance_file in enumerate(np.random.choice(utterances_files, size=num_per_speaker, replace=True)): self.load_into_mat(utterance_file, self.categorical_speakers, speaker_id, max_length, kx, ky, i * num_per_speaker + j) return kx, ky def generate(self, max_length=NUM_FRAMES, counts_per_speaker=(3000, 500)): kx_train, ky_train = self.generate_per_phase(max_length, counts_per_speaker[0], is_test=False) kx_test, ky_test = self.generate_per_phase(max_length, counts_per_speaker[1], is_test=True) logger.info(f'kx_train.shape = {kx_train.shape}') logger.info(f'ky_train.shape = {ky_train.shape}') logger.info(f'kx_test.shape = {kx_test.shape}') logger.info(f'ky_test.shape = {ky_test.shape}') self.kx_train, self.ky_train, self.kx_test, self.ky_test = kx_train, ky_train, kx_test, ky_test @staticmethod def load_into_mat(utterance_file, categorical_speakers, speaker_id, max_length, kx, ky, i): kx[i] = sample_from_mfcc_file(utterance_file, max_length) ky[i] = categorical_speakers.get_index(speaker_id) class SparseCategoricalSpeakers: def __init__(self, speakers_list): self.speaker_ids = sorted(speakers_list) assert len(set(self.speaker_ids)) == len(self.speaker_ids) # all unique. self.map = dict(zip(self.speaker_ids, range(len(self.speaker_ids)))) def get_index(self, speaker_id): return self.map[speaker_id] class OneHotSpeakers: def __init__(self, speakers_list): # pylint: disable=E0611,E0401 from tensorflow.keras.utils import to_categorical self.speaker_ids = sorted(speakers_list) self.int_speaker_ids = list(range(len(self.speaker_ids))) self.map_speakers_to_index = dict([(k, v) for (k, v) in zip(self.speaker_ids, self.int_speaker_ids)]) self.map_index_to_speakers = dict([(v, k) for (k, v) in zip(self.speaker_ids, self.int_speaker_ids)]) self.speaker_categories = to_categorical(self.int_speaker_ids, num_classes=len(self.speaker_ids)) def get_speaker_from_index(self, index): return self.map_index_to_speakers[index] def get_one_hot(self, speaker_id): index = self.map_speakers_to_index[speaker_id] return self.speaker_categories[index] class LazyTripletBatcher: def __init__(self, working_dir: str, max_length: int, model: DeepSpeakerModel): self.working_dir = working_dir self.audio = Audio(cache_dir=working_dir) logger.info(f'Picking audio from {working_dir}.') self.sp_to_utt_train = train_test_sp_to_utt(self.audio, is_test=False) self.sp_to_utt_test = train_test_sp_to_utt(self.audio, is_test=True) self.max_length = max_length self.model = model self.nb_per_speaker = 2 self.nb_speakers = 640 self.history_length = 4 self.history_every = 100 # batches. self.total_history_length = self.nb_speakers * self.nb_per_speaker * self.history_length # 25,600 self.metadata_train_speakers = Counter() self.metadata_output_file = os.path.join(self.working_dir, 'debug_batcher.json') self.history_embeddings_train = deque(maxlen=self.total_history_length) self.history_utterances_train = deque(maxlen=self.total_history_length) self.history_model_inputs_train = deque(maxlen=self.total_history_length) self.history_embeddings = None self.history_utterances = None self.history_model_inputs = None self.batch_count = 0 for _ in tqdm(range(self.history_length), desc='Initializing the batcher'): # init history. self.update_triplets_history() def update_triplets_history(self): model_inputs = [] speakers = list(self.audio.speakers_to_utterances.keys()) np.random.shuffle(speakers) selected_speakers = speakers[: self.nb_speakers] embeddings_utterances = [] for speaker_id in selected_speakers: train_utterances = self.sp_to_utt_train[speaker_id] for selected_utterance in np.random.choice(a=train_utterances, size=self.nb_per_speaker, replace=False): mfcc = sample_from_mfcc_file(selected_utterance, self.max_length) embeddings_utterances.append(selected_utterance) model_inputs.append(mfcc) embeddings = self.model.m.predict(np.array(model_inputs)) assert embeddings.shape[-1] == 512 embeddings = np.reshape(embeddings, (len(selected_speakers), self.nb_per_speaker, 512)) self.history_embeddings_train.extend(list(embeddings.reshape((-1, 512)))) self.history_utterances_train.extend(embeddings_utterances) self.history_model_inputs_train.extend(model_inputs) # reason: can't index a deque with a np.array. self.history_embeddings = np.array(self.history_embeddings_train) self.history_utterances = np.array(self.history_utterances_train) self.history_model_inputs = np.array(self.history_model_inputs_train) with open(self.metadata_output_file, 'w') as w: json.dump(obj=dict(self.metadata_train_speakers), fp=w, indent=2) def get_batch(self, batch_size, is_test=False): return self.get_batch_test(batch_size) if is_test else self.get_random_batch(batch_size, is_test=False) def get_batch_test(self, batch_size): return self.get_random_batch(batch_size, is_test=True) def get_random_batch(self, batch_size, is_test=False): sp_to_utt = self.sp_to_utt_test if is_test else self.sp_to_utt_train speakers = list(self.audio.speakers_to_utterances.keys()) anchor_speakers = np.random.choice(speakers, size=batch_size // 3, replace=False) anchor_utterances = [] positive_utterances = [] negative_utterances = [] for anchor_speaker in anchor_speakers: negative_speaker = np.random.choice(list(set(speakers) - {anchor_speaker}), size=1)[0] assert negative_speaker != anchor_speaker pos_utterances = np.random.choice(sp_to_utt[anchor_speaker], 2, replace=False) neg_utterance = np.random.choice(sp_to_utt[negative_speaker], 1, replace=True)[0] anchor_utterances.append(pos_utterances[0]) positive_utterances.append(pos_utterances[1]) negative_utterances.append(neg_utterance) # anchor and positive should have difference utterances (but same speaker!). anc_pos = np.array([positive_utterances, anchor_utterances]) assert np.all(anc_pos[0, :] != anc_pos[1, :]) assert np.all(np.array([extract_speaker(s) for s in anc_pos[0, :]]) == np.array( [extract_speaker(s) for s in anc_pos[1, :]])) pos_neg = np.array([positive_utterances, negative_utterances]) assert np.all(pos_neg[0, :] != pos_neg[1, :]) assert np.all(np.array([extract_speaker(s) for s in pos_neg[0, :]]) != np.array( [extract_speaker(s) for s in pos_neg[1, :]])) batch_x = np.vstack([ [sample_from_mfcc_file(u, self.max_length) for u in anchor_utterances], [sample_from_mfcc_file(u, self.max_length) for u in positive_utterances], [sample_from_mfcc_file(u, self.max_length) for u in negative_utterances] ]) batch_y = np.zeros(shape=(len(batch_x), 1)) # dummy. sparse softmax needs something. return batch_x, batch_y def get_batch_train(self, batch_size): from deep_speaker.test import batch_cosine_similarity # s1 = time() self.batch_count += 1 if self.batch_count % self.history_every == 0: self.update_triplets_history() all_indexes = range(len(self.history_embeddings_train)) anchor_indexes = np.random.choice(a=all_indexes, size=batch_size // 3, replace=False) # s2 = time() similar_negative_indexes = [] dissimilar_positive_indexes = [] # could be made parallel. for anchor_index in anchor_indexes: # s21 = time() anchor_embedding = self.history_embeddings[anchor_index] anchor_speaker = extract_speaker(self.history_utterances[anchor_index]) # why self.nb_speakers // 2? just random. because it is fast. otherwise it's too much. negative_indexes = [j for (j, a) in enumerate(self.history_utterances) if extract_speaker(a) != anchor_speaker] negative_indexes = np.random.choice(negative_indexes, size=self.nb_speakers // 2) # s22 = time() anchor_embedding_tile = [anchor_embedding] * len(negative_indexes) anchor_cos = batch_cosine_similarity(anchor_embedding_tile, self.history_embeddings[negative_indexes]) # s23 = time() similar_negative_index = negative_indexes[np.argsort(anchor_cos)[-1]] # [-1:] similar_negative_indexes.append(similar_negative_index) # s24 = time() positive_indexes = [j for (j, a) in enumerate(self.history_utterances) if extract_speaker(a) == anchor_speaker and j != anchor_index] # s25 = time() anchor_embedding_tile = [anchor_embedding] * len(positive_indexes) # s26 = time() anchor_cos = batch_cosine_similarity(anchor_embedding_tile, self.history_embeddings[positive_indexes]) dissimilar_positive_index = positive_indexes[np.argsort(anchor_cos)[0]] # [:1] dissimilar_positive_indexes.append(dissimilar_positive_index) # s27 = time() # s3 = time() batch_x = np.vstack([ self.history_model_inputs[anchor_indexes], self.history_model_inputs[dissimilar_positive_indexes], self.history_model_inputs[similar_negative_indexes] ]) # s4 = time() # for anchor, positive, negative in zip(history_utterances[anchor_indexes], # history_utterances[dissimilar_positive_indexes], # history_utterances[similar_negative_indexes]): # print('anchor', os.path.basename(anchor), # 'positive', os.path.basename(positive), # 'negative', os.path.basename(negative)) # print('_' * 80) # assert utterances as well positive != anchor. anchor_speakers = [extract_speaker(a) for a in self.history_utterances[anchor_indexes]] positive_speakers = [extract_speaker(a) for a in self.history_utterances[dissimilar_positive_indexes]] negative_speakers = [extract_speaker(a) for a in self.history_utterances[similar_negative_indexes]] assert len(anchor_indexes) == len(dissimilar_positive_indexes) assert len(similar_negative_indexes) == len(dissimilar_positive_indexes) assert list(self.history_utterances[dissimilar_positive_indexes]) != list( self.history_utterances[anchor_indexes]) assert anchor_speakers == positive_speakers assert negative_speakers != anchor_speakers batch_y = np.zeros(shape=(len(batch_x), 1)) # dummy. sparse softmax needs something. for a in anchor_speakers: self.metadata_train_speakers[a] += 1 for a in positive_speakers: self.metadata_train_speakers[a] += 1 for a in negative_speakers: self.metadata_train_speakers[a] += 1 # s5 = time() # print('1-2', s2 - s1) # print('2-3', s3 - s2) # print('3-4', s4 - s3) # print('4-5', s5 - s4) # print('21-22', (s22 - s21) * (batch_size // 3)) # print('22-23', (s23 - s22) * (batch_size // 3)) # print('23-24', (s24 - s23) * (batch_size // 3)) # print('24-25', (s25 - s24) * (batch_size // 3)) # print('25-26', (s26 - s25) * (batch_size // 3)) # print('26-27', (s27 - s26) * (batch_size // 3)) return batch_x, batch_y def get_speaker_verification_data(self, anchor_speaker, num_different_speakers): speakers = list(self.audio.speakers_to_utterances.keys()) anchor_utterances = [] positive_utterances = [] negative_utterances = [] negative_speakers = np.random.choice(list(set(speakers) - {anchor_speaker}), size=num_different_speakers) assert [negative_speaker != anchor_speaker for negative_speaker in negative_speakers] pos_utterances = np.random.choice(self.sp_to_utt_test[anchor_speaker], 2, replace=False) neg_utterances = [np.random.choice(self.sp_to_utt_test[neg], 1, replace=True)[0] for neg in negative_speakers] anchor_utterances.append(pos_utterances[0]) positive_utterances.append(pos_utterances[1]) negative_utterances.extend(neg_utterances) # anchor and positive should have difference utterances (but same speaker!). anc_pos = np.array([positive_utterances, anchor_utterances]) assert np.all(anc_pos[0, :] != anc_pos[1, :]) assert np.all(np.array([extract_speaker(s) for s in anc_pos[0, :]]) == np.array( [extract_speaker(s) for s in anc_pos[1, :]])) batch_x = np.vstack([ [sample_from_mfcc_file(u, self.max_length) for u in anchor_utterances], [sample_from_mfcc_file(u, self.max_length) for u in positive_utterances], [sample_from_mfcc_file(u, self.max_length) for u in negative_utterances] ]) batch_y = np.zeros(shape=(len(batch_x), 1)) # dummy. sparse softmax needs something. return batch_x, batch_y class TripletBatcher: def __init__(self, kx_train, ky_train, kx_test, ky_test): self.kx_train = kx_train self.ky_train = ky_train self.kx_test = kx_test self.ky_test = ky_test speakers_list = sorted(set(ky_train.argmax(axis=1))) num_different_speakers = len(speakers_list) assert speakers_list == sorted(set(ky_test.argmax(axis=1))) # train speakers = test speakers. assert speakers_list == list(range(num_different_speakers)) self.train_indices_per_speaker = {} self.test_indices_per_speaker = {} for speaker_id in speakers_list: self.train_indices_per_speaker[speaker_id] = list(np.where(ky_train.argmax(axis=1) == speaker_id)[0]) self.test_indices_per_speaker[speaker_id] = list(np.where(ky_test.argmax(axis=1) == speaker_id)[0]) # check. # print(sorted(sum([v for v in self.train_indices_per_speaker.values()], []))) # print(range(len(ky_train))) assert sorted(sum([v for v in self.train_indices_per_speaker.values()], [])) == sorted(range(len(ky_train))) assert sorted(sum([v for v in self.test_indices_per_speaker.values()], [])) == sorted(range(len(ky_test))) self.speakers_list = speakers_list def select_speaker_data(self, speaker, n, is_test): x = self.kx_test if is_test else self.kx_train indices_per_speaker = self.test_indices_per_speaker if is_test else self.train_indices_per_speaker indices = np.random.choice(indices_per_speaker[speaker], size=n) return x[indices] def get_batch(self, batch_size, is_test=False): # y = self.ky_test if is_test else self.ky_train two_different_speakers = np.random.choice(self.speakers_list, size=2, replace=False) anchor_positive_speaker = two_different_speakers[0] negative_speaker = two_different_speakers[1] assert negative_speaker != anchor_positive_speaker batch_x = np.vstack([ self.select_speaker_data(anchor_positive_speaker, batch_size // 3, is_test), self.select_speaker_data(anchor_positive_speaker, batch_size // 3, is_test), self.select_speaker_data(negative_speaker, batch_size // 3, is_test) ]) batch_y = np.zeros(shape=(len(batch_x), len(self.speakers_list))) return batch_x, batch_y class TripletBatcherMiner(TripletBatcher): def __init__(self, kx_train, ky_train, kx_test, ky_test, model: DeepSpeakerModel): super().__init__(kx_train, ky_train, kx_test, ky_test) self.model = model self.num_evaluations_to_find_best_batch = 10 def get_batch(self, batch_size, is_test=False): if is_test: return super().get_batch(batch_size, is_test) max_loss = 0 max_batch = None, None for i in range(self.num_evaluations_to_find_best_batch): bx, by = super().get_batch(batch_size, is_test=False) # only train here. loss = self.model.m.evaluate(bx, by, batch_size=batch_size, verbose=0) if loss > max_loss: max_loss = loss max_batch = bx, by return max_batch class TripletBatcherSelectHardNegatives(TripletBatcher): def __init__(self, kx_train, ky_train, kx_test, ky_test, model: DeepSpeakerModel): super().__init__(kx_train, ky_train, kx_test, ky_test) self.model = model def get_batch(self, batch_size, is_test=False, predict=None): if predict is None: predict = self.model.m.predict from deep_speaker.test import batch_cosine_similarity num_triplets = batch_size // 3 inputs = [] k = 2 # do not change this. for speaker in self.speakers_list: inputs.append(self.select_speaker_data(speaker, n=k, is_test=is_test)) inputs = np.array(inputs) # num_speakers * [k, num_frames, num_fbanks, 1]. embeddings = predict(np.vstack(inputs)) assert embeddings.shape[-1] == 512 # (speaker, utterance, 512) embeddings = np.reshape(embeddings, (len(self.speakers_list), k, 512)) cs = batch_cosine_similarity(embeddings[:, 0], embeddings[:, 1]) arg_sort = np.argsort(cs) assert len(arg_sort) > num_triplets anchor_speakers = arg_sort[0:num_triplets] anchor_embeddings = embeddings[anchor_speakers, 0] negative_speakers = sorted(set(self.speakers_list) - set(anchor_speakers)) negative_embeddings = embeddings[negative_speakers, 0] selected_negative_speakers = [] for anchor_embedding in anchor_embeddings: cs_negative = [batch_cosine_similarity([anchor_embedding], neg) for neg in negative_embeddings] selected_negative_speakers.append(negative_speakers[int(np.argmax(cs_negative))]) # anchor with frame 0. # positive with frame 1. # negative with frame 0. assert len(set(selected_negative_speakers).intersection(anchor_speakers)) == 0 negative = inputs[selected_negative_speakers, 0] positive = inputs[anchor_speakers, 1] anchor = inputs[anchor_speakers, 0] batch_x = np.vstack([anchor, positive, negative]) batch_y = np.zeros(shape=(len(batch_x), len(self.speakers_list))) return batch_x, batch_y class TripletEvaluator: def __init__(self, kx_test, ky_test): self.kx_test = kx_test self.ky_test = ky_test speakers_list = sorted(set(ky_test.argmax(axis=1))) num_different_speakers = len(speakers_list) assert speakers_list == list(range(num_different_speakers)) self.test_indices_per_speaker = {} for speaker_id in speakers_list: self.test_indices_per_speaker[speaker_id] = list(np.where(ky_test.argmax(axis=1) == speaker_id)[0]) assert sorted(sum([v for v in self.test_indices_per_speaker.values()], [])) == sorted(range(len(ky_test))) self.speakers_list = speakers_list def _select_speaker_data(self, speaker): indices = np.random.choice(self.test_indices_per_speaker[speaker], size=1) return self.kx_test[indices] def get_speaker_verification_data(self, positive_speaker, num_different_speakers): all_negative_speakers = list(set(self.speakers_list) - {positive_speaker}) assert len(self.speakers_list) - 1 == len(all_negative_speakers) negative_speakers = np.random.choice(all_negative_speakers, size=num_different_speakers, replace=False) assert positive_speaker not in negative_speakers anchor = self._select_speaker_data(positive_speaker) positive = self._select_speaker_data(positive_speaker) data = [anchor, positive] data.extend([self._select_speaker_data(n) for n in negative_speakers]) return np.vstack(data) if __name__ == '__main__': np.random.seed(123) ltb = LazyTripletBatcher(working_dir='/Users/premy/deep-speaker/', max_length=NUM_FRAMES, model=DeepSpeakerModel()) for i in range(1000): print(i) start = time() ltb.get_batch_train(batch_size=9) print(time() - start) # ltb.get_batch(batch_size=96)
23,984
46.401186
119
py
deep-speaker
deep-speaker-master/deep_speaker/triplet_loss.py
# pylint: disable=E0611,E0401 import tensorflow.keras.backend as K # ALPHA = 0.2 # used in FaceNet https://arxiv.org/pdf/1503.03832.pdf ALPHA = 0.1 # used in Deep Speaker. def batch_cosine_similarity(x1, x2): # https://en.wikipedia.org/wiki/Cosine_similarity # 1 = equal direction ; -1 = opposite direction dot = K.squeeze(K.batch_dot(x1, x2, axes=1), axis=1) # as values have have length 1, we don't need to divide by norm (as it is 1) return dot def deep_speaker_loss(y_true, y_pred, alpha=ALPHA): # y_true is not used. we respect this convention: # y_true.shape = (batch_size, embedding_size) [not used] # y_pred.shape = (batch_size, embedding_size) # EXAMPLE: # _____________________________________________________ # ANCHOR 1 (512,) # ANCHOR 2 (512,) # POS EX 1 (512,) # POS EX 2 (512,) # NEG EX 1 (512,) # NEG EX 2 (512,) # _____________________________________________________ split = K.shape(y_pred)[0] // 3 anchor = y_pred[0:split] positive_ex = y_pred[split:2 * split] negative_ex = y_pred[2 * split:] # If the loss does not decrease below ALPHA then the model does not learn anything. # If all anchor = positive = negative (model outputs the same vector always). # Then sap = san = 1. and loss = max(alpha,0) = alpha. # On the contrary if anchor = positive = [1] and negative = [-1]. # Then sap = 1 and san = -1. loss = max(-1-1+0.1,0) = max(-1.9, 0) = 0. sap = batch_cosine_similarity(anchor, positive_ex) san = batch_cosine_similarity(anchor, negative_ex) loss = K.maximum(san - sap + alpha, 0.0) total_loss = K.mean(loss) return total_loss if __name__ == '__main__': import numpy as np print(deep_speaker_loss(alpha=0.1, y_true=0, y_pred=np.array([[0.9], [1.0], [-1.0]]))) print(deep_speaker_loss(alpha=1, y_true=0, y_pred=np.array([[0.9], [1.0], [-1.0]]))) print(deep_speaker_loss(alpha=2, y_true=0, y_pred=np.array([[0.9], [1.0], [-1.0]]))) print('--------------') print(deep_speaker_loss(alpha=2, y_true=0, y_pred=np.array([[0.6], [1.0], [0.0]]))) print(deep_speaker_loss(alpha=1, y_true=0, y_pred=np.array([[0.6], [1.0], [0.0]]))) print(deep_speaker_loss(alpha=0.1, y_true=0, y_pred=np.array([[0.6], [1.0], [0.0]]))) print(deep_speaker_loss(alpha=0.2, y_true=0, y_pred=np.array([[0.6], [1.0], [0.0]]))) print('--------------') print(deep_speaker_loss(alpha=2, y_true=0, y_pred=np.array([[0.9], [1.0], [-1.0]]))) print(deep_speaker_loss(alpha=1, y_true=0, y_pred=np.array([[0.9], [1.0], [-1.0]]))) print(deep_speaker_loss(alpha=0.1, y_true=0, y_pred=np.array([[0.9], [1.0], [-1.0]]))) print(deep_speaker_loss(alpha=0.2, y_true=0, y_pred=np.array([[0.9], [1.0], [-1.0]])))
2,782
42.484375
90
py
deep-speaker
deep-speaker-master/deep_speaker/conv_models.py
import logging import os import numpy as np import tensorflow as tf # pylint: disable=E0611,E0401 import tensorflow.keras.backend as K # pylint: disable=E0611,E0401 from tensorflow.keras import layers, regularizers # pylint: disable=E0611,E0401 from tensorflow.keras.layers import ( BatchNormalization, Conv2D, Dense, Dropout, Input, Lambda, Reshape, ) # pylint: disable=E0611,E0401 from tensorflow.keras.models import Model # pylint: disable=E0611,E0401 from tensorflow.keras.optimizers import Adam from deep_speaker.constants import NUM_FBANKS, SAMPLE_RATE, NUM_FRAMES from deep_speaker.triplet_loss import deep_speaker_loss logger = logging.getLogger(__name__) @tf.function def tf_normalize(data, ndims, eps=0, adjusted=False): data = tf.convert_to_tensor(data, name='data') reduce_dims = [-i - 1 for i in range(ndims)] # pylint: disable=E1123,E1120 data = tf.cast(data, dtype=tf.dtypes.float32) data_num = tf.reduce_prod(data.shape[-ndims:]) data_mean = tf.reduce_mean(data, axis=reduce_dims, keepdims=True) # Apply a minimum normalization that protects us against uniform images. stddev = tf.math.reduce_std(data, axis=reduce_dims, keepdims=True) adjusted_stddev = stddev if adjusted: min_stddev = tf.math.rsqrt(tf.cast(data_num, tf.dtypes.float32)) eps = tf.maximum(eps, min_stddev) if eps > 0: adjusted_stddev = tf.maximum(adjusted_stddev, eps) return (data - data_mean) / adjusted_stddev @tf.function def tf_fbank(samples): """ Compute Mel-filterbank energy features from an audio signal. See python_speech_features.fbank """ frame_length = int(0.025 * SAMPLE_RATE) frame_step = int(0.01 * SAMPLE_RATE) fft_length = 512 fft_bins = fft_length // 2 + 1 pre_emphasis = samples[:, 1:] - 0.97 * samples[:, :-1] # Original implementation from python_speech_features # frames = tf.expand_dims(sigproc.framesig(preemphasis[0], frame_length, # frame_step, winfunc=lambda x: np.ones((x,))), 0) # powspec = sigproc.powspec(frames, fft_length) # Tensorflow impl #1, using manually-split frames and rfft # spec = tf.abs(tf.signal.rfft(frames, [fft_length])) # powspec = tf.square(spec) / fft_length # Tensorflow impl #2, using stft to handle framing automatically # (There is a one-off mismatch on the number of frames on the resulting tensor, but I guess this is ok) spec = tf.abs(tf.signal.stft(pre_emphasis, frame_length, frame_step, fft_length, window_fn=tf.ones)) powspec = tf.square(spec) / fft_length # Matrix to transform spectrum to mel-frequencies # Original implementation from python_speech_features # linear_to_mel_weight_matrix = get_filterbanks(NUM_FBANKS, fft_length, # SAMPLE_RATE, 0, SAMPLE_RATE/2).astype(np.float32).T linear_to_mel_weight_matrix = tf.signal.linear_to_mel_weight_matrix( num_mel_bins=NUM_FBANKS, num_spectrogram_bins=fft_bins, sample_rate=SAMPLE_RATE, lower_edge_hertz=0, upper_edge_hertz=SAMPLE_RATE / 2, ) feat = tf.matmul(powspec, linear_to_mel_weight_matrix) # feat = tf.where(feat == 0, np.finfo(np.float32).eps, feat) return feat class DeepSpeakerModel: # I thought it was 3 but maybe energy is added at a 4th dimension. # would be better to have 4 dimensions: # MFCC, DIFF(MFCC), DIFF(DIFF(MFCC)), ENERGIES (probably tiled across the frequency domain). # this seems to help match the parameter counts. def __init__( self, batch_input_shape=(None, NUM_FRAMES, NUM_FBANKS, 1), include_softmax=False, num_speakers_softmax=None, pcm_input=False ): if pcm_input: batch_input_shape = None self.include_softmax = include_softmax if self.include_softmax: assert num_speakers_softmax > 0 self.clipped_relu_count = 0 # http://cs231n.github.io/convolutional-networks/ # conv weights # #params = ks * ks * nb_filters * num_channels_input # Conv128-s # 5*5*128*128/2+128 # ks*ks*nb_filters*channels/strides+bias(=nb_filters) # take 100 ms -> 4 frames. # if signal is 3 seconds, then take 100ms per 100ms and average out this network. # 8*8 = 64 features. # used to share all the layers across the inputs # num_frames = K.shape() - do it dynamically after. if pcm_input: batch_input_shape = batch_input_shape or (None, None) # Batch-size, num-samples inputs = Input(batch_shape=batch_input_shape, name='raw_inputs') x = inputs x = Lambda(tf_fbank)(x) x = Lambda(lambda x_: tf_normalize(x_, 1, 1e-12))(x) x = Lambda(lambda x_: tf.expand_dims(x_, axis=-1))(x) else: batch_input_shape = batch_input_shape or (None, None, NUM_FBANKS, 1) inputs = Input(batch_shape=batch_input_shape, name='input') x = inputs x = self.cnn_component(x) x = Reshape((-1, 2048))(x) # Temporal average layer. axis=1 is time. x = Lambda(lambda y: K.mean(y, axis=1), name='average')(x) if include_softmax: logger.info('Including a Dropout layer to reduce overfitting.') # used for softmax because the dataset we pre-train on might be too small. easy to overfit. x = Dropout(0.5)(x) x = Dense(512, name='affine')(x) if include_softmax: # Those weights are just when we train on softmax. x = Dense(num_speakers_softmax, activation='softmax')(x) else: # Does not contain any weights. x = Lambda(lambda y: K.l2_normalize(y, axis=1), name='ln')(x) self.m = Model(inputs, x, name='ResCNN') def keras_model(self): return self.m def get_weights(self): w = self.m.get_weights() if self.include_softmax: w.pop() # last 2 are the W_softmax and b_softmax. w.pop() return w def clipped_relu(self, inputs): relu = Lambda(lambda y: K.minimum(K.maximum(y, 0), 20), name=f'clipped_relu_{self.clipped_relu_count}')(inputs) self.clipped_relu_count += 1 return relu def identity_block(self, input_tensor, kernel_size, filters, stage, block): conv_name_base = f'res{stage}_{block}_branch' x = Conv2D(filters, kernel_size=kernel_size, strides=1, activation=None, padding='same', kernel_initializer='glorot_uniform', kernel_regularizer=regularizers.l2(l=0.0001), name=conv_name_base + '_2a')(input_tensor) x = BatchNormalization(name=conv_name_base + '_2a_bn')(x) x = self.clipped_relu(x) x = Conv2D( filters, kernel_size=kernel_size, strides=1, activation=None, padding='same', kernel_initializer='glorot_uniform', kernel_regularizer=regularizers.l2(l=0.0001), name=conv_name_base + '_2b', )(x) x = BatchNormalization(name=conv_name_base + '_2b_bn')(x) x = self.clipped_relu(x) x = layers.add([x, input_tensor]) x = self.clipped_relu(x) return x def conv_and_res_block(self, inp, filters, stage): conv_name = 'conv{}-s'.format(filters) # TODO: why kernel_regularizer? o = Conv2D(filters, kernel_size=5, strides=2, activation=None, padding='same', kernel_initializer='glorot_uniform', kernel_regularizer=regularizers.l2(l=0.0001), name=conv_name)(inp) o = BatchNormalization(name=conv_name + '_bn')(o) o = self.clipped_relu(o) for i in range(3): o = self.identity_block(o, kernel_size=3, filters=filters, stage=stage, block=i) return o def cnn_component(self, inp): x = self.conv_and_res_block(inp, 64, stage=1) x = self.conv_and_res_block(x, 128, stage=2) x = self.conv_and_res_block(x, 256, stage=3) x = self.conv_and_res_block(x, 512, stage=4) return x def set_weights(self, w): for layer, layer_w in zip(self.m.layers, w): layer.set_weights(layer_w) logger.info(f'Setting weights for [{layer.name}]...') def main(): # Looks correct to me. # I have 37K but paper reports 41K. which is not too far. dsm = DeepSpeakerModel() dsm.m.summary() # I suspect num frames to be 32. # Then fbank=64, then total would be 32*64 = 2048. # plot_model(dsm.m, to_file='model.png', dpi=300, show_shapes=True, expand_nested=True) def _train(): # x = np.random.uniform(size=(6, 32, 64, 4)) # 6 is multiple of 3. # y_softmax = np.random.uniform(size=(6, 100)) # dsm = DeepSpeakerModel(batch_input_shape=(None, 32, 64, 4), include_softmax=True, num_speakers_softmax=100) # dsm.m.compile(optimizer=Adam(lr=0.01), loss='categorical_crossentropy') # print(dsm.m.predict(x).shape) # print(dsm.m.evaluate(x, y_softmax)) # w = dsm.get_weights() dsm = DeepSpeakerModel(batch_input_shape=(None, 32, 64, 4), include_softmax=False) # dsm.m.set_weights(w) dsm.m.compile(optimizer=Adam(lr=0.01), loss=deep_speaker_loss) # it works!!!!!!!!!!!!!!!!!!!! # unit_batch_size = 20 # anchor = np.ones(shape=(unit_batch_size, 32, 64, 4)) # positive = np.array(anchor) # negative = np.ones(shape=(unit_batch_size, 32, 64, 4)) * (-1) # batch = np.vstack((anchor, positive, negative)) # x = batch # y = np.zeros(shape=(len(batch), 512)) # not important. # print('Starting to fit...') # while True: # print(dsm.m.train_on_batch(x, y)) # should not work... and it does not work! unit_batch_size = 20 negative = np.ones(shape=(unit_batch_size, 32, 64, 4)) * (-1) batch = np.vstack((negative, negative, negative)) x = batch y = np.zeros(shape=(len(batch), 512)) # not important. print('Starting to fit...') while True: print(dsm.m.train_on_batch(x, y)) def _test_checkpoint_compatibility(): dsm = DeepSpeakerModel(batch_input_shape=(None, 32, 64, 4), include_softmax=True, num_speakers_softmax=10) dsm.m.save_weights('test.h5') dsm = DeepSpeakerModel(batch_input_shape=(None, 32, 64, 4), include_softmax=False) dsm.m.load_weights('test.h5', by_name=True) os.remove('test.h5') if __name__ == '__main__': _test_checkpoint_compatibility()
10,741
35.16835
119
py
deep-speaker
deep-speaker-master/deep_speaker/train.py
import logging import os # pylint: disable=E0611,E0401 from tensorflow.keras.callbacks import ReduceLROnPlateau, EarlyStopping, ModelCheckpoint # pylint: disable=E0611,E0401 from tensorflow.keras.optimizers import SGD from tqdm import tqdm from deep_speaker.batcher import KerasFormatConverter, LazyTripletBatcher from deep_speaker.constants import BATCH_SIZE, CHECKPOINTS_SOFTMAX_DIR, CHECKPOINTS_TRIPLET_DIR, NUM_FRAMES, NUM_FBANKS from deep_speaker.conv_models import DeepSpeakerModel from deep_speaker.triplet_loss import deep_speaker_loss from deep_speaker.utils import load_best_checkpoint, ensures_dir logger = logging.getLogger(__name__) # Otherwise it's just too much logging from Tensorflow... os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' def fit_model(dsm: DeepSpeakerModel, working_dir: str, max_length: int = NUM_FRAMES, batch_size=BATCH_SIZE): batcher = LazyTripletBatcher(working_dir, max_length, dsm) # build small test set. test_batches = [] for _ in tqdm(range(200), desc='Build test set'): test_batches.append(batcher.get_batch_test(batch_size)) def test_generator(): while True: for bb in test_batches: yield bb def train_generator(): while True: yield batcher.get_random_batch(batch_size, is_test=False) checkpoint_name = dsm.m.name + '_checkpoint' checkpoint_filename = os.path.join(CHECKPOINTS_TRIPLET_DIR, checkpoint_name + '_{epoch}.h5') checkpoint = ModelCheckpoint(monitor='val_loss', filepath=checkpoint_filename, save_best_only=True) dsm.m.fit(x=train_generator(), y=None, steps_per_epoch=2000, shuffle=False, epochs=1000, validation_data=test_generator(), validation_steps=len(test_batches), callbacks=[checkpoint]) def fit_model_softmax(dsm: DeepSpeakerModel, kx_train, ky_train, kx_test, ky_test, batch_size=BATCH_SIZE, max_epochs=1000, initial_epoch=0): checkpoint_name = dsm.m.name + '_checkpoint' checkpoint_filename = os.path.join(CHECKPOINTS_SOFTMAX_DIR, checkpoint_name + '_{epoch}.h5') checkpoint = ModelCheckpoint(monitor='val_accuracy', filepath=checkpoint_filename, save_best_only=True) # if the accuracy does not increase by 0.1% over 20 epochs, we stop the training. early_stopping = EarlyStopping(monitor='val_accuracy', min_delta=0.001, patience=20, verbose=1, mode='max') # if the accuracy does not increase over 10 epochs, we reduce the learning rate by half. reduce_lr = ReduceLROnPlateau(monitor='val_accuracy', factor=0.5, patience=10, min_lr=0.0001, verbose=1) max_len_train = len(kx_train) - len(kx_train) % batch_size kx_train = kx_train[0:max_len_train] ky_train = ky_train[0:max_len_train] max_len_test = len(kx_test) - len(kx_test) % batch_size kx_test = kx_test[0:max_len_test] ky_test = ky_test[0:max_len_test] dsm.m.fit(x=kx_train, y=ky_train, batch_size=batch_size, epochs=initial_epoch + max_epochs, initial_epoch=initial_epoch, verbose=1, shuffle=True, validation_data=(kx_test, ky_test), callbacks=[early_stopping, reduce_lr, checkpoint]) def start_training(working_dir, pre_training_phase=True): ensures_dir(CHECKPOINTS_SOFTMAX_DIR) ensures_dir(CHECKPOINTS_TRIPLET_DIR) batch_input_shape = [None, NUM_FRAMES, NUM_FBANKS, 1] if pre_training_phase: logger.info('Softmax pre-training.') kc = KerasFormatConverter(working_dir) num_speakers_softmax = len(kc.categorical_speakers.speaker_ids) dsm = DeepSpeakerModel(batch_input_shape, include_softmax=True, num_speakers_softmax=num_speakers_softmax) dsm.m.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy']) pre_training_checkpoint = load_best_checkpoint(CHECKPOINTS_SOFTMAX_DIR) if pre_training_checkpoint is not None: initial_epoch = int(pre_training_checkpoint.split('/')[-1].split('.')[0].split('_')[-1]) logger.info(f'Initial epoch is {initial_epoch}.') logger.info(f'Loading softmax checkpoint: {pre_training_checkpoint}.') dsm.m.load_weights(pre_training_checkpoint) # latest one. else: initial_epoch = 0 fit_model_softmax(dsm, kc.kx_train, kc.ky_train, kc.kx_test, kc.ky_test, initial_epoch=initial_epoch) else: logger.info('Training with the triplet loss.') dsm = DeepSpeakerModel(batch_input_shape, include_softmax=False) triplet_checkpoint = load_best_checkpoint(CHECKPOINTS_TRIPLET_DIR) pre_training_checkpoint = load_best_checkpoint(CHECKPOINTS_SOFTMAX_DIR) if triplet_checkpoint is not None: logger.info(f'Loading triplet checkpoint: {triplet_checkpoint}.') dsm.m.load_weights(triplet_checkpoint) elif pre_training_checkpoint is not None: logger.info(f'Loading pre-training checkpoint: {pre_training_checkpoint}.') # If `by_name` is True, weights are loaded into layers only if they share the # same name. This is useful for fine-tuning or transfer-learning models where # some of the layers have changed. dsm.m.load_weights(pre_training_checkpoint, by_name=True) dsm.m.compile(optimizer=SGD(), loss=deep_speaker_loss) fit_model(dsm, working_dir, NUM_FRAMES)
5,455
47.714286
119
py
samurai
samurai-master/doc/source/conf.py
# -*- 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 sphinx_rtd_theme # -- Project information ----------------------------------------------------- project = 'Samurai' author = 'samurai developers' # The short X.Y version version = '0.0' # The full version, including alpha/beta/rc tags release = '0.0.0' # -- 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.todo', 'sphinx.ext.mathjax', 'breathe', 'sphinx_rtd_theme', ] breathe_projects = { 'samurai': '../xml' } # 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'] source_suffix = '.rst' rst_epilog = f""" .. |project| replace:: {project} .. role:: cpp_code(code) :language: c++ """ # 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 = None # 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 = [] # 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' # 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 = { 'collapse_navigation': False, } # 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'] # 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 = {} html_css_files = [ 'css/custom.css', ] # -- Options for HTMLHelp output --------------------------------------------- # Output file base name for HTML help builder. htmlhelp_basename = 'Samuraidoc' # -- Options for LaTeX output ------------------------------------------------ latex_elements = { # The paper size ('letterpaper' or 'a4paper'). # # 'papersize': 'letterpaper', # The font size ('10pt', '11pt' or '12pt'). # 'pointsize': '11pt', # Additional stuff for the LaTeX preamble. # 'preamble': ''' \\fvset{fontsize=\\scriptsize} ''', # 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, 'samurai.tex', 'Samurai Documentation', author, '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, 'samurai', 'Samurai 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, 'Samurai', 'Samurai Documentation', author, 'Samurai developers', '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 todo extension ---------------------------------------------- # If true, `todo` and `todoList` produce output, else they produce nothing. todo_include_todos = True # -- Breathe configuration for readthedocs ----------------------------------- import subprocess, os read_the_docs_build = os.environ.get('READTHEDOCS', None) == 'True' if read_the_docs_build: subprocess.call('cd ..; doxygen', shell=True)
5,971
27.037559
79
py
LAMP
LAMP-main/loop_closure/src/loop_closure_batcher.py
from gnn.gnn_model import LoopClosureGNN import torch from pose_graph_msgs.msg import PoseGraph, LoopCandidateArray, LoopComputationStatus import rospy from copy import deepcopy import traceback import os from subset_algorithms.heuristic import MaximallySeperatedNodes, MaximumCovarianceNodes, Random from subset_algorithms.gnn import BinomialCEM, SimulatedAnnealing, DiscretePSO, DiscreteTabuSearch import threading from concurrent.futures import ThreadPoolExecutor import rospkg def edge_selection_algorithm_dispatch(algorithm_name, lock): if algorithm_name == "cem": return BinomialCEM(lock) elif algorithm_name == "random": return Random(lock) elif algorithm_name == "heuristic_maximally_separated_nodes": return MaximallySeperatedNodes(lock) elif algorithm_name == "heuristic_maximum_covariance": return MaximumCovarianceNodes(lock) elif algorithm_name == "simulated_annealing": return SimulatedAnnealing(lock) elif algorithm_name == "pso": return DiscretePSO(lock) elif algorithm_name == "tabu": return DiscreteTabuSearch(lock) else: raise Exception("Edge subset selection algorithm unknown {algorithm_name}") class LoopClosureBatcher: def __init__(self, params): rospack = rospkg.RosPack() cur_path = rospack.get_path("loop_closure") self.loop_closures_to_choose = params["number_of_loop_closures"] self.min_queue_size = params["min_queue_size"] self.min_processing_time_seconds = params["min_processing_time_seconds"] self.verbose = params["verbose"] try: with open(os.path.join(cur_path, "model", "current_gnn_model.pkl"), "rb") as f: mdict = torch.load(f) except Exception as e: self.log_warn("Couldn't load GPU model due to " + str(e) + ", loading on CPU") with open(os.path.join(cur_path, "model", "current_gnn_model.pkl"), "rb") as f: mdict = torch.load(f, map_location=torch.device('cpu')) self.model = LoopClosureGNN(64, 7, 4, 1) self.model.load_state_dict(mdict) self.model.eval() self.pg_raw = None self.current_pose_graph_nodes = None self.current_pose_graph_edges = None self.current_pose_graph_edge_atrrs = None self.solution_lock = threading.RLock() self.queue_update_lock = threading.RLock() self.edge_subset_algorithm = edge_selection_algorithm_dispatch(params["algorithm"], self.solution_lock) self.edge_subset_algorithm.set_min_processing_seconds(self.min_processing_time_seconds) self.edge_subset_algorithm.set_verbose(self.verbose) self.edge_subset_algorithm.set_params(params) self.keys_to_loop_candidates = dict() self.thread_pool = ThreadPoolExecutor(max_workers=2) self.current_edge_future = None self.previous_batch = None self.current_pose_graph = None def log_debug(self, s): if self.verbose >= 3: rospy.loginfo("Batcher Node: " + s) def log_info(self, s): if self.verbose >= 2: rospy.loginfo("Batcher Node: " + s) def log_warn(self, s): if self.verbose >= 2: rospy.logwarn("Batcher Node: " + s) def log_error(self, s): if self.verbose >= 1: rospy.logerr("Batcher Node: " + s) def create_publishers(self): self.loop_publisher = rospy.Publisher(rospy.resolve_name("~prioritized_loop_candidates"), LoopCandidateArray, queue_size=100) def register_callbacks(self): """ Handle setting up ROS callbacks and services """ rospy.Subscriber(rospy.resolve_name("~loop_candidates"), LoopCandidateArray, self.add_loop_closure_to_queue) rospy.Subscriber(rospy.resolve_name("~pose_graph"), PoseGraph, self.handle_pose_graph_message) rospy.Subscriber(rospy.resolve_name("~loop_computation_status"), LoopComputationStatus, self.handle_status) rospy.Subscriber(rospy.resolve_name("~output_loop_closures"), LoopCandidateArray, self.remove_loop_closures_from_queue) def make_key(self, loop_candidate): return (loop_candidate.key_to, loop_candidate.key_from, loop_candidate.type) def remove_loop_closures_from_queue(self, loop_candidate_array): with self.queue_update_lock: size_before = len(self.keys_to_loop_candidates) for loop_candidate in loop_candidate_array.candidates: key= self.make_key(loop_candidate) try: del self.keys_to_loop_candidates[key] except KeyError: pass size_after = len(self.keys_to_loop_candidates) if size_before - size_after != 0: self.log_debug("Batcher Removed " + str(size_before - size_after) + " elements.") def add_loop_closure_to_queue(self, loop_candidate_array): with self.queue_update_lock: for loop_candidate in loop_candidate_array.candidates: key = self.make_key(loop_candidate) self.keys_to_loop_candidates[key] = loop_candidate self.log_info("Batcher received :" + str(len(loop_candidate_array.candidates)) + " Queue Size: " + str( len(self.keys_to_loop_candidates))) if self.solution_lock.acquire(): have_enough_candidates = len(self.keys_to_loop_candidates) > self.min_queue_size # If we have enough loop closures then start processing if self.current_edge_future is None and have_enough_candidates: self.start_finding_new_batch() self.current_edge_future = self.thread_pool.submit(self.edge_subset_algorithm.run) # If there were previously not enough edges to do selection but now there are, restart processing elif not self.edge_subset_algorithm.is_processing and len(self.keys_to_loop_candidates) > self.min_queue_size: self.start_finding_new_batch() elif have_enough_candidates and self.edge_subset_algorithm.stale_solution: self.start_finding_new_batch() self.solution_lock.release() def handle_pose_graph_message(self, pose_graph): """ update internal pose graph representation when a new pose graph is published :return: """ if self.current_pose_graph is None or pose_graph.header.stamp >= self.current_pose_graph.header.stamp: self.current_pose_graph = pose_graph def start_finding_new_batch(self): working_loop_candidates = deepcopy(self.keys_to_loop_candidates.values()) if (self.loop_closures_to_choose < 1): number_of_loop_closures_to_choose = max(1, int( float(len(working_loop_candidates)) * self.loop_closures_to_choose)) else: number_of_loop_closures_to_choose = min(len(working_loop_candidates), self.loop_closures_to_choose) if len(working_loop_candidates) < self.min_queue_size: self.edge_subset_algorithm.stop_processing() else: self.log_debug("Going to find " + str(number_of_loop_closures_to_choose) + ", Queue Size: " + str(len(working_loop_candidates)) + ", # to choose " + str(self.loop_closures_to_choose)) self.edge_subset_algorithm.initialize(self.model, number_of_loop_closures_to_choose, self.current_pose_graph, working_loop_candidates) def handle_status(self, status): """ Handle the loop closure computation status """ if (status.type == status.COMPLETED_ALL): with self.solution_lock: if self.edge_subset_algorithm.stale_solution: #self.log_debug("Stale Solution") return None outbound_edge_indexes = self.edge_subset_algorithm.get_current_solution() if outbound_edge_indexes is None: return None else: self.previous_batch = outbound_edge_indexes try: self.edge_subset_algorithm.stop_processing() out_edges = [] out_bundle = LoopCandidateArray() out_bundle.originator = 1 for outbound_edge_index in outbound_edge_indexes: try: edge = self.edge_subset_algorithm.loop_candidates[outbound_edge_index] out_edges.append(edge) except IndexError: self.log_warn("Outbound edge index is not in loop candidates") out_bundle.candidates = out_edges for edge in out_edges: key = self.make_key(edge) try: del self.keys_to_loop_candidates[key] except KeyError: pass self.log_info("Out " + str(len(out_edges)) + ", Queue Size: " + str(len(self.keys_to_loop_candidates))) self.start_finding_new_batch() self.loop_publisher.publish(out_bundle) except Exception as e: self.log_error("Failed to make batch {}".format(traceback.format_exc())) else: self.log_error("Unkown status {}".format(status))
9,763
44.840376
196
py
LAMP
LAMP-main/loop_closure/src/subset_algorithms/gnn.py
from __future__ import division from copy import deepcopy import math import time from subset_algorithms.util import constuct_pytorch_geometric_graph, run_model_on_list_of_graphs, choices, fast_deepcopy import rospy import numpy as np from numpy.random import binomial import random from subset_algorithms.heuristic import MaximallySeperatedNodes, MaximumCovarianceNodes, Random from subset_algorithms.base import GNNBasedLoopClosureBatchSelector class BinomialCEM(GNNBasedLoopClosureBatchSelector): def __init__(self, r_lock, min_processing_seconds=30): super(BinomialCEM, self).__init__(r_lock, min_processing_seconds) self.initialized = False def make_histogram_string(self, probabilities, max_bins=10): hist, bin_edges = np.histogram(probabilities, bins=min(probabilities.shape[0], max_bins)) s = "" for value, left_edge, right_edge in zip(hist, bin_edges, bin_edges[1:]): s += "%.2f : %d " % (((right_edge + left_edge) / 2.0 * 100), value) return s def set_params(self, parameters): self.samples = parameters["cem"]["samples"] self.alpha = parameters["cem"]["alpha"] self.ptile = parameters["cem"]["ptile"] def initialize_method(self): self.probabilities = np.ones(len(self.possible_new_edges)) * .5 self.p_quantile_error = float("inf") self.running_iterations = 0 self.initialized = True def step(self): # Sample graphs cur_samples = [] for _ in range(self.samples): cur_samples.append(binomial(1, self.probabilities)) # Make sure samples have exactly num_to_select entries which are 1 for i in range(self.samples): flip_direction = 0 if np.sum(cur_samples[i]) < self.number_of_loop_closures_to_select: flip_direction = 1 num_samples_to_make_1 = self.number_of_loop_closures_to_select - np.sum(cur_samples[i]) idxs_which_are_zero = [idx for idx in range(cur_samples[i].shape[0]) if cur_samples[i][idx] == 0] probs_for_idxs = [self.probabilities[idx] for idx in idxs_which_are_zero] idxs_to_flip_to_one = np.random.choice(idxs_which_are_zero, num_samples_to_make_1, replace=False, p=probs_for_idxs / np.sum(probs_for_idxs)) for idx in idxs_to_flip_to_one: cur_samples[i][idx] = 1 if np.sum(cur_samples[i]) > self.number_of_loop_closures_to_select: flip_direction = -1 num_samples_to_make_0 = np.sum(cur_samples[i]) - self.number_of_loop_closures_to_select idxs_which_are_one = [idx for idx in range(cur_samples[i].shape[0]) if cur_samples[i][idx] == 1] probs_for_idxs = [1 - self.probabilities[idx] for idx in idxs_which_are_one] idxs_to_flip_to_zero = np.random.choice(idxs_which_are_one, num_samples_to_make_0, replace=False, p=probs_for_idxs / np.sum(probs_for_idxs)) for idx in idxs_to_flip_to_zero: cur_samples[i][idx] = 0 assert np.sum( cur_samples[i]) == self.number_of_loop_closures_to_select, "Sum: %d Should be %d Flip Direction %d" % (np.sum(cur_samples[i]), self.number_of_loop_closures_to_select, flip_direction) # Compute fitness graphs_to_run = [] for cur_sample in cur_samples: edges = [] edge_attrs = [] included = 0 for edge_idx,include in enumerate(cur_sample): if include == 1: try: edges.append(self.possible_new_edges[edge_idx]) edge_attrs.append(self.possible_edge_attrs[edge_idx]) included += 1 except IndexError: self.log_debug("IndexError") edges = np.concatenate((self.current_pose_graph_edges,np.array(edges))) edge_attrs = np.concatenate((self.current_pose_graph_edge_atrrs,np.array(edge_attrs))) assert included == self.number_of_loop_closures_to_select, "Included {}, Required {}".format(included,self.number_of_loop_closures_to_select) graphs_to_run.append( constuct_pytorch_geometric_graph(self.current_pose_graph_nodes, edges, edge_attrs)) errors = run_model_on_list_of_graphs(self.gnn_model, graphs_to_run) new_probabilities = np.zeros(len(self.possible_new_edges)) self.p_quantile_error = min(np.percentile(errors, self.ptile), self.p_quantile_error) number_of_included_samples = 0 for fitness, sample_idx in zip(errors, list(range(self.samples))): if fitness <= self.p_quantile_error: new_probabilities += np.array(cur_samples[sample_idx]) number_of_included_samples += 1 if number_of_included_samples != 0: new_probabilities /= number_of_included_samples new_probabilities = (1 - self.alpha) * self.probabilities + self.alpha * new_probabilities self.probabilities = new_probabilities hist_string = self.make_histogram_string(self.probabilities) self.log_debug( "CEM: {} bound {} effectivity {}, Errors N({:e},{:e}) Range: {:e} - {}".format( self.running_iterations,self.p_quantile_error, number_of_included_samples / self.samples,np.mean(errors),np.std(errors, dtype=np.float64), np.max(errors)-np.min(errors),hist_string)) self.running_iterations += 1 def currently_computed_solution_(self): if not self.initialized and self.running_iterations == 0: return None else: # Sample final output (ML Sample) best_edges = sorted(zip(self.probabilities, list(range(len(self.possible_new_edges)))), reverse=True)[ :self.number_of_loop_closures_to_select] out_edges = [] for p, new_edge_idx in best_edges: out_edges.append(new_edge_idx) return out_edges class SimulatedAnnealing(GNNBasedLoopClosureBatchSelector): def __init__(self, r_lock, min_processing_seconds=30): super(SimulatedAnnealing, self).__init__(r_lock, min_processing_seconds) # This fallback is done because convex hulling can take a long time for a lot of closures self.initial_solution_finder = MaximallySeperatedNodes(r_lock, use_hull=True) self.fallback_initial_solution_finder = MaximallySeperatedNodes(r_lock, use_hull=False) self.max_number_of_loop_closures_to_use_initial_solution = 200 self.initial_solution_finder_iterations = 30 self.fallback_solution_finder_iterations = 100 self.steps = 0 self.min_temp = 1e-4 def set_params(self, params): self.initial_solution_finder_iterations = params["simulated_annealing"]["initial_solution_finder_iterations"] self.fallback_solution_finder_iterations = params["simulated_annealing"]["fallback_solution_finder_iterations"] self.min_temp = params["simulated_annealing"]["min_temp"] self.Tmax = params["simulated_annealing"]["max_temp"] self.cooling = params["simulated_annealing"]["cooling"] self.max_number_of_loop_closures_to_use_initial_solution = params["simulated_annealing"]["max_number_of_loop_closures_to_use_initial_solution"] def initialize_method(self): self.initial_solution_finder.initialize(self.gnn_model, self.number_of_loop_closures_to_select, self.pose_graph, self.loop_candidates) self.fallback_initial_solution_finder.initialize(self.gnn_model, self.number_of_loop_closures_to_select, self.pose_graph, self.loop_candidates) self.found_initial_solution = False self.steps = 0 def set_verbose(self, verbose): super(SimulatedAnnealing, self).set_verbose(verbose) self.initial_solution_finder.set_verbose(verbose) self.fallback_initial_solution_finder.set_verbose(verbose) def find_initial_solution(self): self.log_debug("Finding initial solution") if len(self.loop_candidates) <= self.max_number_of_loop_closures_to_use_initial_solution: for i in range(self.initial_solution_finder_iterations): self.initial_solution_finder.step() starting_solution = self.initial_solution_finder.currently_computed_solution_() method = "Base" else: for i in range(self.fallback_solution_finder_iterations): self.fallback_initial_solution_finder.step() starting_solution = self.fallback_initial_solution_finder.currently_computed_solution_() method = "Fallback" if starting_solution is None: self.log_warn("Couldn't find initial solution, falling back") starting_solution = np.random.randint(0, len(self.loop_candidates)-1, size=self.number_of_loop_closures_to_select).tolist() method = "Random" self.log_debug("%s : Starting solution length: %d, Supposed to be %d" % (method, len(starting_solution),self.number_of_loop_closures_to_select)) self.all_points = [] keys_to_nodes = dict() for node in self.pose_graph.nodes: keys_to_nodes[node.key] = node for idx, edge in enumerate(self.loop_candidates): to_node = keys_to_nodes[edge.key_to] from_node = keys_to_nodes[edge.key_from] average_x = (to_node.pose.position.x + from_node.pose.position.x) / 2 average_y = (to_node.pose.position.y + from_node.pose.position.y) / 2 average_z = (to_node.pose.position.z + from_node.pose.position.z) / 2 self.all_points.append(np.array([average_x, average_y, average_z])) self.all_points = np.array(self.all_points) # self.distances = np.zeros((len(self.all_points), len(self.all_points))) # for i in range(len(self.all_points)): # for j in range(len(self.all_points)): # if i != j: # self.distances[i, j] = 1 / (np.linalg.norm(self.all_points[i] - self.all_points[j]) + 1e-4) self.log_debug("Found initial solution") self.state = starting_solution self.T = self.Tmax self.E = self.energy() self.prevState = fast_deepcopy(self.state) self.prevEnergy = self.E self.best_state = fast_deepcopy(self.state) self.best_energy = self.E self.edge_idxs_not_in_solution = list(set(range(len(self.possible_new_edges))).difference(self.state)) self.steps = 1 self.accepts = 0 self.improves = 0 def step(self): if not self.found_initial_solution: self.find_initial_solution() self.take_annealing_step() self.found_initial_solution = True def take_annealing_step(self): self.steps += 1 T = max(self.Tmax * ((1 - self.cooling) ** self.steps), self.min_temp) self.move() self.E = self.energy() dE = self.E - self.prevEnergy if dE > 0.0 and math.exp(-dE / T) < random.random(): # Restore previous state self.state = fast_deepcopy(self.prevState) self.E = self.prevEnergy else: # Accept new state and compare to best state self.accepts += 1 if dE < 0.0: self.improves += 1 self.prevState = fast_deepcopy(self.state) self.prevEnergy = self.E if self.E < self.best_energy: self.best_state = fast_deepcopy(self.state) self.best_energy = self.E if (self.steps % 100) == 0: self.log_progress( self.steps, T, self.E, self.accepts / 100.0, self.improves / 100.0) self.accepts, self.improves = 0, 0 def currently_computed_solution_(self): if self.found_initial_solution and self.steps >= 1: T = max(self.Tmax * ((1 - self.cooling) ** self.steps), self.min_temp) self.log_debug("Final Step: {}, E: {}, T:{}".format(self.steps,self.best_energy,T)) return self.best_state else: return None def move(self, pow=1): swapped_out_idx = random.randint(0, len(self.state) - 1) old_edge = self.state[swapped_out_idx] # Inverse distance weighted if len(self.edge_idxs_not_in_solution) < 500: jump_idxs = self.edge_idxs_not_in_solution else: jump_idxs = random.sample(self.edge_idxs_not_in_solution,500) distance_weights = 1/ (np.linalg.norm(self.all_points[jump_idxs,:] - self.all_points[old_edge,:],ord=2,axis=1) + 1e-7) new_edge = choices(jump_idxs, distance_weights, k=1)[0] # Perform the swap self.state.pop(swapped_out_idx) self.edge_idxs_not_in_solution.remove(new_edge) self.state.append(new_edge) self.edge_idxs_not_in_solution.append(old_edge) #assert len(self.state) == self.num_to_select, f"State size didn't change, M {len(self.state) }, # choose {self.num_to_select}" return (new_edge, old_edge) def energy(self): edges = [] edge_attrs = [] for edge_idx in self.state: try: edges.append(self.possible_new_edges[edge_idx]) edge_attrs.append(self.possible_edge_attrs[edge_idx]) except IndexError: self.log_debug("IndexError") edges = np.concatenate((self.current_pose_graph_edges,np.array(edges))) edge_attrs = np.concatenate((self.current_pose_graph_edge_atrrs,np.array(edge_attrs))) return run_model_on_list_of_graphs(self.gnn_model, [constuct_pytorch_geometric_graph(self.current_pose_graph_nodes, edges, edge_attrs)])[0] def log_progress(self, step, T, E, acceptance, improvement): self.log_debug( 'Temp: {Temp:12.5f} Energy: {Energy:12.2f} Accept: {Accept:7.2%} Improve: {Improve:7.2%} S: {Steps:f} ' .format(Temp=T, Energy=E, Accept=acceptance, Improve=improvement, Steps=step, )) class DiscretePSO(GNNBasedLoopClosureBatchSelector): """ Discrete Particle Swarm Optimization from "A New Discrete Particle Swarm Optimization Algorithm" by Strasser, Sheppard, and Butcher """ def set_params(self, params): self.number_of_particles = params["pso"]["number_of_particles"] self.max_velocity = params["pso"]["max_velocity"] self.scaling_factor = params["pso"]["scaling_factor"] self.interia = params["pso"]["inertia"] self.cognitive = params["pso"]["cognitive"] self.social = params["pso"]["social"] self.topology = params["pso"]["topology"] assert self.topology in ["global","ring"] def initialize_method(self): self.initialize_particles_and_velocities_and_bests() self.i = 0 def step(self): #Calculate new velocities and particles for p,particle in enumerate(self.particles): cognitive_mix = random.uniform(0,self.cognitive) social_mix = random.uniform(0,self.social) if self.topology == "global": topological_best = self.global_best_position elif self.topology == "ring": next_idx = (p + 1) % self.number_of_particles prev_idx = (p - 1) % self.number_of_particles if self.particle_best_fitnesses[next_idx] < self.particle_best_fitnesses[prev_idx]: topological_best = self.particles[next_idx] else: topological_best = self.particles[prev_idx] else: raise Exception() self.velocities[p] = np.clip(self.velocities[p] * self.interia + cognitive_mix * (self.particle_best_position[p] - particle) + social_mix * (topological_best - particle),-self.max_velocity,self.max_velocity) self.particles[p] = self.particles[p] + self.velocities[p] self.particles = self.normalize_particles(self.particles) #get samples samples = self.sample(self.particles) #calculate the fitnesses fitnesses = self.evaluate_fitness(samples) #Find new global best (if it exists) current_best = np.argmin(fitnesses) if fitnesses[current_best] < self.global_best_fitness: new_global_best_found = True self.global_best_position = self.particles[current_best] self.global_best_fitness = fitnesses[current_best] self.global_best_sample = samples[current_best] else: new_global_best_found = False personal_bests_found = 0 #Find new personal bests (if they exist) for i, particle in enumerate(self.particles): if fitnesses[i] < self.particle_best_fitnesses[i]: self.particle_best_position[i] = self.calculate_new_best(self.particles[i], samples[i], scaling_factor=self.scaling_factor) personal_bests_found += 1 self.i += 1 self.log_debug("I: {} Global Best: {} New Global {} New Personals {}/{}, Fitnesses N({},{}), Velocity Mag: {}".format(self.i,self.global_best_fitness,new_global_best_found,personal_bests_found,self.number_of_particles,np.mean(fitnesses),np.std(fitnesses),np.linalg.norm(self.velocities))) def currently_computed_solution_(self): if self.global_best_sample is not None: return self.global_best_sample else: return None def initialize_particles_and_velocities_and_bests(self): self.global_best_sample = None #Particles dimensions are particles x output edge dimension x edge probability #for example, self.particles[0,1,2] is the probability of edge 2 being the first output edge for particle 0 #Velocities are the same dimensions self.particles = np.random.rand(self.number_of_particles, self.number_of_loop_closures_to_select,len(self.possible_new_edges)) self.velocities = np.clip(np.random.normal(0,scale=1.0,size=self.particles.shape),a_min=-self.max_velocity,a_max=self.max_velocity) self.particles = self.normalize_particles(self.particles) self.particle_best_position = self.particles.copy() initial_samples = self.sample(self.particle_best_position) self.particle_best_fitnesses = self.evaluate_fitness(initial_samples) best_particle_idx = np.argmin(self.particle_best_fitnesses) self.global_best_position = self.particle_best_position[best_particle_idx].copy() self.global_best_fitness = self.particle_best_fitnesses[best_particle_idx] self.global_best_sample = initial_samples[best_particle_idx] def normalize_particles(self, particles): for i,particle in enumerate(particles): for j in range(particle.shape[0]): #TODO may not need to clip particles[i,j] = np.clip(particles[i,j],0,1) particles[i,j] = particles[i,j] / np.sum(particles[i,j]) #assert np.sum(particles[i,j]) == 1.0, "Particles should be normalized to 1, they are {}".format(np.sum(particles[i,j])) return particles def sample(self, particles): sampled_solutions = [] for i, particle in enumerate(particles): solution = [] for j in range(particle.shape[0]): chosen_edge = choices(range(particle.shape[1]),weights=particles[i,j],k=1)[0] solution.append(chosen_edge) assert len(solution) == self.number_of_loop_closures_to_select sampled_solutions.append(solution) return np.array(sampled_solutions) def evaluate_fitness(self, samples): assert len(samples.shape) == 2, "Samples should be 2d but it's {}, samples: {}".format(samples.shape,samples) graphs_to_run = [] for cur_sample in samples: edges = [] edge_attrs = [] for edge_idx in cur_sample: try: edges.append(self.possible_new_edges[edge_idx]) edge_attrs.append(self.possible_edge_attrs[edge_idx]) except IndexError: self.log_debug("IndexError") edges = np.concatenate((self.current_pose_graph_edges,np.array(edges))) edge_attrs = np.concatenate((self.current_pose_graph_edge_atrrs,np.array(edge_attrs))) graphs_to_run.append( constuct_pytorch_geometric_graph(self.current_pose_graph_nodes, edges, edge_attrs)) return run_model_on_list_of_graphs(self.gnn_model, graphs_to_run) def calculate_new_best(self,particle, sample, scaling_factor): out_best = np.zeros(shape=particle.shape) for d, dimension in enumerate(particle): dimension_sum = np.sum(dimension) for j,probability in enumerate(dimension): if sample[d] == j: out_best[d,j] = scaling_factor * probability else: out_best[d,j] = probability + (1-scaling_factor) * (dimension_sum - probability) assert 0 <= out_best[d,j] < 1 return out_best class DiscreteTabuSearch(SimulatedAnnealing): def set_params(self, params): self.initial_solution_finder_iterations = params["tabu"]["initial_solution_finder_iterations"] self.fallback_solution_finder_iterations = params["tabu"]["fallback_solution_finder_iterations"] self.max_number_of_loop_closures_to_use_initial_solution = params["tabu"]["max_number_of_loop_closures_to_use_initial_solution"] self.num_neighbors = params["tabu"]["num_neighbors"] self.max_tabu_length = params["tabu"]["max_tabu_length"] def initialize_method(self): super(DiscreteTabuSearch, self).initialize_method() self.tabu_list = dict() #Not used self.Tmax = 0 self.find_initial_solution() self.found_initial_solution = True def undo_move(self, move): new_edge,old_edge = move self.state.remove(new_edge) self.state.append(old_edge) self.edge_idxs_not_in_solution.append(new_edge) self.edge_idxs_not_in_solution.remove(old_edge) def reapply_move(self,move): new_edge,old_edge = move self.state.remove(old_edge) self.state.append(new_edge) self.edge_idxs_not_in_solution.append(old_edge) self.edge_idxs_not_in_solution.remove(new_edge) def step(self): best_candidate_move = None best_candidate_fitness = float("inf") tabu_rejections = 0 for i in range(self.num_neighbors): cur_move = self.move(pow=3) cur_energy = self.energy() if (cur_move not in self.tabu_list): if (cur_energy < best_candidate_fitness): best_candidate_fitness = cur_energy best_candidate_move = cur_move else: tabu_rejections += 1 self.undo_move(cur_move) if best_candidate_move is None: self.log_warn("Tabu Search Couldn't Make Move") return None self.reapply_move(best_candidate_move) found_new_best = False if best_candidate_fitness <= self.best_energy: self.best_state = deepcopy(self.state) self.best_energy = best_candidate_fitness found_new_best = True self.tabu_list[best_candidate_move] = 0 for move, count in self.tabu_list.items(): self.tabu_list[move] += 1 if self.tabu_list[move] > self.max_tabu_length: del self.tabu_list[move] self.steps += 1 self.log_debug("S: {} Best: {} Bound: {} Rejected: {}".format(self.steps,found_new_best,self.best_energy,tabu_rejections))
24,494
45.043233
296
py
LAMP
LAMP-main/loop_closure/src/subset_algorithms/util.py
import random import torch import numpy as np from torch_geometric.data import Data, DataLoader import rospy import operator as op from functools import reduce from itertools import repeat as _repeat from bisect import bisect as _bisect import operator import json def constuct_pytorch_geometric_graph(nodes, edges, edge_attrs=None): nodes = torch.from_numpy(np.array(nodes, dtype=np.float32)) edges = torch.from_numpy(np.array(edges, dtype=np.int64).T) #if edge_attrs is not None: edge_attrs = torch.from_numpy(np.array(edge_attrs, dtype=np.float32)) return Data(x=nodes, edge_index=edges, edge_attr=edge_attrs) # else: # return Data(x=nodes, edge_index=edges) def run_model_on_list_of_graphs(model, graphs): all_batches = [] for graph in graphs: batch = torch.zeros(graph.x.shape[0], dtype=int) all_batches.append(torch.sum(model(graph.x, graph.edge_index, graph.edge_attr, batch)).cpu().detach().numpy()) return np.array(all_batches) def ncr(n, r): r = min(r, n - r) numer = reduce(op.mul, range(n, n - r, -1), 1) denom = reduce(op.mul, range(1, r + 1), 1) return numer // denom # or / in Python 2 #Backport from python 3.7 itertools.accumulate def accumulate(iterable, func=operator.add, initial=None): 'Return running totals' # accumulate([1,2,3,4,5]) --> 1 3 6 10 15 # accumulate([1,2,3,4,5], initial=100) --> 100 101 103 106 110 115 # accumulate([1,2,3,4,5], operator.mul) --> 1 2 6 24 120 it = iter(iterable) total = initial if initial is None: try: total = next(it) except StopIteration: return yield total for element in it: total = func(total, element) yield total #Backport from python 3.7 random.choices def choices(population, weights=None, cum_weights=None, k=1): """Return a k sized list of population elements chosen with replacement. If the relative weights or cumulative weights are not specified, the selections are made with equal probability. """ n = len(population) if cum_weights is None: if weights is None: _int = int n += 0.0 # convert to float for a small speed improvement return [population[_int(random.random() * n)] for i in _repeat(None, k)] cum_weights = list(accumulate(weights)) elif weights is not None: raise TypeError('Cannot specify both weights and cumulative weights') if len(cum_weights) != n: raise ValueError('The number of weights does not match the population') bisect = _bisect total = cum_weights[-1] + 0.0 # convert to float hi = n - 1 return [population[bisect(cum_weights, random.random() * total, 0, hi)] for i in _repeat(None, k)] def fast_deepcopy(a): return json.loads(json.dumps(a))
2,853
33.804878
118
py
LAMP
LAMP-main/loop_closure/src/gnn/gnn_model.py
import torch from torch.nn import Linear, ModuleList, BatchNorm1d import torch.nn.functional as F from torch_geometric.nn import GCNConv, GraphConv, GATConv, SAGEConv, GMMConv, NNConv from torch_geometric.nn import global_mean_pool, global_add_pool, GlobalAttention from torch.nn import ModuleList, Embedding,Sequential, ReLU from torch.nn import Sequential as Seq, Linear as Lin, ReLU class LoopClosureGNN(torch.nn.Module): def __init__(self, hidden_channels, num_input_features, num_edge_attr, num_output_features, ends_in_sigmoid=False): super(LoopClosureGNN, self).__init__() #torch.manual_seed(12345) self.node_encoder = Linear(num_input_features, hidden_channels) self.edge_encoder = Linear(num_edge_attr, hidden_channels) self.conv1 = GMMConv(hidden_channels, hidden_channels, 4,5) self.gcs = ModuleList([GMMConv(hidden_channels,hidden_channels,4, 5) for i in range(2)]) #self.conv1 = SAGEConv(num_input_features, hidden_channels) #self.gcs = ModuleList([SAGEConv(hidden_channels, hidden_channels) for i in range(4)]) num_linear = 3 self.lins = ModuleList([Linear(hidden_channels, hidden_channels) for i in range(num_linear)]) self.batch_norms = ModuleList([BatchNorm1d(hidden_channels) for i in range(num_linear + 2)]) self.lin = Linear(hidden_channels, num_output_features) self.ends_in_sigmoid = ends_in_sigmoid def forward(self, x, edge_index,edge_attr, batch): # 1. Obtain node embeddings #x = self.input_norm(x) x = self.node_encoder(x) #edge_attr = self.edge_encoder(edge_attr) x = self.conv1(x, edge_index,edge_attr) for gc in self.gcs: x = x.tanh() x = gc(x, edge_index,edge_attr) # 2. Readout layer x = global_add_pool(x, batch) # [batch_size, hidden_channels] # 3. Apply a final classifier for lin,batch_norm in zip(self.lins,self.batch_norms[1:]): #x = F.dropout(x, p=0.5, training=self.training) x = batch_norm(x) x = lin(x) x = torch.tanh(x) #x = F.dropout(x, p=0.5, training=self.training) x = self.batch_norms[-1](x) x = self.lin(x) if self.ends_in_sigmoid: x = torch.sigmoid(x) return x class LoopClosureGNNGlobalAttention(torch.nn.Module): def __init__(self, hidden_channels, num_input_features, num_output_features, ends_in_sigmoid=False): super(LoopClosureGNNGlobalAttention, self).__init__() #torch.manual_seed(12345) #self.input_norm = BatchNorm1d(num_input_features) self.conv1 = GENConv(num_input_features, hidden_channels,msg_norm=True,learn_t=True,learn_p=True,learn_msg_scale=True) #self.gcs = ModuleList([GraphConv(hidden_channels, hidden_channels) for i in range(4)]) self.gcs = ModuleList([GENConv(hidden_channels,hidden_channels,msg_norm=True,learn_t=True,learn_p=True,learn_msg_scale=True) for i in range(2)]) #self.conv1 = SAGEConv(num_input_features, hidden_channels) #self.gcs = ModuleList([SAGEConv(hidden_channels, hidden_channels) for i in range(4)]) num_linear = 3 self.lins = ModuleList([Linear(hidden_channels, hidden_channels) for i in range(num_linear)]) self.batch_norms = ModuleList([BatchNorm1d(hidden_channels) for i in range(num_linear + 2)]) self.lin = Linear(hidden_channels, num_output_features) self.ends_in_sigmoid = ends_in_sigmoid gate_nn = Seq(Lin(hidden_channels, hidden_channels), ReLU(), Lin(hidden_channels, 1)) self.glob = GlobalAttention(gate_nn) def forward(self, x, edge_index, batch): # 1. Obtain node embeddings #x = self.input_norm(x) x = self.conv1(x, edge_index) for gc in self.gcs: x = x.tanh() x = gc(x, edge_index) # 2. Readout layer #x = global_add_pool(x, batch) # [batch_size, hidden_channels] x = self.glob(x,batch) # 3. Apply a final classifier for lin,batch_norm in zip(self.lins,self.batch_norms[1:]): #x = F.dropout(x, p=0.5, training=self.training) x = batch_norm(x) x = lin(x) x = torch.tanh(x) #x = F.dropout(x, p=0.5, training=self.training) x = self.batch_norms[-1](x) x = self.lin(x) if self.ends_in_sigmoid: x = torch.sigmoid(x) return x class LoopClosureUNet(torch.nn.Module): def __init__(self,hidden_channels,num_features,num_output_features,unet_depth=3): super(LoopClosureUNet, self).__init__() #pool_ratios = [2000 / data.num_nodes, 0.5] self.unet = GraphUNet(num_features, hidden_channels, hidden_channels, depth=unet_depth) num_linear = 3 self.lins = ModuleList([Linear(hidden_channels, hidden_channels) for i in range(num_linear)]) self.batch_norms = ModuleList([BatchNorm1d(hidden_channels) for i in range(num_linear + 1)]) self.lin = Linear(hidden_channels, num_output_features) def forward(self,x,edge_index,batch): # edge_index, _ = dropout_adj(data.edge_index, p=0.2, # force_undirected=True, # num_nodes=data.num_nodes, # training=self.training) x = F.dropout(x, p=0.92, training=self.training) x = self.unet(x, edge_index) # 2. Readout layer x = global_add_pool(x, batch) # [batch_size, hidden_channels] # 3. Apply a final classifier for lin,batch_norm in zip(self.lins,self.batch_norms): #x = F.dropout(x, p=0.5, training=self.training) x = batch_norm(x) x = lin(x) x = torch.tanh(x) #x = F.dropout(x, p=0.5, training=self.training) x = self.batch_norms[-1](x) x = self.lin(x) return x class LoopClosurePNA(torch.nn.Module): def __init__(self,node_features,output_features,deg): super(LoopClosurePNA, self).__init__() hidden_dim = 32 self.node_emb = Linear(node_features, hidden_dim) aggregators = ['mean', 'min', 'max', 'std'] scalers = ['identity', 'amplification', 'attenuation'] self.convs = [] self.batch_norms = [] for _ in range(2): conv = PNAConv(in_channels=hidden_dim, out_channels=hidden_dim, aggregators=aggregators, scalers=scalers, deg=deg, towers=4, pre_layers=1, post_layers=1, divide_input=False) self.convs.append(conv) self.batch_norms.append(BatchNorm1d(hidden_dim)) self.convs = ModuleList(self.convs) self.batch_norms = ModuleList(self.convs) # self.mlp = Sequential(Linear(hidden_dim, int(hidden_dim/2)), ReLU(),Linear(int(hidden_dim/2),int(hidden_dim/4)), ReLU(), # Linear(int(hidden_dim/4), output_features)) self.mlp = Sequential(Linear(hidden_dim, int(hidden_dim/2)), ReLU(), Linear(int(hidden_dim/2), output_features)) def forward(self, x, edge_index, batch): x = self.node_emb(x) #edge_attr = self.edge_emb(edge_attr) for conv, batch_norm in zip(self.convs, self.batch_norms): #x = F.relu(batch_norm(conv(x, edge_index, edge_attr))) x = F.relu(batch_norm(conv(x, edge_index),edge_index)) x = global_add_pool(x, batch) return self.mlp(x) class DeeperGCN(torch.nn.Module): def __init__(self, hidden_channels, input_features,output_features,num_layers=2): super(DeeperGCN, self).__init__() self.node_encoder = Linear(input_features, hidden_channels) self.layers = torch.nn.ModuleList() for i in range(1, num_layers + 1): conv = GENConv(hidden_channels, hidden_channels, aggr='softmax', t=1.0, learn_t=True, num_layers=2, norm='layer') norm = LayerNorm(hidden_channels) act = ReLU(inplace=True) layer = DeepGCNLayer(conv, norm, act, block='res+', dropout=0.1, ckpt_grad=i % 3) self.layers.append(layer) self.mlp = Sequential(Linear(hidden_channels, int(hidden_channels/2)), ReLU(), Linear(int(hidden_channels/2), output_features)) def forward(self, x, edge_index,batch ): x = self.node_encoder(x) #edge_attr = self.edge_encoder(edge_attr) x = self.layers[0].conv(x, edge_index) for layer in self.layers[1:]: x = layer(x, edge_index) x = self.layers[0].act(self.layers[0].norm(x)) x = F.dropout(x, p=0.1, training=self.training) x = global_add_pool(x, batch) return self.mlp(x)
9,031
43.27451
152
py
LAMP
LAMP-main/loop_closure/script/dataset.py
import itertools import os import numpy as np import torch from torch_geometric.data import InMemoryDataset, Data from tqdm import tqdm import random class LoopClosureDataset(InMemoryDataset): def __init__(self, root, transform=None, pre_transform=None, delete_bad_files=True): super(LoopClosureDataset, self).__init__(root, transform, pre_transform) self.data, self.slices = torch.load(self.processed_paths[0]) @property def raw_file_names(self): return [] @property def processed_file_names(self): return ['closures.dataset'] def download(self): pass def process(self): node_types = set() bad_files = [] for run in os.listdir(self.root): if not run in ["raw", "processed"]: for file in os.listdir(os.path.join(self.root, run, "post")): if "node" in file: with open(os.path.join(self.root, run, "post", file)) as f: for line in f.readlines()[1:]: try: node_id, node_type, x, y, z, qx, qy, qz, qw, cx, cy, cz, cRX, cRY, cRZ, newline = line.split( ",") node_types.add(node_type) except Exception as e: print(e) print(os.path.join(self.root, run, file)) data_list = [] num_distinct_features = 6 num_node_features = (len(node_types) + num_distinct_features) node_types = sorted(list(node_types)) g_nodes = [] g_edges = [] g_targets = [] for run in tqdm(os.listdir(self.root), "Loading Graphs"): if run in ["raw", "processed"]: continue nodes_pre = dict() edges_pre = dict() nodes_post = dict() edges_post = dict() for file in os.listdir(os.path.join(self.root, run, "pre")): epoch = file.split("-")[0] if "node" in file: nodes_pre[epoch] = file else: edges_pre[epoch] = file for file in os.listdir(os.path.join(self.root, run, "post")): epoch = file.split("-")[0] if "node" in file: nodes_post[epoch] = file else: edges_post[epoch] = file for graph_epoch, predicted_epoch in tqdm(list(zip(sorted(nodes_pre.keys()), sorted(nodes_post.keys()))), desc="Loading graphs from run %s" % (run)): cx_after_solve = 0 cy_after_solve = 0 cz_after_solve = 0 with open(os.path.join(self.root, run, "post", nodes_post[predicted_epoch])) as f: for line in f.readlines(): if "odom_node" in line: # TODO this should be the sum of all the errors node_id, node_type, x, y, z, qx, qy, qz, qw, cx, cy, cz, cRX, cRY, cRZ, newline = line.split( ",") cx = float(cx) cy = float(cy) cz = float(cz) cx_after_solve += cx cy_after_solve += cy cz_after_solve += cz target_node = node_id # assert all(map(lambda x: not math.isnan(x), target)) # target = [cx_after_solve, cy_after_solve, # cz_after_solve] target = [cx_after_solve + cy_after_solve + cz_after_solve] g_targets.append(target) cur_nodes = [] node_ids_to_seq_id = dict() with open(os.path.join(self.root, run, "pre", nodes_pre[graph_epoch])) as f: for i, line in enumerate(f.readlines()[1:]): # try: node_id, node_type, x, y, z, qx, qy, qz, qw, cx, cy, cz, cRX, cRY, cRZ, newline = line.split( ",") if "an" in node_type: continue # Number of node types + covariance # node_features = np.zeros((len(node_types) + 6)) node_features = np.zeros((len(node_types) + num_distinct_features)) # node_features = np.zeros((3)) # node_features[6 + node_types.index(node_type)] = 1 node_features[num_distinct_features + node_types.index(node_type)] = 1 # cx = (float(cx)-empirical_node_mean[0])/empirical_node_std[0] # cy = (float(cy)-empirical_node_mean[1])/empirical_node_std[1] # cz = (float(cz)-empirical_node_mean[2])/empirical_node_std[2] cx = float(cx) cy = float(cy) cz = float(cz) node_features[0] = float(cx) node_features[1] = float(cy) node_features[2] = float(cz) node_features[3] = float(x) node_features[4] = float(y) node_features[5] = float(z) cur_nodes.append(node_features) node_ids_to_seq_id[node_id] = i # except Exception as e: # print(f"Node Couldn't parse line {line} due to {e}") # pass g_nodes += cur_nodes cur_edges = [] edge_attrs = [] with open(os.path.join(self.root, run, "post", edges_post[predicted_epoch])) as f: try: for line in f.readlines()[1:]: from_key, to_key, edge_type = line.split(",") edge_1 = node_ids_to_seq_id[from_key] edge_2 = node_ids_to_seq_id[to_key] cur_edges.append((edge_1, edge_2)) edge_attr = np.zeros(4) edge_attr[int(edge_type) - 1] = 1 edge_attrs.append(edge_attr) except: print(run) g_edges.append(cur_edges) cur_nodes = torch.from_numpy(np.array(cur_nodes, dtype=np.float32)) cur_edges = torch.from_numpy(np.array(cur_edges, dtype=np.int64).T) target = torch.from_numpy(np.array([target], dtype=np.float32)) edge_attrs = torch.from_numpy(np.array(edge_attrs, dtype=np.float32)) try: data = Data(x=cur_nodes, edge_index=cur_edges, y=target, edge_attr=edge_attrs) #data.run_name = run data.debug() data_list.append(data) except RuntimeError as e: #print("Skipping dataset {run},{graph_epoch} due to {e}") print("Skipping %s" % (run)) random.shuffle(data_list) data, slices = self.collate(data_list) torch.save((data, slices), self.processed_paths[0]) class LoopClosureNodeLabelDataset(InMemoryDataset): def __init__(self, root, transform=None, pre_transform=None, delete_bad_files=True): super(LoopClosureNodeLabelDataset, self).__init__(root, transform, pre_transform) self.data, self.slices = torch.load(self.processed_paths[0]) @property def raw_file_names(self): return [] @property def processed_file_names(self): return ['closures_nodelevel.dataset'] def download(self): pass def process(self): node_types = set() bad_files = [] for run in os.listdir(self.root): if not run in ["raw", "processed"]: for file in os.listdir(os.path.join(self.root, run, "post")): if "node" in file: with open(os.path.join(self.root, run, "post", file)) as f: for line in f.readlines()[1:]: try: node_id, node_type, x, y, z, qx, qy, qz, qw, cx, cy, cz, cRX, cRY, cRZ, newline = line.split( ",") node_types.add(node_type) except Exception as e: print(e) print(os.path.join(self.root, run, file)) data_list = [] num_distinct_features = 6 num_node_features = (len(node_types) + num_distinct_features) node_types = sorted(list(node_types)) g_nodes = [] g_edges = [] g_targets = [] for run in tqdm(os.listdir(self.root), "Loading Graphs"): if run in ["raw", "processed"]: continue nodes_pre = dict() edges_pre = dict() nodes_post = dict() edges_post = dict() for file in os.listdir(os.path.join(self.root, run, "pre")): epoch = file.split("-")[0] if "node" in file: nodes_pre[epoch] = file else: edges_pre[epoch] = file for file in os.listdir(os.path.join(self.root, run, "post")): epoch = file.split("-")[0] if "node" in file: nodes_post[epoch] = file else: edges_post[epoch] = file for graph_epoch, predicted_epoch in tqdm(list(zip(sorted(nodes_pre.keys()), sorted(nodes_post.keys()))), desc="Loading graphs from run %s" % (run)): target = [] with open(os.path.join(self.root, run, "post", nodes_post[predicted_epoch])) as f: for line in f.readlines(): if "odom_node" in line: # TODO this should be the sum of all the errors node_id, node_type, x, y, z, qx, qy, qz, qw, cx, cy, cz, cRX, cRY, cRZ, newline = line.split( ",") cx = float(cx) cy = float(cy) cz = float(cz) target.append([cx,cy,cz]) #target = [cx_after_solve + cy_after_solve + cz_after_solve] g_targets.append(target) cur_nodes = [] node_ids_to_seq_id = dict() with open(os.path.join(self.root, run, "pre", nodes_pre[graph_epoch])) as f: for i, line in enumerate(f.readlines()[1:]): # try: node_id, node_type, x, y, z, qx, qy, qz, qw, cx, cy, cz, cRX, cRY, cRZ, newline = line.split( ",") if "an" in node_type: continue # Number of node types + covariance # node_features = np.zeros((len(node_types) + 6)) node_features = np.zeros((len(node_types) + num_distinct_features)) # node_features = np.zeros((3)) # node_features[6 + node_types.index(node_type)] = 1 node_features[num_distinct_features + node_types.index(node_type)] = 1 # cx = (float(cx)-empirical_node_mean[0])/empirical_node_std[0] # cy = (float(cy)-empirical_node_mean[1])/empirical_node_std[1] # cz = (float(cz)-empirical_node_mean[2])/empirical_node_std[2] cx = float(cx) cy = float(cy) cz = float(cz) node_features[0] = float(cx) node_features[1] = float(cy) node_features[2] = float(cz) node_features[3] = float(x) node_features[4] = float(y) node_features[5] = float(z) cur_nodes.append(node_features) node_ids_to_seq_id[node_id] = i # except Exception as e: # print(f"Node Couldn't parse line {line} due to {e}") # pass g_nodes += cur_nodes cur_edges = [] edge_attrs = [] with open(os.path.join(self.root, run, "post", edges_post[predicted_epoch])) as f: try: for line in f.readlines()[1:]: from_key, to_key, edge_type = line.split(",") edge_1 = node_ids_to_seq_id[from_key] edge_2 = node_ids_to_seq_id[to_key] cur_edges.append((edge_1, edge_2)) edge_attr = np.zeros(4) edge_attr[int(edge_type) - 1] = 1 edge_attrs.append(edge_attr) except: print(run) g_edges.append(cur_edges) cur_nodes = torch.from_numpy(np.array(cur_nodes, dtype=np.float32)) cur_edges = torch.from_numpy(np.array(cur_edges, dtype=np.int64).T) target = torch.from_numpy(np.array(target, dtype=np.float32)) edge_attrs = torch.from_numpy(np.array(edge_attrs, dtype=np.float32)) try: data = Data(x=cur_nodes, edge_index=cur_edges, y=target, edge_attr=edge_attrs) data.run_name = run data.debug() data_list.append(data) except RuntimeError as e: #print(f"Skipping dataset {run},{graph_epoch} due to {e}") print("Skipping dataset %s" % (run)) random.shuffle(data_list) data, slices = self.collate(data_list) torch.save((data, slices), self.processed_paths[0])
14,544
44.311526
129
py
LAMP
LAMP-main/loop_closure/script/convert_rosbags_to_gnn_dataset.py
#!/usr/bin/env python # This script converts bagfiles stored at fname to readable datasets for pytorch geometric in out_location # Need pose_graph_opt.bag import csv import rosbag import os import yaml import tqdm fname = "/media/chris/hdd3/more_bags" out_location = "/media/chris/hdd3/more_training_data" def dump_edges(edges, filename): with open(filename, 'w') as csvfile: writer = csv.writer(csvfile) row = ["key_from", "key_to", "type"] writer.writerow(row) for edge in edges: row = [edge.key_from, edge.key_to, edge.type] writer.writerow(row) def dump_nodes(nodes, filename): with open(filename, 'w') as csvfile: writer = csv.writer(csvfile) row = ['key', 'id', 'pose_x', 'pose_y', 'pose_z', 'pose_q_x', 'pose_q_y', 'pose_q_z', 'pose_q_w', 'cov_x', 'cov_y', 'cov_z', 'cov_rX', 'cov_rY', 'cov_rZ'] writer.writerow(row) for node in nodes: row = [node.key, node.ID, node.pose.position.x, node.pose.position.y, node.pose.position.z, node.pose.orientation.x, node.pose.orientation.y, node.pose.orientation.z, node.pose.orientation.w, node.covariance[0], node.covariance[7], node.covariance[14], node.covariance[21], node.covariance[28], node.covariance[35],"aaaaaaaa"] writer.writerow(row) def makedirs(path): cur_path = "" directories = path.split("/") for directory in directories[1:]: cur_path = os.path.join(cur_path,directory) print(cur_path) try: os.makedirs(cur_path) except OSError as e: print(e) def closest_time_before(dict,t): best_time_diff = None best_msg = None for time, msg in dict.items(): if best_time_diff is None or (time < t and t < best_time_diff): best_time_diff = t best_msg = msg return best_msg for run in tqdm.tqdm(os.listdir(fname), "Bags"): bags = list(os.listdir(os.path.join(fname,run))) if len(bags) == 1: split_bags = False elif len(bags) == 2: split_bags = True else: print(run) raise Exception() # Make file locations pre_files_location = os.path.join(out_location, run,"pre") post_files_location = os.path.join(out_location, run,"post") if os.path.exists(pre_files_location): print("Skipping: " + fname) continue try: os.removedirs(pre_files_location) except OSError: pass os.makedirs(pre_files_location) try: os.removedirs(post_files_location) except: pass os.makedirs(post_files_location) if not split_bags: last_unoptimized_message = None optimized_topic = "/base1/lamp_pgo/optimized_values" unoptimized_topic = "/base1/lamp/pose_graph_to_optimize" bag_location = os.path.join(fname,run,bags[0]) try: pgo_info_dict = yaml.load(rosbag.Bag(bag_location, 'r')._get_yaml_info()) except rosbag.bag.ROSBagUnindexedException: continue if not pgo_info_dict['indexed']: continue print pgo_info_dict for topic, msg, t in tqdm.tqdm(rosbag.Bag(bag_location).read_messages(topics=[unoptimized_topic, optimized_topic]), desc="Messages",total=pgo_info_dict["messages"]): if topic == unoptimized_topic: last_unoptimized_message = msg if topic == optimized_topic: # Dump pre message dump_edges(last_unoptimized_message.edges, os.path.join(pre_files_location, str(t) + "-edges.csv")) dump_nodes(last_unoptimized_message.nodes, os.path.join(pre_files_location, str(t) + "-nodes.csv")) # Dump post messages dump_edges(msg.edges, os.path.join(post_files_location, str(t) + "-edges.csv")) dump_nodes(msg.nodes, os.path.join(post_files_location, str(t) + "-nodes.csv")) else: pose_graph_bag = os.path.join(fname,run,"pose_graph.bag") pose_graph_opt_bag = os.path.join(fname,run,"pose_graph_opt.bag") pg_info_dict = yaml.load(rosbag.Bag(pose_graph_bag, 'r')._get_yaml_info()) print pg_info_dict pgo_info_dict = yaml.load(rosbag.Bag(pose_graph_opt_bag, 'r')._get_yaml_info()) print pgo_info_dict last_time = None for topic,msg,t in tqdm.tqdm(rosbag.Bag(pose_graph_opt_bag).read_messages(),desc="Reading Messages", total=pgo_info_dict["messages"]): pose_graph_msg = None for topic_pg,msg_pg,t_pg in rosbag.Bag(pose_graph_bag).read_messages(start_time=last_time): if t_pg <= t: pose_graph_msg = msg_pg last_time = t_pg else: break dump_edges(pose_graph_msg.edges, os.path.join(pre_files_location, str(t) + "-edges.csv")) dump_nodes(pose_graph_msg.nodes, os.path.join(pre_files_location, str(t) + "-nodes.csv")) # Dump post messages dump_edges(msg.edges, os.path.join(post_files_location, str(t) + "-edges.csv")) dump_nodes(msg.nodes, os.path.join(post_files_location, str(t) + "-nodes.csv"))
5,335
36.577465
173
py
LAMP
LAMP-main/loop_closure/script/gnn_model.py
import torch from torch.nn import Linear, ModuleList, BatchNorm1d import torch.nn.functional as F from torch_geometric.nn import GCNConv, GraphConv, GATConv, SAGEConv, GMMConv from torch_geometric.nn import global_mean_pool, global_add_pool, GlobalAttention from torch.nn import ModuleList, Embedding,Sequential, ReLU from torch.nn import Sequential as Seq, Linear as Lin, ReLU class LoopClosureGNN(torch.nn.Module): def __init__(self, hidden_channels, num_input_features, num_edge_attr, num_output_features, ends_in_sigmoid=False): super(LoopClosureGNN, self).__init__() #torch.manual_seed(12345) self.node_encoder = Linear(num_input_features, hidden_channels) self.edge_encoder = Linear(num_edge_attr, hidden_channels) #self.input_norm = BatchNorm1d(num_input_features) #self.conv1 = SAGEConv(hidden_channels, hidden_channels) #self.conv1 = GENConv(hidden_channels, hidden_channels,msg_norm=True,learn_t=True,learn_p=True,learn_msg_scale=True) self.conv1 = GMMConv(hidden_channels, hidden_channels, 4,5) #self.gcs = ModuleList([SAGEConv(hidden_channels, hidden_channels) for i in range(4)]) #self.gcs = ModuleList([GENConv(hidden_channels,hidden_channels,msg_norm=True,learn_t=True,learn_p=True,learn_msg_scale=True) for i in range(2)]) self.gcs = ModuleList([GMMConv(hidden_channels,hidden_channels,4, 5) for i in range(2)]) #self.conv1 = SAGEConv(num_input_features, hidden_channels) #self.gcs = ModuleList([SAGEConv(hidden_channels, hidden_channels) for i in range(4)]) num_linear = 3 self.lins = ModuleList([Linear(hidden_channels, hidden_channels) for i in range(num_linear)]) self.batch_norms = ModuleList([BatchNorm1d(hidden_channels) for i in range(num_linear + 2)]) self.lin = Linear(hidden_channels, num_output_features) self.ends_in_sigmoid = ends_in_sigmoid def forward(self, x, edge_index,edge_attr, batch): # 1. Obtain node embeddings #x = self.input_norm(x) x = self.node_encoder(x) #edge_attr = self.edge_encoder(edge_attr) x = self.conv1(x, edge_index,edge_attr) for gc in self.gcs: x = x.tanh() x = gc(x, edge_index,edge_attr) # 2. Readout layer x = global_add_pool(x, batch) # [batch_size, hidden_channels] # 3. Apply a final classifier for lin,batch_norm in zip(self.lins,self.batch_norms[1:]): #x = F.dropout(x, p=0.5, training=self.training) x = batch_norm(x) x = lin(x) x = torch.tanh(x) #x = F.dropout(x, p=0.5, training=self.training) x = self.batch_norms[-1](x) x = self.lin(x) if self.ends_in_sigmoid: x = torch.sigmoid(x) return x class LoopClosureGNNGlobalAttention(torch.nn.Module): def __init__(self, hidden_channels, num_input_features, num_output_features, ends_in_sigmoid=False): super(LoopClosureGNNGlobalAttention, self).__init__() #torch.manual_seed(12345) #self.input_norm = BatchNorm1d(num_input_features) self.conv1 = GENConv(num_input_features, hidden_channels,msg_norm=True,learn_t=True,learn_p=True,learn_msg_scale=True) #self.gcs = ModuleList([GraphConv(hidden_channels, hidden_channels) for i in range(4)]) self.gcs = ModuleList([GENConv(hidden_channels,hidden_channels,msg_norm=True,learn_t=True,learn_p=True,learn_msg_scale=True) for i in range(2)]) #self.conv1 = SAGEConv(num_input_features, hidden_channels) #self.gcs = ModuleList([SAGEConv(hidden_channels, hidden_channels) for i in range(4)]) num_linear = 3 self.lins = ModuleList([Linear(hidden_channels, hidden_channels) for i in range(num_linear)]) self.batch_norms = ModuleList([BatchNorm1d(hidden_channels) for i in range(num_linear + 2)]) self.lin = Linear(hidden_channels, num_output_features) self.ends_in_sigmoid = ends_in_sigmoid gate_nn = Seq(Lin(hidden_channels, hidden_channels), ReLU(), Lin(hidden_channels, 1)) self.glob = GlobalAttention(gate_nn) def forward(self, x, edge_index, batch): # 1. Obtain node embeddings #x = self.input_norm(x) x = self.conv1(x, edge_index) for gc in self.gcs: x = x.tanh() x = gc(x, edge_index) # 2. Readout layer #x = global_add_pool(x, batch) # [batch_size, hidden_channels] x = self.glob(x,batch) # 3. Apply a final classifier for lin,batch_norm in zip(self.lins,self.batch_norms[1:]): #x = F.dropout(x, p=0.5, training=self.training) x = batch_norm(x) x = lin(x) x = torch.tanh(x) #x = F.dropout(x, p=0.5, training=self.training) x = self.batch_norms[-1](x) x = self.lin(x) if self.ends_in_sigmoid: x = torch.sigmoid(x) return x class LoopClosureUNet(torch.nn.Module): def __init__(self,hidden_channels,num_features,num_output_features,unet_depth=3): super(LoopClosureUNet, self).__init__() #pool_ratios = [2000 / data.num_nodes, 0.5] self.unet = GraphUNet(num_features, hidden_channels, hidden_channels, depth=unet_depth) num_linear = 3 self.lins = ModuleList([Linear(hidden_channels, hidden_channels) for i in range(num_linear)]) self.batch_norms = ModuleList([BatchNorm1d(hidden_channels) for i in range(num_linear + 1)]) self.lin = Linear(hidden_channels, num_output_features) def forward(self,x,edge_index,batch): # edge_index, _ = dropout_adj(data.edge_index, p=0.2, # force_undirected=True, # num_nodes=data.num_nodes, # training=self.training) x = F.dropout(x, p=0.92, training=self.training) x = self.unet(x, edge_index) # 2. Readout layer x = global_add_pool(x, batch) # [batch_size, hidden_channels] # 3. Apply a final classifier for lin,batch_norm in zip(self.lins,self.batch_norms): #x = F.dropout(x, p=0.5, training=self.training) x = batch_norm(x) x = lin(x) x = torch.tanh(x) #x = F.dropout(x, p=0.5, training=self.training) x = self.batch_norms[-1](x) x = self.lin(x) return x class LoopClosurePNA(torch.nn.Module): def __init__(self,node_features,output_features,deg): super(LoopClosurePNA, self).__init__() hidden_dim = 32 self.node_emb = Linear(node_features, hidden_dim) aggregators = ['mean', 'min', 'max', 'std'] scalers = ['identity', 'amplification', 'attenuation'] self.convs = [] self.batch_norms = [] for _ in range(2): conv = PNAConv(in_channels=hidden_dim, out_channels=hidden_dim, aggregators=aggregators, scalers=scalers, deg=deg, towers=4, pre_layers=1, post_layers=1, divide_input=False) self.convs.append(conv) self.batch_norms.append(BatchNorm1d(hidden_dim)) self.convs = ModuleList(self.convs) self.batch_norms = ModuleList(self.convs) # self.mlp = Sequential(Linear(hidden_dim, int(hidden_dim/2)), ReLU(),Linear(int(hidden_dim/2),int(hidden_dim/4)), ReLU(), # Linear(int(hidden_dim/4), output_features)) self.mlp = Sequential(Linear(hidden_dim, int(hidden_dim/2)), ReLU(), Linear(int(hidden_dim/2), output_features)) def forward(self, x, edge_index, batch): x = self.node_emb(x) #edge_attr = self.edge_emb(edge_attr) for conv, batch_norm in zip(self.convs, self.batch_norms): #x = F.relu(batch_norm(conv(x, edge_index, edge_attr))) x = F.relu(batch_norm(conv(x, edge_index),edge_index)) x = global_add_pool(x, batch) return self.mlp(x) class DeeperGCN(torch.nn.Module): def __init__(self, hidden_channels, input_features,output_features,num_layers=2): super(DeeperGCN, self).__init__() self.node_encoder = Linear(input_features, hidden_channels) self.layers = torch.nn.ModuleList() for i in range(1, num_layers + 1): conv = GENConv(hidden_channels, hidden_channels, aggr='softmax', t=1.0, learn_t=True, num_layers=2, norm='layer') norm = LayerNorm(hidden_channels) act = ReLU(inplace=True) layer = DeepGCNLayer(conv, norm, act, block='res+', dropout=0.1, ckpt_grad=i % 3) self.layers.append(layer) self.mlp = Sequential(Linear(hidden_channels, int(hidden_channels/2)), ReLU(), Linear(int(hidden_channels/2), output_features)) def forward(self, x, edge_index,batch ): x = self.node_encoder(x) #edge_attr = self.edge_encoder(edge_attr) x = self.layers[0].conv(x, edge_index) for layer in self.layers[1:]: x = layer(x, edge_index) x = self.layers[0].act(self.layers[0].norm(x)) x = F.dropout(x, p=0.1, training=self.training) x = global_add_pool(x, batch) return self.mlp(x)
9,517
44.54067
153
py
LAMP
LAMP-main/loop_closure/script/offline_refit.py
#!/usr/bin/env python # This script fits the gnn model used in batching to new training data from __future__ import division import pickle import os import math import torch from torch_geometric.datasets import TUDataset from torch_geometric.data import InMemoryDataset, Data, DataLoader import numpy as np import torch from torch_geometric.transforms import Compose, LocalDegreeProfile, OneHotDegree, Cartesian, Distance from tqdm import tqdm import matplotlib.pyplot as plt from dataset import LoopClosureDataset, LoopClosureNodeLabelDataset from gnn_model import LoopClosureGNN #from torch.utils.tensorboard import SummaryWriter from torch_geometric.utils import degree from sklearn.metrics import mean_squared_error root = "/home/chris/current_training_data/" save_name = "./model.pkl" cpu_save_name = "./model-cpu.pkl" device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') # device = "cpu" dataset = LoopClosureDataset("/media/chris/hdd3/more_training_data") val_holdout = int(len(dataset) / 4) train_dataset = dataset train_dataset = dataset[:-val_holdout] validation_dataset = dataset[-val_holdout:] train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True) validation_loader = DataLoader(validation_dataset, batch_size=32, shuffle=True) print("Dataset Size: %d, Train Size %d, Val Size %d" % (len(dataset), len(train_dataset), len(validation_dataset))) def train(model): model.train() for data in train_loader: # Iterate in batches over the training dataset. data.to(device) out = model(data.x, data.edge_index,data.edge_attr, data.batch) # Perform a single forward pass. loss = criterion(out, data.y) # Compute the loss. loss.backward() # Derive gradients. optimizer.step() # Update parameters based on gradients. optimizer.zero_grad() # Clear gradients. def test(loader, model): model.eval() predictions, actuals = list(), list() for i, data in enumerate(loader): data.to(device) # evaluate the model on the test set yhat = model(data.x, data.edge_index,data.edge_attr,data.batch) # retrieve numpy array yhat = yhat.cpu().detach().numpy() actual = data.y.cpu().detach().numpy() # actual = actual.reshape((len(actual), 3)) # store predictions.append(yhat) actuals.append(actual) predictions, actuals = np.vstack(predictions), np.vstack(actuals) # calculate mse mse = mean_squared_error(actuals, predictions, multioutput='raw_values') return np.sqrt(mse) if __name__ == "__main__": model = LoopClosureGNN(64, dataset.num_node_features,dataset.num_edge_features,dataset.num_classes).to(device) continue_training = True should_train = False val_skip = 1 try: if continue_training: model.load_state_dict(torch.load(save_name)) except Exception as e: print(e) if should_train: optimizer = torch.optim.Adam(model.parameters(), lr=0.0005, weight_decay=0.005) criterion = torch.nn.MSELoss() pbar = tqdm(range(1, 5 * (10 ** 2)), "Training Epochs") #writer = SummaryWriter() best_test_accuracy = float("inf") saving = False for epoch in pbar: train(model) train_acc = test(train_loader, model) test_acc = test(validation_loader, model) pbar.set_postfix({"Training Accuracy": train_acc, "Testing Accuracy": test_acc, "Saving": saving, "Best Test Accuracy": best_test_accuracy}) # for x in range(test_acc.shape[0]): # writer.add_scalar("Accuracy/train" + str(x), train_acc[x], epoch) # writer.add_scalar("Accuracy/test" + str(x), test_acc[x], epoch) if sum(test_acc) < best_test_accuracy: saving = True best_test_accuracy = sum(test_acc) with open(save_name, "wb") as f: torch.save(model.state_dict(), f) else: saving = False model.load_state_dict(torch.load(save_name)) model.eval() correct = 0 total = 0 differences = [] predictions = [] for i, graph_1 in tqdm(list(enumerate(validation_dataset)), desc="Compute Ordering"): # Iterate in batches over the training for graph_2 in validation_dataset[i::val_skip]: #if graph_1.run_name == graph_2.run_name: target_sum_1 = torch.sum(graph_1.y) target_sum_2 = torch.sum(graph_2.y) #graphs that are super close who cares if np.abs((target_sum_1 - target_sum_2).cpu().detach().numpy()) < 0.0001: continue graph_1.to(device) batch = torch.zeros(graph_1.x.shape[0], dtype=int).to(device) pred_sum_1 = torch.sum(model(graph_1.x, graph_1.edge_index, graph_1.edge_attr,batch)) prediction = pred_sum_1.cpu().detach().numpy() graph_2.to(device) batch = torch.zeros(graph_2.x.shape[0], dtype=int).to(device) pred_sum_2 = torch.sum(model(graph_2.x, graph_2.edge_index,graph_2.edge_attr,batch)) diff = (target_sum_1 - target_sum_2) - (pred_sum_1 - pred_sum_2) differences.append(diff.cpu().detach().numpy()) if (target_sum_1 <= target_sum_2 and pred_sum_1 <= pred_sum_2) or ( target_sum_1 > target_sum_2 and pred_sum_1 > pred_sum_2): correct += 1 total += 1 print("Ordering: Correct %f Total %f Accuracy %f" % (correct, total, correct/total)) print("MSE: %f" % (test(validation_loader,model))) plt.figure() plt.hist(np.array(differences),bins=50) plt.title("Difference histogram") plt.figure() plt.hist(np.array(predictions),bins=50) plt.title("Predictions") plt.show()
5,887
36.987097
152
py
CMUDeepLens
CMUDeepLens-master/deeplens/base.py
# This file contains the base class for a Lasagne model from sklearn.base import BaseEstimator, ClassifierMixin from sklearn.utils.validation import check_is_fitted from random import shuffle import time import numpy as np import cPickle as pickle import theano import theano.tensor as T import lasagne from keras.preprocessing.image import ImageDataGenerator from lasagne.utils import floatX from lasagne.layers import InputLayer, NonlinearityLayer, DenseLayer from lasagne.layers import get_output, get_all_params, get_output_shape from lasagne.nonlinearities import sigmoid, rectify, elu, tanh, identity from lasagne.updates import adam, total_norm_constraint from lasagne.objectives import binary_accuracy, binary_crossentropy from .objectives import weighted_sigmoid_binary_crossentropy class BaseLasagneClassifier(BaseEstimator, ClassifierMixin): """ Base class for a Lasagne classifier model Expects implementations of the following functions and attributes: self._build(n_x,n_y): Builds the network structure self._fit(X, y): Fit the model """ def __init__(self, pos_weight=1, n_epochs=100, batch_size=32, learning_rate=0.001, learning_rate_drop=0.1, learning_rate_steps=3, output_nbatch=100, val_nepoch=5): """ Initialisation """ self.batch_size = batch_size self.n_epochs = n_epochs self.learning_rate = learning_rate self.output_nbatch = output_nbatch self.pos_weight = pos_weight self.val_nepoch = val_nepoch self.learning_rate_drop = learning_rate_drop self.learning_rate_steps = learning_rate_steps self._x = T.tensor4('x') self._y = T.matrix('y') self._lr = T.scalar(name='learning_rate') self._network = None def _model_definition(self, net): """ Function which defines the model from the provided source layer to the output of the DenseLayer, before the dense layer ! """ return net def _build(self, X, y): """ Builds the network and associated training functions, for the specific shapes of the inputs """ n_x = X.shape[-1] n_y = y.shape[-1] n_c = X.shape[1] # Defining input layers self.l_x = InputLayer(shape=(self.batch_size, n_c, n_x, n_x), input_var=self._x, name='x') self.l_y = InputLayer(shape=(self.batch_size, n_y), input_var=self._y, name='y') net = self._model_definition(self.l_x) # Output classifier out = DenseLayer(net, num_units=n_y, nonlinearity=identity) self._network = NonlinearityLayer(out, nonlinearity=sigmoid) # Compute network loss q, p = get_output([out, self.l_y], inputs={self.l_x:self._x, self.l_y:self._y}) # Define loss function loss = weighted_sigmoid_binary_crossentropy(q, p, self.pos_weight) # Average over batch loss = loss.mean() # Get trainable parameters and generate updates params = get_all_params([self._network], trainable=True) grads = T.grad(loss, params) updates = adam(grads, params, learning_rate=self._lr) self._trainer = theano.function([self._x, self._y, self._lr], [loss], updates=updates) # Get detection probability from the network qdet = get_output(self._network, inputs={self.l_x: self._x}, deterministic=True) self._output = theano.function([self._x], qdet) def fit(self, X, y, Xval=None, yval=None): """ Fit the model to the data. Parameters ---------- X: array_like of shape (n_samples, n_channels, n_x, n_x) Training data. y: array_like (n_samples, n_conditional_features), optional Conditional data. Returns ------- self : Generator The fitted model """ # Creates a new network and associated functions if not incremental if self._network is None: self._build(X, y) niter = 0 train_err = 0 # Defines a preprocessing step datagen = ImageDataGenerator(rotation_range=90, zoom_range=[0.9,1], horizontal_flip=True, vertical_flip=True, fill_mode='wrap', dim_ordering='th') lr = self.learning_rate # Loop over training epochs for i in range(self.n_epochs): print("Starting Epoch : %d"%i) if (Xval is not None) and (yval is not None) and (i % self.val_nepoch == 0) and (i > 0): pur, comp = self.eval_purity_completeness(Xval, yval) print("Iteration : %d -> [Validation] Purity: %f ; Completeness: %f"%(niter, pur, comp)) nval = Xval.shape[0] pur, comp = self.eval_purity_completeness(X[0:nval], y[0:nval]) print("Iteration : %d -> [Training] Purity: %f ; Completeness: %f"%(niter, pur, comp)) start_time = time.time() batches = datagen.flow(X, y, batch_size=self.batch_size, shuffle=True) # Loop over batches for b in range(X.shape[0] / self.batch_size): xdata, ydata = batches.next() # One iteration of training err, = self._trainer(floatX(xdata), floatX(ydata), floatX(lr)) train_err += err niter += 1 if (niter % self.output_nbatch) == 0: print("Iteration : %d -> Training loss: %f"%(niter, train_err / (self.output_nbatch))) train_err = 0 print("Epoch took %f s"%(time.time() - start_time)) start_time = time.time() # Lower the learning rate if required if i % (self.n_epochs / self.learning_rate_steps) == 0 and i > 0: lr *= self.learning_rate_drop print("Decreasing learning rate to:" + str(lr)) return self def save(self, filename): """ Exports the model parameters to file """ check_is_fitted(self, "_network") all_values = lasagne.layers.get_all_param_values(self._network) params =[self.batch_size, self.n_epochs, self.learning_rate, self.pos_weight] all_params = [params, all_values] f = file(filename, 'wb') print("saving to " + filename + "...") pickle.dump(all_params, f, protocol=pickle.HIGHEST_PROTOCOL) f.close() def load(self, filename, X,y ): """ Load the network parameter from file """ print("loading from " + filename + "...") f = file(filename, 'rb') all_params = pickle.load(f) f.close() p, all_values = all_params # Extracts parameters self.batch_size, self.n_epochs, self.learning_rate, self.pos_weight = p self._build(X,y) # Rebuild the network and set the weights lasagne.layers.set_all_param_values(self._network, all_values) print("Model loaded") def predict_proba(self, X): """ Returns probability estimates for X """ check_is_fitted(self, "_network") res = [] nsamples = X.shape[0] # Process data using batches, for optimisation and memory constraints for i in range(int(nsamples/self.batch_size)): q = self._output(floatX(X[i*self.batch_size:(i+1)*self.batch_size])) res.append(q) if nsamples % (self.batch_size) > 0 : i = int(nsamples/self.batch_size) ni = nsamples % (self.batch_size) xdata = np.zeros((self.batch_size,) + X.shape[1:]) xdata[:ni] = X[i*self.batch_size:] q = self._output(floatX(xdata)) res.append(q[:ni]) # Concatenate processed data q = np.concatenate(res) return q def predict(self, X, threshold=0.5): """ Predict class of X """ check_is_fitted(self, "_network") q = self.predict_proba(X) upper, lower = 1, 0 return np.where(q > threshold, upper, lower) def eval_purity_completeness(self, X, y, threshold=0.5): """ Evaluate the model purity and completeness using the following definitions Purity = N(true positive) / [N(true positive) + N(false positive)] Compl. = N(true positive) / [N(true positive) + N(false negative)] """ check_is_fitted(self, "_network") p = self.predict(X, threshold) n_fp = np.sum(p * (y == 0)).astype('float32') n_tp = np.sum(p * y).astype('float32') n_fn = np.sum((p == 0) * (y == 1)).astype('float32') pur = n_tp / ( n_tp + n_fp ) comp= n_tp / ( n_tp + n_fn ) return pur, comp def eval_tpr_fpr(self, X, y, threshold=0.5): """ Evaluates the performance of the model using the true and false positive rates as defined by the challenge TPR = N(true positive) / [N(true positive) + N(false negative)] FPR = N(false positive) / [N(false positive) + N(true negative)] """ check_is_fitted(self, "_network") p = self.predict(X, threshold) n_fp = np.sum(p * (y == 0)).astype('float32') n_tp = np.sum(p * y).astype('float32') n_tn = np.sum((p == 0) * (y == 0)).astype('float32') n_fn = np.sum((p == 0) * (y == 1)).astype('float32') n_p = np.sum(y).astype('float32') n_f = np.sum(y == 0).astype('float32') tpr = n_tp / n_p fpr = n_fp / n_f print n_fp, n_tp, n_p, n_f return tpr, fpr def eval_ROC(self, X, y): """ Computes the ROC curve of model """ check_is_fitted(self, "_network") q = np.reshape(self.predict_proba(X), (-1, 1)) t = np.linspace(0,1,1000) upper, lower = 1, 0 p = np.where(q > t, upper, lower) n_fp = np.sum(p * (y == 0), axis=0).astype('float32') n_tp = np.sum(p * y, axis=0).astype('float32') tpr = n_tp / np.sum(y).astype('float32') fpr = n_fp / np.sum( y == 0).astype('float32') return tpr, fpr, t
10,704
31.34139
106
py
xcit
xcit-main/main.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ The main training/evaluation loop Modified from: https://github.com/facebookresearch/deit """ import argparse import datetime import numpy as np import time import torch import torch.backends.cudnn as cudnn import json import os from pathlib import Path from timm.data import Mixup from timm.models import create_model from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy from timm.scheduler import create_scheduler from timm.optim import create_optimizer from timm.utils import NativeScaler, get_state_dict, ModelEma from datasets import build_dataset from engine import train_one_epoch, evaluate from losses import DistillationLoss from samplers import RASampler import utils import xcit def get_args_parser(): parser = argparse.ArgumentParser('XCiT training and evaluation script', add_help=False) parser.add_argument('--batch-size', default=64, type=int) parser.add_argument('--epochs', default=400, type=int) # Model parameters parser.add_argument('--model', default='xcit_s_12', type=str, metavar='MODEL', help='Name of model to train') parser.add_argument('--input-size', default=224, type=int, help='images input size') parser.add_argument('--drop', type=float, default=0.0, metavar='PCT', help='Dropout rate (default: 0.)') parser.add_argument('--drop-path', type=float, default=0.1, metavar='PCT', help='Drop path rate (default: 0.1)') parser.add_argument('--model-ema', action='store_true') parser.add_argument('--no-model-ema', action='store_false', dest='model_ema') parser.set_defaults(model_ema=True) parser.add_argument('--model-ema-decay', type=float, default=0.99996, help='') parser.add_argument('--model-ema-force-cpu', action='store_true', default=False, help='') # Optimizer parameters parser.add_argument('--opt', default='adamw', type=str, metavar='OPTIMIZER', help='Optimizer (default: "adamw"') parser.add_argument('--opt-eps', default=1e-8, type=float, metavar='EPSILON', help='Optimizer Epsilon (default: 1e-8)') parser.add_argument('--opt-betas', default=None, type=float, nargs='+', metavar='BETA', help='Optimizer Betas (default: None, use opt default)') parser.add_argument('--clip-grad', type=float, default=None, metavar='NORM', help='Clip gradient norm (default: None, no clipping)') parser.add_argument('--momentum', type=float, default=0.9, metavar='M', help='SGD momentum (default: 0.9)') parser.add_argument('--weight-decay', type=float, default=0.05, help='weight decay (default: 0.05)') # Learning rate schedule parameters parser.add_argument('--sched', default='cosine', type=str, metavar='SCHEDULER', help='LR scheduler (default: "cosine"') parser.add_argument('--lr', type=float, default=5e-4, metavar='LR', help='learning rate (default: 5e-4)') parser.add_argument('--lr-noise', type=float, nargs='+', default=None, metavar='pct, pct', help='learning rate noise on/off epoch percentages') parser.add_argument('--lr-noise-pct', type=float, default=0.67, metavar='PERCENT', help='learning rate noise limit percent (default: 0.67)') parser.add_argument('--lr-noise-std', type=float, default=1.0, metavar='STDDEV', help='learning rate noise std-dev (default: 1.0)') parser.add_argument('--warmup-lr', type=float, default=1e-6, metavar='LR', help='warmup learning rate (default: 1e-6)') parser.add_argument('--min-lr', type=float, default=1e-5, metavar='LR', help='lower lr bound for cyclic schedulers that hit 0 (1e-5)') parser.add_argument('--decay-epochs', type=float, default=30, metavar='N', help='epoch interval to decay LR') parser.add_argument('--warmup-epochs', type=int, default=5, metavar='N', help='epochs to warmup LR, if scheduler supports') parser.add_argument('--cooldown-epochs', type=int, default=10, metavar='N', help='epochs to cooldown LR at min_lr, after cyclic schedule ends') parser.add_argument('--patience-epochs', type=int, default=10, metavar='N', help='patience epochs for Plateau LR scheduler (default: 10') parser.add_argument('--decay-rate', '--dr', type=float, default=0.1, metavar='RATE', help='LR decay rate (default: 0.1)') # Augmentation parameters parser.add_argument('--color-jitter', type=float, default=0.4, metavar='PCT', help='Color jitter factor (default: 0.4)') parser.add_argument('--aa', type=str, default='rand-m9-mstd0.5-inc1', metavar='NAME', help='Use AutoAugment policy. "v0" or "original". " + \ "(default: rand-m9-mstd0.5-inc1)'), parser.add_argument('--smoothing', type=float, default=0.1, help='Label smoothing (default: 0.1)') parser.add_argument('--train-interpolation', type=str, default='bicubic', help='Training interpolation (random, bilinear, bicubic default: "bicubic")') parser.add_argument('--repeated-aug', action='store_true') parser.add_argument('--no-repeated-aug', action='store_false', dest='repeated_aug') parser.set_defaults(repeated_aug=True) # * Random Erase params parser.add_argument('--reprob', type=float, default=0.25, metavar='PCT', help='Random erase prob (default: 0.25)') parser.add_argument('--remode', type=str, default='pixel', help='Random erase mode (default: "pixel")') parser.add_argument('--recount', type=int, default=1, help='Random erase count (default: 1)') parser.add_argument('--resplit', action='store_true', default=False, help='Do not random erase first (clean) augmentation split') # * Mixup params parser.add_argument('--mixup', type=float, default=0.8, help='mixup alpha, mixup enabled if > 0. (default: 0.8)') parser.add_argument('--cutmix', type=float, default=1.0, help='cutmix alpha, cutmix enabled if > 0. (default: 1.0)') parser.add_argument('--cutmix-minmax', type=float, nargs='+', default=None, help='cutmix min/max ratio, overrides alpha and enables cutmix if set (default: None)') parser.add_argument('--mixup-prob', type=float, default=1.0, help='Probability of performing mixup or cutmix when either/both is enabled') parser.add_argument('--mixup-switch-prob', type=float, default=0.5, help='Probability of switching to cutmix when both mixup and cutmix enabled') parser.add_argument('--mixup-mode', type=str, default='batch', help='How to apply mixup/cutmix params. Per "batch", "pair", or "elem"') # Distillation parameters parser.add_argument('--teacher-model', default='regnety_160', type=str, metavar='MODEL', help='Name of teacher model to train (default: "regnety_160"') parser.add_argument('--teacher-path', type=str, default='') parser.add_argument('--distillation-type', default='none', choices=['none', 'soft', 'hard'], type=str, help="") parser.add_argument('--distillation-alpha', default=0.5, type=float, help="") parser.add_argument('--distillation-tau', default=1.0, type=float, help="") # Dataset parameters parser.add_argument('--data-path', default='/datasets01/imagenet_full_size/061417/', type=str, help='dataset path') parser.add_argument('--data-set', default='IMNET', choices=['CIFAR10', 'CIFAR100', 'IMNET', 'INAT', 'INAT19', 'CARS', 'FLOWERS', 'IMNET22k'], type=str, help='Image Net dataset path') parser.add_argument('--inat-category', default='name', choices=['kingdom', 'phylum', 'class', 'order', 'supercategory', 'family', 'genus', 'name'], type=str, help='semantic granularity') parser.add_argument('--output_dir', default='', help='path where to save, empty for no saving') parser.add_argument('--device', default='cuda', help='device to use for training / testing') parser.add_argument('--seed', default=0, type=int) parser.add_argument('--resume', default='', help='resume from checkpoint') parser.add_argument('--start_epoch', default=0, type=int, metavar='N', help='start epoch') parser.add_argument('--eval', action='store_true', help='Perform evaluation only') parser.add_argument('--dist-eval', action='store_true', default=False, help='Enabling distributed evaluation') parser.add_argument('--num_workers', default=10, type=int) parser.add_argument('--pin-mem', action='store_true', help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.') parser.add_argument('--no-pin-mem', action='store_false', dest='pin_mem', help='') parser.set_defaults(pin_mem=True) # distributed training parameters parser.add_argument('--world_size', default=1, type=int, help='number of distributed processes') parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training') parser.add_argument('--test-freq', default=1, type=int, help='Number of epochs between \ validation runs.') parser.add_argument('--full_crop', action='store_true', help='use crop_ratio=1.0 instead of the\ default 0.875 (Used by CaiT).') parser.add_argument("--pretrained", default=None, type=str, help='Path to pre-trained checkpoint') parser.add_argument('--surgery', default=None, type=str, help='Path to checkpoint to copy the \ patch projection from. \ Can improve stability for very \ large models.') return parser def main(args): utils.init_distributed_mode(args) print(args) device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) cudnn.benchmark = True dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) if True: # args.distributed: num_tasks = utils.get_world_size() global_rank = utils.get_rank() if args.repeated_aug: sampler_train = RASampler( dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True ) else: sampler_train = torch.utils.data.DistributedSampler( dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True ) if args.dist_eval: if len(dataset_val) % num_tasks != 0: print('Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. ' 'This will slightly alter validation results as extra duplicate entries are added to achieve ' 'equal num of samples per-process.') sampler_val = torch.utils.data.DistributedSampler( dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False) else: sampler_val = torch.utils.data.SequentialSampler(dataset_val) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) sampler_val = torch.utils.data.SequentialSampler(dataset_val) data_loader_train = torch.utils.data.DataLoader( dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, ) data_loader_val = torch.utils.data.DataLoader( dataset_val, sampler=sampler_val, batch_size=int(1.5 * args.batch_size), num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False ) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup( mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating model: {args.model}") model = create_model( args.model, pretrained=False, num_classes=args.nb_classes, drop_rate=args.drop, drop_path_rate=args.drop_path, drop_block_rate=None ) if args.pretrained: if args.pretrained.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url( args.pretrained, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.pretrained, map_location='cpu') checkpoint_model = checkpoint['model'] state_dict = model.state_dict() for k in ['head.weight', 'head.bias']: if k in checkpoint_model and checkpoint_model[k].shape != state_dict[k].shape: print(f"Removing key {k} from pretrained checkpoint") del checkpoint_model[k] model.load_state_dict(checkpoint_model, strict=True) model.to(device) if args.surgery: checkpoint = torch.load(args.surgery, map_location='cpu') checkpoint_model = checkpoint['model'] patch_embed_weights = {key.replace("patch_embed.", ""): value for key, value in checkpoint['model'].items() if 'patch_embed' in key} model.patch_embed.load_state_dict(patch_embed_weights) for p in model.patch_embed.parameters(): p.requires_grad = False model_ema = None if args.model_ema: # Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper model_ema = ModelEma( model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else '', resume='') model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size() / 512.0 args.lr = linear_scaled_lr optimizer = create_optimizer(args, model_without_ddp) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) criterion = LabelSmoothingCrossEntropy() if args.mixup > 0.: # smoothing is handled with mixup label transform criterion = SoftTargetCrossEntropy() elif args.smoothing: criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: criterion = torch.nn.CrossEntropyLoss() teacher_model = None if args.distillation_type != 'none': assert args.teacher_path, 'need to specify teacher-path when using distillation' print(f"Creating teacher model: {args.teacher_model}") teacher_model = create_model( args.teacher_model, pretrained=False, num_classes=args.nb_classes, global_pool='avg', ) if args.teacher_path.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url( args.teacher_path, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.teacher_path, map_location='cpu') teacher_model.load_state_dict(checkpoint['model']) teacher_model.to(device) teacher_model.eval() # wrap the criterion in our custom DistillationLoss, which # just dispatches to the original criterion if args.distillation_type is 'none' criterion = DistillationLoss( criterion, teacher_model, args.distillation_type, args.distillation_alpha, args.distillation_tau ) output_dir = Path(args.output_dir) if not os.path.exists(output_dir): os.mkdir(output_dir) resume_path = os.path.join(output_dir, 'checkpoint.pth') if args.resume and os.path.exists(resume_path): if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url( args.resume, map_location='cpu', check_hash=True) else: print("Loading from checkpoint ...") checkpoint = torch.load(resume_path, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if args.model_ema: utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) if 'scaler' in checkpoint: loss_scaler.load_state_dict(checkpoint['scaler']) if args.eval: test_stats = evaluate(data_loader_val, model, device) print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%") return print(f"Start training for {args.epochs} epochs") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) train_stats = train_one_epoch( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn, surgery=args.surgery ) lr_scheduler.step(epoch) if args.output_dir: checkpoint_paths = [output_dir / 'checkpoint.pth'] for checkpoint_path in checkpoint_paths: utils.save_on_master({ 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'model_ema': get_state_dict(model_ema), 'scaler': loss_scaler.state_dict(), 'args': args, }, checkpoint_path) if (epoch % args.test_freq == 0) or (epoch == args.epochs - 1): test_stats = evaluate(data_loader_val, model, device) if test_stats["acc1"] >= max_accuracy: utils.save_on_master({ 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'model_ema': get_state_dict(model_ema), 'args': args, }, os.path.join(output_dir, 'best_model.pth')) print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%") max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = {**{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters} if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str)) if __name__ == '__main__': parser = argparse.ArgumentParser('XCiT training and evaluation script', parents=[get_args_parser()]) args = parser.parse_args() if args.output_dir: Path(args.output_dir).mkdir(parents=True, exist_ok=True) main(args)
21,140
46.507865
119
py
xcit
xcit-main/losses.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Implements the knowledge distillation loss Modified from: https://github.com/facebookresearch/deit """ import torch from torch.nn import functional as F import torch.nn as nn class DistillationLoss(torch.nn.Module): """ This module wraps a standard criterion and adds an extra knowledge distillation loss by taking a teacher model prediction and using it as additional supervision. """ def __init__(self, base_criterion: torch.nn.Module, teacher_model: torch.nn.Module, distillation_type: str, alpha: float, tau: float): super().__init__() self.base_criterion = base_criterion self.teacher_model = teacher_model assert distillation_type in ['none', 'soft', 'hard'] self.distillation_type = distillation_type self.alpha = alpha self.tau = tau def forward(self, inputs, outputs, labels): """ Args: inputs: The original inputs that are feed to the teacher model outputs: the outputs of the model to be trained. It is expected to be either a Tensor, or a Tuple[Tensor, Tensor], with the original output in the first position and the distillation predictions as the second output labels: the labels for the base criterion """ outputs_kd = None if not isinstance(outputs, torch.Tensor): # assume that the model outputs a tuple of [outputs, outputs_kd] outputs, outputs_kd = outputs if isinstance(outputs, tuple): base_loss = self.base_criterion(outputs, labels, inputs) else: base_loss = self.base_criterion(outputs, labels) if self.distillation_type == 'none': return base_loss if outputs_kd is None: raise ValueError("When knowledge distillation is enabled, the model is " "expected to return a Tuple[Tensor, Tensor] with the output of the " "class_token and the dist_token") # don't backprop throught the teacher with torch.no_grad(): teacher_outputs = self.teacher_model(inputs) if self.distillation_type == 'soft': T = self.tau # taken from https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100 # with slight modifications distillation_loss = F.kl_div( F.log_softmax(outputs_kd / T, dim=1), F.log_softmax(teacher_outputs / T, dim=1), reduction='sum', log_target=True ) * (T * T) / outputs_kd.numel() elif self.distillation_type == 'hard': distillation_loss = F.cross_entropy(outputs_kd, teacher_outputs.argmax(dim=1)) loss = base_loss * (1 - self.alpha) + distillation_loss * self.alpha return loss
2,977
40.361111
114
py
xcit
xcit-main/engine.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Train and eval functions used in main.py Modified from: https://github.com/facebookresearch/deit """ import math import sys from typing import Iterable, Optional import torch from timm.data import Mixup from timm.utils import accuracy, ModelEma from losses import DistillationLoss import utils def train_one_epoch(model: torch.nn.Module, criterion: DistillationLoss, data_loader: Iterable, optimizer: torch.optim.Optimizer, device: torch.device, epoch: int, loss_scaler, max_norm: float = 0, model_ema: Optional[ModelEma] = None, mixup_fn: Optional[Mixup] = None, set_training_mode=True, surgery=None): model.train(set_training_mode) if surgery: model.module.patch_embed.eval() metric_logger = utils.MetricLogger(delimiter=" ") metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}')) header = 'Epoch: [{}]'.format(epoch) print_freq = 10 for batch in metric_logger.log_every(data_loader, print_freq, header): samples, targets = batch[0], batch[1] samples = samples.to(device, non_blocking=True) targets = targets.to(device, non_blocking=True) if mixup_fn is not None: samples, targets = mixup_fn(samples, targets) with torch.cuda.amp.autocast(): outputs = model(samples) loss = criterion(samples, outputs, targets) loss_value = loss.item() if not math.isfinite(loss_value): print("Loss is {}, stopping training".format(loss_value)) sys.exit(1) optimizer.zero_grad() # this attribute is added by timm on one optimizer (adahessian) is_second_order = hasattr(optimizer, 'is_second_order') and optimizer.is_second_order loss_scaler(loss, optimizer, clip_grad=max_norm, parameters=model.parameters(), create_graph=is_second_order) torch.cuda.synchronize() if model_ema is not None: model_ema.update(model) metric_logger.update(loss=loss_value) metric_logger.update(lr=optimizer.param_groups[0]["lr"]) # gather the stats from all processes metric_logger.synchronize_between_processes() print("Averaged stats:", metric_logger) return {k: meter.global_avg for k, meter in metric_logger.meters.items()} @torch.no_grad() def evaluate(data_loader, model, device): criterion = torch.nn.CrossEntropyLoss() metric_logger = utils.MetricLogger(delimiter=" ") header = 'Test:' # switch to evaluation mode model.eval() for batch in metric_logger.log_every(data_loader, 10, header): images, target = batch[0], batch[1] images = images.to(device, non_blocking=True) target = target.to(device, non_blocking=True) # compute output with torch.cuda.amp.autocast(): output = model(images) loss = criterion(output, target) acc1, acc5 = accuracy(output, target, topk=(1, 5)) batch_size = images.shape[0] metric_logger.update(loss=loss.item()) metric_logger.meters['acc1'].update(acc1.item(), n=batch_size) metric_logger.meters['acc5'].update(acc5.item(), n=batch_size) # gather the stats from all processes metric_logger.synchronize_between_processes() print('* Acc@1 {top1.global_avg:.3f} Acc@5 {top5.global_avg:.3f} loss {losses.global_avg:.3f}' .format(top1=metric_logger.acc1, top5=metric_logger.acc5, losses=metric_logger.loss)) return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
3,705
33.635514
98
py
xcit
xcit-main/hubconf.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. from xcit import * dependencies = ["torch", "torchvision", "timm"]
136
21.833333
47
py
xcit
xcit-main/utils.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Misc functions, including distributed helpers. Mostly copy-paste from torchvision references. """ import io import os import time from collections import defaultdict, deque import datetime import torch import torch.distributed as dist class SmoothedValue(object): """Track a series of values and provide access to smoothed values over a window or the global series average. """ def __init__(self, window_size=20, fmt=None): if fmt is None: fmt = "{median:.4f} ({global_avg:.4f})" self.deque = deque(maxlen=window_size) self.total = 0.0 self.count = 0 self.fmt = fmt def update(self, value, n=1): self.deque.append(value) self.count += n self.total += value * n def synchronize_between_processes(self): """ Warning: does not synchronize the deque! """ if not is_dist_avail_and_initialized(): return t = torch.tensor([self.count, self.total], dtype=torch.float64, device='cuda') dist.barrier() dist.all_reduce(t) t = t.tolist() self.count = int(t[0]) self.total = t[1] @property def median(self): d = torch.tensor(list(self.deque)) return d.median().item() @property def avg(self): d = torch.tensor(list(self.deque), dtype=torch.float32) return d.mean().item() @property def global_avg(self): return self.total / self.count @property def max(self): return max(self.deque) @property def value(self): return self.deque[-1] def __str__(self): return self.fmt.format( median=self.median, avg=self.avg, global_avg=self.global_avg, max=self.max, value=self.value) class MetricLogger(object): def __init__(self, delimiter="\t"): self.meters = defaultdict(SmoothedValue) self.delimiter = delimiter def update(self, **kwargs): for k, v in kwargs.items(): if isinstance(v, torch.Tensor): v = v.item() assert isinstance(v, (float, int)) self.meters[k].update(v) def __getattr__(self, attr): if attr in self.meters: return self.meters[attr] if attr in self.__dict__: return self.__dict__[attr] raise AttributeError("'{}' object has no attribute '{}'".format( type(self).__name__, attr)) def __str__(self): loss_str = [] for name, meter in self.meters.items(): loss_str.append( "{}: {}".format(name, str(meter)) ) return self.delimiter.join(loss_str) def synchronize_between_processes(self): for meter in self.meters.values(): meter.synchronize_between_processes() def add_meter(self, name, meter): self.meters[name] = meter def log_every(self, iterable, print_freq, header=None): i = 0 if not header: header = '' start_time = time.time() end = time.time() iter_time = SmoothedValue(fmt='{avg:.4f}') data_time = SmoothedValue(fmt='{avg:.4f}') space_fmt = ':' + str(len(str(len(iterable)))) + 'd' log_msg = [ header, '[{0' + space_fmt + '}/{1}]', 'eta: {eta}', '{meters}', 'time: {time}', 'data: {data}' ] if torch.cuda.is_available(): log_msg.append('max mem: {memory:.0f}') log_msg = self.delimiter.join(log_msg) MB = 1024.0 * 1024.0 for obj in iterable: data_time.update(time.time() - end) yield obj iter_time.update(time.time() - end) if i % print_freq == 0 or i == len(iterable) - 1: eta_seconds = iter_time.global_avg * (len(iterable) - i) eta_string = str(datetime.timedelta(seconds=int(eta_seconds))) if torch.cuda.is_available(): print(log_msg.format( i, len(iterable), eta=eta_string, meters=str(self), time=str(iter_time), data=str(data_time), memory=torch.cuda.max_memory_allocated() / MB)) else: print(log_msg.format( i, len(iterable), eta=eta_string, meters=str(self), time=str(iter_time), data=str(data_time))) i += 1 end = time.time() total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('{} Total time: {} ({:.4f} s / it)'.format( header, total_time_str, total_time / len(iterable))) def _load_checkpoint_for_ema(model_ema, checkpoint): """ Workaround for ModelEma._load_checkpoint to accept an already-loaded object """ mem_file = io.BytesIO() torch.save(checkpoint, mem_file) mem_file.seek(0) model_ema._load_checkpoint(mem_file) def setup_for_distributed(is_master): """ This function disables printing when not in master process """ import builtins as __builtin__ builtin_print = __builtin__.print def print(*args, **kwargs): force = kwargs.pop('force', False) if is_master or force: builtin_print(*args, **kwargs) __builtin__.print = print def is_dist_avail_and_initialized(): if not dist.is_available(): return False if not dist.is_initialized(): return False return True def get_world_size(): if not is_dist_avail_and_initialized(): return 1 return dist.get_world_size() def get_rank(): if not is_dist_avail_and_initialized(): return 0 return dist.get_rank() def is_main_process(): return get_rank() == 0 def save_on_master(*args, **kwargs): if is_main_process(): torch.save(*args, **kwargs) def _find_free_port(): import socket sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) # Binding to port 0 will cause the OS to find an available port for us sock.bind(("", 0)) port = sock.getsockname()[1] sock.close() # NOTE: there is still a chance the port could be taken by other processes. return port def init_distributed_mode(args): if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ: args.rank = int(os.environ["RANK"]) args.world_size = int(os.environ['WORLD_SIZE']) args.gpu = int(os.environ['LOCAL_RANK']) elif 'SLURM_PROCID' in os.environ: args.rank = int(os.environ['SLURM_PROCID']) args.gpu = args.rank % torch.cuda.device_count() else: print('Not using distributed mode') args.distributed = False return args.distributed = True torch.cuda.set_device(args.gpu) args.dist_backend = 'nccl' print('| distributed init (rank {}): {}'.format( args.rank, args.dist_url), flush=True) torch.distributed.init_process_group(backend=args.dist_backend, init_method=args.dist_url, world_size=args.world_size, rank=args.rank) torch.distributed.barrier() setup_for_distributed(args.rank == 0)
7,417
28.436508
94
py
xcit
xcit-main/xcit.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Implementation of Cross-Covariance Image Transformer (XCiT) Based on timm and DeiT code bases https://github.com/rwightman/pytorch-image-models/tree/master/timm https://github.com/facebookresearch/deit/ """ import math import torch import torch.nn as nn from functools import partial from timm.models.vision_transformer import _cfg, Mlp from timm.models.registry import register_model from timm.models.layers import DropPath, trunc_normal_, to_2tuple class PositionalEncodingFourier(nn.Module): """ Positional encoding relying on a fourier kernel matching the one used in the "Attention is all of Need" paper. The implementation builds on DeTR code https://github.com/facebookresearch/detr/blob/master/models/position_encoding.py """ def __init__(self, hidden_dim=32, dim=768, temperature=10000): super().__init__() self.token_projection = nn.Conv2d(hidden_dim * 2, dim, kernel_size=1) self.scale = 2 * math.pi self.temperature = temperature self.hidden_dim = hidden_dim self.dim = dim def forward(self, B, H, W): mask = torch.zeros(B, H, W).bool().to(self.token_projection.weight.device) not_mask = ~mask y_embed = not_mask.cumsum(1, dtype=torch.float32) x_embed = not_mask.cumsum(2, dtype=torch.float32) eps = 1e-6 y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale dim_t = torch.arange(self.hidden_dim, dtype=torch.float32, device=mask.device) dim_t = self.temperature ** (2 * (dim_t // 2) / self.hidden_dim) pos_x = x_embed[:, :, :, None] / dim_t pos_y = y_embed[:, :, :, None] / dim_t pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3) pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3) pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) pos = self.token_projection(pos) return pos def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return torch.nn.Sequential( nn.Conv2d( in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False ), nn.SyncBatchNorm(out_planes) ) class ConvPatchEmbed(nn.Module): """ Image to Patch Embedding using multiple convolutional layers """ def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): super().__init__() img_size = to_2tuple(img_size) patch_size = to_2tuple(patch_size) num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) self.img_size = img_size self.patch_size = patch_size self.num_patches = num_patches if patch_size[0] == 16: self.proj = torch.nn.Sequential( conv3x3(3, embed_dim // 8, 2), nn.GELU(), conv3x3(embed_dim // 8, embed_dim // 4, 2), nn.GELU(), conv3x3(embed_dim // 4, embed_dim // 2, 2), nn.GELU(), conv3x3(embed_dim // 2, embed_dim, 2), ) elif patch_size[0] == 8: self.proj = torch.nn.Sequential( conv3x3(3, embed_dim // 4, 2), nn.GELU(), conv3x3(embed_dim // 4, embed_dim // 2, 2), nn.GELU(), conv3x3(embed_dim // 2, embed_dim, 2), ) else: raise("For convolutional projection, patch size has to be in [8, 16]") def forward(self, x, padding_size=None): B, C, H, W = x.shape x = self.proj(x) Hp, Wp = x.shape[2], x.shape[3] x = x.flatten(2).transpose(1, 2) return x, (Hp, Wp) class LPI(nn.Module): """ Local Patch Interaction module that allows explicit communication between tokens in 3x3 windows to augment the implicit communcation performed by the block diagonal scatter attention. Implemented using 2 layers of separable 3x3 convolutions with GeLU and BatchNorm2d """ def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0., kernel_size=3): super().__init__() out_features = out_features or in_features padding = kernel_size // 2 self.conv1 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size, padding=padding, groups=out_features) self.act = act_layer() self.bn = nn.SyncBatchNorm(in_features) self.conv2 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size, padding=padding, groups=out_features) def forward(self, x, H, W): B, N, C = x.shape x = x.permute(0, 2, 1).reshape(B, C, H, W) x = self.conv1(x) x = self.act(x) x = self.bn(x) x = self.conv2(x) x = x.reshape(B, C, N).permute(0, 2, 1) return x class ClassAttention(nn.Module): """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239 """ def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) qkv = qkv.permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) qc = q[:, :, 0:1] # CLS token attn_cls = (qc * k).sum(dim=-1) * self.scale attn_cls = attn_cls.softmax(dim=-1) attn_cls = self.attn_drop(attn_cls) cls_tkn = (attn_cls.unsqueeze(2) @ v).transpose(1, 2).reshape(B, 1, C) cls_tkn = self.proj(cls_tkn) x = torch.cat([self.proj_drop(cls_tkn), x[:, 1:]], dim=1) return x class ClassAttentionBlock(nn.Module): """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239 """ def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, eta=None, tokens_norm=False): super().__init__() self.norm1 = norm_layer(dim) self.attn = ClassAttention( dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop ) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) if eta is not None: # LayerScale Initialization (no layerscale when None) self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) else: self.gamma1, self.gamma2 = 1.0, 1.0 # FIXME: A hack for models pre-trained with layernorm over all the tokens not just the CLS self.tokens_norm = tokens_norm def forward(self, x, H, W, mask=None): x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x))) if self.tokens_norm: x = self.norm2(x) else: x[:, 0:1] = self.norm2(x[:, 0:1]) x_res = x cls_token = x[:, 0:1] cls_token = self.gamma2 * self.mlp(cls_token) x = torch.cat([cls_token, x[:, 1:]], dim=1) x = x_res + self.drop_path(x) return x class XCA(nn.Module): """ Cross-Covariance Attention (XCA) operation where the channels are updated using a weighted sum. The weights are obtained from the (softmax normalized) Cross-covariance matrix (Q^T K \\in d_h \\times d_h) """ def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() self.num_heads = num_heads self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1)) self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) qkv = qkv.permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) q = q.transpose(-2, -1) k = k.transpose(-2, -1) v = v.transpose(-2, -1) q = torch.nn.functional.normalize(q, dim=-1) k = torch.nn.functional.normalize(k, dim=-1) attn = (q @ k.transpose(-2, -1)) * self.temperature attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x @torch.jit.ignore def no_weight_decay(self): return {'temperature'} class XCABlock(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, num_tokens=196, eta=None): super().__init__() self.norm1 = norm_layer(dim) self.attn = XCA( dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop ) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) self.norm3 = norm_layer(dim) self.local_mp = LPI(in_features=dim, act_layer=act_layer) self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) self.gamma3 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) def forward(self, x, H, W): x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x))) x = x + self.drop_path(self.gamma3 * self.local_mp(self.norm3(x), H, W)) x = x + self.drop_path(self.gamma2 * self.mlp(self.norm2(x))) return x class XCiT(nn.Module): """ Based on timm and DeiT code bases https://github.com/rwightman/pytorch-image-models/tree/master/timm https://github.com/facebookresearch/deit/ """ def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None, cls_attn_layers=2, use_pos=True, patch_proj='linear', eta=None, tokens_norm=False): """ Args: img_size (int, tuple): input image size patch_size (int, tuple): patch size in_chans (int): number of input channels num_classes (int): number of classes for classification head embed_dim (int): embedding dimension depth (int): depth of transformer num_heads (int): number of attention heads mlp_ratio (int): ratio of mlp hidden dim to embedding dim qkv_bias (bool): enable bias for qkv if True qk_scale (float): override default qk scale of head_dim ** -0.5 if set drop_rate (float): dropout rate attn_drop_rate (float): attention dropout rate drop_path_rate (float): stochastic depth rate norm_layer: (nn.Module): normalization layer cls_attn_layers: (int) Depth of Class attention layers use_pos: (bool) whether to use positional encoding eta: (float) layerscale initialization value tokens_norm: (bool) Whether to normalize all tokens or just the cls_token in the CA """ super().__init__() self.num_classes = num_classes self.num_features = self.embed_dim = embed_dim norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6) self.patch_embed = ConvPatchEmbed(img_size=img_size, embed_dim=embed_dim, patch_size=patch_size) num_patches = self.patch_embed.num_patches self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) self.pos_drop = nn.Dropout(p=drop_rate) dpr = [drop_path_rate for i in range(depth)] self.blocks = nn.ModuleList([ XCABlock( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, num_tokens=num_patches, eta=eta) for i in range(depth)]) self.cls_attn_blocks = nn.ModuleList([ ClassAttentionBlock( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, norm_layer=norm_layer, eta=eta, tokens_norm=tokens_norm) for i in range(cls_attn_layers)]) self.norm = norm_layer(embed_dim) self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity() self.pos_embeder = PositionalEncodingFourier(dim=embed_dim) self.use_pos = use_pos # Classifier head trunc_normal_(self.cls_token, std=.02) self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) @torch.jit.ignore def no_weight_decay(self): return {'pos_embed', 'cls_token', 'dist_token'} def forward_features(self, x): B, C, H, W = x.shape x, (Hp, Wp) = self.patch_embed(x) if self.use_pos: pos_encoding = self.pos_embeder(B, Hp, Wp).reshape(B, -1, x.shape[1]).permute(0, 2, 1) x = x + pos_encoding x = self.pos_drop(x) for blk in self.blocks: x = blk(x, Hp, Wp) cls_tokens = self.cls_token.expand(B, -1, -1) x = torch.cat((cls_tokens, x), dim=1) for blk in self.cls_attn_blocks: x = blk(x, Hp, Wp) x = self.norm(x)[:, 0] return x def forward(self, x): x = self.forward_features(x) x = self.head(x) if self.training: return x, x else: return x # Patch size 16x16 models @register_model def xcit_nano_12_p16(pretrained=False, **kwargs): model = XCiT( patch_size=16, embed_dim=128, depth=12, num_heads=4, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=False, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_tiny_12_p16(pretrained=False, **kwargs): model = XCiT( patch_size=16, embed_dim=192, depth=12, num_heads=4, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_small_12_p16(pretrained=False, **kwargs): model = XCiT( patch_size=16, embed_dim=384, depth=12, num_heads=8, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_tiny_24_p16(pretrained=False, **kwargs): model = XCiT( patch_size=16, embed_dim=192, depth=24, num_heads=4, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_small_24_p16(pretrained=False, **kwargs): model = XCiT( patch_size=16, embed_dim=384, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_medium_24_p16(pretrained=False, **kwargs): model = XCiT( patch_size=16, embed_dim=512, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_large_24_p16(pretrained=False, **kwargs): model = XCiT( patch_size=16, embed_dim=768, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model # Patch size 8x8 models @register_model def xcit_nano_12_p8(pretrained=False, **kwargs): model = XCiT( patch_size=8, embed_dim=128, depth=12, num_heads=4, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=False, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_tiny_12_p8(pretrained=False, **kwargs): model = XCiT( patch_size=8, embed_dim=192, depth=12, num_heads=4, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_small_12_p8(pretrained=False, **kwargs): model = XCiT( patch_size=8, embed_dim=384, depth=12, num_heads=8, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_tiny_24_p8(pretrained=False, **kwargs): model = XCiT( patch_size=8, embed_dim=192, depth=24, num_heads=4, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_small_24_p8(pretrained=False, **kwargs): model = XCiT( patch_size=8, embed_dim=384, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_medium_24_p8(pretrained=False, **kwargs): model = XCiT( patch_size=8, embed_dim=512, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model @register_model def xcit_large_24_p8(pretrained=False, **kwargs): model = XCiT( patch_size=8, embed_dim=768, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) model.default_cfg = _cfg() return model
20,085
36.473881
100
py
xcit
xcit-main/datasets.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Modified from: https://github.com/facebookresearch/deit """ import os import json from torchvision import datasets, transforms from torchvision.datasets.folder import ImageFolder, default_loader from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD from timm.data import create_transform class INatDataset(ImageFolder): def __init__(self, root, train=True, year=2018, transform=None, target_transform=None, category='name', loader=default_loader): self.transform = transform self.loader = loader self.target_transform = target_transform self.year = year # assert category in ['kingdom','phylum','class','order','supercategory','family','genus','name'] path_json = os.path.join(root, f'{"train" if train else "val"}{year}.json') with open(path_json) as json_file: data = json.load(json_file) with open(os.path.join(root, 'categories.json')) as json_file: data_catg = json.load(json_file) path_json_for_targeter = os.path.join(root, f"train{year}.json") with open(path_json_for_targeter) as json_file: data_for_targeter = json.load(json_file) targeter = {} indexer = 0 for elem in data_for_targeter['annotations']: king = [] king.append(data_catg[int(elem['category_id'])][category]) if king[0] not in targeter.keys(): targeter[king[0]] = indexer indexer += 1 self.nb_classes = len(targeter) self.samples = [] for elem in data['images']: cut = elem['file_name'].split('/') target_current = int(cut[2]) path_current = os.path.join(root, cut[0], cut[2], cut[3]) categors = data_catg[target_current] target_current_true = targeter[categors[category]] self.samples.append((path_current, target_current_true)) # __getitem__ and __len__ inherited from ImageFolder def build_dataset(is_train, args): transform = build_transform(is_train, args) if args.data_set == 'CIFAR100': dataset = datasets.CIFAR100(args.data_path, train=is_train, transform=transform) nb_classes = 100 elif args.data_set == 'CIFAR10': dataset = datasets.CIFAR10(args.data_path, train=is_train, transform=transform) nb_classes = 10 elif args.data_set == 'IMNET': root = os.path.join(args.data_path, 'train' if is_train else 'val') dataset = datasets.ImageFolder(root, transform=transform) nb_classes = 1000 elif args.data_set == 'INAT': dataset = INatDataset(args.data_path, train=is_train, year=2018, category=args.inat_category, transform=transform) nb_classes = dataset.nb_classes elif args.data_set == 'INAT19': dataset = INatDataset(args.data_path, train=is_train, year=2019, category=args.inat_category, transform=transform) nb_classes = dataset.nb_classes elif args.data_set == 'CARS': root = os.path.join(args.data_path, 'train' if is_train else 'test') dataset = datasets.ImageFolder(root, transform=transform) nb_classes = 196 elif args.data_set == 'FLOWERS': root = os.path.join(args.data_path, 'train' if is_train else 'test') dataset = datasets.ImageFolder(root, transform=transform) nb_classes = 102 return dataset, nb_classes def build_transform(is_train, args): resize_im = args.input_size > 32 if is_train: # this should always dispatch to transforms_imagenet_train transform = create_transform( input_size=args.input_size, is_training=True, color_jitter=args.color_jitter, auto_augment=args.aa, interpolation=args.train_interpolation, re_prob=args.reprob, re_mode=args.remode, re_count=args.recount, ) if not resize_im: # replace RandomResizedCropAndInterpolation with # RandomCrop transform.transforms[0] = transforms.RandomCrop( args.input_size, padding=4) return transform t = [] if args.full_crop: mean, std = [0.485, 0.456, 0.406], [0.229, 0.224, 0.225] transformations = {} transformations = transforms.Compose( [transforms.Resize(args.input_size, interpolation=3), transforms.CenterCrop(args.input_size), transforms.ToTensor(), transforms.Normalize(mean, std)]) return transformations if resize_im: size = int((256 / 224) * args.input_size) t.append( transforms.Resize(size, interpolation=3), # to maintain same ratio w.r.t. 224 images ) t.append(transforms.CenterCrop(args.input_size)) t.append(transforms.ToTensor()) t.append(transforms.Normalize(IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD)) return transforms.Compose(t)
5,145
36.838235
105
py
xcit
xcit-main/samplers.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Copied from: https://github.com/facebookresearch/deit """ import torch import torch.distributed as dist import math class RASampler(torch.utils.data.Sampler): """Sampler that restricts data loading to a subset of the dataset for distributed, with repeated augmentation. It ensures that different each augmented version of a sample will be visible to a different process (GPU) Heavily based on torch.utils.data.DistributedSampler """ def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True): if num_replicas is None: if not dist.is_available(): raise RuntimeError("Requires distributed package to be available") num_replicas = dist.get_world_size() if rank is None: if not dist.is_available(): raise RuntimeError("Requires distributed package to be available") rank = dist.get_rank() self.dataset = dataset self.num_replicas = num_replicas self.rank = rank self.epoch = 0 self.num_samples = int(math.ceil(len(self.dataset) * 3.0 / self.num_replicas)) self.total_size = self.num_samples * self.num_replicas # self.num_selected_samples = int(math.ceil(len(self.dataset) / self.num_replicas)) self.num_selected_samples = int(math.floor(len(self.dataset) // 256 * 256 / self.num_replicas)) self.shuffle = shuffle def __iter__(self): # deterministically shuffle based on epoch g = torch.Generator() g.manual_seed(self.epoch) if self.shuffle: indices = torch.randperm(len(self.dataset), generator=g).tolist() else: indices = list(range(len(self.dataset))) # add extra samples to make it evenly divisible indices = [ele for ele in indices for i in range(3)] indices += indices[:(self.total_size - len(indices))] assert len(indices) == self.total_size # subsample indices = indices[self.rank:self.total_size:self.num_replicas] assert len(indices) == self.num_samples return iter(indices[:self.num_selected_samples]) def __len__(self): return self.num_selected_samples def set_epoch(self, epoch): self.epoch = epoch
2,354
36.380952
103
py
xcit
xcit-main/semantic_segmentation/backbone/xcit.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Object detection and instance segmentation with XCiT backbone Based on mmseg, timm and DeiT code bases https://github.com/open-mmlab/mmsegmentation https://github.com/rwightman/pytorch-image-models/tree/master/timm https://github.com/facebookresearch/deit/ """ import math from functools import partial import torch import torch.nn as nn import torch.nn.functional as F from timm.models.vision_transformer import _cfg, Mlp from timm.models.registry import register_model from timm.models.layers import DropPath, trunc_normal_, to_2tuple from mmcv.runner import load_checkpoint from mmseg.utils import get_root_logger from mmseg.models.builder import BACKBONES class PositionalEncodingFourier(nn.Module): """ Positional encoding relying on a fourier kernel matching the one used in the "Attention is all of Need" paper. The implementation builds on DeTR code https://github.com/facebookresearch/detr/blob/master/models/position_encoding.py """ def __init__(self, hidden_dim=32, dim=768, temperature=10000): super().__init__() self.token_projection = nn.Conv2d(hidden_dim * 2, dim, kernel_size=1) self.scale = 2 * math.pi self.temperature = temperature self.hidden_dim = hidden_dim self.dim = dim def forward(self, B, H, W): mask = torch.zeros(B, H, W).bool().to(self.token_projection.weight.device) not_mask = ~mask y_embed = not_mask.cumsum(1, dtype=torch.float32) x_embed = not_mask.cumsum(2, dtype=torch.float32) eps = 1e-6 y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale dim_t = torch.arange(self.hidden_dim, dtype=torch.float32, device=mask.device) dim_t = self.temperature ** (2 * (dim_t // 2) / self.hidden_dim) pos_x = x_embed[:, :, :, None] / dim_t pos_y = y_embed[:, :, :, None] / dim_t pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3) pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3) pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) pos = self.token_projection(pos) return pos def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return torch.nn.Sequential( nn.Conv2d( in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False ), nn.SyncBatchNorm(out_planes) ) class ConvPatchEmbed(nn.Module): """ Image to Patch Embedding using multiple convolutional layers """ def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): super().__init__() img_size = to_2tuple(img_size) patch_size = to_2tuple(patch_size) num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) self.img_size = img_size self.patch_size = patch_size self.num_patches = num_patches if patch_size[0] == 16: self.proj = torch.nn.Sequential( conv3x3(3, embed_dim // 8, 2), nn.GELU(), conv3x3(embed_dim // 8, embed_dim // 4, 2), nn.GELU(), conv3x3(embed_dim // 4, embed_dim // 2, 2), nn.GELU(), conv3x3(embed_dim // 2, embed_dim, 2), ) elif patch_size[0] == 8: self.proj = torch.nn.Sequential( conv3x3(3, embed_dim // 4, 2), nn.GELU(), conv3x3(embed_dim // 4, embed_dim // 2, 2), nn.GELU(), conv3x3(embed_dim // 2, embed_dim, 2), ) else: raise("For convolutional projection, patch size has to be in [8, 16]") def forward(self, x, padding_size=None): B, C, H, W = x.shape x = self.proj(x) Hp, Wp = x.shape[2], x.shape[3] x = x.flatten(2).transpose(1, 2) return x, (Hp, Wp) class LPI(nn.Module): """ Local Patch Interaction module that allows explicit communication between tokens in 3x3 windows to augment the implicit communcation performed by the block diagonal scatter attention. Implemented using 2 layers of separable 3x3 convolutions with GeLU and BatchNorm2d """ def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0., kernel_size=3): super().__init__() out_features = out_features or in_features padding = kernel_size // 2 self.conv1 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size, padding=padding, groups=out_features) self.act = act_layer() self.bn = nn.SyncBatchNorm(in_features) self.conv2 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size, padding=padding, groups=out_features) def forward(self, x, H, W): B, N, C = x.shape x = x.permute(0, 2, 1).reshape(B, C, H, W) x = self.conv1(x) x = self.act(x) x = self.bn(x) x = self.conv2(x) x = x.reshape(B, C, N).permute(0, 2, 1) return x class ClassAttention(nn.Module): """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239 """ def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) qkv = qkv.permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) qc = q[:, :, 0:1] # CLS token attn_cls = (qc * k).sum(dim=-1) * self.scale attn_cls = attn_cls.softmax(dim=-1) attn_cls = self.attn_drop(attn_cls) cls_tkn = (attn_cls.unsqueeze(2) @ v).transpose(1, 2).reshape(B, 1, C) cls_tkn = self.proj(cls_tkn) x = torch.cat([self.proj_drop(cls_tkn), x[:, 1:]], dim=1) return x class ClassAttentionBlock(nn.Module): """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239 """ def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, eta=None, tokens_norm=False): super().__init__() self.norm1 = norm_layer(dim) self.attn = ClassAttention( dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop ) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) if eta is not None: # LayerScale Initialization (no layerscale when None) self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) else: self.gamma1, self.gamma2 = 1.0, 1.0 # FIXME: A hack for models pre-trained with layernorm over all the tokens not just the CLS self.tokens_norm = tokens_norm def forward(self, x, H, W, mask=None): x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x))) if self.tokens_norm: x = self.norm2(x) else: x[:, 0:1] = self.norm2(x[:, 0:1]) x_res = x cls_token = x[:, 0:1] cls_token = self.gamma2 * self.mlp(cls_token) x = torch.cat([cls_token, x[:, 1:]], dim=1) x = x_res + self.drop_path(x) return x class XCA(nn.Module): """ Cross-Covariance Attention (XCA) operation where the channels are updated using a weighted sum. The weights are obtained from the (softmax normalized) Cross-covariance matrix (Q^T K \\in d_h \\times d_h) """ def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() self.num_heads = num_heads self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1)) self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) qkv = qkv.permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) q = q.transpose(-2, -1) k = k.transpose(-2, -1) v = v.transpose(-2, -1) q = torch.nn.functional.normalize(q, dim=-1) k = torch.nn.functional.normalize(k, dim=-1) attn = (q @ k.transpose(-2, -1)) * self.temperature attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x @torch.jit.ignore def no_weight_decay(self): return {'temperature'} class XCABlock(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, num_tokens=196, eta=None): super().__init__() self.norm1 = norm_layer(dim) self.attn = XCA( dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop ) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) self.norm3 = norm_layer(dim) self.local_mp = LPI(in_features=dim, act_layer=act_layer) self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) self.gamma3 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) def forward(self, x, H, W): x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x))) x = x + self.drop_path(self.gamma3 * self.local_mp(self.norm3(x), H, W)) x = x + self.drop_path(self.gamma2 * self.mlp(self.norm2(x))) return x @BACKBONES.register_module() class XCiT(nn.Module): """ Based on timm and DeiT code bases https://github.com/rwightman/pytorch-image-models/tree/master/timm https://github.com/facebookresearch/deit/ """ def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=80, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None, cls_attn_layers=2, use_pos=True, eta=None, tokens_norm=False, out_indices=[3, 5, 7, 11]): """ Args: img_size (int, tuple): input image size patch_size (int, tuple): patch size in_chans (int): number of input channels num_classes (int): number of classes for classification head embed_dim (int): embedding dimension depth (int): depth of transformer num_heads (int): number of attention heads mlp_ratio (int): ratio of mlp hidden dim to embedding dim qkv_bias (bool): enable bias for qkv if True qk_scale (float): override default qk scale of head_dim ** -0.5 if set drop_rate (float): dropout rate attn_drop_rate (float): attention dropout rate drop_path_rate (float): stochastic depth rate norm_layer: (nn.Module): normalization layer cls_attn_layers: (int) Depth of Class attention layers use_pos: (bool) whether to use positional encoding eta: (float) layerscale initialization value tokens_norm: (bool) Whether to normalize all tokens or just the cls_token in the CA out_indices: (list) Indices of layers from which FPN features are extracted """ super().__init__() self.num_classes = num_classes self.num_features = self.embed_dim = embed_dim norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6) self.patch_embed = ConvPatchEmbed(img_size=img_size, embed_dim=embed_dim, patch_size=patch_size) num_patches = self.patch_embed.num_patches self.pos_drop = nn.Dropout(p=drop_rate) dpr = [drop_path_rate for i in range(depth)] self.blocks = nn.ModuleList([ XCABlock( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, num_tokens=num_patches, eta=eta) for i in range(depth)]) self.pos_embeder = PositionalEncodingFourier(dim=embed_dim) self.use_pos = use_pos self.out_indices = out_indices if patch_size == 16: self.fpn1 = nn.Sequential( nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2), nn.SyncBatchNorm(embed_dim), nn.GELU(), nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2), ) self.fpn2 = nn.Sequential( nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2), ) self.fpn3 = nn.Identity() self.fpn4 = nn.MaxPool2d(kernel_size=2, stride=2) elif patch_size == 8: self.fpn1 = nn.Sequential( nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2), ) self.fpn2 = nn.Identity() self.fpn3 = nn.Sequential( nn.MaxPool2d(kernel_size=2, stride=2), ) self.fpn4 = nn.Sequential( nn.MaxPool2d(kernel_size=4, stride=4), ) @torch.jit.ignore def no_weight_decay(self): return {'pos_embed', 'cls_token', 'dist_token'} def init_weights(self, pretrained=None): """Initialize the weights in backbone. Args: pretrained (str, optional): Path to pre-trained weights. Defaults to None. """ def _init_weights(m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) if isinstance(pretrained, str): self.apply(_init_weights) logger = get_root_logger() load_checkpoint(self, pretrained, strict=False, logger=logger) elif pretrained is None: self.apply(_init_weights) else: raise TypeError('pretrained must be a str or None') def forward_features(self, x): B, C, H, W = x.shape x, (Hp, Wp) = self.patch_embed(x) pos_encoding = self.pos_embeder(B, Hp, Wp).reshape(B, -1, x.shape[1]).permute(0, 2, 1) if self.use_pos: x = x + pos_encoding x = self.pos_drop(x) features = [] for i, blk in enumerate(self.blocks): x = blk(x, Hp, Wp) if i in self.out_indices: xp = x.permute(0, 2, 1).reshape(B, -1, Hp, Wp) features.append(xp) ops = [self.fpn1, self.fpn2, self.fpn3, self.fpn4] for i in range(len(features)): features[i] = ops[i](features[i]) return tuple(features) def forward(self, x): x = self.forward_features(x) return x
16,958
36.854911
100
py
xcit
xcit-main/semantic_segmentation/tools/test.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Testing script modified from https://github.com/open-mmlab/mmsegmentation """ import argparse import os import mmcv import torch from mmcv.parallel import MMDataParallel, MMDistributedDataParallel from mmcv.runner import get_dist_info, init_dist, load_checkpoint from mmcv.utils import DictAction from mmseg.apis import multi_gpu_test, single_gpu_test from mmseg.datasets import build_dataloader, build_dataset from mmseg.models import build_segmentor from backbone import xcit def parse_args(): parser = argparse.ArgumentParser( description='mmseg test (and eval) a model') parser.add_argument('config', help='test config file path') parser.add_argument('checkpoint', help='checkpoint file') parser.add_argument( '--aug-test', action='store_true', help='Use Flip and Multi scale aug') parser.add_argument('--out', help='output result file in pickle format') parser.add_argument( '--format-only', action='store_true', help='Format the output results without perform evaluation. It is' 'useful when you want to format the result to a specific format and ' 'submit it to the test server') parser.add_argument( '--eval', type=str, nargs='+', help='evaluation metrics, which depends on the dataset, e.g., "mIoU"' ' for generic datasets, and "cityscapes" for Cityscapes') parser.add_argument('--show', action='store_true', help='show results') parser.add_argument( '--show-dir', help='directory where painted images will be saved') parser.add_argument( '--gpu-collect', action='store_true', help='whether to use gpu to collect results.') parser.add_argument( '--tmpdir', help='tmp directory used for collecting results from multiple ' 'workers, available when gpu_collect is not specified') parser.add_argument( '--options', nargs='+', action=DictAction, help='custom options') parser.add_argument( '--eval-options', nargs='+', action=DictAction, help='custom options for evaluation') parser.add_argument( '--launcher', choices=['none', 'pytorch', 'slurm', 'mpi'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() if 'LOCAL_RANK' not in os.environ: os.environ['LOCAL_RANK'] = str(args.local_rank) return args def main(): args = parse_args() assert args.out or args.eval or args.format_only or args.show \ or args.show_dir, \ ('Please specify at least one operation (save/eval/format/show the ' 'results / save the results) with the argument "--out", "--eval"' ', "--format-only", "--show" or "--show-dir"') if args.eval and args.format_only: raise ValueError('--eval and --format_only cannot be both specified') if args.out is not None and not args.out.endswith(('.pkl', '.pickle')): raise ValueError('The output file must be a pkl file.') cfg = mmcv.Config.fromfile(args.config) if args.options is not None: cfg.merge_from_dict(args.options) # set cudnn_benchmark if cfg.get('cudnn_benchmark', False): torch.backends.cudnn.benchmark = True if args.aug_test: # hard code index cfg.data.test.pipeline[1].img_ratios = [ 0.5, 0.75, 1.0, 1.25, 1.5, 1.75 ] cfg.data.test.pipeline[1].flip = True cfg.model.pretrained = None cfg.data.test.test_mode = True # init distributed env first, since logger depends on the dist info. if args.launcher == 'none': distributed = False else: distributed = True init_dist(args.launcher, **cfg.dist_params) # build the dataloader # TODO: support multiple images per gpu (only minor changes are needed) dataset = build_dataset(cfg.data.test) data_loader = build_dataloader( dataset, samples_per_gpu=1, workers_per_gpu=cfg.data.workers_per_gpu, dist=distributed, shuffle=False) # build the model and load checkpoint cfg.model.train_cfg = None model = build_segmentor(cfg.model, test_cfg=cfg.get('test_cfg')) checkpoint = load_checkpoint(model, args.checkpoint, map_location='cpu') model.CLASSES = checkpoint['meta']['CLASSES'] model.PALETTE = checkpoint['meta']['PALETTE'] efficient_test = False if args.eval_options is not None: efficient_test = args.eval_options.get('efficient_test', False) if not distributed: model = MMDataParallel(model, device_ids=[0]) outputs = single_gpu_test(model, data_loader, args.show, args.show_dir, efficient_test) else: model = MMDistributedDataParallel( model.cuda(), device_ids=[torch.cuda.current_device()], broadcast_buffers=False) outputs = multi_gpu_test(model, data_loader, args.tmpdir, args.gpu_collect, efficient_test) rank, _ = get_dist_info() if rank == 0: if args.out: print(f'\nwriting results to {args.out}') mmcv.dump(outputs, args.out) kwargs = {} if args.eval_options is None else args.eval_options if args.format_only: dataset.format_results(outputs, **kwargs) if args.eval: dataset.evaluate(outputs, args.eval, **kwargs) if __name__ == '__main__': main()
5,618
34.789809
79
py
xcit
xcit-main/semantic_segmentation/tools/train.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Training script modified from https://github.com/open-mmlab/mmsegmentation """ import argparse import copy import os import os.path as osp import time import mmcv import torch from mmcv.runner import init_dist from mmcv.utils import Config, DictAction, get_git_hash from mmseg import __version__ from mmseg.apis import set_random_seed, train_segmentor from mmseg.datasets import build_dataset from mmseg.models import build_segmentor from mmseg.utils import collect_env, get_root_logger from backbone import xcit def parse_args(): parser = argparse.ArgumentParser(description='Train a segmentor') parser.add_argument('config', help='train config file path') parser.add_argument('--work-dir', help='the dir to save logs and models') parser.add_argument( '--load-from', help='the checkpoint file to load weights from') parser.add_argument( '--resume-from', help='the checkpoint file to resume from') parser.add_argument( '--no-validate', action='store_true', help='whether not to evaluate the checkpoint during training') group_gpus = parser.add_mutually_exclusive_group() group_gpus.add_argument( '--gpus', type=int, help='number of gpus to use ' '(only applicable to non-distributed training)') group_gpus.add_argument( '--gpu-ids', type=int, nargs='+', help='ids of gpus to use ' '(only applicable to non-distributed training)') parser.add_argument('--seed', type=int, default=None, help='random seed') parser.add_argument( '--deterministic', action='store_true', help='whether to set deterministic options for CUDNN backend.') parser.add_argument( '--options', nargs='+', action=DictAction, help='custom options') parser.add_argument( '--launcher', choices=['none', 'pytorch', 'slurm', 'mpi'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() if 'LOCAL_RANK' not in os.environ: os.environ['LOCAL_RANK'] = str(args.local_rank) return args def main(): args = parse_args() cfg = Config.fromfile(args.config) if args.options is not None: cfg.merge_from_dict(args.options) # set cudnn_benchmark if cfg.get('cudnn_benchmark', False): torch.backends.cudnn.benchmark = True # work_dir is determined in this priority: CLI > segment in file > filename if args.work_dir is not None: # update configs according to CLI args if args.work_dir is not None cfg.work_dir = args.work_dir elif cfg.get('work_dir', None) is None: # use config filename as default work_dir if cfg.work_dir is None cfg.work_dir = osp.join('./work_dirs', osp.splitext(osp.basename(args.config))[0]) if args.load_from is not None: cfg.load_from = args.load_from if args.resume_from is not None: cfg.resume_from = args.resume_from if args.gpu_ids is not None: cfg.gpu_ids = args.gpu_ids else: cfg.gpu_ids = range(1) if args.gpus is None else range(args.gpus) # init distributed env first, since logger depends on the dist info. if args.launcher == 'none': distributed = False else: distributed = True init_dist(args.launcher, **cfg.dist_params) # create work_dir mmcv.mkdir_or_exist(osp.abspath(cfg.work_dir)) # dump config cfg.dump(osp.join(cfg.work_dir, osp.basename(args.config))) # init the logger before other steps timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime()) log_file = osp.join(cfg.work_dir, f'{timestamp}.log') logger = get_root_logger(log_file=log_file, log_level=cfg.log_level) # init the meta dict to record some important information such as # environment info and seed, which will be logged meta = dict() # log env info env_info_dict = collect_env() env_info = '\n'.join([f'{k}: {v}' for k, v in env_info_dict.items()]) dash_line = '-' * 60 + '\n' logger.info('Environment info:\n' + dash_line + env_info + '\n' + dash_line) meta['env_info'] = env_info # log some basic info logger.info(f'Distributed training: {distributed}') logger.info(f'Config:\n{cfg.pretty_text}') # set random seeds if args.seed is not None: logger.info(f'Set random seed to {args.seed}, deterministic: ' f'{args.deterministic}') set_random_seed(args.seed, deterministic=args.deterministic) cfg.seed = args.seed meta['seed'] = args.seed meta['exp_name'] = osp.basename(args.config) model = build_segmentor( cfg.model, train_cfg=cfg.get('train_cfg'), test_cfg=cfg.get('test_cfg')) logger.info(model) datasets = [build_dataset(cfg.data.train)] if len(cfg.workflow) == 2: val_dataset = copy.deepcopy(cfg.data.val) val_dataset.pipeline = cfg.data.train.pipeline datasets.append(build_dataset(val_dataset)) if cfg.checkpoint_config is not None: # save mmseg version, config file content and class names in # checkpoints as meta data cfg.checkpoint_config.meta = dict( mmseg_version=f'{__version__}+{get_git_hash()[:7]}', config=cfg.pretty_text, CLASSES=datasets[0].CLASSES, PALETTE=datasets[0].PALETTE) # add an attribute for visualization convenience model.CLASSES = datasets[0].CLASSES train_segmentor( model, datasets, cfg, distributed=distributed, validate=(not args.no_validate), timestamp=timestamp, meta=meta) if __name__ == '__main__': main()
5,885
33.22093
79
py
xcit
xcit-main/detection/backbone/xcit.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Object detection and instance segmentation with XCiT backbone Based on mmdet, timm and DeiT code bases https://github.com/open-mmlab/mmdetection https://github.com/rwightman/pytorch-image-models/tree/master/timm https://github.com/facebookresearch/deit/ """ import math from functools import partial import torch import torch.nn as nn import torch.nn.functional as F from timm.models.vision_transformer import _cfg, Mlp from timm.models.registry import register_model from timm.models.layers import DropPath, trunc_normal_, to_2tuple from mmcv.runner import load_checkpoint from mmdet.utils import get_root_logger from mmdet.models.builder import BACKBONES class PositionalEncodingFourier(nn.Module): """ Positional encoding relying on a fourier kernel matching the one used in the "Attention is all of Need" paper. The implementation builds on DeTR code https://github.com/facebookresearch/detr/blob/master/models/position_encoding.py """ def __init__(self, hidden_dim=32, dim=768, temperature=10000): super().__init__() self.token_projection = nn.Conv2d(hidden_dim * 2, dim, kernel_size=1) self.scale = 2 * math.pi self.temperature = temperature self.hidden_dim = hidden_dim self.dim = dim def forward(self, B, H, W): mask = torch.zeros(B, H, W).bool().to(self.token_projection.weight.device) not_mask = ~mask y_embed = not_mask.cumsum(1, dtype=torch.float32) x_embed = not_mask.cumsum(2, dtype=torch.float32) eps = 1e-6 y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale dim_t = torch.arange(self.hidden_dim, dtype=torch.float32, device=mask.device) dim_t = self.temperature ** (2 * (dim_t // 2) / self.hidden_dim) pos_x = x_embed[:, :, :, None] / dim_t pos_y = y_embed[:, :, :, None] / dim_t pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3) pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3) pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) pos = self.token_projection(pos) return pos def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return torch.nn.Sequential( nn.Conv2d( in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False ), nn.SyncBatchNorm(out_planes) ) class ConvPatchEmbed(nn.Module): """ Image to Patch Embedding using multiple convolutional layers """ def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): super().__init__() img_size = to_2tuple(img_size) patch_size = to_2tuple(patch_size) num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) self.img_size = img_size self.patch_size = patch_size self.num_patches = num_patches if patch_size[0] == 16: self.proj = torch.nn.Sequential( conv3x3(3, embed_dim // 8, 2), nn.GELU(), conv3x3(embed_dim // 8, embed_dim // 4, 2), nn.GELU(), conv3x3(embed_dim // 4, embed_dim // 2, 2), nn.GELU(), conv3x3(embed_dim // 2, embed_dim, 2), ) elif patch_size[0] == 8: self.proj = torch.nn.Sequential( conv3x3(3, embed_dim // 4, 2), nn.GELU(), conv3x3(embed_dim // 4, embed_dim // 2, 2), nn.GELU(), conv3x3(embed_dim // 2, embed_dim, 2), ) else: raise("For convolutional projection, patch size has to be in [8, 16]") def forward(self, x, padding_size=None): B, C, H, W = x.shape x = self.proj(x) Hp, Wp = x.shape[2], x.shape[3] x = x.flatten(2).transpose(1, 2) return x, (Hp, Wp) class LPI(nn.Module): """ Local Patch Interaction module that allows explicit communication between tokens in 3x3 windows to augment the implicit communcation performed by the block diagonal scatter attention. Implemented using 2 layers of separable 3x3 convolutions with GeLU and BatchNorm2d """ def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0., kernel_size=3): super().__init__() out_features = out_features or in_features padding = kernel_size // 2 self.conv1 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size, padding=padding, groups=out_features) self.act = act_layer() self.bn = nn.SyncBatchNorm(in_features) self.conv2 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size, padding=padding, groups=out_features) def forward(self, x, H, W): B, N, C = x.shape x = x.permute(0, 2, 1).reshape(B, C, H, W) x = self.conv1(x) x = self.act(x) x = self.bn(x) x = self.conv2(x) x = x.reshape(B, C, N).permute(0, 2, 1) return x class ClassAttention(nn.Module): """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239 """ def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) qkv = qkv.permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) qc = q[:, :, 0:1] # CLS token attn_cls = (qc * k).sum(dim=-1) * self.scale attn_cls = attn_cls.softmax(dim=-1) attn_cls = self.attn_drop(attn_cls) cls_tkn = (attn_cls.unsqueeze(2) @ v).transpose(1, 2).reshape(B, 1, C) cls_tkn = self.proj(cls_tkn) x = torch.cat([self.proj_drop(cls_tkn), x[:, 1:]], dim=1) return x class ClassAttentionBlock(nn.Module): """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239 """ def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, eta=None, tokens_norm=False): super().__init__() self.norm1 = norm_layer(dim) self.attn = ClassAttention( dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop ) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) if eta is not None: # LayerScale Initialization (no layerscale when None) self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) else: self.gamma1, self.gamma2 = 1.0, 1.0 # FIXME: A hack for models pre-trained with layernorm over all the tokens not just the CLS self.tokens_norm = tokens_norm def forward(self, x, H, W, mask=None): x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x))) if self.tokens_norm: x = self.norm2(x) else: x[:, 0:1] = self.norm2(x[:, 0:1]) x_res = x cls_token = x[:, 0:1] cls_token = self.gamma2 * self.mlp(cls_token) x = torch.cat([cls_token, x[:, 1:]], dim=1) x = x_res + self.drop_path(x) return x class XCA(nn.Module): """ Cross-Covariance Attention (XCA) operation where the channels are updated using a weighted sum. The weights are obtained from the (softmax normalized) Cross-covariance matrix (Q^T K \\in d_h \\times d_h) """ def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() self.num_heads = num_heads self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1)) self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) qkv = qkv.permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) q = q.transpose(-2, -1) k = k.transpose(-2, -1) v = v.transpose(-2, -1) q = torch.nn.functional.normalize(q, dim=-1) k = torch.nn.functional.normalize(k, dim=-1) attn = (q @ k.transpose(-2, -1)) * self.temperature attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x @torch.jit.ignore def no_weight_decay(self): return {'temperature'} class XCABlock(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, num_tokens=196, eta=None): super().__init__() self.norm1 = norm_layer(dim) self.attn = XCA( dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop ) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) self.norm3 = norm_layer(dim) self.local_mp = LPI(in_features=dim, act_layer=act_layer) self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) self.gamma3 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) def forward(self, x, H, W): x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x))) x = x + self.drop_path(self.gamma3 * self.local_mp(self.norm3(x), H, W)) x = x + self.drop_path(self.gamma2 * self.mlp(self.norm2(x))) return x @BACKBONES.register_module() class XCiT(nn.Module): """ Based on timm and DeiT code bases https://github.com/rwightman/pytorch-image-models/tree/master/timm https://github.com/facebookresearch/deit/ """ def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=80, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None, cls_attn_layers=2, use_pos=True, eta=None, tokens_norm=False, out_indices=[3, 5, 7, 11]): """ Args: img_size (int, tuple): input image size patch_size (int, tuple): patch size in_chans (int): number of input channels num_classes (int): number of classes for classification head embed_dim (int): embedding dimension depth (int): depth of transformer num_heads (int): number of attention heads mlp_ratio (int): ratio of mlp hidden dim to embedding dim qkv_bias (bool): enable bias for qkv if True qk_scale (float): override default qk scale of head_dim ** -0.5 if set drop_rate (float): dropout rate attn_drop_rate (float): attention dropout rate drop_path_rate (float): stochastic depth rate norm_layer: (nn.Module): normalization layer cls_attn_layers: (int) Depth of Class attention layers use_pos: (bool) whether to use positional encoding eta: (float) layerscale initialization value tokens_norm: (bool) Whether to normalize all tokens or just the cls_token in the CA out_indices: (list) Indices of layers from which FPN features are extracted """ super().__init__() self.num_classes = num_classes self.num_features = self.embed_dim = embed_dim norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6) self.patch_embed = ConvPatchEmbed(img_size=img_size, embed_dim=embed_dim, patch_size=patch_size) num_patches = self.patch_embed.num_patches self.pos_drop = nn.Dropout(p=drop_rate) dpr = [drop_path_rate for i in range(depth)] self.blocks = nn.ModuleList([ XCABlock( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, num_tokens=num_patches, eta=eta) for i in range(depth)]) self.pos_embeder = PositionalEncodingFourier(dim=embed_dim) self.use_pos = use_pos self.out_indices = out_indices if patch_size == 16: self.fpn1 = nn.Sequential( nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2), nn.SyncBatchNorm(embed_dim), nn.GELU(), nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2), ) self.fpn2 = nn.Sequential( nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2), ) self.fpn3 = nn.Identity() self.fpn4 = nn.MaxPool2d(kernel_size=2, stride=2) elif patch_size == 8: self.fpn1 = nn.Sequential( nn.ConvTranspose2d(embed_dim, embed_dim, kernel_size=2, stride=2), ) self.fpn2 = nn.Identity() self.fpn3 = nn.Sequential( nn.MaxPool2d(kernel_size=2, stride=2), ) self.fpn4 = nn.Sequential( nn.MaxPool2d(kernel_size=4, stride=4), ) @torch.jit.ignore def no_weight_decay(self): return {'pos_embed', 'cls_token', 'dist_token'} def init_weights(self, pretrained=None): """Initialize the weights in backbone. Args: pretrained (str, optional): Path to pre-trained weights. Defaults to None. """ def _init_weights(m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) if isinstance(pretrained, str): self.apply(_init_weights) logger = get_root_logger() load_checkpoint(self, pretrained, strict=False, logger=logger) elif pretrained is None: self.apply(_init_weights) else: raise TypeError('pretrained must be a str or None') def forward_features(self, x): B, C, H, W = x.shape x, (Hp, Wp) = self.patch_embed(x) pos_encoding = self.pos_embeder(B, Hp, Wp).reshape(B, -1, x.shape[1]).permute(0, 2, 1) if self.use_pos: x = x + pos_encoding x = self.pos_drop(x) features = [] for i, blk in enumerate(self.blocks): x = blk(x, Hp, Wp) if i in self.out_indices: xp = x.permute(0, 2, 1).reshape(B, -1, Hp, Wp) features.append(xp) ops = [self.fpn1, self.fpn2, self.fpn3, self.fpn4] for i in range(len(features)): features[i] = ops[i](features[i]) return tuple(features) def forward(self, x): x = self.forward_features(x) return x
16,955
36.848214
100
py
xcit
xcit-main/detection/tools/test.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Testing script modified from https://github.com/open-mmlab/mmdetection """ import argparse import os import warnings import mmcv import torch from mmcv import Config, DictAction from mmcv.cnn import fuse_conv_bn from mmcv.parallel import MMDataParallel, MMDistributedDataParallel from mmcv.runner import (get_dist_info, init_dist, load_checkpoint, wrap_fp16_model) from mmdet.apis import multi_gpu_test, single_gpu_test from mmdet.datasets import (build_dataloader, build_dataset, replace_ImageToTensor) from mmdet.models import build_detector, build_backbone from backbone import xcit def parse_args(): parser = argparse.ArgumentParser( description='MMDet test (and eval) a model') parser.add_argument('config', help='test config file path') parser.add_argument('checkpoint', help='checkpoint file') parser.add_argument('--out', help='output result file in pickle format') parser.add_argument( '--fuse-conv-bn', action='store_true', help='Whether to fuse conv and bn, this will slightly increase' 'the inference speed') parser.add_argument( '--format-only', action='store_true', help='Format the output results without perform evaluation. It is' 'useful when you want to format the result to a specific format and ' 'submit it to the test server') parser.add_argument( '--eval', type=str, nargs='+', help='evaluation metrics, which depends on the dataset, e.g., "bbox",' ' "segm", "proposal" for COCO, and "mAP", "recall" for PASCAL VOC') parser.add_argument('--show', action='store_true', help='show results') parser.add_argument( '--show-dir', help='directory where painted images will be saved') parser.add_argument( '--show-score-thr', type=float, default=0.3, help='score threshold (default: 0.3)') parser.add_argument( '--gpu-collect', action='store_true', help='whether to use gpu to collect results.') parser.add_argument( '--tmpdir', help='tmp directory used for collecting results from multiple ' 'workers, available when gpu-collect is not specified') parser.add_argument( '--cfg-options', nargs='+', action=DictAction, help='override some settings in the used config, the key-value pair ' 'in xxx=yyy format will be merged into config file. If the value to ' 'be overwritten is a list, it should be like key="[a,b]" or key=a,b ' 'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" ' 'Note that the quotation marks are necessary and that no white space ' 'is allowed.') parser.add_argument( '--options', nargs='+', action=DictAction, help='custom options for evaluation, the key-value pair in xxx=yyy ' 'format will be kwargs for dataset.evaluate() function (deprecate), ' 'change to --eval-options instead.') parser.add_argument( '--eval-options', nargs='+', action=DictAction, help='custom options for evaluation, the key-value pair in xxx=yyy ' 'format will be kwargs for dataset.evaluate() function') parser.add_argument( '--launcher', choices=['none', 'pytorch', 'slurm', 'mpi'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() if 'LOCAL_RANK' not in os.environ: os.environ['LOCAL_RANK'] = str(args.local_rank) if args.options and args.eval_options: raise ValueError( '--options and --eval-options cannot be both ' 'specified, --options is deprecated in favor of --eval-options') if args.options: warnings.warn('--options is deprecated in favor of --eval-options') args.eval_options = args.options return args def main(): args = parse_args() assert args.out or args.eval or args.format_only or args.show \ or args.show_dir, \ ('Please specify at least one operation (save/eval/format/show the ' 'results / save the results) with the argument "--out", "--eval"' ', "--format-only", "--show" or "--show-dir"') if args.eval and args.format_only: raise ValueError('--eval and --format_only cannot be both specified') if args.out is not None and not args.out.endswith(('.pkl', '.pickle')): raise ValueError('The output file must be a pkl file.') cfg = Config.fromfile(args.config) if args.cfg_options is not None: cfg.merge_from_dict(args.cfg_options) # import modules from string list. if cfg.get('custom_imports', None): from mmcv.utils import import_modules_from_strings import_modules_from_strings(**cfg['custom_imports']) # set cudnn_benchmark if cfg.get('cudnn_benchmark', False): torch.backends.cudnn.benchmark = True cfg.model.pretrained = None if cfg.model.get('neck'): if isinstance(cfg.model.neck, list): for neck_cfg in cfg.model.neck: if neck_cfg.get('rfp_backbone'): if neck_cfg.rfp_backbone.get('pretrained'): neck_cfg.rfp_backbone.pretrained = None elif cfg.model.neck.get('rfp_backbone'): if cfg.model.neck.rfp_backbone.get('pretrained'): cfg.model.neck.rfp_backbone.pretrained = None # in case the test dataset is concatenated samples_per_gpu = 1 if isinstance(cfg.data.test, dict): cfg.data.test.test_mode = True samples_per_gpu = cfg.data.test.pop('samples_per_gpu', 1) if samples_per_gpu > 1: # Replace 'ImageToTensor' to 'DefaultFormatBundle' cfg.data.test.pipeline = replace_ImageToTensor( cfg.data.test.pipeline) elif isinstance(cfg.data.test, list): for ds_cfg in cfg.data.test: ds_cfg.test_mode = True samples_per_gpu = max( [ds_cfg.pop('samples_per_gpu', 1) for ds_cfg in cfg.data.test]) if samples_per_gpu > 1: for ds_cfg in cfg.data.test: ds_cfg.pipeline = replace_ImageToTensor(ds_cfg.pipeline) # init distributed env first, since logger depends on the dist info. if args.launcher == 'none': distributed = False else: distributed = True init_dist(args.launcher, **cfg.dist_params) # build the dataloader dataset = build_dataset(cfg.data.test) data_loader = build_dataloader( dataset, samples_per_gpu=samples_per_gpu, workers_per_gpu=cfg.data.workers_per_gpu, dist=distributed, shuffle=False) # build the model and load checkpoint cfg.model.train_cfg = None model = build_detector(cfg.model, test_cfg=cfg.get('test_cfg')) fp16_cfg = cfg.get('fp16', None) if fp16_cfg is not None: wrap_fp16_model(model) checkpoint = load_checkpoint(model, args.checkpoint, map_location='cpu') if args.fuse_conv_bn: model = fuse_conv_bn(model) # old versions did not save class info in checkpoints, this walkaround is # for backward compatibility if 'CLASSES' in checkpoint.get('meta', {}): model.CLASSES = checkpoint['meta']['CLASSES'] else: model.CLASSES = dataset.CLASSES if not distributed: model = MMDataParallel(model, device_ids=[0]) outputs = single_gpu_test(model, data_loader, args.show, args.show_dir, args.show_score_thr) else: model = MMDistributedDataParallel( model.cuda(), device_ids=[torch.cuda.current_device()], broadcast_buffers=False) outputs = multi_gpu_test(model, data_loader, args.tmpdir, args.gpu_collect) rank, _ = get_dist_info() if rank == 0: if args.out: print(f'\nwriting results to {args.out}') mmcv.dump(outputs, args.out) kwargs = {} if args.eval_options is None else args.eval_options if args.format_only: dataset.format_results(outputs, **kwargs) if args.eval: eval_kwargs = cfg.get('evaluation', {}).copy() # hard-code way to remove EvalHook args for key in [ 'interval', 'tmpdir', 'start', 'gpu_collect', 'save_best', 'rule' ]: eval_kwargs.pop(key, None) eval_kwargs.update(dict(metric=args.eval, **kwargs)) print(dataset.evaluate(outputs, **eval_kwargs)) if __name__ == '__main__': main()
8,846
37.633188
79
py
xcit
xcit-main/detection/tools/train.py
# Copyright (c) 2015-present, Facebook, Inc. # All rights reserved. """ Training script modified from https://github.com/open-mmlab/mmdetection """ import argparse import copy import os import os.path as osp import time import warnings import mmcv import torch from mmcv import Config, DictAction from mmcv.runner import get_dist_info, init_dist from mmcv.utils import get_git_hash from mmdet import __version__ from mmdet.apis import set_random_seed, train_detector from mmdet.datasets import build_dataset from mmdet.models import build_detector from mmdet.utils import collect_env, get_root_logger from backbone import xcit def parse_args(): parser = argparse.ArgumentParser(description='Train a detector') parser.add_argument('config', help='train config file path') parser.add_argument('--work-dir', help='the dir to save logs and models') parser.add_argument( '--resume-from', help='the checkpoint file to resume from') parser.add_argument( '--no-validate', action='store_true', help='whether not to evaluate the checkpoint during training') group_gpus = parser.add_mutually_exclusive_group() group_gpus.add_argument( '--gpus', type=int, help='number of gpus to use ' '(only applicable to non-distributed training)') group_gpus.add_argument( '--gpu-ids', type=int, nargs='+', help='ids of gpus to use ' '(only applicable to non-distributed training)') parser.add_argument('--seed', type=int, default=None, help='random seed') parser.add_argument( '--deterministic', action='store_true', help='whether to set deterministic options for CUDNN backend.') parser.add_argument( '--options', nargs='+', action=DictAction, help='override some settings in the used config, the key-value pair ' 'in xxx=yyy format will be merged into config file (deprecate), ' 'change to --cfg-options instead.') parser.add_argument( '--cfg-options', nargs='+', action=DictAction, help='override some settings in the used config, the key-value pair ' 'in xxx=yyy format will be merged into config file. If the value to ' 'be overwritten is a list, it should be like key="[a,b]" or key=a,b ' 'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" ' 'Note that the quotation marks are necessary and that no white space ' 'is allowed.') parser.add_argument( '--launcher', choices=['none', 'pytorch', 'slurm', 'mpi'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() if 'LOCAL_RANK' not in os.environ: os.environ['LOCAL_RANK'] = str(args.local_rank) if args.options and args.cfg_options: raise ValueError( '--options and --cfg-options cannot be both ' 'specified, --options is deprecated in favor of --cfg-options') if args.options: warnings.warn('--options is deprecated in favor of --cfg-options') args.cfg_options = args.options return args def main(): args = parse_args() cfg = Config.fromfile(args.config) if args.cfg_options is not None: cfg.merge_from_dict(args.cfg_options) # import modules from string list. if cfg.get('custom_imports', None): from mmcv.utils import import_modules_from_strings import_modules_from_strings(**cfg['custom_imports']) # set cudnn_benchmark if cfg.get('cudnn_benchmark', False): torch.backends.cudnn.benchmark = True # work_dir is determined in this priority: CLI > segment in file > filename if args.work_dir is not None: # update configs according to CLI args if args.work_dir is not None cfg.work_dir = args.work_dir elif cfg.get('work_dir', None) is None: # use config filename as default work_dir if cfg.work_dir is None cfg.work_dir = osp.join('./work_dirs', osp.splitext(osp.basename(args.config))[0]) if args.resume_from is not None: cfg.resume_from = args.resume_from if args.gpu_ids is not None: cfg.gpu_ids = args.gpu_ids else: cfg.gpu_ids = range(1) if args.gpus is None else range(args.gpus) # init distributed env first, since logger depends on the dist info. if args.launcher == 'none': distributed = False else: distributed = True init_dist(args.launcher, **cfg.dist_params) # re-set gpu_ids with distributed training mode _, world_size = get_dist_info() cfg.gpu_ids = range(world_size) # create work_dir mmcv.mkdir_or_exist(osp.abspath(cfg.work_dir)) # dump config cfg.dump(osp.join(cfg.work_dir, osp.basename(args.config))) # init the logger before other steps timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime()) log_file = osp.join(cfg.work_dir, f'{timestamp}.log') logger = get_root_logger(log_file=log_file, log_level=cfg.log_level) # init the meta dict to record some important information such as # environment info and seed, which will be logged meta = dict() # log env info env_info_dict = collect_env() env_info = '\n'.join([(f'{k}: {v}') for k, v in env_info_dict.items()]) dash_line = '-' * 60 + '\n' logger.info('Environment info:\n' + dash_line + env_info + '\n' + dash_line) meta['env_info'] = env_info meta['config'] = cfg.pretty_text # log some basic info logger.info(f'Distributed training: {distributed}') logger.info(f'Config:\n{cfg.pretty_text}') # set random seeds if args.seed is not None: logger.info(f'Set random seed to {args.seed}, ' f'deterministic: {args.deterministic}') set_random_seed(args.seed, deterministic=args.deterministic) cfg.seed = args.seed meta['seed'] = args.seed meta['exp_name'] = osp.basename(args.config) model = build_detector( cfg.model, train_cfg=cfg.get('train_cfg'), test_cfg=cfg.get('test_cfg')) datasets = [build_dataset(cfg.data.train)] if len(cfg.workflow) == 2: val_dataset = copy.deepcopy(cfg.data.val) val_dataset.pipeline = cfg.data.train.pipeline datasets.append(build_dataset(val_dataset)) if cfg.checkpoint_config is not None: # save mmdet version, config file content and class names in # checkpoints as meta data cfg.checkpoint_config.meta = dict( mmdet_version=__version__ + get_git_hash()[:7], CLASSES=datasets[0].CLASSES) # add an attribute for visualization convenience model.CLASSES = datasets[0].CLASSES train_detector( model, datasets, cfg, distributed=distributed, validate=(not args.no_validate), timestamp=timestamp, meta=meta) if __name__ == '__main__': main()
7,064
35.045918
79
py
psi_2022
psi_2022-main/plotting.py
""" Use with saved model. Produces the current sweep and latent space plots """ from src.models import DIVA from src.data import MemMapDataset_O from src.data._utils import get_dataloaders from src.common.utils import load_model from src.common.physics_approximations import * import torch import numpy as np import matplotlib.pyplot as plt import matplotlib as mpl MP_names_JET = ['BTF', 'D_tot', 'IpiFP', 'PNBI_TOT', 'P_OH', 'PICR_TOT', 'k', 'delRoben', 'delRuntn', 'ahor', 'Rgeo', 'q95', 'Vol', 'elm_timings'] def main(model_name): state_dict, hparams, dataset = load_model(model_name) model = DIVA(**hparams) model.load_state_dict(state_dict) model.double() current_sweep_plot(model, dataset) latent_space_plot(model, dataset) def latent_space_plot(model, dataset): Z_MACH, Z_STOCH = [], [] for idx in range(dataset.total_num_pulses // 2): sample_profs, sample_mps = torch.from_numpy(dataset.data['profs'][idx].copy()), torch.from_numpy(dataset.data['mps'][idx].copy()) sample_profs_norm, sample_mps_norm = dataset.norm_profiles(sample_profs), dataset.norm_mps(sample_mps) with torch.no_grad(): _,z_mach, z_stoch, *_ = model.prof2z(sample_profs_norm) Z_MACH.extend(z_mach) Z_STOCH.extend(z_stoch) Z_MACH, Z_STOCH = torch.vstack(Z_MACH), torch.vstack(Z_STOCH) image_res = 512 sample_size = image_res ** 2 # 2D r1, r2 = -5, 5 a, b = sample_size, 2 ld_1, ld_2 = 0, 2 range_xy = torch.linspace(start=r1, end=r2, steps=image_res) range_xy = torch.cartesian_prod(range_xy, range_xy) range_imagecoord = torch.linspace(0, image_res-1, steps=image_res, dtype=torch.int32) # so we can easily go back range_imagecoord = torch.cartesian_prod(range_imagecoord, range_imagecoord) z_mach_mean, z_stoch_mean = Z_MACH.mean(0), Z_STOCH.mean(0) z_mach_sample, z_stoch_sample = torch.tile(z_mach_mean, (sample_size, 1)), torch.tile(z_stoch_mean, (sample_size, 1)) z_mach_sample[:, ld_1] = range_xy[:, 0] z_mach_sample[:, ld_2] = range_xy[:, 1] image_array = np.zeros((image_res, image_res)) with torch.no_grad(): z_conditional = torch.cat((z_mach_sample, z_stoch_sample), 1) out_profs = model.z2prof(z_conditional) out_mps = model.z2mp(z_mach_sample) sample_profs, sample_mps = dataset.denorm_profiles(out_profs), dataset.denorm_mps(out_mps) sample_teseps, sample_neseps, sample_rseps = find_tesep(sample_profs) image_array = np.zeros((image_res, image_res)) for i in range(range_imagecoord.shape[0]): _x, _y = range_imagecoord[i] _y = image_res - 1 - _y # (0, 0) for img are on top left so reverse image_array[_y, _x] = sample_neseps[i] sample_zx, sample_zy = 2, 0 #The star on th graph data_sample = torch.tensor([sample_zx, sample_zy]) min_i = -1 min_dist = 100000 for i in range(range_xy.shape[0]): a = data_sample.cpu().numpy() b = range_xy[i].cpu().numpy() dist = np.linalg.norm(a-b) if dist < min_dist: min_dist = dist min_i = i sample_1 = min_i fig, ls_ax,= plt.subplots(constrained_layout=True) cmap = mpl.cm.plasma norm = mpl.colors.Normalize(vmin=0, vmax=1e20) """ LATENT SPACE PLOT """ cax = ls_ax.imshow(np.rot90(image_array, 3), extent=[r1, r2, r1, r2], cmap=cmap, norm=norm, interpolation='spline36') fig.colorbar(cax, ax=ls_ax, label='Inferred $n_e^{sep}$ [m$^{-3}$]', location='left') ls_ax.set_xlabel('Latent Dimension 4') ls_ax.set_ylabel('Latent Dimension 6') plt.show() def current_sweep_plot(model, dataset): MEAN_MP = [] for idx in range(dataset.total_num_pulses // 2): sample_profs, sample_mps = torch.from_numpy(dataset.data['profs'][idx].copy()), torch.from_numpy(dataset.data['mps'][idx].copy()) MEAN_MP.append(sample_mps) MEAN_MP = torch.vstack(MEAN_MP).mean(0) N_SAMPLES = 1000 current_sweep = torch.linspace(1e6, 5e6, N_SAMPLES) MP_IN = torch.tile(MEAN_MP, (N_SAMPLES, 1)) MP_IN[:, 2] = current_sweep cmap = mpl.cm.viridis norm = mpl.colors.Normalize(vmin=0, vmax=6e6) with torch.no_grad(): out_profs_norm, _, _ = model.inference(dataset.norm_mps(MP_IN), from_mean=False) out_profs = dataset.denorm_profiles(out_profs_norm) fig = plt.figure() for k, sample in enumerate(out_profs): plt.plot(sample[0], color=cmap(norm(current_sweep[k]))) fig.colorbar(mpl.cm.ScalarMappable(norm=norm, cmap=cmap),orientation='horizontal', label='$I_P$ [MA]') plt.show() if __name__ == '__main__': model_name = 'DIVA.pth' main(model_name)
4,731
36.555556
146
py
psi_2022
psi_2022-main/train.py
from src.models import DIVA from src.data import MemMapDataset_O from src.data._utils import initialize_dataloaders from src.common.utils import save_model from typing import List import os import timeit import torch import numpy as np from tqdm import tqdm """ Training Paramters """ EPOCHS: int = 100 # Around 50 it starts getting okay BATCH_SIZE: int = 512 # 512 LR: float = 0.01 # 0.01 """ Model Paramters """ MACH_LATENT_DIM:int = 10 # 8 STOCH_LATENT_DIM:int = 3 # 3 CONV_FILTER_SIZES: List[int] = [2, 4, 6] # [2, 4, 6] MP_REG_LAYER_WIDTH: int = 20 # 40 MP_REG_LAYER_DEPTH: int = 4 # 6 MP_REG_LAYER_SIZES: List[int] = [MP_REG_LAYER_WIDTH]*MP_REG_LAYER_DEPTH """ DIVA Specific KL-divergences _STOCH : applied to Z_stoch _COND_SUP : applied to mach_cond vs mach_enc _COND_UNSUP : applied to mach_enc when unsupervised """ BETA_STOCH: float = 0.5 # 0.8 BETA_COND_SUP: float = 1.0 # 1.0 BETA_COND_UNSUP: float = 0.001 # 0.0086 """ Loss function parameterization GAMMA : applied to reconstructions of profiles and machine parameters LAMBDA : applied to the physics losses """ PHYSICS: bool = True GAMMA_PROF: float = 1200.0 # 550 GAMMA_MP: float = 20.0 # 2.6618 LAMBDA_PRESSURE: float = 500.0 # 55.0 LAMBDA_BETA: float= 1000.0 # 10.0 LAMDBA_BPOL: float = 100.0 # 31.0 """ Scheduler paramterizations """ SCHEDULER_STEP_SIZE: int = 10 # 10 SCHEDULER_GAMMA: int = 0.9 # 0.99 """ Save Information """ MODEL_NAME = 'DIVA' # ['BTF', 'D_tot', 'N_tot', 'IpiFP', 'PNBI_TOT', 'PICR_TOT', 'PECR_TOT', 'k', 'delRoben', 'delRuntn', 'ahor', 'Rgeo', 'q95', 'Vol'] hparams = dict(mach_latent_dim=MACH_LATENT_DIM, stoch_latent_dim=STOCH_LATENT_DIM, conv_filter_sizes=CONV_FILTER_SIZES, mp_layer_sizes=MP_REG_LAYER_SIZES, GAMMA_PROF=GAMMA_PROF, GAMMA_MP=GAMMA_MP, BETA_KLD_COND=BETA_COND_SUP, BETA_KLD_STOCH=BETA_STOCH, BETA_KLD_MACH=BETA_COND_UNSUP, LAMBDA_PRESSURE=LAMBDA_PRESSURE, LAMBDA_BETA=LAMBDA_BETA, LAMDBA_BPOL=LAMDBA_BPOL, physics=PHYSICS, model_name=MODEL_NAME) def main(base_data_path: str = None): dataset = MemMapDataset_O(data_path=base_data_path) hparams['out_length'] = dataset.prof_length hparams['action_dim'] = dataset.mp_dim model = DIVA(**hparams) model.double() optimizer = torch.optim.Adam(model.parameters(9), lr=LR) train_dl, val_dl, test_dl = initialize_dataloaders(dataset, batch_size=BATCH_SIZE) training_loss, best_val_loss = [], np.inf # epoch_iter = tqdm(range(EPOCHS)) epoch_iter = range(EPOCHS) for epoch in epoch_iter: epoch_loss, epoch_loss_dict = train_epoch(model, optimizer, train_dl, dataset) training_loss.append(epoch_loss) if epoch % 5 == 0: val_loss, _ = test_epoch(model, val_dl, dataset) if val_loss < best_val_loss: print(f'Epoch: {epoch}\nNew best val loss: {val_loss:.4}, saving model') save_model(model, hparams, dataset) def train_epoch(model, optimizer, loader, dataset): epoch_loss = 0.0 for steps, batch in enumerate(loader): batch_profs, batch_mps = batch preds = model.forward(batch_profs, batch_mps) loss_dict = model.loss_function(batch, preds, dataset, step=steps) loss = loss_dict.pop('loss') / len(loader) loss.backward() optimizer.step() optimizer.zero_grad() epoch_loss += loss.item() return epoch_loss, loss_dict def test_epoch(model, loader, dataset): epoch_loss = 0.0 with torch.no_grad(): for steps, batch in enumerate(loader): batch_profs, batch_mps = batch preds = model.forward(batch_profs, batch_mps) loss_dict = model.loss_function(batch, preds, dataset, step=steps) loss = loss_dict.pop('loss') / len(loader) epoch_loss += loss.item() return epoch_loss, loss_dict if __name__ == '__main__': data_path: str = os.path.join(os.getenv('PROC_DIR'), 'JET_PDB_PULSE_ARRAYS') main(data_path)
4,108
35.362832
132
py
psi_2022
psi_2022-main/src/common/utils.py
import torch def save_model(model, hparams, dataset): model_name = hparams['model_name'] save_dict = {'state_dict': model.state_dict(), 'hparams': hparams, 'dataset': dataset} torch.save(save_dict, './' + model_name + '.pth') def load_model(model_name): save_dict = torch.load(f'./{model_name}') return save_dict['state_dict'], save_dict['hparams'], save_dict['dataset']
393
42.777778
90
py
psi_2022
psi_2022-main/src/common/physics_approximations.py
import torch import numpy as np boltzmann_constant = 1.380e-23 mu_0 = 1.256e-6 def find_tesep(profs): ne, te = profs[:, 0], profs[:, 1] teseps, neseps, rseps = np.empty(te.shape[0]), np.empty(te.shape[0]), np.empty(te.shape[0]) for k, (ne_slice, te_slice) in enumerate(zip(ne, te)): l_idx, r_idx = 0, 1 while te_slice[r_idx] > 100: l_idx += 1 r_idx += 1 if r_idx == 50:# or (len(te_slice) == 60 and r_idx == 60): break if r_idx == 50: continue weights_r, weights_l = get_weights(te_slice, l_idx, r_idx) tesep_estimation = weights_l*te_slice[l_idx] + weights_r*te_slice[r_idx] nesep_estimation = weights_l*ne_slice[l_idx] + weights_r*ne_slice[r_idx] # rsep_estimation = weights_l*r_slice[l_idx] + weights_r*r_slice[r_idx] teseps[k] = tesep_estimation neseps[k] = nesep_estimation rseps[k] = l_idx # rsep_estimation = weights_l*x[idx_l] + weights_r*x[idx_r] return teseps, neseps, rseps def get_weights(te, idx_l, idx_r, query=100): # Gets weighst as usual dist = te[idx_r] - query + query - te[idx_l] weights = (1-(te[idx_r] - query)/dist, 1-(query - te[idx_l])/dist) return weights def static_pressure_stored_energy_approximation(profs): if not isinstance(profs, torch.Tensor): profs = torch.from_numpy(profs) return boltzmann_constant*torch.prod(profs, 1).sum(1) def torch_shaping_approx(minor_radius, tri_u, tri_l, elongation): triangularity = (tri_u + tri_l) / 2.0 b = elongation*minor_radius gamma_top = -(minor_radius + triangularity) gamma_bot = minor_radius - triangularity alpha_top = -gamma_top / (b*b) alpha_bot = -gamma_bot / (b*b) top_int = (torch.arcsinh(2*torch.abs(alpha_top)*b) + 2*torch.abs(alpha_top)*b*torch.sqrt(4*alpha_top*alpha_top*b*b+1)) / (2*torch.abs(alpha_top)) bot_int = (torch.arcsinh(2*torch.abs(alpha_bot)*b) + 2*torch.abs(alpha_bot)*b*torch.sqrt(4*alpha_bot*alpha_bot*b*b+1)) / (2*torch.abs(alpha_bot)) return bot_int + top_int def bpol_approx(minor_radius, tri_u, tri_l, elongation, current): shaping = torch_shaping_approx(minor_radius, tri_u, tri_l, elongation) return mu_0*current / shaping def calculate_peped(profs): if not isinstance(profs, torch.Tensor): profs = torch.from_numpy(profs) ne = profs[:, 0:1, :] te = profs[:, 1:, :] p = boltzmann_constant*ne*te second_diff = torch.diff(p, n=2, dim=-1) min_diff_val, min_diff_idx = torch.min(second_diff, dim=-1) ped_loc = min_diff_idx -2 peped = torch.zeros((len(p))) for k in range(len(peped)): peped[k] = torch.select(p[k], dim=-1, index=ped_loc[k].item()) return peped, ped_loc def beta_approximation(profiles_tensors, minor_radius, tri_u, tri_l, elongation, current, bt, beta_pol=False): """ To approximate beta! The factor of 2 at the front is to compensate the ions which are nowhere to be found in this analysis. The additional factor of 100 is to get it in percent form. """ e_c = 1.602e-19 bpol = bpol_approx(minor_radius, tri_u, tri_l, elongation, current) if beta_pol: pressure_ped, _ = calculate_peped(profiles_tensors) beta_pol_approx = 2*mu_0*pressure_ped / (bpol*bpol) return beta_pol_approx density, temperature = profiles_tensors[:, 0, :], profiles_tensors[:, 1, :] pressure_prof = density*temperature pressure_average = pressure_prof.mean(-1) # TODO: This beta average is not really realistic I find... but am interested to see how it impacts return (100*2)*e_c*2*mu_0 * pressure_average / (bt*bt + bpol*bpol) def pressure_calculation(profs: torch.Tensor, dataset=None, normalize=True): if normalize and dataset is not None: profs = dataset.denorm_profiles(profs.copy()) return boltzmann_constant*torch.prod(profs, 1) def calculate_physics_constraints(profiles_og, mps_og, train_set): # Denormalize everything! profiles = torch.clone(profiles_og) profiles = train_set.denorm_profiles(profiles, to_torch=True) mps = torch.clone(mps_og) mps = train_set.denorm_mps(mps, to_torch=True) sp = static_pressure_stored_energy_approximation(profiles) minor_radius, tri_u, tri_l, elongation, current, bt = mps[:, 2], mps[:, 4],mps[:, 5],mps[:, 6], mps[:, 8], mps[:, 9] bpol = bpol_approx(minor_radius, tri_u, tri_l, elongation, current) beta = beta_approximation(profiles, minor_radius, tri_u, tri_l, elongation, current, bt) pressure = pressure_calculation(profiles, normalize=False) return sp, beta, bpol, pressure
4,682
44.028846
149
py
psi_2022
psi_2022-main/src/models/DIVA.py
import torch import torch.nn as nn import torch.nn.functional as F from typing import Dict, List, Tuple from .model_utils import PRIORREG, AUXREG, ENCODER, DECODER from src.common.physics_approximations import * MIN_STDDEV = 1e-15 class DIVA(nn.Module): def __init__(self, mach_latent_dim: int, stoch_latent_dim: int, conv_filter_sizes: List[int], mp_layer_sizes: List[int], input_dim: int = 2, out_length: int = 50, action_dim: int = 15, clamping_zero_tensor: torch.Tensor = None, BETA_KLD_COND: float=0.001, BETA_KLD_STOCH: float=0.001, BETA_KLD_MACH: float=0.001, GAMMA_PROF: float=100.0, GAMMA_MP: float=100.0, LAMBDA_PRESSURE: float=10.0, LAMBDA_BETA: float=1.0, LAMDBA_BPOL: float = 1.0, physics: bool = True, **kwargs): super(DIVA, self).__init__() self.physics = physics self.mach_latent_dim, self.stoch_latent_dim = mach_latent_dim, stoch_latent_dim self.conv_filter_sizes, self.mp_encoder_layer_sizes = conv_filter_sizes, mp_layer_sizes self.trans_conv_filter_sizes, self.mp_decoder_layer_sizes = conv_filter_sizes[::-1], mp_layer_sizes[::-1] self.prof_len, self.mp_dim = out_length, action_dim self.prof_encoder = ENCODER(filter_sizes=self.conv_filter_sizes, in_ch=input_dim, in_length=self.prof_len) self.prof_decoder = DECODER(filter_sizes=self.trans_conv_filter_sizes, end_conv_size=self.prof_encoder.end_conv_size, clamping_zero_tensor=clamping_zero_tensor) self.mp_encoder = PRIORREG(in_dim=self.mp_dim, out_dim=self.mach_latent_dim, hidden_dims=self.mp_encoder_layer_sizes, make_prior=True) self.mp_decoder = AUXREG(in_dim = self.mach_latent_dim, out_dim=self.mp_dim, hidden_dims=self.mp_decoder_layer_sizes) in_prior_size = self.prof_encoder.end_conv_size*self.conv_filter_sizes[-1] self.z_mu_mach = nn.Linear(in_prior_size, self.mach_latent_dim) self.z_var_mach = nn.Linear(in_prior_size, self.mach_latent_dim) self.z_mu_stoch = nn.Linear(in_prior_size, self.stoch_latent_dim) self.z_var_stoch = nn.Linear(in_prior_size, self.stoch_latent_dim) self.decoder_input = nn.Linear(self.stoch_latent_dim + self.mach_latent_dim, self.trans_conv_filter_sizes[0]*self.prof_encoder.end_conv_size) self.output_layer = nn.Linear(self.prof_decoder.final_size, out_length) self.GAMMA_PROF, self.GAMMA_MP = GAMMA_PROF, GAMMA_MP self.BETA_KLD_COND, self.BETA_KLD_STOCH , self.BETA_KLD_MACH = BETA_KLD_COND, BETA_KLD_STOCH, BETA_KLD_MACH self.LAMBDA_PRESSURE, self.LAMBDA_BETA, self.LAMDBA_BPOL = LAMBDA_PRESSURE, LAMBDA_BETA, LAMDBA_BPOL def forward(self, prof_t: torch.Tensor, mp_t: torch.Tensor, **kwargs) -> List[torch.Tensor]: z_enc, z_mach_enc, z_stoch_enc, mu_mach_enc, var_mach_enc, mu_stoch_enc, var_stoch_enc = self.prof2z(prof_t) z_cond, mu_mach_cond, var_mach_cond = self.mp2z(mp_t) prof_out, mp_out = self.z2prof(z_enc), self.z2mp(z_mach_enc) return [z_enc, z_mach_enc, z_stoch_enc, mu_mach_enc, var_mach_enc, mu_stoch_enc, var_stoch_enc, z_cond, mu_mach_cond, var_mach_cond, prof_out, mp_out] def prof2z(self, prof): """ Encode the profile to z_stoch and z_mach""" enc = self.prof_encoder(prof) mu_mach, var_mach, mu_stoch, var_stoch = self.z_mu_mach(enc), torch.clamp(torch.nn.functional.softplus(self.z_var_mach(enc)), min=MIN_STDDEV), self.z_mu_stoch(enc), torch.clamp(torch.nn.functional.softplus(self.z_var_stoch(enc)), min=MIN_STDDEV) z_mach, z_stoch = self.reparameterize(mu_mach, var_mach), self.reparameterize(mu_stoch, var_stoch) z = torch.cat([z_mach, z_stoch], 1) return [z, z_mach, z_stoch, mu_mach, var_mach, mu_stoch, var_stoch] def mp2z(self, mp): """ Conditional prior on z_mach via machine paramters """ mu_mach_cond, var_mach_cond = self.mp_encoder(mp) z_cond = self.reparameterize(mu_mach_cond, var_mach_cond) return [z_cond, mu_mach_cond, var_mach_cond] def z2prof(self, z): """ Decode z_mach and z_stoch to profile """ z = self.decoder_input(z) dec = self.prof_decoder(z) prof = self.output_layer(dec) return prof def z2mp(self, z_mach): """ Decode z_mach to machine parameters """ mp_out = self.mp_decoder(z_mach) return mp_out def reparameterize(self, mu: torch.Tensor, var: torch.Tensor) -> torch.Tensor: """Reparameterization trick to sample N(mu, var) from N(0, 1) Parameters ---------- mu : torch.Tensor Mean of the latent distribution var : torch.Tensor standard deviation of the latent distribution Returns ------- torch.Tensor mu + e^(var/2) + randn() """ std = torch.exp(0.5 * var) eps = torch.randn_like(std) z = mu + eps*std return z def inference(self, mp_in, prof_in=None, from_mean: bool = True): """ DO CONDITIONAL INFERENCE! """ # z_enc, z_mach_enc, z_stoch_enc, mu_mach_enc, var_mach_enc, mu_stoch_enc, var_stoch_enc = # self.prof2z(prof_t) z_mach_cond, mu_mach_cond, var_mach_cond = self.mp2z(mp_in) if prof_in is not None: _, _, z_stoch_rand, _, _, _, _ = self.prof2z(prof_in) else: z_stoch_rand = torch.normal(0, 1, size=(len(z_mach_cond), self.stoch_latent_dim)) if not from_mean: z = torch.cat([z_mach_cond, z_stoch_rand], 1) out_mps = self.z2mp(z_mach_cond) else: z = torch.cat([mu_mach_cond, z_stoch_rand], 1) out_mps = self.z2mp(mu_mach_cond) out_profs = self.z2prof(z) return out_profs, out_mps, mu_mach_cond def loss_function(self, inputs, outputs, train_set=None, step=0): prof_in, mp_in = inputs _, _, _, mu_mach_enc, var_mach_enc, mu_stoch_enc, var_stoch_enc, _, mu_mach_cond, var_mach_cond, prof_out, mp_out = outputs recon_prof, recon_mp = F.mse_loss(prof_in, prof_out), F.mse_loss(mp_in, mp_out) recon_loss = self.GAMMA_MP*recon_mp + self.GAMMA_PROF*recon_prof physics, sp_loss, beta_loss, bpol_loss = torch.Tensor([0]), torch.Tensor([0]), torch.Tensor([0]), torch.Tensor([0]) """ # kld_stoch measures the KL-div between the stochastic latent space priors given by the profile encoding against a normal distribution # kld_cond measures KL-div between the machine latent space priors given by machine parameters against that given by profile encoding # TODO: kld_unsup measures KL-div between mach latent space priors given by profile encoding against a normal distribution """ kld_stoch = torch.distributions.kl.kl_divergence(torch.distributions.normal.Normal(mu_stoch_enc, torch.exp(0.5*var_stoch_enc)), torch.distributions.normal.Normal(0, 1)).mean(0).sum() kld_mach_sup = torch.distributions.kl.kl_divergence(torch.distributions.normal.Normal(mu_mach_enc, torch.exp(0.5*var_mach_enc)), torch.distributions.normal.Normal(mu_mach_cond, torch.exp(0.5*var_mach_cond))).mean(0).sum() kld_mach_unsup = torch.distributions.kl.kl_divergence(torch.distributions.normal.Normal(mu_mach_enc, torch.exp(0.5*var_mach_enc)), torch.distributions.normal.Normal(0, 1)).mean(0).sum() # add all_kld together if step % 2 == 0: kld_loss = self.BETA_KLD_STOCH*kld_stoch + self.BETA_KLD_COND*kld_mach_sup else: kld_loss = self.BETA_KLD_STOCH*kld_stoch + self.BETA_KLD_MACH*kld_mach_unsup sp_in, beta_in, bpol_in, pressure_in = calculate_physics_constraints(prof_in, mp_in, train_set) sp_out, beta_out, bpol_out, pressure_out = calculate_physics_constraints(prof_out, mp_out, train_set) sp_loss, beta_loss, bpol_loss = F.mse_loss(sp_in, sp_out), F.mse_loss(pressure_in, pressure_out), F.mse_loss(bpol_in, bpol_out) physics = self.LAMBDA_PRESSURE*sp_loss + self.LAMBDA_BETA*beta_loss + self.LAMDBA_BPOL*bpol_loss loss = recon_loss + kld_loss + physics return dict(loss=loss, recon_loss=recon_loss, recon_mp=recon_mp, recon_prof=recon_prof, kld_loss=kld_loss, kld_cond=kld_mach_sup, kld_mach=kld_mach_unsup, kld_stoch=kld_stoch, physics=physics, sp_loss=sp_loss, beta_loss=beta_loss, bpol_loss=bpol_loss) """ def loss_function(self, inputs, outputs, train_set=None, step=0): prof_in, mp_in = inputs _, _, _, mu_mach_enc, var_mach_enc, mu_stoch_enc, var_stoch_enc, _, mu_mach_cond, var_mach_cond, prof_out, mp_out = outputs recon_prof, recon_mp = F.mse_loss(prof_in, prof_out), F.mse_loss(mp_in, mp_out) recon_loss = self.GAMMA_MP*recon_mp + self.GAMMA_PROF*recon_prof physics, sp_loss, beta_loss, bpol_loss = torch.Tensor([0]), torch.Tensor([0]), torch.Tensor([0]), torch.Tensor([0]) """ # kld_stoch measures the KL-div between the stochastic latent space priors given by the profile encoding against a normal distribution # kld_cond measures KL-div between the machine latent space priors given by machine parameters against that given by profile encoding # TODO: kld_unsup measures KL-div between mach latent space priors given by profile encoding against a normal distribution """ kld_stoch = torch.distributions.kl.kl_divergence( torch.distributions.normal.Normal(mu_stoch_enc, torch.exp(0.5*var_stoch_enc)), torch.distributions.normal.Normal(0, 1) ).mean(0).sum() kld_mach_sup = torch.distributions.kl.kl_divergence( torch.distributions.normal.Normal(mu_mach_enc, torch.exp(0.5*var_mach_enc)), torch.distributions.normal.Normal(mu_mach_cond, torch.exp(0.5*var_mach_cond)), ).mean(0).sum() kld_mach_unsup = torch.distributions.kl.kl_divergence( torch.distributions.normal.Normal(mu_mach_enc, torch.exp(0.5*var_mach_enc)), torch.distributions.normal.Normal(0, 1) ).mean(0).sum() # add all_kld together if step % 2 == 0: kld_loss = self.BETA_KLD_STOCH*kld_stoch + self.BETA_KLD_COND*kld_mach_sup else: kld_loss = self.BETA_KLD_STOCH*kld_stoch + self.BETA_KLD_MACH*kld_mach_unsup # add together if train_set is not None and self.physics: sp_in, beta_in, bpol_in, pressure_in = calculate_physics_constraints(prof_in, mp_in, train_set) sp_out, beta_out, bpol_out, pressure_out = calculate_physics_constraints(prof_out, mp_out, train_set) sp_loss, beta_loss, bpol_loss = F.mse_loss(sp_in, sp_out), F.mse_loss(pressure_in, pressure_out), F.mse_loss(bpol_in, bpol_out) physics = self.LAMBDA_SP*sp_loss + self.LAMBDA_BETA*beta_loss + self.LAMDBA_BPOL*bpol_loss loss = recon_loss + kld_loss + physics return dict(loss=loss, recon_loss=recon_loss, recon_mp=recon_mp, recon_prof=recon_prof, kld_loss=kld_loss, kld_cond=kld_mach_sup, kld_mach=kld_mach_unsup, kld_stoch=kld_stoch, physics=physics, sp_loss=sp_loss, beta_loss=beta_loss, bpol_loss=bpol_loss) """
11,083
53.871287
253
py
psi_2022
psi_2022-main/src/models/model_utils.py
import torch import torch.nn as nn import torch.nn.functional as F from typing import List import numpy as np def get_conv_output_size(initial_input_size, number_blocks, max_pool=True): """ The conv blocks we use keep the same size but use max pooling, so the output of all convolution blocks will be of length input_size / 2""" if max_pool==False: return initial_input_size out_size = initial_input_size for i in range(number_blocks): out_size = int(out_size / 2) return out_size def get_trans_output_size(input_size, stride, padding, kernel_size): """ A function to get the output length of a vector of length input_size after a tranposed convolution layer""" return (input_size -1)*stride - 2*padding + (kernel_size - 1) +1 def get_final_output(initial_input_size, number_blocks, number_trans_per_block, stride, padding, kernel_size): """A function to get the final output size after tranposed convolution blocks""" out_size = initial_input_size for i in range(number_blocks): for k in range(number_trans_per_block): out_size = get_trans_output_size(out_size, stride, padding, kernel_size) return out_size class Flatten(nn.Module): def forward(self, input): return input.view(input.size(0), -1) class UnFlatten(nn.Module): def __init__(self, size, length): super(UnFlatten, self).__init__() self.size = size self.length = length def forward(self, input): out = input.view(input.size(0), self.size, self.length) return out class AUXREG(nn.Module): """Module block for predicing mps from the latent dimensions""" def __init__(self, in_dim: int, out_dim: int, hidden_dims: List[int], ): super().__init__() self.in_dim = in_dim # Latent space size self.out_dim = out_dim # Machine parameter size self.hidden_dims = hidden_dims self.block = nn.ModuleList() last_dim = in_dim for dim in self.hidden_dims: self.block.append(nn.Linear(last_dim, dim)) self.block.append(nn.ReLU()) last_dim = dim self.block.append(nn.Linear(last_dim, out_dim)) def forward(self, z): for lay in self.block: z = lay(z) return z class PRIORREG(nn.Module): """Module block for conditional prior via machine paramteers""" def __init__(self, in_dim: int, hidden_dims: List[int], out_dim: int = None, make_prior=False): super().__init__() self.in_dim = in_dim # Machine parameter size self.out_dim = out_dim # Latent dimension size self.hidden_dims = hidden_dims self.make_prior = make_prior self.block = nn.ModuleList() last_dim = in_dim for dim in self.hidden_dims: self.block.append(nn.Linear(last_dim, dim)) self.block.append(nn.ReLU()) last_dim = dim # self.block.append(nn.Linear(last_dim, out_dim)) if self.make_prior: self.fc_mu = nn.Linear(last_dim, self.out_dim) self.fc_var = nn.Linear(last_dim, self.out_dim) else: self.block.append(nn.Linear(last_dim, self.out_dim)) def forward(self, z): for lay in self.block: z = lay(z) if self.make_prior: mu, var = self.fc_mu(z), self.fc_var(z) return mu, var else: return z class ENCODER(nn.Module): def __init__(self, filter_sizes: List[int], in_length: int = 75, in_ch: int = 2): super().__init__() hidden_channels = filter_sizes self.kernel_size = 5 self.padding = 0 self.stride = 1 self.pool_padding = 0 self.pool_dilation = 1 self.pool_kernel_size = 2 self.pool_stride = self.pool_kernel_size self.block = nn.ModuleList() self.end_conv_size = in_length # [(W - K + 2P) / S] + 1 for dim in hidden_channels: self.block.append(nn.Conv1d(in_ch, dim, kernel_size=self.kernel_size, padding=self.padding)) self.end_conv_size = ((self.end_conv_size - self.kernel_size + 2*self.padding) / self.stride) + 1 self.block.append(nn.ReLU()) self.block.append(nn.MaxPool1d(self.pool_kernel_size, padding=self.pool_padding, dilation=self.pool_dilation, )) self.end_conv_size = ((self.end_conv_size + 2*self.pool_padding - self.pool_dilation*(self.pool_kernel_size -1)-1) / self.pool_stride) + 1 in_ch = dim self.block.append(Flatten()) self.end_conv_size = int(self.end_conv_size) def forward(self, x): # print('Encoder in shape', x.shape) for lay in self.block: x = lay(x) # print(x.shape) # print('encoder out shape', x.shape) return x class DECODER(nn.Module): def __init__(self, filter_sizes: List[int], end_conv_size: int, clamping_zero_tensor: torch.Tensor): super().__init__() in_ch = 2 self.hidden_channels = filter_sizes self.end_conv_size = end_conv_size self.num_trans_conv_blocks = 1 self.trans_stride = 1 self.trans_padding = 0 self.trans_kernel_size = 2 self.block = nn.ModuleList() if self.hidden_channels[-1] != in_ch: self.hidden_channels.append(in_ch) self.block.append(UnFlatten(size=self.hidden_channels[0], length=self.end_conv_size)) # Needs trasnpose kernel instead for i in range(len(self.hidden_channels) - 1): self.block.append(nn.ConvTranspose1d(self.hidden_channels[i], self.hidden_channels[i+1], kernel_size=self.trans_kernel_size)) self.block.append(nn.ReLU()) self.final_size = get_final_output(end_conv_size, len(self.hidden_channels) -1, self.num_trans_conv_blocks, self.trans_stride, self.trans_padding, self.trans_kernel_size) self.clamping_zero_tensor = clamping_zero_tensor if self.clamping_zero_tensor is not None: print('Applying a clamping tensor to reconstruct only non-negative profiles!') def forward(self, x): # print('Decoder in shape', x.shape) for lay in self.block: x = lay(x) if self.clamping_zero_tensor is not None: torch.clamp(x, self.clamping_zero_tensor) # print('Decoder out shape', x.shape) return x
6,482
38.530488
178
py
psi_2022
psi_2022-main/src/data/_utils.py
import numpy as np import torch def initialize_dataloaders(dataset, val_split: float = 0.3, test_split: float = 0.1, batch_size: int = 256, nw: int = 4): pulse_idxs = np.arange(dataset.total_num_pulses) make_vals = True # Need to randomly shuffle the pulses, as they are sorted originally from 30000 -> onward! np.random.shuffle(pulse_idxs) num_val_pulses = int(val_split * dataset.total_num_pulses) num_test_pulses = int(test_split * dataset.total_num_pulses) train_pulse_idxs = pulse_idxs[:-(num_val_pulses + num_test_pulses)] val_pulse_idxs = pulse_idxs[-(num_val_pulses + num_test_pulses):-num_test_pulses] test_pulse_idxs = pulse_idxs[-num_test_pulses:] train_norms = dataset.get_norms(train_pulse_idxs, return_norms=True) # Need to then get the relevant slices from the list of pulses... train_slice_idxs, val_slice_idxs, test_slice_idxs = [], [], [] for set_pulse_idxs, set_slice_idxs in zip([train_pulse_idxs, val_pulse_idxs, test_pulse_idxs], [train_slice_idxs, val_slice_idxs, test_slice_idxs]): for num in set_pulse_idxs: if num == 0: start_slice_idx = 0 else: start_slice_idx = dataset.cumsum_num_slices[num-1] + 1 # the plus one comes from the cumsum_num_slices subtracting one at the begninggn end_slice_idx = dataset.cumsum_num_slices[num] + 1 set_slice_idxs.extend(np.arange(start_slice_idx, end_slice_idx)) dataset.train_slice_idxs = train_slice_idxs dataset.val_slice_idxs = train_slice_idxs dataset.test_slice_idxs = train_slice_idxs train_dataset = torch.utils.data.Subset(dataset, train_slice_idxs) val_dataset = torch.utils.data.Subset(dataset, val_slice_idxs) test_dataset = torch.utils.data.Subset(dataset, test_slice_idxs) pin_memory = True dataloader_kwargs = dict( batch_size=batch_size, shuffle=True, num_workers=nw, pin_memory=pin_memory ) train_loader = torch.utils.data.DataLoader(train_dataset, **dataloader_kwargs) valid_loader = torch.utils.data.DataLoader(val_dataset, **dataloader_kwargs) test_loader = torch.utils.data.DataLoader(test_dataset, **dataloader_kwargs) return train_loader, valid_loader, test_loader def get_dataloaders(dataset, batch_size: int = 256): train_dataset = torch.utils.data.Subset(dataset, dataset.train_slice_idxs) val_dataset = torch.utils.data.Subset(dataset, dataset.val_slice_idxs) test_dataset = torch.utils.data.Subset(dataset, dataset.test_slice_idxs) dataloader_kwargs = dict( batch_size=batch_size, shuffle=False, ) train_loader = torch.utils.data.DataLoader(train_dataset, **dataloader_kwargs) valid_loader = torch.utils.data.DataLoader(val_dataset, **dataloader_kwargs) test_loader = torch.utils.data.DataLoader(test_dataset, **dataloader_kwargs) return train_loader, valid_loader, test_loader
3,007
47.516129
153
py
psi_2022
psi_2022-main/src/data/memmap_dataset_observational.py
from torch.utils.data import Dataset from mmap_ninja.ragged import RaggedMmap import os import numpy as np from typing import Tuple import torch def compute_raggedmmap(data_dir: str, data_string: str, batch_size: int = 128) -> RaggedMmap: """Compute the ragged map Parameters ---------- data_dir : str Base dir where arrays are stored and where mmaps will be stored data_string : str PROFS, MP, RADII, etc., , batch_size : int, optional I think this is actually irrelevant... by default 128 Returns ------- RaggedMmap Shout out to mmap_ninja """ def load_data(paths): for path in paths: yield np.load(path) # TODO: Change the string to include interface Memmap save_name = os.path.join(data_dir, f'{data_string}_MMAP') if os.path.exists(save_name): ragged_mmap = RaggedMmap(save_name) print(f'A ragged map for {data_string} exists at {save_name} with length: {len(ragged_mmap)}') else: relevant_paths = sorted([os.path.join(data_dir, fname) for fname in os.listdir(data_dir) if fname.endswith(f"_{data_string}.npy")]) ragged_mmap = RaggedMmap.from_generator(out_dir=save_name, sample_generator=load_data(relevant_paths), batch_size=batch_size, verbose=True) print(f'Ragged mmap for {data_string} saved to {save_name}') return ragged_mmap class MemMapDataset_O(Dataset): """A mememory mapped dataset. """ def __init__(self, data_path: str, device: str = 'cpu', filter_mps = None): self.data = {} self.data['profs'] = compute_raggedmmap(data_path, data_string='PROFS') self.data['mps'] = compute_raggedmmap(data_path, data_string='MP') self.total_num_pulses = len(self.data['profs']) self.list_num_slices: list = [len(pulse) for pulse in self.data['profs']] self.cumsum_num_slices: np.ndarray = np.cumsum(self.list_num_slices) - 1 self.total_num_slices = sum(self.list_num_slices) self.prof_length = self.data['profs'][0].shape[-1] self.mp_dim = self.data['mps'][0].shape[-1] self.factor = np.ones((2,self.prof_length)) self.factor[0, :] *= 1e-19 self.device = None self.filter_mps = filter_mps def get_pulse_idx(self, idx: int) -> Tuple[int, int]: """A fun function that calculates which slice to take from the mmap Since the mmap is (imo) a list of pulses, we need to find which pulse the queried idx is coming from. This is calculated by looking at the,m minimum value of the cumulative sum of all the slices across pulses that are >= idx Then the internal pulse slice idx is just queried idx subtracted by the cumulative sum up to that pulse ( + 1) Parameters ---------- idx : int a given slice, which can take on values of [0 -> total_num_slices] Returns ------- Tuple[int, int] : (pulse_idx, slice_idx) """ # list_num_slices: list = [len(pulse) for pulse in self.prof_mmap] # cumsum_num_slices: np.ndarray = np.cumsum(self.list_num_slices) - 1 pulse_idx: int = np.where(self.cumsum_num_slices >= idx)[0][0] slice_idx: int = (idx - (self.cumsum_num_slices[pulse_idx] + 1)) return pulse_idx, slice_idx def get_norms(self, relevant_idxs, return_norms=True) -> None: profs = self.data['profs'][relevant_idxs] mps = self.data['mps'][relevant_idxs] self.prof_means = np.mean(np.stack([np.mean(prof*self.factor, axis=0) for prof in profs], 0), 0) self.prof_stds = np.mean(np.stack([np.std(prof*self.factor, axis=0) for prof in profs], 0), 0) self.mp_means = np.mean(np.stack([np.mean(mp, axis=0) for mp in mps], 0), 0) self.mp_stds = np.mean(np.stack([np.std(mp, axis=0) for mp in mps], 0), 0) if self.device is not None: self.prof_means, self.prof_stds, self.mp_means, self.mp_stds = self.prof_means.to(self.device), self.prof_stds.to(self.device), self.mp_means.to(self.device), self.mp_stds.to(self.device) return self.return_norms() def set_norms(self, prof_means, prof_stds, mp_means, mp_stds) -> None: self.prof_means, self.prof_stds, self.mp_means, self.mp_stds = prof_means, prof_stds, mp_means, mp_stds def return_norms(self) -> dict: return dict(prof_means=self.prof_means, prof_stds=self.prof_stds, mp_means=self.mp_means, mp_stds=self.mp_stds) def norm_profiles(self, profiles): return (profiles*self.factor - self.prof_means) / self.prof_stds def denorm_profiles(self, profiles, to_torch=False): if not to_torch: return ((profiles*self.prof_stds) + self.prof_means) / self.factor else: return ((profiles*torch.from_numpy(self.prof_stds)) + torch.from_numpy(self.prof_means)) / torch.from_numpy(self.factor) def norm_mps(self, mps): return (mps - self.mp_means) / self.mp_stds def denorm_mps(self, mps, to_torch=False): if to_torch: return (mps*torch.from_numpy(self.mp_stds)) + torch.from_numpy(self.mp_means) if self.filter_mps is not None: # mps = self.filter_mps(mps) return (mps*self.filter_mps(self.mp_stds)) + self.filter_mps(self.mp_means) else: return (mps*self.mp_stds) + self.mp_means def __len__(self): return self.total_num_slices def __getitem__(self, idx): pulse_idx_to_take_from, slice_idx_to_take_from = self.get_pulse_idx(idx) sample_pulse_profs = self.data['profs'][pulse_idx_to_take_from] sample_pulse_mps = self.data['mps'][pulse_idx_to_take_from] profs = sample_pulse_profs[slice_idx_to_take_from] mps = sample_pulse_mps[slice_idx_to_take_from] profs, mps = self.norm_profiles(profs), self.norm_mps(mps) if self.filter_mps is not None: mps = self.filter_mps(mps) return profs, mps
6,190
43.221429
199
py
LLPFC
LLPFC-main/main.py
import json import sys import numpy as np # set the random seed for torchvision import logging import torch from torch.utils.data.sampler import SubsetRandomSampler from llpfc import llpfc from kl import kl from llpvat import llpvat from llpgan import llpgan from utils import set_optimizer, set_device, set_reproducibility, set_data_and_model, set_dataset_class, get_args, set_generator def main(args): set_reproducibility(args) logging.basicConfig(level=logging.INFO, filename=args.logging_filename, filemode="a+", format="%(asctime)-15s %(levelname)-8s %(message)s") logger = logging.getLogger() device = set_device(args) logger.info("\n\n") logger.info("program starts") logger.info("running arguments %s" % sys.argv) json_data = dict() json_data['args'] = sys.argv json_data['epoch_vs_test_accuracy'] = [] llp_data, transform_train, num_classes, model, test_loader = set_data_and_model(args) model = model.to(device) total_epochs = args.total_epochs optimizer, scheduler = set_optimizer(args, model, total_epochs) if args.algorithm == "llpfc": dataset_class = set_dataset_class(args) llpfc(llp_data, transform_train, scheduler, model, optimizer, test_loader, dataset_class, device, args, logger, json_data) elif args.algorithm == "kl": dataset_class = set_dataset_class(args) training_data, bag2indices, bag2size, bag2prop = llp_data kl_train_dataset = dataset_class(training_data, bag2indices, bag2prop, transform_train) train_loader = torch.utils.data.DataLoader(dataset=kl_train_dataset, batch_size=args.train_batch_size, shuffle=True) val_loader = None kl(model, optimizer, train_loader, scheduler, total_epochs, val_loader, test_loader, device, logger, json_data) elif args.algorithm == "llpvat": dataset_class = set_dataset_class(args) training_data, bag2indices, bag2size, bag2prop = llp_data kl_train_dataset = dataset_class(training_data, bag2indices, bag2prop, transform_train) llpvat(kl_train_dataset, scheduler, model, optimizer, test_loader, device, args, logger, json_data) elif args.algorithm == "llpgan": dataset_class = set_dataset_class(args) training_data, bag2indices, bag2size, bag2prop = llp_data kl_train_dataset = dataset_class(training_data, bag2indices, bag2prop, transform_train) gen = set_generator(args) gen = gen.to(device) gen_opt, gen_sch = set_optimizer(args, gen, total_epochs) llpgan(kl_train_dataset, model, gen, optimizer, gen_opt, scheduler, gen_sch, test_loader, device, args, logger, json_data) if args.save_path is not None: torch.save(model.state_dict(), args.save_path) if args.path_to_json is not None: with open(args.path_to_json, 'w') as f: json.dump(json_data, f) logger.info("training completed") logger.info("") if __name__ == "__main__": parser = get_args() main(parser)
3,444
35.648936
128
py
LLPFC
LLPFC-main/llpvat.py
import torch import torch.nn as nn from llpvatlib.train_fun import llpvat_train_by_bag from llpfclib.train_fun import test_model from llpvatlib.utils import VATLoss def loss_f_test(x, y, device, epsilon=1e-8): x = torch.clamp(x, epsilon, 1 - epsilon) return nn.functional.nll_loss(torch.log(x), y, reduction='sum') def llpvat(kl_train_dataset, scheduler, model, optimizer, test_loader, device, args, logger, json_data): train_loader = torch.utils.data.DataLoader(dataset=kl_train_dataset, batch_size=args.train_batch_size, shuffle=True) vat_loss_f = VATLoss(xi=args.vat_xi, eps=args.vat_eps, ip=args.vat_ip).to(device) for epoch in range(args.total_epochs): logger.info(f"Epoch-{epoch}") logger.info(f" lr: {optimizer.param_groups[0]['lr']}") llpvat_train_by_bag(model, optimizer, train_loader, vat_loss_f, epoch, device, scheduler, logger) if test_loader is not None: acc, test_error = test_model(model, test_loader, loss_f_test, device) logger.info(f" test_error = {test_error}, accuracy = {100 * acc}%") json_data['epoch_vs_test_accuracy'].append({'epoch': epoch, 'test_acc': acc, 'test_error': test_error})
1,215
49.666667
120
py
LLPFC
LLPFC-main/llpfc.py
import torch import torch.nn as nn from torch.distributions.constraints import simplex from torch.utils.data import SubsetRandomSampler import numpy as np from llpfclib.make_groups import make_groups_forward from llpfclib.train_fun import train_model_forward_one_epoch, test_model, validate_model_forward def loss_f(x, y, weights, device, epsilon=1e-8): assert torch.all(simplex.check(x)) x = torch.clamp(x, epsilon, 1 - epsilon) unweighted = nn.functional.nll_loss(torch.log(x), y, reduction='none') weights /= weights.sum() return (unweighted * weights).sum() def loss_f_val(x, y, weights, device, epsilon=1e-8): assert torch.all(simplex.check(x)) x = torch.clamp(x, epsilon, 1 - epsilon) unweighted = nn.functional.nll_loss(torch.log(x), y, reduction='none') return (unweighted * weights).sum() def loss_f_test(x, y, device, epsilon=1e-8): x = torch.clamp(x, epsilon, 1 - epsilon) return nn.functional.nll_loss(torch.log(x), y, reduction='sum') def llpfc(llp_data, transform_train, scheduler, model, optimizer, test_loader, dataset_class, device, args, logger, json_data): training_data, bag2indices, bag2size, bag2prop = llp_data num_regroup = -1 train_sampler = None valid_sampler = None llp_valid_loader = None for epoch in range(args.total_epochs): if epoch % args.num_epoch_regroup == 0: instance2group, group2transition, instance2weight, noisy_y = make_groups_forward(args.num_classes, bag2indices, bag2size, bag2prop, args.noisy_prior_choice, args.weights, logger) fc_train_dataset = dataset_class(training_data, noisy_y, group2transition, instance2weight, instance2group, transform_train) if (llp_valid_loader is None) and args.validate: # always use the first group assigment to validate VAL_PROP = 0.1 num_data_points = len(fc_train_dataset) split = int(np.floor(VAL_PROP * num_data_points)) indices = list(range(num_data_points)) np.random.shuffle(indices) train_indices, val_indices = indices[split:], indices[:split] train_sampler = SubsetRandomSampler(train_indices) valid_sampler = SubsetRandomSampler(val_indices) llp_valid_loader = torch.utils.data.DataLoader(dataset=fc_train_dataset, sampler=valid_sampler, batch_size=args.train_batch_size) if train_sampler is None: llp_train_loader = torch.utils.data.DataLoader(dataset=fc_train_dataset, shuffle=True, batch_size=args.train_batch_size) else: llp_train_loader = torch.utils.data.DataLoader(dataset=fc_train_dataset, sampler=train_sampler, batch_size=args.train_batch_size) num_regroup += 1 logger.info(f"Regroup-{num_regroup} Epoch-{epoch}") logger.info(f" lr: {optimizer.param_groups[0]['lr']}") train_model_forward_one_epoch(model, loss_f, optimizer, llp_train_loader, device, epoch, scheduler, logger) if test_loader is not None: acc, test_error = test_model(model, test_loader, loss_f_test, device) logger.info(f" test_error = {test_error}, accuracy = {100 * acc}%") json_data['epoch_vs_test_accuracy'].append({'epoch': epoch, 'test_acc': acc, 'test_error': test_error}) if args.validate: assert llp_valid_loader is not None val_loss = validate_model_forward(model, loss_f_val, llp_valid_loader, device) logger.info(f" valid_loss = {val_loss}")
4,654
48.521277
117
py
LLPFC
LLPFC-main/llpgan.py
import torch import torch.nn as nn from llpganlib.train_fun import llpgan_train_by_bag, test_llpgan def loss_f_test(x, y, device, epsilon=1e-8): x = torch.clamp(x, epsilon, 1 - epsilon) return nn.functional.nll_loss(torch.log(x), y, reduction='sum') def llpgan(kl_train_dataset, dis, gen, dis_opt, gen_opt, dis_sch, gen_sch, test_loader, device, args, logger, json_data): train_loader = torch.utils.data.DataLoader(dataset=kl_train_dataset, batch_size=args.train_batch_size, shuffle=True) for epoch in range(args.total_epochs): logger.info(f"Epoch-{epoch}") logger.info(f" dis lr: {dis_opt.param_groups[0]['lr']}") logger.info(f" gen lr: {gen_opt.param_groups[0]['lr']}") llpgan_train_by_bag(gen, dis, gen_opt, dis_opt, dis_sch, gen_sch, args.noise_dim, train_loader, epoch, device, logger) if test_loader is not None: acc, test_error = test_llpgan(dis, test_loader, loss_f_test, device) logger.info(f" test_error = {test_error}, accuracy = {100 * acc}%") json_data['epoch_vs_test_accuracy'].append({'epoch': epoch, 'test_acc': acc, 'test_error': test_error})
1,574
35.627907
120
py
LLPFC
LLPFC-main/utils.py
import argparse import pickle import random # set the random seed for torchvision import numpy as np import torch from models.NIN import NIN from models.WideRes import wide_resnet_d_w from models.ResNet import resnet18 from models.vgg import vgg19_bn, vgg16_bn from models.densenet import densenet121 from models.LLPGAN_GEN import LLPGAN_GEN_MNIST, LLPGAN_GEN_COLOR from llpfclib.utils import FORWARD_CORRECT_MNIST, FORWARD_CORRECT_CIFAR10, FORWARD_CORRECT_SVHN from kllib.utils import KL_CIFAR10, KL_SVHN, KL_EMNIST import torch.optim as optim import torchvision import torchvision.transforms as transforms class InvalidArguments(Exception): pass def get_args(): parser = argparse.ArgumentParser(description="train a model on LLP data") # required: parser.add_argument("-d", "--dataset", nargs='?', choices=["cifar10", "svhn", "emnist_letters"], required=True, help="name of the dataset, the program uses torchvision.datasets") parser.add_argument("-p", "--path_lp", nargs='?', required=True, help="path to the label proportion dataset generated by make_data.py") parser.add_argument("-c", "--num_classes", nargs='?', type=int, required=True, help="number of classes") parser.add_argument("-f", "--data_folder_labeled", nargs='?', required=True, help="path to the folder of labeled test data, if not exists, the dataset will be downloaded") parser.add_argument("-log", "--logging_filename", nargs='?', required=True, help="path to save the log file") # optional: parser.add_argument("-a", "--algorithm", nargs='?', choices=["llpfc", "kl", "llpvat", "llpgan"], default="llpfc", help="choose a training algorithm") parser.add_argument("-n", "--network", nargs='?', choices=["wide_resnet_d_w", "nin", "ResNet18", "vgg19_bn", "vgg16_bn", "densenet121"], default="wide_resnet_d_w", help="the neural network model") parser.add_argument("-wrnd", "--WideResNet_depth", nargs='?', type=int, default=28) parser.add_argument("-wrnw", "--WideResNet_width", nargs='?', type=int, default=2) parser.add_argument("-dr", "--drop_rate", nargs="?", type=float, default=0.3, help="the drop rate in dropout layers, for wide resnet") # add more to this parser.add_argument("-o", "--optimizer", nargs="?", default="Adamax", choices=["Adamax", "LBFGS", "Adagrad", "nesterov", "AdamW", "SGD"], help="optimizer of the neural network") parser.add_argument("-ams", "--amsgrad", nargs="?", type=int, default=0, choices=[0, 1], help="whether to use the AMSGrad variant of this algorithm") parser.add_argument("-l", "--lr", nargs="?", type=float, default=1e-3, help="learning rate") parser.add_argument("-m", "--momentum", nargs="?", type=float, default=0.9, help="momentum") parser.add_argument("-wd", "--weight_decay", nargs="?", type=float, default=0, help="weight decay") parser.add_argument("-e", "--total_epochs", nargs="?", type=int, default=200, help="total number of epochs to train") parser.add_argument("-r", "--num_epoch_regroup", nargs="?", type=int, default=20, help="groups will be regenerated every this number of epochs, " "only effective if the algorithm is llpfc") parser.add_argument("-np", "--noisy_prior_choice", nargs="?", type=str, default="approx", choices=["approx", "uniform", "merge"], help="the heuristics to estimate the noisy prior for each group, " "approx solves the constrained optimization and uniform assigns uniform noisy priors") parser.add_argument("-v", "--validate", nargs='?', type=int, default=0, choices=[0, 1], help="if True, then validate on 10%% of the training data set; " "if False, output testing loss and accuracy while training" ) parser.add_argument("-b", "--train_batch_size", nargs='?', type=int, default=128, help="training batch size") parser.add_argument("-t", "--test_batch_size", nargs="?", type=int, default=256, help="test batch size") parser.add_argument("-s", "--save_path", nargs='?', default=None, help="path to save the trained model, model will not be saved if the path is None") parser.add_argument("-dv", "--device", nargs='?', default="check", choices=["cuda", "cpu", "check"], help="device to train network; if it's check, use cuda whenever it's available") parser.add_argument("-w", "--weights", nargs='?', choices=["uniform", "ch_vol"], default="uniform", help="set the weights for each group in llpfc") parser.add_argument("-sc", "--scheduler", nargs='?', choices=["drop", "CAWR"], default="drop", help="set the scheduler of training lr") parser.add_argument("-ms", "--milestones", nargs='+', type=int, default=[], help="number of epochs to drop lr if --scheduler is set to be 'drop'") parser.add_argument("-ga", "--gamma", nargs='?', type=float, default=0.1, help="drop the learning rate by this factor if --scheduler is set to be 'drop'") parser.add_argument("-T0", "--T_0", nargs='?', type=int, default=10, help="parameter of the CAWR scheduler") parser.add_argument("-Tm", "--T_mult", nargs='?', type=int, default=1, help="parameter of the CAWR scheduler") parser.add_argument("--seed", nargs='?', type=int, help="seed for all RNG") parser.add_argument("-fr", "--full_reproducibility", nargs='?', type=int, default=0, choices=[0, 1], help="choose to disable all nondeterministic algorithms, may at the cost of performance, " " decrypted from now") # xi=1e-6, eps=6.0, ip=1 parser.add_argument("-xi", "--vat_xi", nargs='?', type=float, default=1e-6, help="parameter for vat loss, effective only algorithm=llpvat") parser.add_argument("-eps", "--vat_eps", nargs='?', type=float, default=6.0, help="parameter for vat loss, effective only algorithm=llpvat") parser.add_argument("-ip", "--vat_ip", nargs='?', type=float, default=1, help="parameter for vat loss, effective only algorithm=llpvat") parser.add_argument("-nd", "--noise_dim", nargs='?', type=int, default=500, help="parameter for llpgan, the input dimension of the generator") parser.add_argument("-js", "--path_to_json", nargs='?', type=str, default=None, help="will write the training results to this path if provided, write nothing if is none") return parser.parse_args() def set_reproducibility(args): # random seed: if args.seed is not None: np.random.seed(args.seed) torch.random.manual_seed(args.seed) random.seed(args.seed) # use deterministic algorithms if args.full_reproducibility: # look https://pytorch.org/docs/stable/notes/randomness.html for references torch.backends.cudnn.benchmark = False torch.use_deterministic_algorithms(True) def set_device(args): # configure device if args.device == "check": device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') else: device = torch.device(args.device) return device def set_data_and_model(args): # read the training data llp_data = pickle.load(open(args.path_lp, "rb")) if args.dataset == "cifar10": transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)), # mean-std of cifar10 ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)), # mean-std of cifar10 ]) test_dataset = torchvision.datasets.CIFAR10(root=args.data_folder_labeled, train=False, transform=transform_test, download=True) image_size = 32 in_channel = 3 elif args.dataset == "svhn": transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4376821, 0.4437697, 0.47280442), (0.19803012, 0.20101562, 0.19703614)), # mean-std of svhn ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4376821, 0.4437697, 0.47280442), (0.19803012, 0.20101562, 0.19703614)), # mean-std of svhn ]) test_dataset = torchvision.datasets.SVHN(root=args.data_folder_labeled, split='test', transform=transform_test, download=True) image_size = 32 in_channel = 3 elif args.dataset == "emnist_letters": image_size = 28 in_channel = 1 transform_train = transforms.Compose([ transforms.RandomCrop(28, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.1736, ), (0.3317, )), # mean-std of emnist ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.1736, ), (0.3317, )), # mean-std of emnist ]) if (args.network == "densenet121") or (len(args.network) >=3 and args.network[:3] == "vgg"): transform_train = transforms.Compose([ transforms.Resize(32), # resize the image for dense net as it has too many pool layers transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.1736,), (0.3317,)), # mean-std of emnist ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.1736,), (0.3317,)), # mean-std of emnist ]) image_size = 32 test_dataset = torchvision.datasets.EMNIST(root=args.data_folder_labeled, split="letters", train=False, transform=transform_test, download=True) test_dataset.targets = test_dataset.targets - 1 # the labels range originally from 1 to 26 else: raise InvalidArguments("Unknown dataset name: ", args.dataset) test_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=args.test_batch_size, shuffle=False) num_classes = args.num_classes return_features = False if args.algorithm == "llpgan": return_features = True if args.network == "wide_resnet_d_w": model = wide_resnet_d_w(d=args.WideResNet_depth, w=args.WideResNet_width, dropout_rate=args.drop_rate, num_classes=num_classes, in_channel=in_channel, image_size=image_size, return_features=return_features ) elif args.network == "nin": model = NIN(num_classes=num_classes, image_size=image_size, in_channel=in_channel, ) if args.algorithm == "llpgan": raise InvalidArguments("NIN is not compatible with LLPGAN as it has no fully connected layer") elif args.network == "ResNet18": model = resnet18(num_classes, in_channel, return_features=return_features) elif args.network == "vgg19_bn": model = vgg19_bn(num_classes, in_channel, return_features=return_features) elif args.network == "vgg16_bn": model = vgg16_bn(num_classes, in_channel, return_features=return_features) elif args.network == "densenet121": model = densenet121(num_classes, in_channel, memory_efficient=False, return_features=return_features) else: raise InvalidArguments("Unknown selection of network: ", args.network) return llp_data, transform_train, num_classes, model, test_loader def set_optimizer(args, model, total_epochs): if args.optimizer == "Adamax": optimizer = optim.Adamax(model.parameters(), lr=args.lr, betas=(0.9, 0.999), eps=1e-08, weight_decay=args.weight_decay) elif args.optimizer == "Adagrad": optimizer = optim.Adagrad(model.parameters(), lr=args.lr, weight_decay=args.weight_decay) elif args.optimizer == "LBFGS": optimizer = optim.LBFGS(model.parameters(), lr=args.lr) # ToDo: implement the closure function and test elif args.optimizer == "nesterov": optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay, nesterov=True) elif args.optimizer == "SGD": optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay, nesterov=False) elif args.optimizer == "AdamW": optimizer = optim.AdamW(model.parameters(), lr=args.lr, betas=(0.9, 0.999), eps=1e-08, weight_decay=args.weight_decay, amsgrad=bool(args.amsgrad)) else: raise InvalidArguments("Unknown selection of optimizer: ", args.optimizer) if args.scheduler == "drop": scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.milestones, gamma=args.gamma) elif args.scheduler == "CAWR": scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=args.T_0, T_mult=args.T_mult) else: raise InvalidArguments("Unknown selection of scheduler: ", args.scheduler) return optimizer, scheduler def set_dataset_class(args): competitors = ["kl", "llpvat", "llpgan"] if args.algorithm == "llpfc" and args.dataset == "cifar10": return FORWARD_CORRECT_CIFAR10 elif args.algorithm == "llpfc" and args.dataset == "svhn": return FORWARD_CORRECT_SVHN elif args.algorithm == "llpfc" and args.dataset == "emnist_letters": return FORWARD_CORRECT_MNIST elif (args.algorithm in competitors) and args.dataset == "cifar10": return KL_CIFAR10 elif (args.algorithm in competitors) and args.dataset == "svhn": return KL_SVHN elif (args.algorithm in competitors) and args.dataset == "emnist_letters": return KL_EMNIST raise InvalidArguments("Unknown combination of llp algorithm and dataset" ": (%s, %s)" % (args.algorithm, args.dataset)) def set_generator(args): # return a tuple of (generator, optimizer of generator, ) if args.dataset in ["cifar10", "svhn"]: return LLPGAN_GEN_COLOR(args.noise_dim) elif args.dataset == "emnist_letters": if (args.network == "densenet121") or (len(args.network) >= 3 and args.network[:3] == "vgg"): return LLPGAN_GEN_MNIST(args.noise_dim, 32, 32) return LLPGAN_GEN_MNIST(args.noise_dim, 28, 28) else: raise InvalidArguments("Unknown choice of dataset: %s" % args.dataset)
18,095
46.248042
127
py
LLPFC
LLPFC-main/kl.py
import torch import torch.nn as nn from kllib.train_fun import kl_train_by_bag, validate_model_kl from llpfclib.train_fun import test_model def loss_f_test(x, y, device, epsilon=1e-8): x = torch.clamp(x, epsilon, 1 - epsilon) return nn.functional.nll_loss(torch.log(x), y, reduction='sum') def kl(model, optimizer, train_loader, scheduler, num_epochs, val_loader, test_loader, device, logger, json_data): for epoch in range(num_epochs): logger.info(f"Epoch-{epoch}") logger.info(f" lr: {optimizer.param_groups[0]['lr']}") kl_train_by_bag(model, optimizer, train_loader, epoch, device, scheduler, logger) if test_loader is not None: acc, test_error = test_model(model, test_loader, loss_f_test, device) logger.info(f" test_error = {test_error}, accuracy = {100 * acc}%") json_data['epoch_vs_test_accuracy'].append({'epoch': epoch, 'test_acc': acc, 'test_error': test_error}) if val_loader is not None: val_error = validate_model_kl(model, val_loader, device) logger.info(f" val_error = {val_error}")
1,134
46.291667
115
py
LLPFC
LLPFC-main/make_data.py
import argparse from llpfclib.make_bags import make_bags_dirichlet, InsufficientDataPoints, make_bags_uniform, truncate_data import torchvision import pickle import os import random # set random seed import numpy as np # set random seed class InvalidArguments(Exception): pass def get_args(): parser = argparse.ArgumentParser(description="Partition data into bags for LLP") # required: parser.add_argument("-d", "--dataset", nargs='?', choices=["cifar10", "svhn", "emnist_letters"], required=True, help="name of the dataset, the program uses torchvision.datasets") # ToDo: add more data sets later parser.add_argument("-c", "--num_classes", nargs='?', type=int, required=True, metavar="10", help="number of classes") parser.add_argument("-s", "--data_save_name", nargs='?', required=True, metavar="cifar10_1024_0", help="name of the file to save") # optional: parser.add_argument("-l", "--data_folder_labeled", nargs='?', default="../data/labeled_data/", metavar="../data/labeled_data/", help="path to the folder of original data, if not exists, the dataset will be downloaded") parser.add_argument("-p", "--data_folder_llp", nargs='?', default="../data/llp_data/", metavar="../data/llp_data/", help="path to save the training data for llp") parser.add_argument("-m", "--method", nargs='?', default="dirichlet", choices=["dirichlet", "uniform"], help="method to generate bags") # dirichlet, uniform parser.add_argument("-a", "--alpha", nargs="?", default="equal", choices=["equal"], help="parameter of dirichlet distribution; required if use dirichlet to generate bags") # Todo: add more to this parser.add_argument("-b", "--bag_size", nargs='?', type=int, metavar="1024", help="size of bag; note not all bag sizes will equal to this number if use dirichlet") parser.add_argument("-n", "--num_bags", nargs='?', type=int, metavar="100", help="number of bags to generate; it too large, the dataset may have insufficient data points") parser.add_argument("-r", "--seed", nargs='?', type=int, metavar="0", help="seed for all RNG") # both random and numpy.random will use this seed return parser.parse_args() def main(args): if args.dataset == "cifar10": train_dataset = torchvision.datasets.CIFAR10(root=args.data_folder_labeled, train=True, download=True) labels = train_dataset.targets elif args.dataset == "svhn": train_dataset = torchvision.datasets.SVHN(root=args.data_folder_labeled, split="train", download=True) labels = train_dataset.labels elif args.dataset == "emnist_letters": train_dataset = torchvision.datasets.EMNIST(root=args.data_folder_labeled, split="letters", train=True, download=True) labels = train_dataset.targets - 1 # the labels range originally from 1 to 26 else: raise InvalidArguments("Unknown dataset name: ", args.dataset) if args.method == "dirichlet": if args.alpha == "equal": alpha = tuple([1 for _ in range(args.num_classes)]) else: raise InvalidArguments("Unknown choice of alpha: ", args.alpha) flag = 1 fail_counter = 0 while flag: try: bag2indices, bag2size, bag2prop = make_bags_dirichlet(labels, num_classes=args.num_classes, bag_size=args.bag_size, num_bags=args.num_bags, alpha=alpha) flag = 0 except InsufficientDataPoints: flag = 1 fail_counter += 1 if fail_counter >= 100: raise InsufficientDataPoints("THE DATA GENERATION PROCESS FAILS FOR 100 TIMES CONSECUTIVELY. " "PLEASE CHECK ARGUMENTS OF --alpha %s, --bag_size %d, --num_bags %d" % (args.alpha, args.bag_size, args.num_bags)) continue elif args.method == "uniform": bag2indices, bag2size, bag2prop = make_bags_uniform(train_dataset.targets, args.num_classes, args.bag_size, args.num_bags) else: raise InvalidArguments("Unknown method to generate bags: ", args.method) print("%d of bags generated, each bag has size %d, the random seed is %d, data is saved as %s" % (len(bag2indices.keys()), len(bag2indices[0]), args.seed, args.data_save_name)) training_data, bag2indices = truncate_data(train_dataset.data, bag2indices) to_save = [training_data, bag2indices, bag2size, bag2prop] with open(os.path.join(args.data_folder_llp, args.data_save_name), 'wb') as f: pickle.dump(to_save, f) if __name__ == "__main__": args = get_args() if args.seed is not None: random.seed(args.seed) np.random.seed(args.seed) main(args)
5,251
52.050505
149
py
LLPFC
LLPFC-main/llpfclib/utils.py
import torch from PIL import Image from torch.utils.data import Sampler import numpy as np def truncate_data_group(x, y, instance2group): idx_list = [] for i in range(x.shape[0]): if instance2group[i] != -1: idx_list.append(i) x_truncated = x[idx_list] y_truncated = y[idx_list] idx2new = {idx_list[i]: i for i in range(len(idx_list))} instance2group_new = {} for old, new in idx2new.items(): instance2group_new[new] = instance2group[old] new2idx = {idx2new[idx]: idx for idx in idx2new.keys()} return x_truncated, y_truncated, instance2group_new, new2idx class LLPFC_DATASET_BASE(torch.utils.data.Dataset): def __init__(self, data, noisy_y, group2transition, instance2weight, instance2group, transform): self.data, self.noisy_y, self.instance2group, self.new2idx = truncate_data_group(data, noisy_y, instance2group) self.group2transition = group2transition self.instance2weight = instance2weight self.transform = transform def __len__(self): return len(self.data) class FORWARD_CORRECT_CIFAR10(LLPFC_DATASET_BASE): def __getitem__(self, index): img, y_ = self.data[index], self.noisy_y[index] trans_m = self.group2transition[self.instance2group[index]] weight = self.instance2weight[self.new2idx[index]] img = Image.fromarray(img) if self.transform is not None: img = self.transform(img) return img, int(y_), torch.tensor(trans_m, dtype=None), weight class FORWARD_CORRECT_SVHN(LLPFC_DATASET_BASE): def __getitem__(self, index): img, y_ = self.data[index], self.noisy_y[index] img = Image.fromarray(np.transpose(img, (1, 2, 0))) trans_m = self.group2transition[self.instance2group[index]] weight = self.instance2weight[self.new2idx[index]] if self.transform is not None: img = self.transform(img) return img, int(y_), torch.tensor(trans_m, dtype=None), weight class FORWARD_CORRECT_MNIST(LLPFC_DATASET_BASE): # this should work for both EMNIST and MNIST def __getitem__(self, index): img, y_ = self.data[index], self.noisy_y[index] trans_m = self.group2transition[self.instance2group[index]] weight = self.instance2weight[self.new2idx[index]] img = Image.fromarray(img.numpy(), mode='L') if self.transform is not None: img = self.transform(img) return img, int(y_), torch.tensor(trans_m, dtype=None), weight
2,296
34.890625
113
py
LLPFC
LLPFC-main/llpfclib/train_fun.py
import torch import torch.nn as nn def test_model(model, test_loader, criterion, device): # test a model with fully label dataset model.eval() with torch.no_grad(): correct = 0 total = 0 total_loss = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs = model(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() prob = nn.functional.softmax(outputs, dim=1) loss = criterion(prob, labels, device) total_loss += loss.item() return correct / total, total_loss / total def validate_model_forward(model, loss_f_val, val_loader, device): model.eval() total_loss = 0 total = 0 for i, (images, noisy_y, trans_m, weights) in enumerate(val_loader): total_loss += compute_forward_loss_on_minibatch(model, loss_f_val, images, noisy_y, trans_m, weights, device).item() total += noisy_y.size(0) return total_loss / total def train_model_forward_one_epoch(model, loss_f, optimizer, train_loader, device, epoch, scheduler, logger): # train the model one epoch with forward correction # label input of loss_f must be an integer model.train() total_step = len(train_loader) for i, (images, noisy_y, trans_m, weights) in enumerate(train_loader): loss = compute_forward_loss_on_minibatch(model, loss_f, images, noisy_y, trans_m, weights, device) # Backward pass optimizer.zero_grad() loss.backward() optimizer.step() if (i + 1) % 100 == 0: logger.info(' Step [{}/{}], Loss: {:.4f}'.format(i + 1, total_step, loss.item())) if type(scheduler) == torch.optim.lr_scheduler.CosineAnnealingWarmRestarts: scheduler.step(epoch + i / total_step) if type(scheduler) == torch.optim.lr_scheduler.MultiStepLR: scheduler.step() elif type(scheduler) == torch.optim.lr_scheduler.ReduceLROnPlateau: scheduler.step(validate_model_forward(model, loss_f, train_loader, device)) def compute_forward_loss_on_minibatch(model, loss_f, images, noisy_y, trans_m, weights, device): # Move tensors to the configured device images = images.to(device) noisy_y = noisy_y.to(device) trans_m = trans_m.to(device) weights = weights.to(device) # Forward pass outputs = model(images) prob = nn.functional.softmax(outputs, dim=1) prob_corrected = torch.bmm(trans_m.float(), prob.reshape(prob.shape[0], -1, 1)).reshape(prob.shape[0], -1) loss = loss_f(prob_corrected, noisy_y, weights, device) return loss
2,465
34.73913
118
py
LLPFC
LLPFC-main/models/ResNet.py
# code in this file is modified from https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py import torch from torch import Tensor import torch.nn as nn from typing import Type, Any, Callable, Union, List, Optional def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d: """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=dilation, groups=groups, bias=False, dilation=dilation) def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d: """1x1 convolution""" return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) class BasicBlock(nn.Module): expansion: int = 1 def __init__( self, inplanes: int, planes: int, stride: int = 1, downsample: Optional[nn.Module] = None, groups: int = 1, base_width: int = 64, dilation: int = 1, norm_layer: Optional[Callable[..., nn.Module]] = None ) -> None: super(BasicBlock, self).__init__() if norm_layer is None: norm_layer = nn.BatchNorm2d if groups != 1 or base_width != 64: raise ValueError('BasicBlock only supports groups=1 and base_width=64') if dilation > 1: raise NotImplementedError("Dilation > 1 not supported in BasicBlock") # Both self.conv1 and self.downsample layers downsample the input when stride != 1 self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = norm_layer(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = norm_layer(planes) self.downsample = downsample self.stride = stride def forward(self, x: Tensor) -> Tensor: identity = 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: identity = self.downsample(x) out += identity out = self.relu(out) return out class Bottleneck(nn.Module): # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2) # while original implementation places the stride at the first 1x1 convolution(self.conv1) # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385. # This variant is also known as ResNet V1.5 and improves accuracy according to # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch. expansion: int = 4 def __init__( self, inplanes: int, planes: int, stride: int = 1, downsample: Optional[nn.Module] = None, groups: int = 1, base_width: int = 64, dilation: int = 1, norm_layer: Optional[Callable[..., nn.Module]] = None ) -> None: super(Bottleneck, self).__init__() if norm_layer is None: norm_layer = nn.BatchNorm2d width = int(planes * (base_width / 64.)) * groups # Both self.conv2 and self.downsample layers downsample the input when stride != 1 self.conv1 = conv1x1(inplanes, width) self.bn1 = norm_layer(width) self.conv2 = conv3x3(width, width, stride, groups, dilation) self.bn2 = norm_layer(width) self.conv3 = conv1x1(width, planes * self.expansion) self.bn3 = norm_layer(planes * self.expansion) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x: Tensor) -> Tensor: identity = 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: identity = self.downsample(x) out += identity out = self.relu(out) return out class ResNet(nn.Module): def __init__(self, block: Type[Union[BasicBlock, Bottleneck]], layers: List[int], num_classes: int, in_channel: int, # specify color/gray image zero_init_residual: bool = False, groups: int = 1, width_per_group: int = 64, replace_stride_with_dilation: Optional[List[bool]] = None, norm_layer: Optional[Callable[..., nn.Module]] = None, return_features: bool = False, ) -> None: super(ResNet, self).__init__() if norm_layer is None: norm_layer = nn.BatchNorm2d self._norm_layer = norm_layer self.inplanes = 64 self.dilation = 1 if replace_stride_with_dilation is None: # each element in the tuple indicates if we should replace # the 2x2 stride with a dilated convolution instead replace_stride_with_dilation = [False, False, False] if len(replace_stride_with_dilation) != 3: raise ValueError("replace_stride_with_dilation should be None " "or a 3-element tuple, got {}".format(replace_stride_with_dilation)) self.groups = groups self.base_width = width_per_group self.conv1 = nn.Conv2d(in_channel, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = norm_layer(self.inplanes) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0]) self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1]) self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2]) self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) self.fc = nn.Linear(512 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) # Zero-initialize the last BN in each residual branch, # so that the residual branch starts with zeros, and each residual block behaves like an identity. # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 if zero_init_residual: for m in self.modules(): if isinstance(m, Bottleneck): nn.init.constant_(m.bn3.weight, 0) # type: ignore[arg-type] elif isinstance(m, BasicBlock): nn.init.constant_(m.bn2.weight, 0) # type: ignore[arg-type] self.return_features = return_features def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int, stride: int = 1, dilate: bool = False) -> nn.Sequential: norm_layer = self._norm_layer downsample = None previous_dilation = self.dilation if dilate: self.dilation *= stride stride = 1 if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( conv1x1(self.inplanes, planes * block.expansion, stride), norm_layer(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer)) self.inplanes = planes * block.expansion for _ in range(1, blocks): layers.append(block(self.inplanes, planes, groups=self.groups, base_width=self.base_width, dilation=self.dilation, norm_layer=norm_layer)) return nn.Sequential(*layers) def _forward_impl(self, x: Tensor): # See note [TorchScript super()] x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.avgpool(x) features = torch.flatten(x, 1) out = self.fc(features) if self.return_features: return out, features return out def forward(self, x: Tensor): return self._forward_impl(x) def _resnet( arch: str, block: Type[Union[BasicBlock, Bottleneck]], layers: List[int], num_classes: int, in_channel: int, **kwargs: Any ) -> ResNet: model = ResNet(block, layers, num_classes, in_channel, **kwargs) return model def resnet18(num_classes: int, in_channel: int, **kwargs: Any) -> ResNet: r"""ResNet-18 model from `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_. """ return _resnet('resnet18', BasicBlock, [2, 2, 2, 2], num_classes, in_channel, **kwargs)
9,492
36.820717
111
py
LLPFC
LLPFC-main/models/LLPGAN_DIS.py
import torch.nn as nn class LLPGAN_DIS(nn.Module): # use the same discriminator as LLP-GAN paper def __init__(self, num_class, image_size, in_channel=3, return_features=False): super(LLPGAN_DIS, self).__init__() self.conv_layers = nn.Sequential( nn.Dropout(p=0.2, ), nn.Conv2d(in_channel, 64, 3, padding=1, stride=1), nn.ReLU(), nn.Conv2d(64, 64, 3, padding=1, stride=1), nn.ReLU(), nn.Conv2d(64, 64, 3, padding=1, stride=2), nn.ReLU(), nn.Dropout(p=0.5, ), nn.Conv2d(64, 128, 3, padding=1, stride=1), nn.ReLU(), nn.Conv2d(128, 128, 3, padding=1, stride=1), nn.ReLU(), nn.Conv2d(128, 128, 3, padding=1, stride=2), nn.ReLU(), nn.Dropout(p=0.5, ), nn.Conv2d(128, 256, 3, padding=1, stride=1), nn.ReLU(), nn.Conv2d(256, 128, 1, padding=0, stride=1), nn.ReLU(), nn.Conv2d(128, 64, 1, padding=0, stride=1), nn.ReLU(), ) if isinstance(image_size, int): pool_size = round(round(image_size/2.0)/2.0) else: pool_size = (round(round(image_size[0]/2.0)/2.0), round(round(image_size[1]/2.0)/2.0)) self.pool_layer = nn.AvgPool2d(pool_size, stride=pool_size, ) self.fc_layer = nn.Linear(64, num_class, bias=True) self.return_features = return_features def forward(self, x): x = self.conv_layers(x) features = self.pool_layer(x).reshape(-1, 64) out = self.fc_layer(features) if self.return_features: return out, features return out
1,421
29.913043
89
py
LLPFC
LLPFC-main/models/vgg.py
# code in this file is modified from https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py import torch import torch.nn as nn from typing import Union, List, Dict, Any, cast class VGG(nn.Module): def __init__(self, features: nn.Module, num_classes: int = 1000, init_weights: bool = True, return_features: bool = False ) -> None: super(VGG, self).__init__() self.features = features self.avgpool = nn.AdaptiveAvgPool2d((7, 7)) self.classifier = nn.Sequential( nn.Linear(512 * 7 * 7, 4096), nn.ReLU(True), nn.Dropout(), nn.Linear(4096, 4096), nn.ReLU(True), nn.Dropout(), nn.Linear(4096, num_classes), ) if init_weights: self._initialize_weights() self.return_features = return_features def forward(self, x: torch.Tensor): x = self.features(x) x = self.avgpool(x) features = torch.flatten(x, 1) out = self.classifier(features) if self.return_features: return out, features return out def _initialize_weights(self) -> None: for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') if m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.BatchNorm2d): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) elif isinstance(m, nn.Linear): nn.init.normal_(m.weight, 0, 0.01) nn.init.constant_(m.bias, 0) def make_layers(in_channels: int, cfg: List[Union[str, int]], batch_norm: bool = False) -> nn.Sequential: layers: List[nn.Module] = [] in_channels = in_channels for v in cfg: if v == 'M': layers += [nn.MaxPool2d(kernel_size=2, stride=2)] else: v = cast(int, v) 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 nn.Sequential(*layers) cfgs: Dict[str, List[Union[str, int]]] = { 'A': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 'B': [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 'D': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'], 'E': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'], } def _vgg(num_classes: int, in_channels: int, cfg: str, batch_norm: bool, **kwargs: Any) -> VGG: model = VGG(make_layers(in_channels, cfgs[cfg], batch_norm=batch_norm), num_classes=num_classes, **kwargs) return model def vgg19_bn(num_classes: int, in_channels: int, **kwargs: Any) -> VGG: r"""VGG 19-layer model (configuration 'E') with batch normalization `"Very Deep Convolutional Networks For Large-Scale Image Recognition" <https://arxiv.org/pdf/1409.1556.pdf>`_. """ return _vgg(num_classes, in_channels, 'E', True, **kwargs) def vgg16_bn(num_classes: int, in_channels: int, **kwargs: Any) -> VGG: r"""VGG 16-layer model (configuration "D") with batch normalization `"Very Deep Convolutional Networks For Large-Scale Image Recognition" <https://arxiv.org/pdf/1409.1556.pdf>`_. """ return _vgg(num_classes, in_channels, 'D', True, **kwargs)
3,722
37.78125
114
py
LLPFC
LLPFC-main/models/WideRes.py
# The code in this file is modified from https://github.com/kevinorjohn/LLP-VAT # MIT License # # Copyright (c) 2020 Kuen-Han Tsai # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in all # copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. import numpy as np import torch import torch.nn as nn import torch.nn.functional as F def wide_resnet_d_w(d, w, **kwargs): net = WideResNet(d, w, **kwargs) net.apply(conv_init) return net class GaussianNoise(nn.Module): """ add gasussian noise into feature """ def __init__(self, std): super(GaussianNoise, self).__init__() self.std = std def forward(self, x): zeros_ = torch.zeros_like(x) n = torch.normal(zeros_, std=self.std) return x + n def conv3x3(in_planes, out_planes, stride=1): return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=True) def conv_init(m): classname = m.__class__.__name__ if classname.find('Conv') != -1: nn.init.xavier_uniform_(m.weight, gain=np.sqrt(2)) nn.init.constant_(m.bias, 0) elif classname.find('BatchNorm') != -1: nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) class WideBasic(nn.Module): def __init__(self, in_planes, planes, dropout_rate, stride=1): super(WideBasic, self).__init__() self.bn1 = nn.BatchNorm2d(in_planes) self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, bias=True) self.dropout = nn.Dropout(p=dropout_rate) self.bn2 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=True) self.shortcut = nn.Sequential() if stride != 1 or in_planes != planes: self.shortcut = nn.Sequential( nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride, bias=True), ) def forward(self, x): out = self.dropout(self.conv1(F.relu(self.bn1(x)))) out = self.conv2(F.relu(self.bn2(out))) out += self.shortcut(x) return out class WideResNet(nn.Module): def __init__(self, depth, widen_factor, dropout_rate, num_classes, in_channel, image_size, return_features=False): super(WideResNet, self).__init__() self.in_planes = 16 if image_size == 32: # CIFAR10, SVHN self.pool_size = 8 elif image_size == 28: # EMNIST self.pool_size = 7 assert ((depth - 4) % 6 == 0), 'Wide-resnet depth should be 6n+4' n = (depth - 4) // 6 k = widen_factor print('| Wide-Resnet %dx%d' % (depth, k)) nStages = [16, 16 * k, 32 * k, 64 * k] self.conv1 = conv3x3(in_channel, nStages[0]) self.layer1 = self._wide_layer(WideBasic, nStages[1], n, dropout_rate, stride=1) self.layer2 = self._wide_layer(WideBasic, nStages[2], n, dropout_rate, stride=2) self.layer3 = self._wide_layer(WideBasic, nStages[3], n, dropout_rate, stride=2) self.bn1 = nn.BatchNorm2d(nStages[3], momentum=0.9) self.linear = nn.Linear(nStages[3], num_classes) self.return_features = return_features def _wide_layer(self, block, planes, num_blocks, dropout_rate, stride): strides = [stride] + [1] * (num_blocks - 1) layers = [] for stride in strides: layers.append(block(self.in_planes, planes, dropout_rate, stride)) self.in_planes = planes return nn.Sequential(*layers) def forward(self, x): out = self.conv1(x) out = self.layer1(out) out = self.layer2(out) out = self.layer3(out) out = F.relu(self.bn1(out)) out = F.avg_pool2d(out, self.pool_size) features = out.view(out.size(0), -1) out = self.linear(features) if self.return_features: return out, features return out
5,794
35.21875
118
py
LLPFC
LLPFC-main/models/densenet.py
# code in this file is modified from https://github.com/pytorch/vision/blob/master/torchvision/models/densenet.py import re import torch import torch.nn as nn import torch.nn.functional as F import torch.utils.checkpoint as cp from collections import OrderedDict from torch import Tensor from typing import Any, List, Tuple class _DenseLayer(nn.Module): def __init__(self, num_input_features: int, growth_rate: int, bn_size: int, drop_rate: float, memory_efficient: bool = False ) -> None: super(_DenseLayer, self).__init__() self.norm1: nn.BatchNorm2d self.add_module('norm1', nn.BatchNorm2d(num_input_features)) self.relu1: nn.ReLU self.add_module('relu1', nn.ReLU(inplace=True)) self.conv1: nn.Conv2d self.add_module('conv1', nn.Conv2d(num_input_features, bn_size * growth_rate, kernel_size=1, stride=1, bias=False)) self.norm2: nn.BatchNorm2d self.add_module('norm2', nn.BatchNorm2d(bn_size * growth_rate)) self.relu2: nn.ReLU self.add_module('relu2', nn.ReLU(inplace=True)) self.conv2: nn.Conv2d self.add_module('conv2', nn.Conv2d(bn_size * growth_rate, growth_rate, kernel_size=3, stride=1, padding=1, bias=False)) self.drop_rate = float(drop_rate) self.memory_efficient = memory_efficient def bn_function(self, inputs: List[Tensor]) -> Tensor: concated_features = torch.cat(inputs, 1) bottleneck_output = self.conv1(self.relu1(self.norm1(concated_features))) # noqa: T484 return bottleneck_output # todo: rewrite when torchscript supports any def any_requires_grad(self, input: List[Tensor]) -> bool: for tensor in input: if tensor.requires_grad: return True return False @torch.jit.unused # noqa: T484 def call_checkpoint_bottleneck(self, input: List[Tensor]) -> Tensor: def closure(*inputs): return self.bn_function(inputs) return cp.checkpoint(closure, *input) @torch.jit._overload_method # noqa: F811 def forward(self, input: List[Tensor]) -> Tensor: pass @torch.jit._overload_method # noqa: F811 def forward(self, input: Tensor) -> Tensor: pass # torchscript does not yet support *args, so we overload method # allowing it to take either a List[Tensor] or single Tensor def forward(self, input: Tensor) -> Tensor: # noqa: F811 if isinstance(input, Tensor): prev_features = [input] else: prev_features = input if self.memory_efficient and self.any_requires_grad(prev_features): if torch.jit.is_scripting(): raise Exception("Memory Efficient not supported in JIT") bottleneck_output = self.call_checkpoint_bottleneck(prev_features) else: bottleneck_output = self.bn_function(prev_features) new_features = self.conv2(self.relu2(self.norm2(bottleneck_output))) if self.drop_rate > 0: new_features = F.dropout(new_features, p=self.drop_rate, training=self.training) return new_features class _DenseBlock(nn.ModuleDict): _version = 2 def __init__( self, num_layers: int, num_input_features: int, bn_size: int, growth_rate: int, drop_rate: float, memory_efficient: bool = False ) -> None: super(_DenseBlock, self).__init__() for i in range(num_layers): layer = _DenseLayer( num_input_features + i * growth_rate, growth_rate=growth_rate, bn_size=bn_size, drop_rate=drop_rate, memory_efficient=memory_efficient, ) self.add_module('denselayer%d' % (i + 1), layer) def forward(self, init_features: Tensor) -> Tensor: features = [init_features] for name, layer in self.items(): new_features = layer(features) features.append(new_features) return torch.cat(features, 1) class _Transition(nn.Sequential): def __init__(self, num_input_features: int, num_output_features: int) -> None: super(_Transition, self).__init__() self.add_module('norm', nn.BatchNorm2d(num_input_features)) self.add_module('relu', nn.ReLU(inplace=True)) self.add_module('conv', nn.Conv2d(num_input_features, num_output_features, kernel_size=1, stride=1, bias=False)) self.add_module('pool', nn.AvgPool2d(kernel_size=2, stride=2)) class DenseNet(nn.Module): r"""Densenet-BC model class, based on `"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`_. Args: growth_rate (int) - how many filters to add each layer (`k` in paper) block_config (list of 4 ints) - how many layers in each pooling block num_init_features (int) - the number of filters to learn in the first convolution layer bn_size (int) - multiplicative factor for number of bottle neck layers (i.e. bn_size * k features in the bottleneck layer) drop_rate (float) - dropout rate after each dense layer num_classes (int) - number of classification classes memory_efficient (bool) - If True, uses checkpointing. Much more memory efficient, but slower. Default: *False*. See `"paper" <https://arxiv.org/pdf/1707.06990.pdf>`_. """ def __init__(self, num_classes: int, in_channels: int, growth_rate: int = 32, block_config: Tuple[int, int, int, int] = (6, 12, 24, 16), num_init_features: int = 64, bn_size: int = 4, drop_rate: float = 0, memory_efficient: bool = False, return_features: bool = False, ) -> None: super(DenseNet, self).__init__() # First convolution self.features = nn.Sequential(OrderedDict([ ('conv0', nn.Conv2d(in_channels, num_init_features, kernel_size=7, stride=2, padding=3, bias=False)), ('norm0', nn.BatchNorm2d(num_init_features)), ('relu0', nn.ReLU(inplace=True)), ('pool0', nn.MaxPool2d(kernel_size=3, stride=2, padding=1)), ])) # Each denseblock num_features = num_init_features for i, num_layers in enumerate(block_config): block = _DenseBlock( num_layers=num_layers, num_input_features=num_features, bn_size=bn_size, growth_rate=growth_rate, drop_rate=drop_rate, memory_efficient=memory_efficient ) self.features.add_module('denseblock%d' % (i + 1), block) num_features = num_features + num_layers * growth_rate if i != len(block_config) - 1: trans = _Transition(num_input_features=num_features, num_output_features=num_features // 2) self.features.add_module('transition%d' % (i + 1), trans) num_features = num_features // 2 # Final batch norm self.features.add_module('norm5', nn.BatchNorm2d(num_features)) # Linear layer self.classifier = nn.Linear(num_features, num_classes) # Official init from torch repo. for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight) elif isinstance(m, nn.BatchNorm2d): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) elif isinstance(m, nn.Linear): nn.init.constant_(m.bias, 0) self.return_features = return_features def forward(self, x: Tensor) -> Tensor: features = self.features(x) out = F.relu(features, inplace=True) out = F.adaptive_avg_pool2d(out, (1, 1)) out = torch.flatten(out, 1) out = self.classifier(out) if self.return_features: return out, features return out def _densenet(num_classes: int, in_channels: int, growth_rate: int, block_config: Tuple[int, int, int, int], num_init_features: int, **kwargs: Any) -> DenseNet: model = DenseNet(num_classes, in_channels, growth_rate, block_config, num_init_features, **kwargs) return model def densenet121(num_classes: int, in_channels: int, **kwargs: Any) -> DenseNet: r"""Densenet-121 model from `"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`_. memory_efficient (bool) - If True, uses checkpointing. Much more memory efficient, but slower. Default: *False*. See `"paper" <https://arxiv.org/pdf/1707.06990.pdf>`_. """ return _densenet(num_classes, in_channels, 32, (6, 12, 24, 16), 64, **kwargs)
9,286
39.911894
113
py
LLPFC
LLPFC-main/models/NIN.py
# code in this file is modified from https://github.com/yangqiongyongyu/Network-In-Network-Pytorch/blob/master/models/nin.py # hyperparameters selected based on https://worksheets.codalab.org/worksheets/0x7b8f6fbc6b5c49c18ac7ca94aafaa1a7 import torch.nn as nn import math class NIN(nn.Module): def __init__(self, num_classes, image_size, in_channel=3): super(NIN, self).__init__() self.num_classes = num_classes if isinstance(image_size, int): pool_size = math.ceil(math.ceil(image_size/2.0-0.5)/2.0-0.5) else: pool_size = (math.ceil(math.ceil(image_size[0]/2.0-0.5)/2.0-0.5), math.ceil(math.ceil(image_size[1]/2.0-0.5)/2.0-0.5)) self.features = nn.Sequential( nn.Conv2d(in_channel, 192, 5, padding=2), nn.ReLU(inplace=True), nn.Conv2d(192, 160, 1), nn.ReLU(inplace=True), nn.Conv2d(160, 96, 1), nn.ReLU(inplace=True), nn.MaxPool2d(3, stride=2, ceil_mode=True), nn.Dropout(inplace=True), nn.Conv2d(96, 192, 5, padding=2), nn.ReLU(inplace=True), nn.Conv2d(192, 192, 1), nn.ReLU(inplace=True), nn.Conv2d(192, 192, 1), nn.ReLU(inplace=True), nn.AvgPool2d(3, stride=2, ceil_mode=True), nn.Dropout(inplace=True), nn.Conv2d(192, 192, 3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(192, 192, 1), nn.ReLU(inplace=True), nn.Conv2d(192, self.num_classes, 1), nn.ReLU(inplace=True), nn.AvgPool2d(pool_size, stride=1) ) self._initialize_weights() def forward(self, x): x = self.features(x) x = x.view(x.size(0), self.num_classes) return x def _initialize_weights(self): for m in self.modules(): if isinstance(m, nn.Conv2d): m.weight.data.normal_(0, 0.05) if m.bias is not None: m.bias.data.zero_()
1,708
30.648148
124
py
LLPFC
LLPFC-main/models/LLPGAN_GEN.py
import torch.nn as nn class LLPGAN_GEN_MNIST(nn.Module): def __init__(self, noise_size=100, out_h=28, out_w=28): self.out_h, self.out_w = out_h, out_w super(LLPGAN_GEN_MNIST, self).__init__() self.model = nn.Sequential( nn.Linear(noise_size, 500), nn.ReLU(), nn.BatchNorm1d(500, eps=1e-05, momentum=0.1, ), nn.Linear(500, 500), nn.ReLU(), nn.BatchNorm1d(500, eps=1e-05, momentum=0.1, ), nn.Linear(500, self.out_h * self.out_w), nn.ReLU(), nn.BatchNorm1d(self.out_h * self.out_w, eps=1e-05, momentum=0.1, ), ) def forward(self, noise): return self.model(noise).reshape(-1, 1, self.out_h, self.out_w) class LLPGAN_GEN_COLOR(nn.Module): def __init__(self, noise_size=32*32): super(LLPGAN_GEN_COLOR, self).__init__() self.linear = nn.Sequential( nn.Linear(noise_size, 4*4*512), nn.ReLU(), nn.BatchNorm1d(4*4*512, eps=1e-05, momentum=0.1, ) ) self.trans_conv = nn.Sequential( nn.ConvTranspose2d(512, 256, (5, 5)), nn.ReLU(), nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, ), nn.ConvTranspose2d(256, 128, (5, 5)), nn.ReLU(), nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, ), nn.ConvTranspose2d(128, 3, (5, 5)), nn.ReLU(), nn.BatchNorm2d(3, eps=1e-05, momentum=0.1, ), ) def forward(self, noise): out = self.linear(noise) return self.trans_conv(out.reshape((-1, 512, 4, 4)))
1,640
29.962264
79
py
LLPFC
LLPFC-main/llpvatlib/utils.py
# code in this file is modified from https://github.com/kevinorjohn/LLP-VAT/blob/a111d6785e8b0b79761c4d68c5b96288048594d6/llp_vat/ import contextlib import torch import torch.nn as nn import torch.nn.functional as F @contextlib.contextmanager def _disable_tracking_bn_stats(model): def switch_attr(m): if hasattr(m, 'track_running_stats'): m.track_running_stats ^= True model.apply(switch_attr) yield model.apply(switch_attr) def _l2_normalize(d): d_reshaped = d.view(d.shape[0], -1, *(1 for _ in range(d.dim() - 2))) d /= torch.norm(d_reshaped, dim=1, keepdim=True) + 1e-8 return d class VATLoss(nn.Module): def __init__(self, xi, eps, ip): """VAT loss :param xi: hyperparameter of VAT (default: 10.0) :param eps: hyperparameter of VAT (default: 1.0) :param ip: iteration times of computing adv noise (default: 1) """ super(VATLoss, self).__init__() self.xi = xi self.eps = eps self.ip = ip def forward(self, model, x): with torch.no_grad(): pred = F.softmax(model(x), dim=1) # prepare random unit tensor # d = torch.rand(x.shape).sub(0.5).to(x.device) d = torch.randn_like(x) d = _l2_normalize(d) with _disable_tracking_bn_stats(model): # calc adversarial direction for _ in range(self.ip): d.requires_grad_() pred_hat = model(x + self.xi * d) logp_hat = F.log_softmax(pred_hat, dim=1) adv_distance = F.kl_div(logp_hat, pred, reduction='batchmean') adv_distance.backward() d = _l2_normalize(d.grad) model.zero_grad() # calc LDS r_adv = d * self.eps pred_hat = model(x + r_adv) logp_hat = F.log_softmax(pred_hat, dim=1) lds = F.kl_div(logp_hat, pred, reduction='batchmean') return lds
1,989
30.587302
130
py
LLPFC
LLPFC-main/llpvatlib/train_fun.py
import torch from kllib.train_fun import compute_kl_loss_on_bagbatch import numpy as np def sigmoid_rampup(current, rampup_length): # modified from https://github.com/kevinorjohn/LLP-VAT/blob/a111d6785e8b0b79761c4d68c5b96288048594d6/llp_vat/ """Exponential rampup from https://arxiv.org/abs/1610.02242""" if rampup_length == 0: return 1.0 else: current = np.clip(current, 0.0, rampup_length) phase = 1.0 - current / rampup_length return float(np.exp(-5.0 * phase * phase)) def get_rampup_weight(weight, iteration, rampup): # modified from https://github.com/kevinorjohn/LLP-VAT/blob/a111d6785e8b0b79761c4d68c5b96288048594d6/llp_vat/ alpha = weight * sigmoid_rampup(iteration, rampup) return alpha def llp_loss_f(model, images, props, vat_loss_f, iteration, device): prop_loss = compute_kl_loss_on_bagbatch(model, images, props, device) alpha = get_rampup_weight(0.05, iteration, -1) # hard-coded based on tsai and lin's implementation vat_loss = vat_loss_f(model, torch.reshape(images, (-1, images.shape[-3], images.shape[-2], images.shape[-1])).to(device)) return prop_loss, alpha, vat_loss def llpvat_train_by_bag(model, optimizer, train_loader, vat_loss_f, epoch, device, scheduler, logger): model.train() total_step = len(train_loader) for i, (images, props) in enumerate(train_loader): prop_loss, alpha, vat_loss = llp_loss_f(model, images, props, vat_loss_f, i, device) loss = prop_loss + alpha * vat_loss # Backward pass optimizer.zero_grad() loss.backward() optimizer.step() if (i + 1) % 100 == 0: logger.info(' Step [{}/{}], Loss: {:.4f}'.format(i + 1, total_step, loss.item())) logger.info(' VAT Loss: {:.4f}'.format(vat_loss.item())) logger.info(' KL Loss: {:.4f}'.format(prop_loss.item())) logger.info(' alpha = {:.4f}'.format(alpha)) if type(scheduler) == torch.optim.lr_scheduler.CosineAnnealingWarmRestarts: scheduler.step(epoch + i / total_step) if type(scheduler) == torch.optim.lr_scheduler.MultiStepLR: scheduler.step()
2,296
44.94
119
py
LLPFC
LLPFC-main/kllib/utils.py
from PIL import Image import torch import numpy as np class KL_DATASET_BASE(torch.utils.data.Dataset): def __init__(self, data, bag2indices, bag2prop, transform): self.data = data self.bag2indices = bag2indices self.bag2prop = bag2prop self.transform = transform def __len__(self): return len(self.bag2indices.keys()) class KL_CIFAR10(KL_DATASET_BASE): def __getitem__(self, bag_index): indices = self.bag2indices[bag_index] images = torch.zeros((len(indices), self.data[0].shape[2], self.data[0].shape[0], self.data[0].shape[1]), dtype=torch.float32) for i in range(len(indices)): idx = indices[i] img = self.data[idx] img = Image.fromarray(img) if self.transform is not None: img = self.transform(img) images[i] = img return images, self.bag2prop[bag_index] class KL_SVHN(KL_DATASET_BASE): def __getitem__(self, bag_index): indices = self.bag2indices[bag_index] images = torch.zeros((len(indices), self.data[0].shape[0], self.data[0].shape[1], self.data[0].shape[2],), dtype=torch.float32) for i in range(len(indices)): idx = indices[i] img = self.data[idx] img = Image.fromarray(np.transpose(img, (1, 2, 0))) if self.transform is not None: img = self.transform(img) images[i] = img return images, self.bag2prop[bag_index] class KL_EMNIST(KL_DATASET_BASE): def __init__(self, data, bag2indices, bag2prop, transform): super(KL_EMNIST, self).__init__(data, bag2indices, bag2prop, transform) img = self.transform(Image.fromarray(self.data[0].numpy(), mode='L')) self.new_h = img.shape[1] self.new_w = img.shape[2] # need this for resized emnist def __getitem__(self, bag_index): indices = self.bag2indices[bag_index] images = torch.zeros((len(indices), 1, self.new_h, self.new_w,), dtype=torch.float32) for i in range(len(indices)): idx = indices[i] img = self.data[idx] img = Image.fromarray(img.numpy(), mode='L') if self.transform is not None: img = self.transform(img) images[i] = img return images, self.bag2prop[bag_index]
2,417
36.2
114
py
LLPFC
LLPFC-main/kllib/train_fun.py
import torch import torch.nn as nn def compute_kl_loss_on_bagbatch(model, images, props, device, epsilon=1e-8): # Move tensors to the configured device images = images.to(device) props = props.to(device) # Forward pass batch_size, bag_size, channel, height, width = images.shape images = images.reshape((batch_size * bag_size, channel, height, width)) outputs = model(images) prob = nn.functional.softmax(outputs, dim=-1).reshape((batch_size, bag_size, -1)) avg_prob = torch.mean(prob, dim=1) avg_prob = torch.clamp(avg_prob, epsilon, 1 - epsilon) loss = torch.sum(-props * torch.log(avg_prob), dim=-1).mean() return loss def validate_model_kl(model, val_loader, device): model.eval() total_loss = 0 total = 0 for i, (images, props) in enumerate(val_loader): total_loss += compute_kl_loss_on_bagbatch(model, images, props, device).item() total += 1 return total_loss/total def kl_train_by_bag(model, optimizer, train_loader, epoch, device, scheduler, logger): model.train() total_step = len(train_loader) for i, (images, props) in enumerate(train_loader): loss = compute_kl_loss_on_bagbatch(model, images, props, device) # Backward pass optimizer.zero_grad() loss.backward() optimizer.step() if (i + 1) % 100 == 0: logger.info(' Step [{}/{}], Loss: {:.4f}'.format(i + 1, total_step, loss.item())) if type(scheduler) == torch.optim.lr_scheduler.CosineAnnealingWarmRestarts: scheduler.step(epoch + i / total_step) if type(scheduler) == torch.optim.lr_scheduler.MultiStepLR: scheduler.step() elif type(scheduler) == torch.optim.lr_scheduler.ReduceLROnPlateau: scheduler.step(validate_model_kl(model, train_loader, device))
1,840
38.170213
107
py
LLPFC
LLPFC-main/llpganlib/train_fun.py
import numpy as np import torch import torch.nn as nn from torch.nn.functional import mse_loss from torch.autograd import Variable def test_llpgan(model, test_loader, criterion, device): # test a model with fully label dataset model.eval() with torch.no_grad(): correct = 0 total = 0 total_loss = 0 for images, labels in test_loader: images = images.to(device) labels = labels.to(device) outputs, _ = model(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() prob = nn.functional.softmax(outputs, dim=1) loss = criterion(prob, labels, device) total_loss += loss.item() return correct / total, total_loss / total def compute_dis_loss(dis, true_images, fake_images, props, device, lambd=1, epsilon=1e-8): # Forward pass batch_size, bag_size, channel, height, width = true_images.shape true_images = true_images.reshape((batch_size * bag_size, channel, height, width)) true_outputs, _ = dis(true_images) fake_outputs, _ = dis(fake_images) # compute the lower bound of kl prob = nn.functional.softmax(true_outputs, dim=-1).reshape((batch_size, bag_size, -1)) clamped_prob = torch.clamp(prob, epsilon, 1 - epsilon) log_prob = torch.log(clamped_prob) avg_log_prop = torch.mean(log_prob, dim=1) lower_kl_loss = -torch.sum(-props * avg_log_prop, dim=-1).mean() * lambd # compute the true/fake binary loss true_outputs_cat = torch.cat((true_outputs, torch.zeros(true_outputs.shape[0], 1).to(device)), dim=1) true_prob = 1 - nn.functional.softmax(true_outputs_cat, dim=1)[:, -1] clamped_true_prob = torch.clamp(true_prob, epsilon, 1 - epsilon) log_true_prob = torch.log(clamped_true_prob) avg_log_true_prop = -torch.mean(log_true_prob) fake_outputs_cat = torch.cat((fake_outputs, torch.zeros(fake_outputs.shape[0], 1).to(device)), dim=1) fake_prob = nn.functional.softmax(fake_outputs_cat, dim=1)[:, -1] clamped_fake_prob = torch.clamp(fake_prob, epsilon, 1 - epsilon) log_fake_prob = torch.log(clamped_fake_prob) avg_log_fake_prop = -torch.mean(log_fake_prob) return lower_kl_loss + avg_log_true_prop + avg_log_fake_prop def compute_gen_loss(dis, true_images, fake_images): batch_size, bag_size, channel, height, width = true_images.shape true_images = true_images.reshape((batch_size * bag_size, channel, height, width)) true_outputs, true_features = dis(true_images) fake_outputs, fake_features = dis(fake_images) loss = mse_loss(fake_features, true_features) return loss # also return feature_maps to compute generator loss def llpgan_train_by_bag(gen, dis, gen_opt, dis_opt, dis_sch, gen_sch, noise_dim, train_loader, epoch, device, logger ): gen.train() dis.train() total_step = len(train_loader) for i, (images, props) in enumerate(train_loader): true_images = images.to(device) props = props.to(device) batch_data_points = true_images.shape[0] * true_images.shape[1] noise = Variable(torch.FloatTensor(np.random.normal(0, 1, (batch_data_points, noise_dim))).to(device)) dis_opt.zero_grad() dis_loss = compute_dis_loss(dis, true_images, gen(noise).detach(), props, device, lambd=1) dis_loss.backward() dis_opt.step() gen_opt.zero_grad() gen_loss = compute_gen_loss(dis, true_images, gen(noise)) gen_loss.backward() gen_opt.step() if (i + 1) % 100 == 0: logger.info(' Step [{}/{}], dis Loss: {:.4f}'.format(i + 1, total_step, dis_loss.item())) logger.info(' gen Loss: {:.4f}'.format(gen_loss.item())) if type(dis_sch) == torch.optim.lr_scheduler.CosineAnnealingWarmRestarts: dis_sch.step(epoch + i / total_step) if type(gen_sch) == torch.optim.lr_scheduler.CosineAnnealingWarmRestarts: gen_sch.step(epoch + i / total_step) if type(dis_sch) == torch.optim.lr_scheduler.MultiStepLR: dis_sch.step() if type(gen_sch) == torch.optim.lr_scheduler.MultiStepLR: gen_sch.step()
4,536
40.245455
115
py
DehazeFormer
DehazeFormer-main/test.py
import os import argparse import torch import torch.nn as nn import torch.nn.functional as F from pytorch_msssim import ssim from torch.utils.data import DataLoader from collections import OrderedDict from utils import AverageMeter, write_img, chw_to_hwc from datasets.loader import PairLoader from models import * parser = argparse.ArgumentParser() parser.add_argument('--model', default='dehazeformer-s', type=str, help='model name') parser.add_argument('--num_workers', default=16, type=int, help='number of workers') parser.add_argument('--data_dir', default='./data/', type=str, help='path to dataset') parser.add_argument('--save_dir', default='./saved_models/', type=str, help='path to models saving') parser.add_argument('--result_dir', default='./results/', type=str, help='path to results saving') parser.add_argument('--dataset', default='RESIDE-IN', type=str, help='dataset name') parser.add_argument('--exp', default='indoor', type=str, help='experiment setting') args = parser.parse_args() def single(save_dir): state_dict = torch.load(save_dir)['state_dict'] new_state_dict = OrderedDict() for k, v in state_dict.items(): name = k[7:] new_state_dict[name] = v return new_state_dict def test(test_loader, network, result_dir): PSNR = AverageMeter() SSIM = AverageMeter() torch.cuda.empty_cache() network.eval() os.makedirs(os.path.join(result_dir, 'imgs'), exist_ok=True) f_result = open(os.path.join(result_dir, 'results.csv'), 'w') for idx, batch in enumerate(test_loader): input = batch['source'].cuda() target = batch['target'].cuda() filename = batch['filename'][0] with torch.no_grad(): output = network(input).clamp_(-1, 1) # [-1, 1] to [0, 1] output = output * 0.5 + 0.5 target = target * 0.5 + 0.5 psnr_val = 10 * torch.log10(1 / F.mse_loss(output, target)).item() _, _, H, W = output.size() down_ratio = max(1, round(min(H, W) / 256)) # Zhou Wang ssim_val = ssim(F.adaptive_avg_pool2d(output, (int(H / down_ratio), int(W / down_ratio))), F.adaptive_avg_pool2d(target, (int(H / down_ratio), int(W / down_ratio))), data_range=1, size_average=False).item() PSNR.update(psnr_val) SSIM.update(ssim_val) print('Test: [{0}]\t' 'PSNR: {psnr.val:.02f} ({psnr.avg:.02f})\t' 'SSIM: {ssim.val:.03f} ({ssim.avg:.03f})' .format(idx, psnr=PSNR, ssim=SSIM)) f_result.write('%s,%.02f,%.03f\n'%(filename, psnr_val, ssim_val)) out_img = chw_to_hwc(output.detach().cpu().squeeze(0).numpy()) write_img(os.path.join(result_dir, 'imgs', filename), out_img) f_result.close() os.rename(os.path.join(result_dir, 'results.csv'), os.path.join(result_dir, '%.02f | %.04f.csv'%(PSNR.avg, SSIM.avg))) if __name__ == '__main__': network = eval(args.model.replace('-', '_'))() network.cuda() saved_model_dir = os.path.join(args.save_dir, args.exp, args.model+'.pth') if os.path.exists(saved_model_dir): print('==> Start testing, current model name: ' + args.model) network.load_state_dict(single(saved_model_dir)) else: print('==> No existing trained model!') exit(0) dataset_dir = os.path.join(args.data_dir, args.dataset) test_dataset = PairLoader(dataset_dir, 'test', 'test') test_loader = DataLoader(test_dataset, batch_size=1, num_workers=args.num_workers, pin_memory=True) result_dir = os.path.join(args.result_dir, args.dataset, args.model) test(test_loader, network, result_dir)
3,451
30.962963
100
py
DehazeFormer
DehazeFormer-main/train.py
import os import argparse import json import torch import torch.nn as nn import torch.nn.functional as F from torch.cuda.amp import autocast, GradScaler from torch.utils.data import DataLoader from tensorboardX import SummaryWriter from tqdm import tqdm from utils import AverageMeter from datasets.loader import PairLoader from models import * parser = argparse.ArgumentParser() parser.add_argument('--model', default='dehazeformer-s', type=str, help='model name') parser.add_argument('--num_workers', default=16, type=int, help='number of workers') parser.add_argument('--no_autocast', action='store_false', default=True, help='disable autocast') parser.add_argument('--save_dir', default='./saved_models/', type=str, help='path to models saving') parser.add_argument('--data_dir', default='./data/', type=str, help='path to dataset') parser.add_argument('--log_dir', default='./logs/', type=str, help='path to logs') parser.add_argument('--dataset', default='RESIDE-IN', type=str, help='dataset name') parser.add_argument('--exp', default='indoor', type=str, help='experiment setting') parser.add_argument('--gpu', default='0,1,2,3', type=str, help='GPUs used for training') args = parser.parse_args() os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu def train(train_loader, network, criterion, optimizer, scaler): losses = AverageMeter() torch.cuda.empty_cache() network.train() for batch in train_loader: source_img = batch['source'].cuda() target_img = batch['target'].cuda() with autocast(args.no_autocast): output = network(source_img) loss = criterion(output, target_img) losses.update(loss.item()) optimizer.zero_grad() scaler.scale(loss).backward() scaler.step(optimizer) scaler.update() return losses.avg def valid(val_loader, network): PSNR = AverageMeter() torch.cuda.empty_cache() network.eval() for batch in val_loader: source_img = batch['source'].cuda() target_img = batch['target'].cuda() with torch.no_grad(): # torch.no_grad() may cause warning output = network(source_img).clamp_(-1, 1) mse_loss = F.mse_loss(output * 0.5 + 0.5, target_img * 0.5 + 0.5, reduction='none').mean((1, 2, 3)) psnr = 10 * torch.log10(1 / mse_loss).mean() PSNR.update(psnr.item(), source_img.size(0)) return PSNR.avg if __name__ == '__main__': setting_filename = os.path.join('configs', args.exp, args.model+'.json') if not os.path.exists(setting_filename): setting_filename = os.path.join('configs', args.exp, 'default.json') with open(setting_filename, 'r') as f: setting = json.load(f) network = eval(args.model.replace('-', '_'))() network = nn.DataParallel(network).cuda() criterion = nn.L1Loss() if setting['optimizer'] == 'adam': optimizer = torch.optim.Adam(network.parameters(), lr=setting['lr']) elif setting['optimizer'] == 'adamw': optimizer = torch.optim.AdamW(network.parameters(), lr=setting['lr']) else: raise Exception("ERROR: unsupported optimizer") scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=setting['epochs'], eta_min=setting['lr'] * 1e-2) scaler = GradScaler() dataset_dir = os.path.join(args.data_dir, args.dataset) train_dataset = PairLoader(dataset_dir, 'train', 'train', setting['patch_size'], setting['edge_decay'], setting['only_h_flip']) train_loader = DataLoader(train_dataset, batch_size=setting['batch_size'], shuffle=True, num_workers=args.num_workers, pin_memory=True, drop_last=True) val_dataset = PairLoader(dataset_dir, 'test', setting['valid_mode'], setting['patch_size']) val_loader = DataLoader(val_dataset, batch_size=setting['batch_size'], num_workers=args.num_workers, pin_memory=True) save_dir = os.path.join(args.save_dir, args.exp) os.makedirs(save_dir, exist_ok=True) if not os.path.exists(os.path.join(save_dir, args.model+'.pth')): print('==> Start training, current model name: ' + args.model) # print(network) writer = SummaryWriter(log_dir=os.path.join(args.log_dir, args.exp, args.model)) best_psnr = 0 for epoch in tqdm(range(setting['epochs'] + 1)): loss = train(train_loader, network, criterion, optimizer, scaler) writer.add_scalar('train_loss', loss, epoch) scheduler.step() if epoch % setting['eval_freq'] == 0: avg_psnr = valid(val_loader, network) writer.add_scalar('valid_psnr', avg_psnr, epoch) if avg_psnr > best_psnr: best_psnr = avg_psnr torch.save({'state_dict': network.state_dict()}, os.path.join(save_dir, args.model+'.pth')) writer.add_scalar('best_psnr', best_psnr, epoch) else: print('==> Existing trained model') exit(1)
4,879
31.972973
121
py
DehazeFormer
DehazeFormer-main/models/dehazeformer.py
import torch import torch.nn as nn import torch.nn.functional as F import math import numpy as np from torch.nn.init import _calculate_fan_in_and_fan_out from timm.models.layers import to_2tuple, trunc_normal_ class RLN(nn.Module): r"""Revised LayerNorm""" def __init__(self, dim, eps=1e-5, detach_grad=False): super(RLN, self).__init__() self.eps = eps self.detach_grad = detach_grad self.weight = nn.Parameter(torch.ones((1, dim, 1, 1))) self.bias = nn.Parameter(torch.zeros((1, dim, 1, 1))) self.meta1 = nn.Conv2d(1, dim, 1) self.meta2 = nn.Conv2d(1, dim, 1) trunc_normal_(self.meta1.weight, std=.02) nn.init.constant_(self.meta1.bias, 1) trunc_normal_(self.meta2.weight, std=.02) nn.init.constant_(self.meta2.bias, 0) def forward(self, input): mean = torch.mean(input, dim=(1, 2, 3), keepdim=True) std = torch.sqrt((input - mean).pow(2).mean(dim=(1, 2, 3), keepdim=True) + self.eps) normalized_input = (input - mean) / std if self.detach_grad: rescale, rebias = self.meta1(std.detach()), self.meta2(mean.detach()) else: rescale, rebias = self.meta1(std), self.meta2(mean) out = normalized_input * self.weight + self.bias return out, rescale, rebias class Mlp(nn.Module): def __init__(self, network_depth, in_features, hidden_features=None, out_features=None): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.network_depth = network_depth self.mlp = nn.Sequential( nn.Conv2d(in_features, hidden_features, 1), nn.ReLU(True), nn.Conv2d(hidden_features, out_features, 1) ) self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Conv2d): gain = (8 * self.network_depth) ** (-1/4) fan_in, fan_out = _calculate_fan_in_and_fan_out(m.weight) std = gain * math.sqrt(2.0 / float(fan_in + fan_out)) trunc_normal_(m.weight, std=std) if m.bias is not None: nn.init.constant_(m.bias, 0) def forward(self, x): return self.mlp(x) def window_partition(x, window_size): B, H, W, C = x.shape x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size**2, C) return windows def window_reverse(windows, window_size, H, W): B = int(windows.shape[0] / (H * W / window_size / window_size)) x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1) x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) return x def get_relative_positions(window_size): coords_h = torch.arange(window_size) coords_w = torch.arange(window_size) coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww relative_positions = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww relative_positions = relative_positions.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 relative_positions_log = torch.sign(relative_positions) * torch.log(1. + relative_positions.abs()) return relative_positions_log class WindowAttention(nn.Module): def __init__(self, dim, window_size, num_heads): super().__init__() self.dim = dim self.window_size = window_size # Wh, Ww self.num_heads = num_heads head_dim = dim // num_heads self.scale = head_dim ** -0.5 relative_positions = get_relative_positions(self.window_size) self.register_buffer("relative_positions", relative_positions) self.meta = nn.Sequential( nn.Linear(2, 256, bias=True), nn.ReLU(True), nn.Linear(256, num_heads, bias=True) ) self.softmax = nn.Softmax(dim=-1) def forward(self, qkv): B_, N, _ = qkv.shape qkv = qkv.reshape(B_, N, 3, self.num_heads, self.dim // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) q = q * self.scale attn = (q @ k.transpose(-2, -1)) relative_position_bias = self.meta(self.relative_positions) relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww attn = attn + relative_position_bias.unsqueeze(0) attn = self.softmax(attn) x = (attn @ v).transpose(1, 2).reshape(B_, N, self.dim) return x class Attention(nn.Module): def __init__(self, network_depth, dim, num_heads, window_size, shift_size, use_attn=False, conv_type=None): super().__init__() self.dim = dim self.head_dim = int(dim // num_heads) self.num_heads = num_heads self.window_size = window_size self.shift_size = shift_size self.network_depth = network_depth self.use_attn = use_attn self.conv_type = conv_type if self.conv_type == 'Conv': self.conv = nn.Sequential( nn.Conv2d(dim, dim, kernel_size=3, padding=1, padding_mode='reflect'), nn.ReLU(True), nn.Conv2d(dim, dim, kernel_size=3, padding=1, padding_mode='reflect') ) if self.conv_type == 'DWConv': self.conv = nn.Conv2d(dim, dim, kernel_size=5, padding=2, groups=dim, padding_mode='reflect') if self.conv_type == 'DWConv' or self.use_attn: self.V = nn.Conv2d(dim, dim, 1) self.proj = nn.Conv2d(dim, dim, 1) if self.use_attn: self.QK = nn.Conv2d(dim, dim * 2, 1) self.attn = WindowAttention(dim, window_size, num_heads) self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Conv2d): w_shape = m.weight.shape if w_shape[0] == self.dim * 2: # QK fan_in, fan_out = _calculate_fan_in_and_fan_out(m.weight) std = math.sqrt(2.0 / float(fan_in + fan_out)) trunc_normal_(m.weight, std=std) else: gain = (8 * self.network_depth) ** (-1/4) fan_in, fan_out = _calculate_fan_in_and_fan_out(m.weight) std = gain * math.sqrt(2.0 / float(fan_in + fan_out)) trunc_normal_(m.weight, std=std) if m.bias is not None: nn.init.constant_(m.bias, 0) def check_size(self, x, shift=False): _, _, h, w = x.size() mod_pad_h = (self.window_size - h % self.window_size) % self.window_size mod_pad_w = (self.window_size - w % self.window_size) % self.window_size if shift: x = F.pad(x, (self.shift_size, (self.window_size-self.shift_size+mod_pad_w) % self.window_size, self.shift_size, (self.window_size-self.shift_size+mod_pad_h) % self.window_size), mode='reflect') else: x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect') return x def forward(self, X): B, C, H, W = X.shape if self.conv_type == 'DWConv' or self.use_attn: V = self.V(X) if self.use_attn: QK = self.QK(X) QKV = torch.cat([QK, V], dim=1) # shift shifted_QKV = self.check_size(QKV, self.shift_size > 0) Ht, Wt = shifted_QKV.shape[2:] # partition windows shifted_QKV = shifted_QKV.permute(0, 2, 3, 1) qkv = window_partition(shifted_QKV, self.window_size) # nW*B, window_size**2, C attn_windows = self.attn(qkv) # merge windows shifted_out = window_reverse(attn_windows, self.window_size, Ht, Wt) # B H' W' C # reverse cyclic shift out = shifted_out[:, self.shift_size:(self.shift_size+H), self.shift_size:(self.shift_size+W), :] attn_out = out.permute(0, 3, 1, 2) if self.conv_type in ['Conv', 'DWConv']: conv_out = self.conv(V) out = self.proj(conv_out + attn_out) else: out = self.proj(attn_out) else: if self.conv_type == 'Conv': out = self.conv(X) # no attention and use conv, no projection elif self.conv_type == 'DWConv': out = self.proj(self.conv(V)) return out class TransformerBlock(nn.Module): def __init__(self, network_depth, dim, num_heads, mlp_ratio=4., norm_layer=nn.LayerNorm, mlp_norm=False, window_size=8, shift_size=0, use_attn=True, conv_type=None): super().__init__() self.use_attn = use_attn self.mlp_norm = mlp_norm self.norm1 = norm_layer(dim) if use_attn else nn.Identity() self.attn = Attention(network_depth, dim, num_heads=num_heads, window_size=window_size, shift_size=shift_size, use_attn=use_attn, conv_type=conv_type) self.norm2 = norm_layer(dim) if use_attn and mlp_norm else nn.Identity() self.mlp = Mlp(network_depth, dim, hidden_features=int(dim * mlp_ratio)) def forward(self, x): identity = x if self.use_attn: x, rescale, rebias = self.norm1(x) x = self.attn(x) if self.use_attn: x = x * rescale + rebias x = identity + x identity = x if self.use_attn and self.mlp_norm: x, rescale, rebias = self.norm2(x) x = self.mlp(x) if self.use_attn and self.mlp_norm: x = x * rescale + rebias x = identity + x return x class BasicLayer(nn.Module): def __init__(self, network_depth, dim, depth, num_heads, mlp_ratio=4., norm_layer=nn.LayerNorm, window_size=8, attn_ratio=0., attn_loc='last', conv_type=None): super().__init__() self.dim = dim self.depth = depth attn_depth = attn_ratio * depth if attn_loc == 'last': use_attns = [i >= depth-attn_depth for i in range(depth)] elif attn_loc == 'first': use_attns = [i < attn_depth for i in range(depth)] elif attn_loc == 'middle': use_attns = [i >= (depth-attn_depth)//2 and i < (depth+attn_depth)//2 for i in range(depth)] # build blocks self.blocks = nn.ModuleList([ TransformerBlock(network_depth=network_depth, dim=dim, num_heads=num_heads, mlp_ratio=mlp_ratio, norm_layer=norm_layer, window_size=window_size, shift_size=0 if (i % 2 == 0) else window_size // 2, use_attn=use_attns[i], conv_type=conv_type) for i in range(depth)]) def forward(self, x): for blk in self.blocks: x = blk(x) return x class PatchEmbed(nn.Module): def __init__(self, patch_size=4, in_chans=3, embed_dim=96, kernel_size=None): super().__init__() self.in_chans = in_chans self.embed_dim = embed_dim if kernel_size is None: kernel_size = patch_size self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=kernel_size, stride=patch_size, padding=(kernel_size-patch_size+1)//2, padding_mode='reflect') def forward(self, x): x = self.proj(x) return x class PatchUnEmbed(nn.Module): def __init__(self, patch_size=4, out_chans=3, embed_dim=96, kernel_size=None): super().__init__() self.out_chans = out_chans self.embed_dim = embed_dim if kernel_size is None: kernel_size = 1 self.proj = nn.Sequential( nn.Conv2d(embed_dim, out_chans*patch_size**2, kernel_size=kernel_size, padding=kernel_size//2, padding_mode='reflect'), nn.PixelShuffle(patch_size) ) def forward(self, x): x = self.proj(x) return x class SKFusion(nn.Module): def __init__(self, dim, height=2, reduction=8): super(SKFusion, self).__init__() self.height = height d = max(int(dim/reduction), 4) self.avg_pool = nn.AdaptiveAvgPool2d(1) self.mlp = nn.Sequential( nn.Conv2d(dim, d, 1, bias=False), nn.ReLU(), nn.Conv2d(d, dim*height, 1, bias=False) ) self.softmax = nn.Softmax(dim=1) def forward(self, in_feats): B, C, H, W = in_feats[0].shape in_feats = torch.cat(in_feats, dim=1) in_feats = in_feats.view(B, self.height, C, H, W) feats_sum = torch.sum(in_feats, dim=1) attn = self.mlp(self.avg_pool(feats_sum)) attn = self.softmax(attn.view(B, self.height, C, 1, 1)) out = torch.sum(in_feats*attn, dim=1) return out class DehazeFormer(nn.Module): def __init__(self, in_chans=3, out_chans=4, window_size=8, embed_dims=[24, 48, 96, 48, 24], mlp_ratios=[2., 4., 4., 2., 2.], depths=[16, 16, 16, 8, 8], num_heads=[2, 4, 6, 1, 1], attn_ratio=[1/4, 1/2, 3/4, 0, 0], conv_type=['DWConv', 'DWConv', 'DWConv', 'DWConv', 'DWConv'], norm_layer=[RLN, RLN, RLN, RLN, RLN]): super(DehazeFormer, self).__init__() # setting self.patch_size = 4 self.window_size = window_size self.mlp_ratios = mlp_ratios # split image into non-overlapping patches self.patch_embed = PatchEmbed( patch_size=1, in_chans=in_chans, embed_dim=embed_dims[0], kernel_size=3) # backbone self.layer1 = BasicLayer(network_depth=sum(depths), dim=embed_dims[0], depth=depths[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], norm_layer=norm_layer[0], window_size=window_size, attn_ratio=attn_ratio[0], attn_loc='last', conv_type=conv_type[0]) self.patch_merge1 = PatchEmbed( patch_size=2, in_chans=embed_dims[0], embed_dim=embed_dims[1]) self.skip1 = nn.Conv2d(embed_dims[0], embed_dims[0], 1) self.layer2 = BasicLayer(network_depth=sum(depths), dim=embed_dims[1], depth=depths[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], norm_layer=norm_layer[1], window_size=window_size, attn_ratio=attn_ratio[1], attn_loc='last', conv_type=conv_type[1]) self.patch_merge2 = PatchEmbed( patch_size=2, in_chans=embed_dims[1], embed_dim=embed_dims[2]) self.skip2 = nn.Conv2d(embed_dims[1], embed_dims[1], 1) self.layer3 = BasicLayer(network_depth=sum(depths), dim=embed_dims[2], depth=depths[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], norm_layer=norm_layer[2], window_size=window_size, attn_ratio=attn_ratio[2], attn_loc='last', conv_type=conv_type[2]) self.patch_split1 = PatchUnEmbed( patch_size=2, out_chans=embed_dims[3], embed_dim=embed_dims[2]) assert embed_dims[1] == embed_dims[3] self.fusion1 = SKFusion(embed_dims[3]) self.layer4 = BasicLayer(network_depth=sum(depths), dim=embed_dims[3], depth=depths[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], norm_layer=norm_layer[3], window_size=window_size, attn_ratio=attn_ratio[3], attn_loc='last', conv_type=conv_type[3]) self.patch_split2 = PatchUnEmbed( patch_size=2, out_chans=embed_dims[4], embed_dim=embed_dims[3]) assert embed_dims[0] == embed_dims[4] self.fusion2 = SKFusion(embed_dims[4]) self.layer5 = BasicLayer(network_depth=sum(depths), dim=embed_dims[4], depth=depths[4], num_heads=num_heads[4], mlp_ratio=mlp_ratios[4], norm_layer=norm_layer[4], window_size=window_size, attn_ratio=attn_ratio[4], attn_loc='last', conv_type=conv_type[4]) # merge non-overlapping patches into image self.patch_unembed = PatchUnEmbed( patch_size=1, out_chans=out_chans, embed_dim=embed_dims[4], kernel_size=3) def check_image_size(self, x): # NOTE: for I2I test _, _, h, w = x.size() mod_pad_h = (self.patch_size - h % self.patch_size) % self.patch_size mod_pad_w = (self.patch_size - w % self.patch_size) % self.patch_size x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect') return x def forward_features(self, x): x = self.patch_embed(x) x = self.layer1(x) skip1 = x x = self.patch_merge1(x) x = self.layer2(x) skip2 = x x = self.patch_merge2(x) x = self.layer3(x) x = self.patch_split1(x) x = self.fusion1([x, self.skip2(skip2)]) + x x = self.layer4(x) x = self.patch_split2(x) x = self.fusion2([x, self.skip1(skip1)]) + x x = self.layer5(x) x = self.patch_unembed(x) return x def forward(self, x): H, W = x.shape[2:] x = self.check_image_size(x) feat = self.forward_features(x) K, B = torch.split(feat, (1, 3), dim=1) x = K * x - B + x x = x[:, :, :H, :W] return x def dehazeformer_t(): return DehazeFormer( embed_dims=[24, 48, 96, 48, 24], mlp_ratios=[2., 4., 4., 2., 2.], depths=[4, 4, 4, 2, 2], num_heads=[2, 4, 6, 1, 1], attn_ratio=[0, 1/2, 1, 0, 0], conv_type=['DWConv', 'DWConv', 'DWConv', 'DWConv', 'DWConv']) def dehazeformer_s(): return DehazeFormer( embed_dims=[24, 48, 96, 48, 24], mlp_ratios=[2., 4., 4., 2., 2.], depths=[8, 8, 8, 4, 4], num_heads=[2, 4, 6, 1, 1], attn_ratio=[1/4, 1/2, 3/4, 0, 0], conv_type=['DWConv', 'DWConv', 'DWConv', 'DWConv', 'DWConv']) def dehazeformer_b(): return DehazeFormer( embed_dims=[24, 48, 96, 48, 24], mlp_ratios=[2., 4., 4., 2., 2.], depths=[16, 16, 16, 8, 8], num_heads=[2, 4, 6, 1, 1], attn_ratio=[1/4, 1/2, 3/4, 0, 0], conv_type=['DWConv', 'DWConv', 'DWConv', 'DWConv', 'DWConv']) def dehazeformer_d(): return DehazeFormer( embed_dims=[24, 48, 96, 48, 24], mlp_ratios=[2., 4., 4., 2., 2.], depths=[32, 32, 32, 16, 16], num_heads=[2, 4, 6, 1, 1], attn_ratio=[1/4, 1/2, 3/4, 0, 0], conv_type=['DWConv', 'DWConv', 'DWConv', 'DWConv', 'DWConv']) def dehazeformer_w(): return DehazeFormer( embed_dims=[48, 96, 192, 96, 48], mlp_ratios=[2., 4., 4., 2., 2.], depths=[16, 16, 16, 8, 8], num_heads=[2, 4, 6, 1, 1], attn_ratio=[1/4, 1/2, 3/4, 0, 0], conv_type=['DWConv', 'DWConv', 'DWConv', 'DWConv', 'DWConv']) def dehazeformer_m(): return DehazeFormer( embed_dims=[24, 48, 96, 48, 24], mlp_ratios=[2., 4., 4., 2., 2.], depths=[12, 12, 12, 6, 6], num_heads=[2, 4, 6, 1, 1], attn_ratio=[1/4, 1/2, 3/4, 0, 0], conv_type=['Conv', 'Conv', 'Conv', 'Conv', 'Conv']) def dehazeformer_l(): return DehazeFormer( embed_dims=[48, 96, 192, 96, 48], mlp_ratios=[2., 4., 4., 2., 2.], depths=[16, 16, 16, 12, 12], num_heads=[2, 4, 6, 1, 1], attn_ratio=[1/4, 1/2, 3/4, 0, 0], conv_type=['Conv', 'Conv', 'Conv', 'Conv', 'Conv'])
17,111
29.502674
108
py
DehazeFormer
DehazeFormer-main/datasets/loader.py
import os import random import numpy as np import cv2 from torch.utils.data import Dataset from utils import hwc_to_chw, read_img def augment(imgs=[], size=256, edge_decay=0., only_h_flip=False): H, W, _ = imgs[0].shape Hc, Wc = [size, size] # simple re-weight for the edge if random.random() < Hc / H * edge_decay: Hs = 0 if random.randint(0, 1) == 0 else H - Hc else: Hs = random.randint(0, H-Hc) if random.random() < Wc / W * edge_decay: Ws = 0 if random.randint(0, 1) == 0 else W - Wc else: Ws = random.randint(0, W-Wc) for i in range(len(imgs)): imgs[i] = imgs[i][Hs:(Hs+Hc), Ws:(Ws+Wc), :] # horizontal flip if random.randint(0, 1) == 1: for i in range(len(imgs)): imgs[i] = np.flip(imgs[i], axis=1) if not only_h_flip: # bad data augmentations for outdoor rot_deg = random.randint(0, 3) for i in range(len(imgs)): imgs[i] = np.rot90(imgs[i], rot_deg, (0, 1)) return imgs def align(imgs=[], size=256): H, W, _ = imgs[0].shape Hc, Wc = [size, size] Hs = (H - Hc) // 2 Ws = (W - Wc) // 2 for i in range(len(imgs)): imgs[i] = imgs[i][Hs:(Hs+Hc), Ws:(Ws+Wc), :] return imgs class PairLoader(Dataset): def __init__(self, data_dir, sub_dir, mode, size=256, edge_decay=0, only_h_flip=False): assert mode in ['train', 'valid', 'test'] self.mode = mode self.size = size self.edge_decay = edge_decay self.only_h_flip = only_h_flip self.root_dir = os.path.join(data_dir, sub_dir) self.img_names = sorted(os.listdir(os.path.join(self.root_dir, 'GT'))) self.img_num = len(self.img_names) def __len__(self): return self.img_num def __getitem__(self, idx): cv2.setNumThreads(0) cv2.ocl.setUseOpenCL(False) # read image, and scale [0, 1] to [-1, 1] img_name = self.img_names[idx] source_img = read_img(os.path.join(self.root_dir, 'hazy', img_name)) * 2 - 1 target_img = read_img(os.path.join(self.root_dir, 'GT', img_name)) * 2 - 1 if self.mode == 'train': [source_img, target_img] = augment([source_img, target_img], self.size, self.edge_decay, self.only_h_flip) if self.mode == 'valid': [source_img, target_img] = align([source_img, target_img], self.size) return {'source': hwc_to_chw(source_img), 'target': hwc_to_chw(target_img), 'filename': img_name} class SingleLoader(Dataset): def __init__(self, root_dir): self.root_dir = root_dir self.img_names = sorted(os.listdir(self.root_dir)) self.img_num = len(self.img_names) def __len__(self): return self.img_num def __getitem__(self, idx): cv2.setNumThreads(0) cv2.ocl.setUseOpenCL(False) # read image, and scale [0, 1] to [-1, 1] img_name = self.img_names[idx] img = read_img(os.path.join(self.root_dir, img_name)) * 2 - 1 return {'img': hwc_to_chw(img), 'filename': img_name}
2,768
25.122642
109
py
DehazeFormer
DehazeFormer-main/utils/data_parallel.py
from torch.nn.parallel import DataParallel import torch from torch.nn.parallel._functions import Scatter from torch.nn.parallel.parallel_apply import parallel_apply def scatter(inputs, target_gpus, chunk_sizes, dim=0): r""" Slices tensors into approximately equal chunks and distributes them across given GPUs. Duplicates references to objects that are not tensors. """ def scatter_map(obj): if isinstance(obj, torch.Tensor): try: return Scatter.apply(target_gpus, chunk_sizes, dim, obj) except: print('obj', obj.size()) print('dim', dim) print('chunk_sizes', chunk_sizes) quit() if isinstance(obj, tuple) and len(obj) > 0: return list(zip(*map(scatter_map, obj))) if isinstance(obj, list) and len(obj) > 0: return list(map(list, zip(*map(scatter_map, obj)))) if isinstance(obj, dict) and len(obj) > 0: return list(map(type(obj), zip(*map(scatter_map, obj.items())))) return [obj for targets in target_gpus] # After scatter_map is called, a scatter_map cell will exist. This cell # has a reference to the actual function scatter_map, which has references # to a closure that has a reference to the scatter_map cell (because the # fn is recursive). To avoid this reference cycle, we set the function to # None, clearing the cell try: return scatter_map(inputs) finally: scatter_map = None def scatter_kwargs(inputs, kwargs, target_gpus, chunk_sizes, dim=0): r"""Scatter with support for kwargs dictionary""" inputs = scatter(inputs, target_gpus, chunk_sizes, dim) if inputs else [] kwargs = scatter(kwargs, target_gpus, chunk_sizes, dim) if kwargs else [] if len(inputs) < len(kwargs): inputs.extend([() for _ in range(len(kwargs) - len(inputs))]) elif len(kwargs) < len(inputs): kwargs.extend([{} for _ in range(len(inputs) - len(kwargs))]) inputs = tuple(inputs) kwargs = tuple(kwargs) return inputs, kwargs class BalancedDataParallel(DataParallel): def __init__(self, gpu0_bsz, *args, **kwargs): self.gpu0_bsz = gpu0_bsz super().__init__(*args, **kwargs) def forward(self, *inputs, **kwargs): if not self.device_ids: return self.module(*inputs, **kwargs) if len(self.device_ids) == 1: inputs, kwargs = super().scatter(inputs, kwargs, self.device_ids) return self.module(*inputs[0], **kwargs[0]) if self.gpu0_bsz == 0: device_ids = self.device_ids[1:] else: device_ids = self.device_ids inputs, kwargs = self.scatter(inputs, kwargs, device_ids) if self.gpu0_bsz == 0: replicas = self.replicate(self.module, self.device_ids) else: replicas = self.replicate(self.module, self.device_ids[:len(inputs)]) # replicas = self.replicate(self.module, device_ids[:len(inputs)]) if self.gpu0_bsz == 0: replicas = replicas[1:] outputs = self.parallel_apply(replicas, device_ids, inputs, kwargs) return self.gather(outputs, self.output_device) def parallel_apply(self, replicas, device_ids, inputs, kwargs): return parallel_apply(replicas, inputs, kwargs, device_ids[:len(inputs)]) def scatter(self, inputs, kwargs, device_ids): bsz = inputs[0].size(self.dim) num_dev = len(self.device_ids) gpu0_bsz = self.gpu0_bsz bsz_unit = (bsz - gpu0_bsz) // (num_dev - 1) if gpu0_bsz < bsz_unit: chunk_sizes = [gpu0_bsz] + [bsz_unit] * (num_dev - 1) delta = bsz - sum(chunk_sizes) for i in range(delta): chunk_sizes[i + 1] += 1 if gpu0_bsz == 0: chunk_sizes = chunk_sizes[1:] else: return super().scatter(inputs, kwargs, device_ids) return scatter_kwargs(inputs, kwargs, device_ids, chunk_sizes, dim=self.dim)
4,069
38.514563
84
py
vector-quantize-pytorch
vector-quantize-pytorch-master/setup.py
from setuptools import setup, find_packages setup( name = 'vector_quantize_pytorch', packages = find_packages(), version = '1.6.30', license='MIT', description = 'Vector Quantization - Pytorch', long_description_content_type = 'text/markdown', author = 'Phil Wang', author_email = 'lucidrains@gmail.com', url = 'https://github.com/lucidrains/vector-quantizer-pytorch', keywords = [ 'artificial intelligence', 'deep learning', 'pytorch', 'quantization' ], install_requires=[ 'einops>=0.6.1', 'torch' ], classifiers=[ 'Development Status :: 4 - Beta', 'Intended Audience :: Developers', 'Topic :: Scientific/Engineering :: Artificial Intelligence', 'License :: OSI Approved :: MIT License', 'Programming Language :: Python :: 3.6', ], )
810
25.16129
65
py
vector-quantize-pytorch
vector-quantize-pytorch-master/examples/autoencoder.py
# FashionMnist VQ experiment with various settings. # From https://github.com/minyoungg/vqtorch/blob/main/examples/autoencoder.py from tqdm.auto import trange import torch import torch.nn as nn from torchvision import datasets, transforms from torch.utils.data import DataLoader from vector_quantize_pytorch import VectorQuantize lr = 3e-4 train_iter = 1000 num_codes = 256 seed = 1234 device = "cuda" if torch.cuda.is_available() else "cpu" class SimpleVQAutoEncoder(nn.Module): def __init__(self, **vq_kwargs): super().__init__() self.layers = nn.ModuleList( [ nn.Conv2d(1, 16, kernel_size=3, stride=1, padding=1), nn.MaxPool2d(kernel_size=2, stride=2), nn.GELU(), nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1), nn.MaxPool2d(kernel_size=2, stride=2), VectorQuantize(dim=32, **vq_kwargs), nn.Upsample(scale_factor=2, mode="nearest"), nn.Conv2d(32, 16, kernel_size=3, stride=1, padding=1), nn.GELU(), nn.Upsample(scale_factor=2, mode="nearest"), nn.Conv2d(16, 1, kernel_size=3, stride=1, padding=1), ] ) return def forward(self, x): for layer in self.layers: if isinstance(layer, VectorQuantize): x_shape = x.shape[:-1] x_flat = x.view(x.size(0), -1, x.size(1)) x_flat, indices, commit_loss = layer(x_flat) x = x_flat.view(*x_shape, -1) else: x = layer(x) return x.clamp(-1, 1), indices, commit_loss def train(model, train_loader, train_iterations=1000, alpha=10): def iterate_dataset(data_loader): data_iter = iter(data_loader) while True: try: x, y = next(data_iter) except StopIteration: data_iter = iter(data_loader) x, y = next(data_iter) yield x.to(device), y.to(device) for _ in (pbar := trange(train_iterations)): opt.zero_grad() x, _ = next(iterate_dataset(train_loader)) out, indices, cmt_loss = model(x) rec_loss = (out - x).abs().mean() (rec_loss + alpha * cmt_loss).backward() opt.step() pbar.set_description( f"rec loss: {rec_loss.item():.3f} | " + f"cmt loss: {cmt_loss.item():.3f} | " + f"active %: {indices.unique().numel() / num_codes * 100:.3f}" ) return transform = transforms.Compose( [transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))] ) train_dataset = DataLoader( datasets.FashionMNIST( root="~/data/fashion_mnist", train=True, download=True, transform=transform ), batch_size=256, shuffle=True, ) print("baseline") torch.random.manual_seed(seed) model = SimpleVQAutoEncoder(codebook_size=num_codes).to(device) opt = torch.optim.AdamW(model.parameters(), lr=lr) train(model, train_dataset, train_iterations=train_iter)
3,076
31.052083
83
py
vector-quantize-pytorch
vector-quantize-pytorch-master/vector_quantize_pytorch/vector_quantize_pytorch.py
from functools import partial import torch from torch import nn, einsum import torch.nn.functional as F import torch.distributed as distributed from torch.optim import Optimizer from torch.cuda.amp import autocast from einops import rearrange, repeat, reduce, pack, unpack from typing import Callable def exists(val): return val is not None def default(val, d): return val if exists(val) else d def noop(*args, **kwargs): pass def identity(t): return t def l2norm(t): return F.normalize(t, p = 2, dim = -1) def cdist(x, y): x2 = reduce(x ** 2, 'b n d -> b n', 'sum') y2 = reduce(y ** 2, 'b n d -> b n', 'sum') xy = einsum('b i d, b j d -> b i j', x, y) * -2 return (rearrange(x2, 'b i -> b i 1') + rearrange(y2, 'b j -> b 1 j') + xy).sqrt() def log(t, eps = 1e-20): return torch.log(t.clamp(min = eps)) def ema_inplace(old, new, decay): is_mps = str(old.device).startswith('mps:') if not is_mps: old.lerp_(new, 1 - decay) else: old.mul_(decay).add_(new * (1 - decay)) def pack_one(t, pattern): return pack([t], pattern) def unpack_one(t, ps, pattern): return unpack(t, ps, pattern)[0] def uniform_init(*shape): t = torch.empty(shape) nn.init.kaiming_uniform_(t) return t def gumbel_noise(t): noise = torch.zeros_like(t).uniform_(0, 1) return -log(-log(noise)) def gumbel_sample( logits, temperature = 1., stochastic = False, straight_through = False, reinmax = False, dim = -1, training = True ): dtype, size = logits.dtype, logits.shape[dim] if training and stochastic and temperature > 0: sampling_logits = (logits / temperature) + gumbel_noise(logits) else: sampling_logits = logits ind = sampling_logits.argmax(dim = dim) one_hot = F.one_hot(ind, size).type(dtype) assert not (reinmax and not straight_through), 'reinmax can only be turned on if using straight through gumbel softmax' if not straight_through or temperature <= 0. or not training: return ind, one_hot # use reinmax for better second-order accuracy - https://arxiv.org/abs/2304.08612 # algorithm 2 if reinmax: π0 = logits.softmax(dim = dim) π1 = (one_hot + (logits / temperature).softmax(dim = dim)) / 2 π1 = ((log(π1) - logits).detach() + logits).softmax(dim = 1) π2 = 2 * π1 - 0.5 * π0 one_hot = π2 - π2.detach() + one_hot else: π1 = (logits / temperature).softmax(dim = dim) one_hot = one_hot + π1 - π1.detach() return ind, one_hot def laplace_smoothing(x, n_categories, eps = 1e-5, dim = -1): denom = x.sum(dim = dim, keepdim = True) return (x + eps) / (denom + n_categories * eps) def sample_vectors(samples, num): num_samples, device = samples.shape[0], samples.device if num_samples >= num: indices = torch.randperm(num_samples, device = device)[:num] else: indices = torch.randint(0, num_samples, (num,), device = device) return samples[indices] def batched_sample_vectors(samples, num): return torch.stack([sample_vectors(sample, num) for sample in samples.unbind(dim = 0)], dim = 0) def pad_shape(shape, size, dim = 0): return [size if i == dim else s for i, s in enumerate(shape)] def sample_multinomial(total_count, probs): device = probs.device probs = probs.cpu() total_count = probs.new_full((), total_count) remainder = probs.new_ones(()) sample = torch.empty_like(probs, dtype = torch.long) for i, p in enumerate(probs): s = torch.binomial(total_count, p / remainder) sample[i] = s total_count -= s remainder -= p return sample.to(device) def all_gather_sizes(x, dim): size = torch.tensor(x.shape[dim], dtype = torch.long, device = x.device) all_sizes = [torch.empty_like(size) for _ in range(distributed.get_world_size())] distributed.all_gather(all_sizes, size) return torch.stack(all_sizes) def all_gather_variably_sized(x, sizes, dim = 0): rank = distributed.get_rank() all_x = [] for i, size in enumerate(sizes): t = x if i == rank else x.new_empty(pad_shape(x.shape, size, dim)) distributed.broadcast(t, src = i, async_op = True) all_x.append(t) distributed.barrier() return all_x def sample_vectors_distributed(local_samples, num): local_samples = rearrange(local_samples, '1 ... -> ...') rank = distributed.get_rank() all_num_samples = all_gather_sizes(local_samples, dim = 0) if rank == 0: samples_per_rank = sample_multinomial(num, all_num_samples / all_num_samples.sum()) else: samples_per_rank = torch.empty_like(all_num_samples) distributed.broadcast(samples_per_rank, src = 0) samples_per_rank = samples_per_rank.tolist() local_samples = sample_vectors(local_samples, samples_per_rank[rank]) all_samples = all_gather_variably_sized(local_samples, samples_per_rank, dim = 0) out = torch.cat(all_samples, dim = 0) return rearrange(out, '... -> 1 ...') def batched_bincount(x, *, minlength): batch, dtype, device = x.shape[0], x.dtype, x.device target = torch.zeros(batch, minlength, dtype = dtype, device = device) values = torch.ones_like(x) target.scatter_add_(-1, x, values) return target def kmeans( samples, num_clusters, num_iters = 10, use_cosine_sim = False, sample_fn = batched_sample_vectors, all_reduce_fn = noop ): num_codebooks, dim, dtype, device = samples.shape[0], samples.shape[-1], samples.dtype, samples.device means = sample_fn(samples, num_clusters) for _ in range(num_iters): if use_cosine_sim: dists = samples @ rearrange(means, 'h n d -> h d n') else: dists = -torch.cdist(samples, means, p = 2) buckets = torch.argmax(dists, dim = -1) bins = batched_bincount(buckets, minlength = num_clusters) all_reduce_fn(bins) zero_mask = bins == 0 bins_min_clamped = bins.masked_fill(zero_mask, 1) new_means = buckets.new_zeros(num_codebooks, num_clusters, dim, dtype = dtype) new_means.scatter_add_(1, repeat(buckets, 'h n -> h n d', d = dim), samples) new_means = new_means / rearrange(bins_min_clamped, '... -> ... 1') all_reduce_fn(new_means) if use_cosine_sim: new_means = l2norm(new_means) means = torch.where( rearrange(zero_mask, '... -> ... 1'), means, new_means ) return means, bins def batched_embedding(indices, embeds): batch, dim = indices.shape[1], embeds.shape[-1] indices = repeat(indices, 'h b n -> h b n d', d = dim) embeds = repeat(embeds, 'h c d -> h b c d', b = batch) return embeds.gather(2, indices) # regularization losses def orthogonal_loss_fn(t): # eq (2) from https://arxiv.org/abs/2112.00384 h, n = t.shape[:2] normed_codes = l2norm(t) cosine_sim = einsum('h i d, h j d -> h i j', normed_codes, normed_codes) return (cosine_sim ** 2).sum() / (h * n ** 2) - (1 / n) # distance types class EuclideanCodebook(nn.Module): def __init__( self, dim, codebook_size, num_codebooks = 1, kmeans_init = False, kmeans_iters = 10, sync_kmeans = True, decay = 0.8, eps = 1e-5, threshold_ema_dead_code = 2, reset_cluster_size = None, use_ddp = False, learnable_codebook = False, gumbel_sample = gumbel_sample, sample_codebook_temp = 1., ema_update = True, affine_param = False, sync_affine_param = False, affine_param_batch_decay = 0.99, affine_param_codebook_decay = 0.9 ): super().__init__() self.transform_input = identity self.decay = decay self.ema_update = ema_update init_fn = uniform_init if not kmeans_init else torch.zeros embed = init_fn(num_codebooks, codebook_size, dim) self.codebook_size = codebook_size self.num_codebooks = num_codebooks self.kmeans_iters = kmeans_iters self.eps = eps self.threshold_ema_dead_code = threshold_ema_dead_code self.reset_cluster_size = default(reset_cluster_size, threshold_ema_dead_code) assert callable(gumbel_sample) self.gumbel_sample = gumbel_sample self.sample_codebook_temp = sample_codebook_temp assert not (use_ddp and num_codebooks > 1 and kmeans_init), 'kmeans init is not compatible with multiple codebooks in distributed environment for now' self.sample_fn = sample_vectors_distributed if use_ddp and sync_kmeans else batched_sample_vectors self.kmeans_all_reduce_fn = distributed.all_reduce if use_ddp and sync_kmeans else noop self.all_reduce_fn = distributed.all_reduce if use_ddp else noop self.register_buffer('initted', torch.Tensor([not kmeans_init])) self.register_buffer('cluster_size', torch.zeros(num_codebooks, codebook_size)) self.register_buffer('embed_avg', embed.clone()) self.learnable_codebook = learnable_codebook if learnable_codebook: self.embed = nn.Parameter(embed) else: self.register_buffer('embed', embed) # affine related params self.affine_param = affine_param self.sync_affine_param = sync_affine_param if not affine_param: return self.affine_param_batch_decay = affine_param_batch_decay self.affine_param_codebook_decay = affine_param_codebook_decay self.register_buffer('batch_mean', None) self.register_buffer('batch_variance', None) self.register_buffer('codebook_mean_needs_init', torch.Tensor([True])) self.register_buffer('codebook_mean', torch.empty(num_codebooks, 1, dim)) self.register_buffer('codebook_variance_needs_init', torch.Tensor([True])) self.register_buffer('codebook_variance', torch.empty(num_codebooks, 1, dim)) @torch.jit.ignore def init_embed_(self, data, mask = None): if self.initted: return if exists(mask): c = data.shape[0] data = rearrange(data[mask], '(c n) d -> c n d', c = c) embed, cluster_size = kmeans( data, self.codebook_size, self.kmeans_iters, sample_fn = self.sample_fn, all_reduce_fn = self.kmeans_all_reduce_fn ) self.embed.data.copy_(embed) self.embed_avg.data.copy_(embed.clone()) self.cluster_size.data.copy_(cluster_size) self.initted.data.copy_(torch.Tensor([True])) @torch.jit.ignore def update_with_decay(self, buffer_name, new_value, decay): old_value = getattr(self, buffer_name) needs_init = getattr(self, buffer_name + "_needs_init", False) if needs_init: self.register_buffer(buffer_name + "_needs_init", torch.Tensor([False])) if not exists(old_value) or needs_init: self.register_buffer(buffer_name, new_value.detach()) return value = old_value * decay + new_value.detach() * (1 - decay) self.register_buffer(buffer_name, value) @torch.jit.ignore def update_affine(self, data, embed, mask = None): assert self.affine_param var_fn = partial(torch.var, unbiased = False) # calculate codebook mean and variance embed = rearrange(embed, 'h ... d -> h (...) d') if self.training: self.update_with_decay('codebook_mean', reduce(embed, 'h n d -> h 1 d', 'mean'), self.affine_param_codebook_decay) self.update_with_decay('codebook_variance', reduce(embed, 'h n d -> h 1 d', var_fn), self.affine_param_codebook_decay) # prepare batch data, which depends on whether it has masking data = rearrange(data, 'h ... d -> h (...) d') if exists(mask): c = data.shape[0] data = rearrange(data[mask], '(c n) d -> c n d', c = c) # calculate batch mean and variance if not self.sync_affine_param: self.update_with_decay('batch_mean', reduce(data, 'h n d -> h 1 d', 'mean'), self.affine_param_batch_decay) self.update_with_decay('batch_variance', reduce(data, 'h n d -> h 1 d', var_fn), self.affine_param_batch_decay) return num_vectors, device, dtype = data.shape[-2], data.device, data.dtype # number of vectors, for denominator num_vectors = torch.tensor([num_vectors], device = device, dtype = dtype) distributed.all_reduce(num_vectors) # calculate distributed mean batch_sum = reduce(data, 'h n d -> h 1 d', 'sum') distributed.all_reduce(batch_sum) batch_mean = batch_sum / num_vectors self.update_with_decay('batch_mean', batch_mean, self.affine_param_batch_decay) # calculate distributed variance variance_numer = reduce((data - batch_mean) ** 2, 'h n d -> h 1 d', 'sum') distributed.all_reduce(variance_numer) batch_variance = variance_numer / num_vectors self.update_with_decay('batch_variance', batch_variance, self.affine_param_batch_decay) def replace(self, batch_samples, batch_mask): for ind, (samples, mask) in enumerate(zip(batch_samples.unbind(dim = 0), batch_mask.unbind(dim = 0))): if not torch.any(mask): continue sampled = self.sample_fn(rearrange(samples, '... -> 1 ...'), mask.sum().item()) sampled = rearrange(sampled, '1 ... -> ...') self.embed.data[ind][mask] = sampled self.cluster_size.data[ind][mask] = self.reset_cluster_size self.embed_avg.data[ind][mask] = sampled * self.reset_cluster_size def expire_codes_(self, batch_samples): if self.threshold_ema_dead_code == 0: return expired_codes = self.cluster_size < self.threshold_ema_dead_code if not torch.any(expired_codes): return batch_samples = rearrange(batch_samples, 'h ... d -> h (...) d') self.replace(batch_samples, batch_mask = expired_codes) @autocast(enabled = False) def forward( self, x, sample_codebook_temp = None, mask = None ): needs_codebook_dim = x.ndim < 4 sample_codebook_temp = default(sample_codebook_temp, self.sample_codebook_temp) x = x.float() if needs_codebook_dim: x = rearrange(x, '... -> 1 ...') dtype = x.dtype flatten, ps = pack_one(x, 'h * d') if exists(mask): mask = repeat(mask, 'b n -> c (b h n)', c = flatten.shape[0], h = flatten.shape[-2] // (mask.shape[0] * mask.shape[1])) self.init_embed_(flatten, mask = mask) if self.affine_param: self.update_affine(flatten, self.embed, mask = mask) embed = self.embed if self.learnable_codebook else self.embed.detach() if self.affine_param: codebook_std = self.codebook_variance.clamp(min = 1e-5).sqrt() batch_std = self.batch_variance.clamp(min = 1e-5).sqrt() embed = (embed - self.codebook_mean) * (batch_std / codebook_std) + self.batch_mean dist = -cdist(flatten, embed) embed_ind, embed_onehot = self.gumbel_sample(dist, dim = -1, temperature = sample_codebook_temp, training = self.training) embed_ind = unpack_one(embed_ind, ps, 'h *') if self.training: unpacked_onehot = unpack_one(embed_onehot, ps, 'h * c') quantize = einsum('h b n c, h c d -> h b n d', unpacked_onehot, embed) else: quantize = batched_embedding(embed_ind, embed) if self.training and self.ema_update: if self.affine_param: flatten = (flatten - self.batch_mean) * (codebook_std / batch_std) + self.codebook_mean if exists(mask): embed_onehot[~mask] = 0. cluster_size = embed_onehot.sum(dim = 1) self.all_reduce_fn(cluster_size) ema_inplace(self.cluster_size.data, cluster_size, self.decay) embed_sum = einsum('h n d, h n c -> h c d', flatten, embed_onehot) self.all_reduce_fn(embed_sum.contiguous()) ema_inplace(self.embed_avg.data, embed_sum, self.decay) cluster_size = laplace_smoothing(self.cluster_size, self.codebook_size, self.eps) * self.cluster_size.sum(dim = -1, keepdim = True) embed_normalized = self.embed_avg / rearrange(cluster_size, '... -> ... 1') self.embed.data.copy_(embed_normalized) self.expire_codes_(x) if needs_codebook_dim: quantize, embed_ind = map(lambda t: rearrange(t, '1 ... -> ...'), (quantize, embed_ind)) dist = unpack_one(dist, ps, 'h * d') return quantize, embed_ind, dist class CosineSimCodebook(nn.Module): def __init__( self, dim, codebook_size, num_codebooks = 1, kmeans_init = False, kmeans_iters = 10, sync_kmeans = True, decay = 0.8, eps = 1e-5, threshold_ema_dead_code = 2, reset_cluster_size = None, use_ddp = False, learnable_codebook = False, gumbel_sample = gumbel_sample, sample_codebook_temp = 1., ema_update = True ): super().__init__() self.transform_input = l2norm self.ema_update = ema_update self.decay = decay if not kmeans_init: embed = l2norm(uniform_init(num_codebooks, codebook_size, dim)) else: embed = torch.zeros(num_codebooks, codebook_size, dim) self.codebook_size = codebook_size self.num_codebooks = num_codebooks self.kmeans_iters = kmeans_iters self.eps = eps self.threshold_ema_dead_code = threshold_ema_dead_code self.reset_cluster_size = default(reset_cluster_size, threshold_ema_dead_code) assert callable(gumbel_sample) self.gumbel_sample = gumbel_sample self.sample_codebook_temp = sample_codebook_temp self.sample_fn = sample_vectors_distributed if use_ddp and sync_kmeans else batched_sample_vectors self.kmeans_all_reduce_fn = distributed.all_reduce if use_ddp and sync_kmeans else noop self.all_reduce_fn = distributed.all_reduce if use_ddp else noop self.register_buffer('initted', torch.Tensor([not kmeans_init])) self.register_buffer('cluster_size', torch.zeros(num_codebooks, codebook_size)) self.register_buffer('embed_avg', embed.clone()) self.learnable_codebook = learnable_codebook if learnable_codebook: self.embed = nn.Parameter(embed) else: self.register_buffer('embed', embed) @torch.jit.ignore def init_embed_(self, data, mask = None): if self.initted: return if exists(mask): c = data.shape[0] data = rearrange(data[mask], '(c n) d -> c n d', c = c) embed, cluster_size = kmeans( data, self.codebook_size, self.kmeans_iters, use_cosine_sim = True, sample_fn = self.sample_fn, all_reduce_fn = self.kmeans_all_reduce_fn ) self.embed.data.copy_(embed) self.embed_avg.data.copy_(embed.clone()) self.cluster_size.data.copy_(cluster_size) self.initted.data.copy_(torch.Tensor([True])) def replace(self, batch_samples, batch_mask): batch_samples = l2norm(batch_samples) for ind, (samples, mask) in enumerate(zip(batch_samples.unbind(dim = 0), batch_mask.unbind(dim = 0))): if not torch.any(mask): continue sampled = self.sample_fn(rearrange(samples, '... -> 1 ...'), mask.sum().item()) sampled = rearrange(sampled, '1 ... -> ...') self.embed.data[ind][mask] = sampled self.embed_avg.data[ind][mask] = sampled * self.reset_cluster_size self.cluster_size.data[ind][mask] = self.reset_cluster_size def expire_codes_(self, batch_samples): if self.threshold_ema_dead_code == 0: return expired_codes = self.cluster_size < self.threshold_ema_dead_code if not torch.any(expired_codes): return batch_samples = rearrange(batch_samples, 'h ... d -> h (...) d') self.replace(batch_samples, batch_mask = expired_codes) @autocast(enabled = False) def forward( self, x, sample_codebook_temp = None, mask = None ): needs_codebook_dim = x.ndim < 4 sample_codebook_temp = default(sample_codebook_temp, self.sample_codebook_temp) x = x.float() if needs_codebook_dim: x = rearrange(x, '... -> 1 ...') dtype = x.dtype flatten, ps = pack_one(x, 'h * d') if exists(mask): mask = repeat(mask, 'b n -> c (b h n)', c = flatten.shape[0], h = flatten.shape[-2] // (mask.shape[0] * mask.shape[1])) self.init_embed_(flatten, mask = mask) embed = self.embed if self.learnable_codebook else self.embed.detach() dist = einsum('h n d, h c d -> h n c', flatten, embed) embed_ind, embed_onehot = self.gumbel_sample(dist, dim = -1, temperature = sample_codebook_temp, training = self.training) embed_ind = unpack_one(embed_ind, ps, 'h *') if self.training: unpacked_onehot = unpack_one(embed_onehot, ps, 'h * c') quantize = einsum('h b n c, h c d -> h b n d', unpacked_onehot, embed) else: quantize = batched_embedding(embed_ind, embed) if self.training and self.ema_update: if exists(mask): embed_onehot[~mask] = 0. bins = embed_onehot.sum(dim = 1) self.all_reduce_fn(bins) ema_inplace(self.cluster_size.data, bins, self.decay) embed_sum = einsum('h n d, h n c -> h c d', flatten, embed_onehot) self.all_reduce_fn(embed_sum.contiguous()) ema_inplace(self.embed_avg.data, embed_sum, self.decay) cluster_size = laplace_smoothing(self.cluster_size, self.codebook_size, self.eps) * self.cluster_size.sum(dim = -1, keepdim = True) embed_normalized = self.embed_avg / rearrange(cluster_size, '... -> ... 1') embed_normalized = l2norm(embed_normalized) self.embed.data.copy_(l2norm(embed_normalized)) self.expire_codes_(x) if needs_codebook_dim: quantize, embed_ind = map(lambda t: rearrange(t, '1 ... -> ...'), (quantize, embed_ind)) dist = unpack_one(dist, ps, 'h * d') return quantize, embed_ind, dist # main class class VectorQuantize(nn.Module): def __init__( self, dim, codebook_size, codebook_dim = None, heads = 1, separate_codebook_per_head = False, decay = 0.8, eps = 1e-5, kmeans_init = False, kmeans_iters = 10, sync_kmeans = True, use_cosine_sim = False, threshold_ema_dead_code = 0, channel_last = True, accept_image_fmap = False, commitment_weight = 1., commitment_use_cross_entropy_loss = False, orthogonal_reg_weight = 0., orthogonal_reg_active_codes_only = False, orthogonal_reg_max_codes = None, stochastic_sample_codes = False, sample_codebook_temp = 1., straight_through = False, reinmax = False, # using reinmax for improved straight-through, assuming straight through helps at all sync_codebook = False, sync_affine_param = False, ema_update = True, learnable_codebook = False, in_place_codebook_optimizer: Callable[..., Optimizer] = None, # Optimizer used to update the codebook embedding if using learnable_codebook affine_param = False, affine_param_batch_decay = 0.99, affine_param_codebook_decay = 0.9, sync_update_v = 0. # the v that controls optimistic vs pessimistic update for synchronous update rule (21) https://minyoungg.github.io/vqtorch/assets/draft_050523.pdf ): super().__init__() self.dim = dim self.heads = heads self.separate_codebook_per_head = separate_codebook_per_head codebook_dim = default(codebook_dim, dim) codebook_input_dim = codebook_dim * heads requires_projection = codebook_input_dim != dim self.project_in = nn.Linear(dim, codebook_input_dim) if requires_projection else nn.Identity() self.project_out = nn.Linear(codebook_input_dim, dim) if requires_projection else nn.Identity() self.eps = eps self.commitment_weight = commitment_weight self.commitment_use_cross_entropy_loss = commitment_use_cross_entropy_loss # whether to use cross entropy loss to codebook as commitment loss self.learnable_codebook = learnable_codebook has_codebook_orthogonal_loss = orthogonal_reg_weight > 0 self.has_codebook_orthogonal_loss = has_codebook_orthogonal_loss self.orthogonal_reg_weight = orthogonal_reg_weight self.orthogonal_reg_active_codes_only = orthogonal_reg_active_codes_only self.orthogonal_reg_max_codes = orthogonal_reg_max_codes assert not (ema_update and learnable_codebook), 'learnable codebook not compatible with EMA update' assert 0 <= sync_update_v <= 1. assert not (sync_update_v > 0. and not learnable_codebook), 'learnable codebook must be turned on' self.sync_update_v = sync_update_v codebook_class = EuclideanCodebook if not use_cosine_sim else CosineSimCodebook gumbel_sample_fn = partial( gumbel_sample, stochastic = stochastic_sample_codes, reinmax = reinmax, straight_through = straight_through ) codebook_kwargs = dict( dim = codebook_dim, num_codebooks = heads if separate_codebook_per_head else 1, codebook_size = codebook_size, kmeans_init = kmeans_init, kmeans_iters = kmeans_iters, sync_kmeans = sync_kmeans, decay = decay, eps = eps, threshold_ema_dead_code = threshold_ema_dead_code, use_ddp = sync_codebook, learnable_codebook = has_codebook_orthogonal_loss or learnable_codebook, sample_codebook_temp = sample_codebook_temp, gumbel_sample = gumbel_sample_fn, ema_update = ema_update ) if affine_param: assert not use_cosine_sim, 'affine param is only compatible with euclidean codebook' codebook_kwargs = dict( **codebook_kwargs, affine_param = True, sync_affine_param = sync_affine_param, affine_param_batch_decay = affine_param_batch_decay, affine_param_codebook_decay = affine_param_codebook_decay, ) self._codebook = codebook_class(**codebook_kwargs) self.in_place_codebook_optimizer = in_place_codebook_optimizer(self._codebook.parameters()) if exists(in_place_codebook_optimizer) else None self.codebook_size = codebook_size self.accept_image_fmap = accept_image_fmap self.channel_last = channel_last @property def codebook(self): codebook = self._codebook.embed if self.separate_codebook_per_head: return codebook return rearrange(codebook, '1 ... -> ...') def get_codes_from_indices(self, indices): codebook = self.codebook is_multiheaded = codebook.ndim > 2 if not is_multiheaded: codes = codebook[indices] return rearrange(codes, '... h d -> ... (h d)') indices, ps = pack_one(indices, 'b * h') indices = rearrange(indices, 'b n h -> b h n') indices = repeat(indices, 'b h n -> b h n d', d = codebook.shape[-1]) codebook = repeat(codebook, 'h n d -> b h n d', b = indices.shape[0]) codes = codebook.gather(2, indices) codes = rearrange(codes, 'b h n d -> b n (h d)') codes = unpack_one(codes, ps, 'b * d') return codes def forward( self, x, indices = None, mask = None, sample_codebook_temp = None ): orig_input = x only_one = x.ndim == 2 if only_one: assert not exists(mask) x = rearrange(x, 'b d -> b 1 d') shape, device, heads, is_multiheaded, codebook_size, return_loss = x.shape, x.device, self.heads, self.heads > 1, self.codebook_size, exists(indices) need_transpose = not self.channel_last and not self.accept_image_fmap should_inplace_optimize = exists(self.in_place_codebook_optimizer) # rearrange inputs if self.accept_image_fmap: height, width = x.shape[-2:] x = rearrange(x, 'b c h w -> b (h w) c') if need_transpose: x = rearrange(x, 'b d n -> b n d') # project input x = self.project_in(x) # handle multi-headed separate codebooks if is_multiheaded: ein_rhs_eq = 'h b n d' if self.separate_codebook_per_head else '1 (b h) n d' x = rearrange(x, f'b n (h d) -> {ein_rhs_eq}', h = heads) # l2norm for cosine sim, otherwise identity x = self._codebook.transform_input(x) # codebook forward kwargs codebook_forward_kwargs = dict( sample_codebook_temp = sample_codebook_temp, mask = mask ) # quantize quantize, embed_ind, distances = self._codebook(x, **codebook_forward_kwargs) # one step in-place update if should_inplace_optimize and self.training: if exists(mask): loss = F.mse_loss(quantize, x.detach(), reduction = 'none') loss_mask = mask if is_multiheaded: loss_mask = repeat(mask, 'b n -> c (b h) n', c = loss.shape[0], h = loss.shape[1] // mask.shape[0]) loss = loss[loss_mask].mean() else: loss = F.mse_loss(quantize, x.detach()) loss.backward() self.in_place_codebook_optimizer.step() self.in_place_codebook_optimizer.zero_grad() # quantize again quantize, embed_ind, distances = self._codebook(x, **codebook_forward_kwargs) if self.training: # determine code to use for commitment loss maybe_detach = torch.detach if not self.learnable_codebook else identity commit_quantize = maybe_detach(quantize) # straight through quantize = x + (quantize - x).detach() if self.sync_update_v > 0.: # (21) in https://minyoungg.github.io/vqtorch/assets/draft_050523.pdf quantize = quantize + self.sync_update_v * (quantize - quantize.detach()) # function for calculating cross entropy loss to distance matrix # used for (1) naturalspeech2 training residual vq latents to be close to the correct codes and (2) cross-entropy based commitment loss def calculate_ce_loss(codes): if not is_multiheaded: dist_einops_eq = '1 b n l -> b l n' elif self.separate_codebook_per_head: dist_einops_eq = 'c b n l -> b l n c' else: dist_einops_eq = '1 (b h) n l -> b l n h' ce_loss = F.cross_entropy( rearrange(distances, dist_einops_eq, b = shape[0]), codes, ignore_index = -1 ) return ce_loss # if returning cross entropy loss on codes that were passed in if return_loss: return quantize, calculate_ce_loss(indices) # transform embedding indices if is_multiheaded: if self.separate_codebook_per_head: embed_ind = rearrange(embed_ind, 'h b n -> b n h', h = heads) else: embed_ind = rearrange(embed_ind, '1 (b h) n -> b n h', h = heads) if self.accept_image_fmap: embed_ind = rearrange(embed_ind, 'b (h w) ... -> b h w ...', h = height, w = width) if only_one: embed_ind = rearrange(embed_ind, 'b 1 -> b') # aggregate loss loss = torch.tensor([0.], device = device, requires_grad = self.training) if self.training: if self.commitment_weight > 0: if self.commitment_use_cross_entropy_loss: if exists(mask): ce_loss_mask = mask if is_multiheaded: ce_loss_mask = repeat(ce_loss_mask, 'b n -> b n h', h = heads) embed_ind.masked_fill_(~ce_loss_mask, -1) commit_loss = calculate_ce_loss(embed_ind) else: if exists(mask): # with variable lengthed sequences commit_loss = F.mse_loss(commit_quantize, x, reduction = 'none') loss_mask = mask if is_multiheaded: loss_mask = repeat(loss_mask, 'b n -> c (b h) n', c = commit_loss.shape[0], h = commit_loss.shape[1] // mask.shape[0]) commit_loss = commit_loss[loss_mask].mean() else: commit_loss = F.mse_loss(commit_quantize, x) loss = loss + commit_loss * self.commitment_weight if self.has_codebook_orthogonal_loss: codebook = self._codebook.embed # only calculate orthogonal loss for the activated codes for this batch if self.orthogonal_reg_active_codes_only: assert not (is_multiheaded and self.separate_codebook_per_head), 'orthogonal regularization for only active codes not compatible with multi-headed with separate codebooks yet' unique_code_ids = torch.unique(embed_ind) codebook = codebook[:, unique_code_ids] num_codes = codebook.shape[-2] if exists(self.orthogonal_reg_max_codes) and num_codes > self.orthogonal_reg_max_codes: rand_ids = torch.randperm(num_codes, device = device)[:self.orthogonal_reg_max_codes] codebook = codebook[:, rand_ids] orthogonal_reg_loss = orthogonal_loss_fn(codebook) loss = loss + orthogonal_reg_loss * self.orthogonal_reg_weight # handle multi-headed quantized embeddings if is_multiheaded: if self.separate_codebook_per_head: quantize = rearrange(quantize, 'h b n d -> b n (h d)', h = heads) else: quantize = rearrange(quantize, '1 (b h) n d -> b n (h d)', h = heads) # project out quantize = self.project_out(quantize) # rearrange quantized embeddings if need_transpose: quantize = rearrange(quantize, 'b n d -> b d n') if self.accept_image_fmap: quantize = rearrange(quantize, 'b (h w) c -> b c h w', h = height, w = width) if only_one: quantize = rearrange(quantize, 'b 1 d -> b d') # if masking, only return quantized for where mask has True if exists(mask): quantize = torch.where( rearrange(mask, '... -> ... 1'), quantize, orig_input ) return quantize, embed_ind, loss
35,993
33.911736
195
py
vector-quantize-pytorch
vector-quantize-pytorch-master/vector_quantize_pytorch/random_projection_quantizer.py
import torch from torch import nn, einsum import torch.nn.functional as F from vector_quantize_pytorch.vector_quantize_pytorch import VectorQuantize from einops import rearrange, repeat, pack, unpack def exists(val): return val is not None class RandomProjectionQuantizer(nn.Module): """ https://arxiv.org/abs/2202.01855 """ def __init__( self, *, dim, codebook_size, codebook_dim, num_codebooks = 1, norm = True, **kwargs ): super().__init__() self.num_codebooks = num_codebooks rand_projs = torch.empty(num_codebooks, dim, codebook_dim) nn.init.xavier_normal_(rand_projs) self.register_buffer('rand_projs', rand_projs) # in section 3 of https://arxiv.org/abs/2202.01855 # "The input data is normalized to have 0 mean and standard deviation of 1 ... to prevent collapse" self.norm = nn.LayerNorm(dim, elementwise_affine = False) if norm else nn.Identity() self.vq = VectorQuantize( dim = codebook_dim * num_codebooks, heads = num_codebooks, codebook_size = codebook_size, use_cosine_sim = True, separate_codebook_per_head = True, **kwargs ) def forward( self, x, indices = None ): return_loss = exists(indices) x = self.norm(x) x = einsum('b n d, h d e -> b n h e', x, self.rand_projs) x, ps = pack([x], 'b n *') self.vq.eval() out = self.vq(x, indices = indices) if return_loss: _, ce_loss = out return ce_loss _, indices, _ = out return indices
1,723
24.731343
107
py
vector-quantize-pytorch
vector-quantize-pytorch-master/vector_quantize_pytorch/residual_vq.py
from math import ceil from functools import partial from itertools import zip_longest from random import randrange import torch from torch import nn import torch.nn.functional as F from vector_quantize_pytorch.vector_quantize_pytorch import VectorQuantize from einops import rearrange, repeat, pack, unpack # helper functions def exists(val): return val is not None def default(val, d): return val if exists(val) else d def round_up_multiple(num, mult): return ceil(num / mult) * mult # main class class ResidualVQ(nn.Module): """ Follows Algorithm 1. in https://arxiv.org/pdf/2107.03312.pdf """ def __init__( self, *, num_quantizers, shared_codebook = False, heads = 1, quantize_dropout = False, quantize_dropout_cutoff_index = 0, quantize_dropout_multiple_of = 1, accept_image_fmap = False, **kwargs ): super().__init__() assert heads == 1, 'residual vq is not compatible with multi-headed codes' self.num_quantizers = num_quantizers self.accept_image_fmap = accept_image_fmap self.layers = nn.ModuleList([VectorQuantize(accept_image_fmap = accept_image_fmap, **kwargs) for _ in range(num_quantizers)]) self.quantize_dropout = quantize_dropout and num_quantizers > 1 assert quantize_dropout_cutoff_index >= 0 self.quantize_dropout_cutoff_index = quantize_dropout_cutoff_index self.quantize_dropout_multiple_of = quantize_dropout_multiple_of # encodec paper proposes structured dropout, believe this was set to 4 if not shared_codebook: return first_vq, *rest_vq = self.layers codebook = first_vq._codebook for vq in rest_vq: vq._codebook = codebook @property def codebooks(self): codebooks = [layer._codebook.embed for layer in self.layers] codebooks = torch.stack(codebooks, dim = 0) codebooks = rearrange(codebooks, 'q 1 c d -> q c d') return codebooks def get_codes_from_indices(self, indices): batch, quantize_dim = indices.shape[0], indices.shape[-1] # may also receive indices in the shape of 'b h w q' (accept_image_fmap) indices, ps = pack([indices], 'b * q') # because of quantize dropout, one can pass in indices that are coarse # and the network should be able to reconstruct if quantize_dim < self.num_quantizers: assert self.quantize_dropout > 0., 'quantize dropout must be greater than 0 if you wish to reconstruct from a signal with less fine quantizations' indices = F.pad(indices, (0, self.num_quantizers - quantize_dim), value = -1) # get ready for gathering codebooks = repeat(self.codebooks, 'q c d -> q b c d', b = batch) gather_indices = repeat(indices, 'b n q -> q b n d', d = codebooks.shape[-1]) # take care of quantizer dropout mask = gather_indices == -1. gather_indices = gather_indices.masked_fill(mask, 0) # have it fetch a dummy code to be masked out later all_codes = codebooks.gather(2, gather_indices) # gather all codes # mask out any codes that were dropout-ed all_codes = all_codes.masked_fill(mask, 0.) # if (accept_image_fmap = True) then return shape (quantize, batch, height, width, dimension) all_codes, = unpack(all_codes, ps, 'q b * d') return all_codes def forward( self, x, indices = None, return_all_codes = False, sample_codebook_temp = None ): num_quant, quant_dropout_multiple_of, return_loss, device = self.num_quantizers, self.quantize_dropout_multiple_of, exists(indices), x.device assert not (self.accept_image_fmap and exists(indices)) quantized_out = 0. residual = x all_losses = [] all_indices = [] if return_loss: assert not torch.any(indices == -1), 'some of the residual vq indices were dropped out. please use indices derived when the module is in eval mode to derive cross entropy loss' ce_losses = [] should_quantize_dropout = self.training and self.quantize_dropout and not return_loss # sample a layer index at which to dropout further residual quantization # also prepare null indices and loss if should_quantize_dropout: rand_quantize_dropout_index = randrange(self.quantize_dropout_cutoff_index, num_quant) if quant_dropout_multiple_of != 1: rand_quantize_dropout_index = round_up_multiple(rand_quantize_dropout_index + 1, quant_dropout_multiple_of) - 1 null_indices_shape = (x.shape[0], *x.shape[-2:]) if self.accept_image_fmap else tuple(x.shape[:2]) null_indices = torch.full(null_indices_shape, -1., device = device, dtype = torch.long) null_loss = torch.full((1,), 0., device = device, dtype = x.dtype) # go through the layers for quantizer_index, layer in enumerate(self.layers): if should_quantize_dropout and quantizer_index > rand_quantize_dropout_index: all_indices.append(null_indices) all_losses.append(null_loss) continue layer_indices = None if return_loss: layer_indices = indices[..., quantizer_index] quantized, *rest = layer(residual, indices = layer_indices, sample_codebook_temp = sample_codebook_temp) residual = residual - quantized.detach() quantized_out = quantized_out + quantized if return_loss: ce_loss = rest[0] ce_losses.append(ce_loss) continue embed_indices, loss = rest all_indices.append(embed_indices) all_losses.append(loss) # whether to early return the cross entropy loss if return_loss: return quantized_out, sum(ce_losses) # stack all losses and indices all_losses, all_indices = map(partial(torch.stack, dim = -1), (all_losses, all_indices)) ret = (quantized_out, all_indices, all_losses) if return_all_codes: # whether to return all codes from all codebooks across layers all_codes = self.get_codes_from_indices(all_indices) # will return all codes in shape (quantizer, batch, sequence length, codebook dimension) ret = (*ret, all_codes) return ret # grouped residual vq class GroupedResidualVQ(nn.Module): def __init__( self, *, dim, groups = 1, accept_image_fmap = False, **kwargs ): super().__init__() self.dim = dim self.groups = groups assert (dim % groups) == 0 dim_per_group = dim // groups self.accept_image_fmap = accept_image_fmap self.rvqs = nn.ModuleList([]) for _ in range(groups): self.rvqs.append(ResidualVQ( dim = dim_per_group, accept_image_fmap = accept_image_fmap, **kwargs )) @property def codebooks(self): return torch.stack(tuple(rvq.codebooks for rvq in self.rvqs)) def get_codes_from_indices(self, indices): codes = tuple(rvq.get_codes_from_indices(chunk_indices) for rvq, chunk_indices in zip(self.rvqs, indices)) return torch.stack(codes) def forward( self, x, indices = None, return_all_codes = False, sample_codebook_temp = None ): shape = x.shape split_dim = 1 if self.accept_image_fmap else -1 assert shape[split_dim] == self.dim # split the feature dimension into groups x = x.chunk(self.groups, dim = split_dim) indices = default(indices, tuple()) return_ce_loss = len(indices) > 0 assert len(indices) == 0 or len(indices) == self.groups forward_kwargs = dict( return_all_codes = return_all_codes, sample_codebook_temp = sample_codebook_temp ) # invoke residual vq on each group out = tuple(rvq(chunk, indices = chunk_indices, **forward_kwargs) for rvq, chunk, chunk_indices in zip_longest(self.rvqs, x, indices)) out = tuple(zip(*out)) # if returning cross entropy loss to rvq codebooks if return_ce_loss: quantized, ce_losses = out return torch.cat(quantized, dim = split_dim), sum(ce_losses) # otherwise, get all the zipped outputs and combine them quantized, all_indices, commit_losses, *maybe_all_codes = out quantized = torch.cat(quantized, dim = split_dim) all_indices = torch.stack(all_indices) commit_losses = torch.stack(commit_losses) ret = (quantized, all_indices, commit_losses, *maybe_all_codes) return ret
8,989
31.930403
188
py
vector-quantize-pytorch
vector-quantize-pytorch-master/vector_quantize_pytorch/__init__.py
from vector_quantize_pytorch.vector_quantize_pytorch import VectorQuantize from vector_quantize_pytorch.residual_vq import ResidualVQ, GroupedResidualVQ from vector_quantize_pytorch.random_projection_quantizer import RandomProjectionQuantizer
242
80
89
py
smt
smt-master/doc/conf.py
#!/usr/bin/env python3 # -*- coding: utf-8 -*- # # SMT documentation build configuration file, created by # sphinx-quickstart on Sun Aug 6 19:36:14 2017. # # This file is execfile()d with the current directory set to its # containing dir. # # Note that not all possible configuration values are present in this # autogenerated file. # # All configuration values have a default; values that are commented out # serve to show the default. # 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('.')) from smt import __version__ # -- 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.mathjax", "sphinx.ext.viewcode", "matplotlib.sphinxext.plot_directive", "numpydoc", ] # 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'] source_suffix = ".rst" # The master toctree document. master_doc = "index" # General information about the project. project = "SMT" copyright = "2017, John Hwang" author = "John Hwang, Mohamed Amine Bouhlel, Remi Lafage" # The version info for the project you're documenting, acts as replacement for # |version| and |release|, also used in various other places throughout the # built documents. # # The short X.Y version. # Take the full version when no need to distinguish version and release version = __version__ # The full version, including alpha/beta/rc tags. release = version # 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 = None # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. # This patterns also effect to html_static_path and html_extra_path exclude_patterns = ["_build", "Thumbs.db", ".DS_Store"] # The name of the Pygments (syntax highlighting) style to use. pygments_style = "sphinx" # If true, `todo` and `todoList` produce output, else they produce nothing. todo_include_todos = False # -- 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 = "bizstyle" # 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 = { "rightsidebar": False, "sidebarwidth": 250, "body_min_width": 1100, "body_max_width": 1100, } html_logo = "smt_logo.png" # 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"] # -- Options for HTMLHelp output ------------------------------------------ # Output file base name for HTML help builder. htmlhelp_basename = "SMTdoc" # -- Options for LaTeX output --------------------------------------------- latex_elements = { # 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, "SMT.tex", "SMT Documentation", "John Hwang, Mohamed Amine Bouhlel, Remi Lafage", "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, "smt", "SMT 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, "SMT", "SMT Documentation", author, "SMT", "One line description of project.", "Miscellaneous", ) ]
5,197
28.873563
79
py
LongTailCXR
LongTailCXR-main/src/main.py
import os import shutil import argparse import numpy as np import pandas as pd import torch import torchvision from sklearn.utils import class_weight from datasets import * from utils import * from losses import * def main(args): # Set model/output directory name MODEL_NAME = args.dataset MODEL_NAME += f'_{args.model_name}' MODEL_NAME += f'_rand' if args.rand_init else '' MODEL_NAME += f'_bal-mixup-{args.mixup_alpha}' if args.bal_mixup else '' MODEL_NAME += f'_mixup-{args.mixup_alpha}' if args.mixup else '' MODEL_NAME += f'_decoupling-{args.decoupling_method}' if args.decoupling_method != '' else '' MODEL_NAME += f'_rw-{args.rw_method}' if args.rw_method != '' else '' MODEL_NAME += f'_{args.loss}' MODEL_NAME += '-drw' if args.drw else '' MODEL_NAME += f'_cb-beta-{args.cb_beta}' if args.rw_method == 'cb' else '' MODEL_NAME += f'_fl-gamma-{args.fl_gamma}' if args.loss == 'focal' else '' MODEL_NAME += f'_lr-{args.lr}' MODEL_NAME += f'_bs-{args.batch_size}' # Create output directory for model (and delete if already exists) if not os.path.isdir(args.out_dir): os.mkdir(args.out_dir) model_dir = os.path.join(args.out_dir, MODEL_NAME) if os.path.isdir(model_dir): shutil.rmtree(model_dir) os.mkdir(model_dir) # Set all seeds for reproducibility set_seed(args.seed) # Create datasets + loaders if args.dataset == 'nih-cxr-lt': dataset = NIH_CXR_Dataset N_CLASSES = 20 else: dataset = MIMIC_CXR_Dataset N_CLASSES = 19 train_dataset = dataset(data_dir=args.data_dir, label_dir=args.label_dir, split='train') val_dataset = dataset(data_dir=args.data_dir, label_dir=args.label_dir, split='balanced-val') bal_test_dataset = dataset(data_dir=args.data_dir, label_dir=args.label_dir, split='balanced-test') test_dataset = dataset(data_dir=args.data_dir, label_dir=args.label_dir, split='test') if args.bal_mixup: cls_weights = [len(train_dataset) / cls_count for cls_count in train_dataset.cls_num_list] instance_weights = [cls_weights[label] for label in train_dataset.labels] sampler = torch.utils.data.WeightedRandomSampler(torch.Tensor(instance_weights), len(train_dataset)) bal_train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=False, num_workers=8, pin_memory=True, worker_init_fn=worker_init_fn, sampler=sampler) imbal_train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=8, pin_memory=True, worker_init_fn=worker_init_fn) train_loader = ComboLoader([imbal_train_loader, bal_train_loader]) elif args.decoupling_method == 'cRT': cls_weights = [len(train_dataset) / cls_count for cls_count in train_dataset.cls_num_list] instance_weights = [cls_weights[label] for label in train_dataset.labels] sampler = torch.utils.data.WeightedRandomSampler(torch.Tensor(instance_weights), len(train_dataset)) train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=False, num_workers=8, pin_memory=True, worker_init_fn=worker_init_fn, sampler=sampler) else: train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=8, pin_memory=True, worker_init_fn=worker_init_fn) val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=2, pin_memory=True, worker_init_fn=val_worker_init_fn) # Create csv documenting training history history = pd.DataFrame(columns=['epoch', 'phase', 'loss', 'balanced_acc', 'mcc', 'auroc']) history.to_csv(os.path.join(model_dir, 'history.csv'), index=False) # Set device device = torch.device('cuda:0') # Instantiate model model = torchvision.models.resnet50(pretrained=(not args.rand_init)) model.fc = torch.nn.Linear(model.fc.in_features, N_CLASSES) if args.decoupling_method == 'tau_norm': msg = model.load_state_dict(torch.load(args.decoupling_weights, map_location='cpu')['weights']) print(f'Loaded weights from {args.decoupling_weights} with message: {msg}') model.fc.bias.data = torch.zeros_like(model.fc.bias.data) fc_weights = model.fc.weight.data.clone() weight_norms = torch.norm(fc_weights, 2, 1) model.fc.weight.data = torch.stack([fc_weights[i] / torch.pow(weight_norms[i], -4) for i in range(N_CLASSES)], dim=0) elif args.decoupling_method == 'cRT': msg = model.load_state_dict(torch.load(args.decoupling_weights, map_location='cpu')['weights']) print(f'Loaded weights from {args.decoupling_weights} with message: {msg}') model.fc = torch.nn.Linear(model.fc.in_features, N_CLASSES) # re-initialize classifier head model = model.to(device) # Set loss and weighting method if args.rw_method == 'sklearn': weights = class_weight.compute_class_weight(class_weight='balanced', classes=np.unique(train_dataset.labels), y=np.array(train_dataset.labels)) weights = torch.Tensor(weights).to(device) elif args.rw_method == 'cb': weights = get_CB_weights(samples_per_cls=train_dataset.cls_num_list, beta=args.cb_beta) weights = torch.Tensor(weights).to(device) else: weights = None if weights is None: print('No class reweighting') else: print(f'Class weights with rw_method {args.rw_method}:') for i, c in enumerate(train_dataset.CLASSES): print(f'\t{c}: {weights[i]}') loss_fxn = get_loss(args, None if args.drw else weights, train_dataset) # Set optimizer if args.decoupling_method != '': optimizer = torch.optim.Adam(model.fc.parameters(), lr=args.lr) else: optimizer = torch.optim.Adam(model.parameters(), lr=args.lr) # Train with early stopping if args.decoupling_method != 'tau_norm': epoch = 1 early_stopping_dict = {'best_acc': 0., 'epochs_no_improve': 0} best_model_wts = None while epoch <= args.max_epochs and early_stopping_dict['epochs_no_improve'] <= args.patience: if args.bal_mixup: history = bal_mixup_train(model=model, device=device, loss_fxn=loss_fxn, optimizer=optimizer, data_loader=train_loader, history=history, epoch=epoch, model_dir=model_dir, classes=train_dataset.CLASSES, mixup_alpha=args.mixup_alpha) else: history = train(model=model, device=device, loss_fxn=loss_fxn, optimizer=optimizer, data_loader=train_loader, history=history, epoch=epoch, model_dir=model_dir, classes=train_dataset.CLASSES, mixup=args.mixup, mixup_alpha=args.mixup_alpha) history, early_stopping_dict, best_model_wts = validate(model=model, device=device, loss_fxn=loss_fxn, optimizer=optimizer, data_loader=val_loader, history=history, epoch=epoch, model_dir=model_dir, early_stopping_dict=early_stopping_dict, best_model_wts=best_model_wts, classes=val_dataset.CLASSES) if args.drw and epoch == 10: for g in optimizer.param_groups: g['lr'] *= 0.1 # anneal LR loss_fxn = get_loss(args, weights, train_dataset) # get class-weighted loss early_stopping_dict['epochs_no_improve'] = 0 # reset patience epoch += 1 else: best_model_wts = model.state_dict() # Evaluate on balanced test set evaluate(model=model, device=device, loss_fxn=loss_fxn, dataset=bal_test_dataset, split='balanced-test', batch_size=args.batch_size, history=history, model_dir=model_dir, weights=best_model_wts) # Evaluate on imbalanced test set evaluate(model=model, device=device, loss_fxn=loss_fxn, dataset=test_dataset, split='test', batch_size=args.batch_size, history=history, model_dir=model_dir, weights=best_model_wts) if __name__ == '__main__': # Command-line arguments parser = argparse.ArgumentParser() parser.add_argument('--data_dir', default='/ssd1/greg/NIH_CXR/images', type=str) parser.add_argument('--label_dir', default='labels/', type=str) parser.add_argument('--out_dir', default='results/', type=str, help="path to directory where results and model weights will be saved") parser.add_argument('--dataset', required=True, type=str, choices=['nih-lt', 'mimic-cxr-lt']) parser.add_argument('--loss', default='ce', type=str, choices=['ce', 'focal', 'ldam']) parser.add_argument('--drw', action='store_true', default=False) parser.add_argument('--rw_method', default='', choices=['', 'sklearn', 'cb']) parser.add_argument('--cb_beta', default=0.9999, type=float) parser.add_argument('--fl_gamma', default=2., type=float) parser.add_argument('--bal_mixup', action='store_true', default=False) parser.add_argument('--mixup', action='store_true', default=False) parser.add_argument('--mixup_alpha', default=0.2, type=float) parser.add_argument('--decoupling_method', default='', choices=['', 'cRT', 'tau_norm'], type=str) parser.add_argument('--decoupling_weights', type=str) parser.add_argument('--model_name', default='resnet50', type=str, help="CNN backbone to use") parser.add_argument('--max_epochs', default=60, type=int, help="maximum number of epochs to train") parser.add_argument('--batch_size', default=256, type=int, help="batch size for training, validation, and testing (will be lowered if TTA used)") parser.add_argument('--lr', default=1e-4, type=float) parser.add_argument('--patience', default=15, type=int, help="early stopping 'patience' during training") parser.add_argument('--rand_init', action='store_true', default=False) parser.add_argument('--n_TTA', default=0, type=int, help="number of augmented copies to use during test-time augmentation (TTA), default 0") parser.add_argument('--seed', default=0, type=int, help="set random seed") args = parser.parse_args() print(args) main(args)
10,072
52.86631
311
py
LongTailCXR
LongTailCXR-main/src/losses.py
import numpy as np import torch import torch.nn.functional as F def get_loss(args, weights, train_dataset): if args.loss == 'ce': loss_fxn = torch.nn.CrossEntropyLoss(weight=weights, reduction='mean') elif args.loss == 'focal': loss_fxn = torch.hub.load('adeelh/pytorch-multi-class-focal-loss', model='FocalLoss', alpha=weights, gamma=args.fl_gamma, reduction='mean') elif args.loss == 'ldam': loss_fxn = LDAMLoss(cls_num_list=train_dataset.cls_num_list, weight=weights) return loss_fxn def get_CB_weights(samples_per_cls, beta): effective_num = 1.0 - np.power(beta, samples_per_cls) weights = (1.0 - beta) / np.array(effective_num) weights = weights / np.sum(weights) * len(samples_per_cls) return weights ## CREDIT TO https://github.com/kaidic/LDAM-DRW ## class LDAMLoss(torch.nn.Module): def __init__(self, cls_num_list, max_m=0.5, weight=None, s=30): super(LDAMLoss, self).__init__() m_list = 1.0 / np.sqrt(np.sqrt(cls_num_list)) m_list = m_list * (max_m / np.max(m_list)) m_list = torch.cuda.FloatTensor(m_list) self.m_list = m_list assert s > 0 self.s = s self.weight = weight print(self.weight) def forward(self, x, target): index = torch.zeros_like(x, dtype=torch.uint8) index.scatter_(1, target.data.view(-1, 1), 1) index_float = index.type(torch.cuda.FloatTensor) batch_m = torch.matmul(self.m_list[None, :], index_float.transpose(0,1)) batch_m = batch_m.view((-1, 1)) x_m = x - batch_m output = torch.where(index, x_m, x) return F.cross_entropy(self.s*output, target, weight=self.weight)
1,722
34.163265
147
py
LongTailCXR
LongTailCXR-main/src/utils.py
import os import random import matplotlib.pyplot as plt import numpy as np import pandas as pd import torch import tqdm from copy import deepcopy from mlxtend.plotting import plot_confusion_matrix from sklearn.metrics import roc_auc_score, balanced_accuracy_score, classification_report, confusion_matrix, matthews_corrcoef from sklearn.preprocessing import LabelBinarizer def set_seed(seed): """Set all random seeds and settings for reproducibility (deterministic behavior).""" torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = True np.random.seed(seed) random.seed(seed) os.environ['PYTHONHASHSEED'] = str(seed) def worker_init_fn(worker_id): worker_seed = torch.initial_seed() % 2**32 np.random.seed(worker_seed) random.seed(worker_seed) def val_worker_init_fn(worker_id): np.random.seed(worker_id) random.seed(worker_id) def mixup_data(x, y, alpha=1.0, use_cuda=True): if alpha > 0: lam = np.random.beta(alpha, alpha) else: lam = 1 batch_size = x.size()[0] if use_cuda: index = torch.randperm(batch_size).cuda() else: index = torch.randperm(batch_size) mixed_x = lam * x + (1 - lam) * x[index, :] y_a, y_b = y, y[index] return mixed_x, y_a, y_b, lam def mixup_criterion(criterion, pred, y_a, y_b, lam): return lam * criterion(pred, y_a) + (1 - lam) * criterion(pred, y_b) def bal_mixup_train(model, device, loss_fxn, optimizer, data_loader, history, epoch, model_dir, classes, mixup_alpha): """Train PyTorch model for one epoch on NIH ChestXRay14 dataset. Parameters ---------- model : PyTorch model device : PyTorch device loss_fxn : PyTorch loss function optimizer : PyTorch optimizer data_loader : PyTorch data loader history : pandas DataFrame Data frame containing history of training metrics epoch : int Current epoch number (1-K) model_dir : str Path to output directory where metrics, model weights, etc. will be stored classes : list[str] Ordered list of names of output classes Returns ------- history : pandas DataFrame Updated history data frame with metrics from completed training epoch """ pbar = tqdm.tqdm(enumerate(data_loader), total=len(data_loader), desc=f'Epoch {epoch}') running_loss = 0. y_true, y_hat = [], [] for i, batch in pbar: x = batch[0][0].to(device) y = batch[0][1].to(device) bal_x = batch[0][0].to(device) bal_y = batch[0][1].to(device) lam = np.random.beta(mixup_alpha, mixup_alpha) mixed_x = (1 - lam) * x + lam * bal_x out = model(mixed_x) loss = mixup_criterion(loss_fxn, out, y, bal_y, lam) for param in model.parameters(): param.grad = None loss.backward() optimizer.step() running_loss += loss.item() y_hat.append(out.softmax(dim=1).detach().cpu().numpy()) y_true.append(y.detach().cpu().numpy()) pbar.set_postfix({'loss': running_loss / (i + 1)}) # Collect true and predicted labels into flat numpy arrays y_true, y_hat = np.concatenate(y_true), np.concatenate(y_hat) # Compute metrics auc = roc_auc_score(y_true, y_hat, average='macro', multi_class='ovr') b_acc = balanced_accuracy_score(y_true, y_hat.argmax(axis=1)) mcc = matthews_corrcoef(y_true, y_hat.argmax(axis=1)) print('Balanced Accuracy:', round(b_acc, 3), '|', 'MCC:', round(mcc, 3), '|', 'AUC:', round(auc, 3)) current_metrics = pd.DataFrame([[epoch, 'train', running_loss / (i + 1), b_acc, mcc, auc]], columns=history.columns) current_metrics.to_csv(os.path.join(model_dir, 'history.csv'), mode='a', header=False, index=False) return history.append(current_metrics) def train(model, device, loss_fxn, optimizer, data_loader, history, epoch, model_dir, classes, mixup, mixup_alpha): """Train PyTorch model for one epoch on NIH ChestXRay14 dataset. Parameters ---------- model : PyTorch model device : PyTorch device loss_fxn : PyTorch loss function optimizer : PyTorch optimizer data_loader : PyTorch data loader history : pandas DataFrame Data frame containing history of training metrics epoch : int Current epoch number (1-K) model_dir : str Path to output directory where metrics, model weights, etc. will be stored classes : list[str] Ordered list of names of output classes Returns ------- history : pandas DataFrame Updated history data frame with metrics from completed training epoch """ pbar = tqdm.tqdm(enumerate(data_loader), total=len(data_loader), desc=f'Epoch {epoch}') running_loss = 0. y_true, y_hat = [], [] for i, (x, y) in pbar: x = x.to(device) y = y.to(device) if mixup: x, y_a, y_b, lam = mixup_data(x, y, mixup_alpha, True) out = model(x) if mixup: loss = mixup_criterion(loss_fxn, out, y_a, y_b, lam) else: loss = loss_fxn(out, y) for param in model.parameters(): param.grad = None loss.backward() optimizer.step() running_loss += loss.item() y_hat.append(out.softmax(dim=1).detach().cpu().numpy()) y_true.append(y.detach().cpu().numpy()) pbar.set_postfix({'loss': running_loss / (i + 1)}) # Collect true and predicted labels into flat numpy arrays y_true, y_hat = np.concatenate(y_true), np.concatenate(y_hat) # Compute metrics auc = roc_auc_score(y_true, y_hat, average='macro', multi_class='ovr') b_acc = balanced_accuracy_score(y_true, y_hat.argmax(axis=1)) mcc = matthews_corrcoef(y_true, y_hat.argmax(axis=1)) print('Balanced Accuracy:', round(b_acc, 3), '|', 'MCC:', round(mcc, 3), '|', 'AUC:', round(auc, 3)) current_metrics = pd.DataFrame([[epoch, 'train', running_loss / (i + 1), b_acc, mcc, auc]], columns=history.columns) current_metrics.to_csv(os.path.join(model_dir, 'history.csv'), mode='a', header=False, index=False) return history.append(current_metrics) def validate(model, device, loss_fxn, optimizer, data_loader, history, epoch, model_dir, early_stopping_dict, best_model_wts, classes): """Evaluate PyTorch model on validation set of NIH ChestXRay14 dataset. Parameters ---------- model : PyTorch model device : PyTorch device loss_fxn : PyTorch loss function ls : int Ratio of label smoothing to apply during loss computation optimizer : PyTorch optimizer data_loader : PyTorch data loader history : pandas DataFrame Data frame containing history of training metrics epoch : int Current epoch number (1-K) model_dir : str Path to output directory where metrics, model weights, etc. will be stored early_stopping_dict : dict Dictionary of form {'epochs_no_improve': <int>, 'best_loss': <float>} for early stopping best_model_wts : PyTorch state_dict Model weights from best epoch classes : list[str] Ordered list of names of output classes fusion : bool Whether or not fusion is being performed (image + metadata inputs) meta_only : bool Whether or not to train on *only* metadata as input Returns ------- history : pandas DataFrame Updated history data frame with metrics from completed training epoch early_stopping_dict : dict Updated early stopping metrics best_model_wts : PyTorch state_dict (Potentially) updated model weights (if best validation loss achieved) """ model.eval() pbar = tqdm.tqdm(enumerate(data_loader), total=len(data_loader), desc=f'[VAL] Epoch {epoch}') running_loss = 0. y_true, y_hat = [], [] with torch.no_grad(): for i, (x, y) in pbar: x = x.to(device) y = y.to(device) out = model(x) loss = loss_fxn(out, y) running_loss += loss.item() y_hat.append(out.softmax(dim=1).detach().cpu().numpy()) y_true.append(y.detach().cpu().numpy()) pbar.set_postfix({'loss': running_loss / (i + 1)}) # Collect true and predicted labels into flat numpy arrays y_true, y_hat = np.concatenate(y_true), np.concatenate(y_hat) # Compute metrics auc = roc_auc_score(y_true, y_hat, average='macro', multi_class='ovr') b_acc = balanced_accuracy_score(y_true, y_hat.argmax(axis=1)) mcc = matthews_corrcoef(y_true, y_hat.argmax(axis=1)) print('[VAL] Balanced Accuracy:', round(b_acc, 3), '|', 'MCC:', round(mcc, 3), '|', 'AUC:', round(auc, 3)) current_metrics = pd.DataFrame([[epoch, 'val', running_loss / (i + 1), b_acc, mcc, auc]], columns=history.columns) current_metrics.to_csv(os.path.join(model_dir, 'history.csv'), mode='a', header=False, index=False) # Early stopping: save model weights only when val (balanced) accuracy has improved if b_acc > early_stopping_dict['best_acc']: print(f'--- EARLY STOPPING: Accuracy has improved from {round(early_stopping_dict["best_acc"], 3)} to {round(b_acc, 3)}! Saving weights. ---') early_stopping_dict['epochs_no_improve'] = 0 early_stopping_dict['best_acc'] = b_acc best_model_wts = deepcopy(model.state_dict()) torch.save({'weights': best_model_wts, 'optimizer': optimizer.state_dict()}, os.path.join(model_dir, f'chkpt_epoch-{epoch}.pt')) else: print(f'--- EARLY STOPPING: Accuracy has not improved from {round(early_stopping_dict["best_acc"], 3)} ---') early_stopping_dict['epochs_no_improve'] += 1 return history.append(current_metrics), early_stopping_dict, best_model_wts def evaluate(model, device, loss_fxn, dataset, split, batch_size, history, model_dir, weights): """Evaluate PyTorch model on test set of NIH ChestXRay14 dataset. Saves training history csv, summary text file, training curves, etc. Parameters ---------- model : PyTorch model device : PyTorch device loss_fxn : PyTorch loss function ls : int Ratio of label smoothing to apply during loss computation batch_size : int history : pandas DataFrame Data frame containing history of training metrics model_dir : str Path to output directory where metrics, model weights, etc. will be stored weights : PyTorch state_dict Model weights from best epoch n_TTA : int Number of augmented copies to use for test-time augmentation (0-K) fusion : bool Whether or not fusion is being performed (image + metadata inputs) meta_only : bool Whether or not to train on *only* metadata as input """ model.load_state_dict(weights) # load best weights model.eval() ## INFERENCE data_loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=8 if split == 'test' else 2, pin_memory=True, worker_init_fn=val_worker_init_fn) pbar = tqdm.tqdm(enumerate(data_loader), total=len(data_loader), desc=f'[{split.upper()}] EVALUATION') running_loss = 0. y_true, y_hat = [], [] with torch.no_grad(): for i, (x, y) in pbar: x = x.to(device) y = y.to(device) out = model(x) loss = loss_fxn(out, y) running_loss += loss.item() y_hat.append(out.softmax(dim=1).detach().cpu().numpy()) y_true.append(y.detach().cpu().numpy()) pbar.set_postfix({'loss': running_loss / (i + 1)}) # Collect true and predicted labels into flat numpy arrays y_true, y_hat = np.concatenate(y_true), np.concatenate(y_hat) # Compute metrics auc = roc_auc_score(y_true, y_hat, average='macro', multi_class='ovr') b_acc = balanced_accuracy_score(y_true, y_hat.argmax(axis=1)) conf_mat = confusion_matrix(y_true, y_hat.argmax(axis=1)) accuracies = conf_mat.diagonal() / conf_mat.sum(axis=1) mcc = matthews_corrcoef(y_true, y_hat.argmax(axis=1)) cls_report = classification_report(y_true, y_hat.argmax(axis=1), target_names=dataset.CLASSES, digits=3) print(f'[{split.upper()}] Balanced Accuracy: {round(b_acc, 3)} | MCC: {round(mcc, 3)} | AUC: {round(auc, 3)}') # Collect and save true and predicted disease labels for test set pred_df = pd.DataFrame(y_hat, columns=dataset.CLASSES) true_df = pd.DataFrame(LabelBinarizer().fit(range(len(dataset.CLASSES))).transform(y_true), columns=dataset.CLASSES) pred_df.to_csv(os.path.join(model_dir, f'{split}_pred.csv'), index=False) true_df.to_csv(os.path.join(model_dir, f'{split}_true.csv'), index=False) # Plot confusion matrix fig, ax = plot_confusion_matrix(conf_mat, figsize=(24, 24), colorbar=True, show_absolute=True, show_normed=True, class_names=dataset.CLASSES) fig.savefig(os.path.join(model_dir, f'{split}_cm.png'), dpi=300, bbox_inches='tight') # Plot loss curves fig, ax = plt.subplots(1, 1, figsize=(6, 6)) ax.plot(history.loc[history['phase'] == 'train', 'epoch'], history.loc[history['phase'] == 'train', 'loss'], label='train') ax.plot(history.loc[history['phase'] == 'val', 'epoch'], history.loc[history['phase'] == 'val', 'loss'], label='val') ax.set_xlabel('Epoch') ax.set_ylabel('Loss') ax.legend() fig.savefig(os.path.join(model_dir, 'loss.png'), dpi=300, bbox_inches='tight') # Plot accuracy curves fig, ax = plt.subplots(1, 1, figsize=(6, 6)) ax.plot(history.loc[history['phase'] == 'train', 'epoch'], history.loc[history['phase'] == 'train', 'balanced_acc'], label='train') ax.plot(history.loc[history['phase'] == 'val', 'epoch'], history.loc[history['phase'] == 'val', 'balanced_acc'], label='val') ax.set_xlabel('Epoch') ax.set_ylabel('Balanced Accuracy') ax.legend() fig.savefig(os.path.join(model_dir, 'balanced_acc.png'), dpi=300, bbox_inches='tight') # Plot AUROC learning curves fig, ax = plt.subplots(1, 1, figsize=(6, 6)) ax.plot(history.loc[history['phase'] == 'train', 'epoch'], history.loc[history['phase'] == 'train', 'auroc'], label='train') ax.plot(history.loc[history['phase'] == 'val', 'epoch'], history.loc[history['phase'] == 'val', 'auroc'], label='val') ax.set_xlabel('Epoch') ax.set_ylabel('AUROC') ax.legend() fig.savefig(os.path.join(model_dir, 'auc.png'), dpi=300, bbox_inches='tight') # Create summary text file describing final performance summary = f'Balanced Accuracy: {round(b_acc, 3)}\n' summary += f'Matthews Correlation Coefficient: {round(mcc, 3)}\n' summary += f'Mean AUC: {round(auc, 3)}\n\n' summary += 'Class:| Accuracy\n' for i, c in enumerate(dataset.CLASSES): summary += f'{c}:| {round(accuracies[i], 3)}\n' summary += '\n' summary += cls_report f = open(os.path.join(model_dir, f'{split}_summary.txt'), 'w') f.write(summary) f.close()
15,395
38.88601
186
py