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
fuzzyJoiner
fuzzyJoiner-master/old/TripletLossFacenetLSTM-8.31.18.py
from random import shuffle import numpy as np import pandas import tensorflow as tf import random as random import json from keras import backend as K from keras.preprocessing.text import Tokenizer from keras.preprocessing.sequence import pad_sequences from keras.layers import Dense, Input, Flatten, Dropout, Lambda, GRU, Activation from keras.layers.wrappers import Bidirectional from keras.layers import Conv1D, MaxPooling1D, Embedding from keras.models import Model, model_from_json, Sequential from embeddings import KazumaCharEmbedding from annoy import AnnoyIndex from keras.callbacks import ModelCheckpoint, EarlyStopping from names_cleanser import NameDataCleanser, CompanyDataCleanser import sys import statistics from scipy.stats.mstats import gmean import argparse #must fix MAX_NB_WORDS = 140000 EMBEDDING_DIM = 100 MAX_SEQUENCE_LENGTH = 10 MARGIN=10 ALPHA=45 DEBUG = False DEBUG_DATA_LENGTH = 100 DEBUG_ANN = False USE_ANGULAR_LOSS=False LOSS_FUNCTION=None TRAIN_NEIGHBOR_LEN=20 TEST_NEIGHBOR_LEN=20 EMBEDDING_TYPE = 'Kazuma' NUM_LAYERS = 3 USE_L2_NORM = False filepath="weights.best.hdf5" output_file_name_for_hpo = "val_dict_list.json" def f1score(positive, negative): #labels[predictions.ravel() < 0.5].sum() fsocre = 0.0 true_positive = 0.0 false_positive = 0 false_negitive = 0 for i in range(len(positive)): if positive[i] <= negative[i]: true_positive += 1 else: false_negitive += 1 false_positive += 1 print('tp' + str(true_positive)) print('fp' + str(false_positive)) print('fn' + str(false_negitive)) fscore = (2 * true_positive) / ((2 * true_positive) + false_negitive + false_positive) return fscore def get_embedding_layer(tokenizer): word_index = tokenizer.word_index num_words = len(word_index) + 1 embedding_matrix = np.zeros((num_words, EMBEDDING_DIM)) print('about to get kz') kz = KazumaCharEmbedding() print('got kz') for word, i in word_index.items(): if i >= MAX_NB_WORDS: continue embedding_vector = kz.emb(word) if embedding_vector is not None: if sum(embedding_vector) == 0: print(str("failed to find embedding for:" + word).encode('utf-8')) # words not found in embedding index will be all-zeros. embedding_matrix[i] = embedding_vector print("Number of words:" + str(num_words)) embedding_layer = Embedding(num_words, EMBEDDING_DIM, weights=[embedding_matrix], input_length=MAX_SEQUENCE_LENGTH, trainable=False) return embedding_layer def get_sequences(texts, tokenizer): sequences = {} sequences['anchor'] = tokenizer.texts_to_sequences(texts['anchor']) sequences['anchor'] = pad_sequences(sequences['anchor'], maxlen=MAX_SEQUENCE_LENGTH) sequences['negative'] = tokenizer.texts_to_sequences(texts['negative']) sequences['negative'] = pad_sequences(sequences['negative'], maxlen=MAX_SEQUENCE_LENGTH) sequences['positive'] = tokenizer.texts_to_sequences(texts['positive']) sequences['positive'] = pad_sequences(sequences['positive'], maxlen=MAX_SEQUENCE_LENGTH) return sequences def read_entities(filepath): entities = [] with open(filepath, 'r', encoding='utf8') as fl: for line in fl: entities.append(line) return entities def read_file(file_path): texts = {'anchor':[], 'negative':[], 'positive':[]} fl = open(file_path, 'r', encoding='utf8') i = 0 for line in fl: line_array = line.split("|") texts['anchor'].append(line_array[0]) texts['positive'].append(line_array[1]) texts['negative'].append(line_array[2]) i += 1 if i > DEBUG_DATA_LENGTH and DEBUG: break return texts def split(entities, test_split = 0.2): if DEBUG: ents = entities[0:DEBUG_DATA_LENGTH] else: random.shuffle(entities) ents = entities num_validation_samples = int(test_split * len(ents)) return ents[:-num_validation_samples], ents[-num_validation_samples:] """ define a single objective function based on angular loss instead of triplet loss """ def angular_loss(y_true, y_pred): alpha = K.constant(ALPHA) a_p = y_pred[:,0,0] n_c = y_pred[:,1,0] return K.mean(K.maximum(K.constant(0), K.square(a_p) - K.constant(4) * K.square(tf.tan(alpha)) * K.square(n_c))) """ Facenet triplet loss function: https://arxiv.org/pdf/1503.03832.pdf """ def schroff_triplet_loss(y_true, y_pred): margin = K.constant(0.2) return K.mean(K.maximum(K.constant(0), K.square(y_pred[:,0,0]) - K.square(y_pred[:,1,0]) + margin)) def triplet_loss(y_true, y_pred): margin = K.constant(MARGIN) return K.mean(K.square(y_pred[:,0,0]) + K.square(K.maximum(margin - y_pred[:,1,0], K.constant(0)))) def triplet_tanh_loss(y_true, y_pred): return K.mean(K.tanh(y_pred[:,0,0]) + (K.constant(1) - K.tanh(y_pred[:,1,0]))) def triplet_tanh_pn_loss(y_true, y_pred): return K.mean(K.tanh(y_pred[:,0,0]) + ((K.constant(1) - K.tanh(y_pred[:,1,0])) + (K.constant(1) - K.tanh(y_pred[:,2,0]))) / K.constant(2)); # the following triplet loss function is from: Deep Metric Learning with Improved Triplet Loss for # Face clustering in Videos def improved_loss(y_true, y_pred): margin = K.constant(1) lambda_p = K.constant(0.02) threshold = K.constant(0.1) a_p_distance = y_pred[:,0,0] a_n_distance = y_pred[:,1,0] p_n_distance = y_pred[:,2,0] phi = a_p_distance - ((a_n_distance + p_n_distance) / K.constant(2)) + margin psi = a_p_distance - threshold return K.maximum(K.constant(0), phi) + lambda_p * K.maximum(K.constant(0), psi) def accuracy(y_true, y_pred): return K.mean(y_pred[:,0,0] < y_pred[:,1,0]) def l2Norm(x): return K.l2_normalize(x, axis=-1) def euclidean_distance(vects): x, y = vects return K.sqrt(K.maximum(K.sum(K.square(x - y), axis=1, keepdims=True), K.epsilon())) def n_c_angular_distance(vects): x_a, x_p, x_n = vects return K.sqrt(K.maximum(K.sum(K.square(x_n - ((x_a + x_p) / K.constant(2))), axis=1, keepdims=True), K.epsilon())) def a_p_angular_distance(vects): x_a, x_p, x_n = vects return K.sqrt(K.maximum(K.sum(K.square(x_a - x_p), axis=1, keepdims=True), K.epsilon())) def build_unique_entities(entity2same): unique_text = [] entity2index = {} for key in entity2same: entity2index[key] = len(unique_text) unique_text.append(key) vals = entity2same[key] for v in vals: entity2index[v] = len(unique_text) unique_text.append(v) return unique_text, entity2index def generate_semi_hard_triplets_from_ANN(model, sequences, entity2unique, entity2same, unique_text, test): predictions = model.predict(sequences) t = AnnoyIndex(len(predictions[0]), metric='euclidean') # Length of item vector that will be indexed t.set_seed(123) for i in range(len(predictions)): # print(predictions[i]) v = predictions[i] t.add_item(i, v) t.build(100) # 100 trees triplets = {} triplets['anchor'] = [] triplets['positive'] = [] triplets['negative'] = [] if test: NNlen = TEST_NEIGHBOR_LEN else: NNlen = TRAIN_NEIGHBOR_LEN for key in entity2same: index = entity2unique[key] expected_text = set(entity2same[key]) expected_ids = [entity2unique[i] for i in expected_text] for positive in expected_text: k = entity2unique[positive] nearest = t.get_nns_by_vector(predictions[k], NNlen) dist_k = t.get_distance(index, k) semi_hards = [] for n in nearest: if n == index or n in expected_ids or n == k: continue n_dist = t.get_distance(index, n) if n_dist > dist_k: semi_hards.append(unique_text[n]) # shuffle(semi_hards) # semi_hards = semi_hards[0:20] for i in semi_hards: triplets['anchor'].append(key) triplets['positive'].append(unique_text[k]) triplets['negative'].append(i) return triplets def generate_triplets_from_ANN(model, sequences, entity2unique, entity2same, unique_text, test): predictions = model.predict(sequences) t = AnnoyIndex(len(predictions[0]), metric='euclidean') # Length of item vector that will be indexed t.set_seed(123) for i in range(len(predictions)): # print(predictions[i]) v = predictions[i] t.add_item(i, v) t.build(100) # 100 trees match = 0 no_match = 0 ann_accuracy = 0 total = 0 precise = 0 triplets = {} closest_positive_counts = [] pos_distances = [] neg_distances = [] all_pos_distances = [] all_neg_distances = [] triplets['anchor'] = [] triplets['positive'] = [] triplets['negative'] = [] if test: NNlen = TEST_NEIGHBOR_LEN else: NNlen = TRAIN_NEIGHBOR_LEN for key in entity2same: index = entity2unique[key] nearest = t.get_nns_by_vector(predictions[index], NNlen) nearest_text = set([unique_text[i] for i in nearest]) expected_text = set(entity2same[key]) # annoy has this annoying habit of returning the queried item back as a nearest neighbor. Remove it. if key in nearest_text: nearest_text.remove(key) # print("query={} names = {} true_match = {}".format(unique_text[index], nearest_text, expected_text)) overlap = expected_text.intersection(nearest_text) # collect up some statistics on how well we did on the match m = len(overlap) match += m # since we asked for only x nearest neighbors, and we get at most x-1 neighbors that are not the same as key (!) # make sure we adjust our estimate of no match appropriately no_match += min(len(expected_text), NNlen - 1) - m # sample only the negatives that are true negatives # that is, they are not in the expected set - sampling only 'semi-hard negatives is not defined here' # positives = expected_text - nearest_text positives = overlap negatives = nearest_text - expected_text # print(key + str(expected_text) + str(nearest_text)) for i in negatives: for j in positives: dist_pos = t.get_distance(index, entity2unique[j]) pos_distances.append(dist_pos) dist_neg = t.get_distance(index, entity2unique[i]) neg_distances.append(dist_neg) if dist_pos < dist_neg: ann_accuracy += 1 total += 1 # print(key + "|" + j + "|" + i) # print(dist_pos) # print(dist_neg) min_neg_distance = 1000000 for i in negatives: dist_neg = t.get_distance(index, entity2unique[i]) all_neg_distances.append(dist_neg) if dist_neg < min_neg_distance: min_neg_distance = dist_neg for j in expected_text: dist_pos = t.get_distance(index, entity2unique[j]) all_pos_distances.append(dist_pos) closest_pos_count = 0 for p in overlap: dist_pos = t.get_distance(index, entity2unique[p]) if dist_pos < min_neg_distance: closest_pos_count+=1 if closest_pos_count > 0: precise+=1 closest_positive_counts.append(closest_pos_count / min(len(expected_text), NNlen - 1)) for i in negatives: for j in expected_text: triplets['anchor'].append(key) triplets['positive'].append(j) triplets['negative'].append(i) print("mean closest positive count:" + str(statistics.mean(closest_positive_counts))) print("mean positive distance:" + str(statistics.mean(pos_distances))) print("stdev positive distance:" + str(statistics.stdev(pos_distances))) print("max positive distance:" + str(max(pos_distances))) print("mean neg distance:" + str(statistics.mean(neg_distances))) print("stdev neg distance:" + str(statistics.stdev(neg_distances))) print("max neg distance:" + str(max(neg_distances))) print("mean all positive distance:" + str(statistics.mean(all_pos_distances))) print("stdev all positive distance:" + str(statistics.stdev(all_pos_distances))) print("max all positive distance:" + str(max(all_pos_distances))) print("mean all neg distance:" + str(statistics.mean(all_neg_distances))) print("stdev all neg distance:" + str(statistics.stdev(all_neg_distances))) print("max all neg distance:" + str(max(all_neg_distances))) print("Accuracy in the ANN for triplets that obey the distance func:" + str(ann_accuracy / total)) print("Precision at 1: " + str(precise / len(entity2same))) obj = {} obj['accuracy'] = ann_accuracy / total obj['steps'] = 1 with open(output_file_name_for_hpo, 'w', encoding='utf8') as out: json.dump(obj, out) if test: return match/(match + no_match) else: return triplets, match/(match + no_match) def generate_names(entities, people, limit_pairs=False): if people: num_names = 4 generator = NameDataCleanser(0, num_names, limit_pairs=limit_pairs) else: generator = CompanyDataCleanser(limit_pairs) num_names = 2 entity2same = {} for entity in entities: ret = generator.cleanse_data(entity) if ret and len(ret) >= num_names: entity2same[ret[0]] = ret[1:] return entity2same def embedded_representation_model(embedding_layer): seq = Sequential() seq.add(embedding_layer) seq.add(Flatten()) return seq def build_model(embedder): main_input = Input(shape=(MAX_SEQUENCE_LENGTH,)) net = embedder(main_input) for i in range(0, NUM_LAYERS): net = GRU(128, return_sequences=True, activation='relu', name='embed' + str(i))(net) net = GRU(128, activation='relu', name='embed' + str(i+1))(net) if USE_L2_NORM: net = Lambda(l2Norm, output_shape=[128])(net) base_model = Model(embedder.input, net, name='triplet_model') base_model.summary() input_shape=(MAX_SEQUENCE_LENGTH,) input_anchor = Input(shape=input_shape, name='input_anchor') input_positive = Input(shape=input_shape, name='input_pos') input_negative = Input(shape=input_shape, name='input_neg') net_anchor = base_model(input_anchor) net_positive = base_model(input_positive) net_negative = base_model(input_negative) positive_dist = Lambda(euclidean_distance, name='pos_dist', output_shape=(1,))([net_anchor, net_positive]) negative_dist = Lambda(euclidean_distance, name='neg_dist', output_shape=(1,))([net_anchor, net_negative]) if USE_ANGULAR_LOSS: n_c = Lambda(n_c_angular_distance, name='nc_angular_dist')([net_anchor, net_positive, net_negative]) a_p = Lambda(a_p_angular_distance, name='ap_angular_dist')([net_anchor, net_positive, net_negative]) stacked_dists = Lambda( lambda vects: K.stack(vects, axis=1), name='stacked_dists', output_shape=(3, 1) )([a_p, n_c]) model = Model([input_anchor, input_positive, input_negative], stacked_dists, name='triple_siamese') model.compile(optimizer="rmsprop", loss=angular_loss, metrics=[accuracy]) else: exemplar_negative_dist = Lambda(euclidean_distance, name='exemplar_neg_dist', output_shape=(1,))([net_positive, net_negative]) stacked_dists = Lambda( # lambda vects: C.splice(*vects, axis=C.Axis.new_leading_axis()).eval(vects), lambda vects: K.stack(vects, axis=1), name='stacked_dists', output_shape=(3, 1) )([positive_dist, negative_dist, exemplar_negative_dist]) model = Model([input_anchor, input_positive, input_negative], stacked_dists, name='triple_siamese') model.compile(optimizer="rmsprop", loss=LOSS_FUNCTION, metrics=[accuracy]) test_positive_model = Model([input_anchor, input_positive, input_negative], positive_dist) test_negative_model = Model([input_anchor, input_positive, input_negative], negative_dist) inter_model = Model(input_anchor, net_anchor) print("output_shapes") model.summary() # print(positive_dist.output_shape) # print(negative_dist.output_shape) # print(exemplar_negative_dist) # print(neg_dist.output_shape) return model, test_positive_model, test_negative_model, inter_model parser = argparse.ArgumentParser(description='Run fuzzy join algorithm') parser.add_argument('--debug_sample_size', type=int, help='sample size for debug run') parser.add_argument('--margin', type=int, help='margin') parser.add_argument('--loss_function', type=str, help='triplet loss function type: schroff-loss, improved-loss, angular-loss, tanh-loss, improved-tanh-loss') parser.add_argument('--use_l2_norm', type=str, help='whether to add a l2 norm') parser.add_argument('--num_layers', type=int, help='num_layers to use. Minimum is 2') parser.add_argument('--input', type=str, help='Input file') parser.add_argument('--entity_type', type=str, help='people or companies') parser.add_argument('--model', type=str, help='name for model file') args = parser.parse_args() filepath = args.model LOSS_FUNCTION = None if args.loss_function == 'schroff-loss': LOSS_FUNCTION=schroff_triplet_loss elif args.loss_function == 'improved-loss': LOSS_FUNCTION=improved_loss elif args.loss_function == 'our-loss': LOSS_FUNCTION=triplet_loss elif args.loss_function == 'tanh-loss': LOSS_FUNCTION=triplet_tanh_loss elif args.loss_function == 'improved-tanh-loss': LOSS_FUNCTION=triplet_tanh_pn_loss elif args.loss_function == 'angular-loss': USE_ANGULAR_LOSS = True LOSS_FUNCTION = angular_loss print('Loss function: ' + args.loss_function) if args.debug_sample_size: DEBUG=True DEBUG_DATA_LENGTH=args.debug_sample_size print('Debug data length:' + str(DEBUG_DATA_LENGTH)) print('Margin:' + str(MARGIN)) USE_L2_NORM = args.use_l2_norm.lower() in ("yes", "true", "t", "1") print('Use L2Norm: ' + str(USE_L2_NORM)) print('Use L2Norm: ' + str(args.use_l2_norm)) NUM_LAYERS = args.num_layers - 1 print('Num layers: ' + str(NUM_LAYERS)) people = 'people' in args.entity_type # read all entities and create positive parts of a triplet entities = read_entities(args.input) train, test = split(entities, test_split = .20) print("TRAIN") print(str(train).encode('utf-8')) print("TEST") print(str(test).encode('utf-8')) entity2same_train = generate_names(train, people) entity2same_test = generate_names(test, people, limit_pairs=True) print(str(entity2same_train).encode('utf-8')) print(str(entity2same_test).encode('utf-8')) # change the default behavior of the tokenizer to ignore all punctuation except , - and . which are important # clues for entity names tokenizer = Tokenizer(num_words=MAX_NB_WORDS, lower=True, filters='!"#$%&()*+/:;<=>?@[\]^_`{|}~', split=" ") # build a set of data structures useful for annoy, the set of unique entities (unique_text), # a mapping of entities in texts to an index in unique_text, a mapping of entities to other same entities, and the actual # vectorized representation of the text. These structures will be used iteratively as we build up the model # so we need to create them once for re-use unique_text, entity2unique = build_unique_entities(entity2same_train) unique_text_test, entity2unique_test = build_unique_entities(entity2same_test) print("train text len:" + str(len(unique_text))) print("test text len:" + str(len(unique_text_test))) tokenizer.fit_on_texts(unique_text + unique_text_test) sequences = tokenizer.texts_to_sequences(unique_text) sequences = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH) sequences_test = tokenizer.texts_to_sequences(unique_text_test) sequences_test = pad_sequences(sequences_test, maxlen=MAX_SEQUENCE_LENGTH) # build models embedder = get_embedding_layer(tokenizer) model, test_positive_model, test_negative_model, inter_model = build_model(embedder) embedder_model = embedded_representation_model(embedder) if DEBUG_ANN: triplets = generate_triplets_from_ANN(embedder_model, sequences, entity2unique, entity2same_train, unique_text, True) print(len(triplets['anchor'])) sys.exit() test_data, test_match_stats = generate_triplets_from_ANN(embedder_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, False) test_seq = get_sequences(test_data, tokenizer) print("Test stats:" + str(test_match_stats)) counter = 0 current_model = embedder_model prev_match_stats = 0 train_data, train_match_stats = generate_triplets_from_ANN(current_model, sequences, entity2unique, entity2same_train, unique_text, False) number_of_names = len(train_data['anchor']) # print(train_data['anchor']) print("number of names" + str(number_of_names)) Y_train = np.random.randint(2, size=(1,2,number_of_names)).T checkpoint = ModelCheckpoint(filepath, monitor='val_accuracy', verbose=1, save_best_only=True, mode='max') early_stop = EarlyStopping(monitor='val_accuracy', patience=1, mode='max') callbacks_list = [checkpoint, early_stop] train_seq = get_sequences(train_data, tokenizer) # check just for 5 epochs because this gets called many times model.fit([train_seq['anchor'], train_seq['positive'], train_seq['negative']], Y_train, epochs=100, batch_size=40, callbacks=callbacks_list, validation_split=0.2) current_model = inter_model # print some statistics on this epoch print("training data predictions") positives = test_positive_model.predict([train_seq['anchor'], train_seq['positive'], train_seq['negative']]) negatives = test_negative_model.predict([train_seq['anchor'], train_seq['positive'], train_seq['negative']]) print("f1score for train is: {}".format(f1score(positives, negatives))) print("test data predictions") positives = test_positive_model.predict([test_seq['anchor'], test_seq['positive'], test_seq['negative']]) negatives = test_negative_model.predict([test_seq['anchor'], test_seq['positive'], test_seq['negative']]) print("f1score for test is: {}".format(f1score(positives, negatives))) test_match_stats = generate_triplets_from_ANN(current_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, True) print("Test stats:" + str(test_match_stats))
22,999
35.624204
163
py
fuzzyJoiner
fuzzyJoiner-master/old/seq2seqTriplet.py
'''Sequence to sequence example in Keras (character-level). This script demonstrates how to implement a basic character-level sequence-to-sequence model. We apply it to translating short English sentences into short French sentences, character-by-character. Note that it is fairly unusual to do character-level machine translation, as word-level models are more common in this domain. # Summary of the algorithm - We start with input sequences from a domain (e.g. English sentences) and correspding target sequences from another domain (e.g. French sentences). - An encoder LSTM turns input sequences to 2 state vectors (we keep the last LSTM state and discard the outputs). - A decoder LSTM is trained to turn the target sequences into the same sequence but offset by one timestep in the future, a training process called "teacher forcing" in this context. Is uses as initial state the state vectors from the encoder. Effectively, the decoder learns to generate `targets[t+1...]` given `targets[...t]`, conditioned on the input sequence. - In inference mode, when we want to decode unknown input sequences, we: - Encode the input sequence into state vectors - Start with a target sequence of size 1 (just the start-of-sequence character) - Feed the state vectors and 1-char target sequence to the decoder to produce predictions for the next character - Sample the next character using these predictions (we simply use argmax). - Append the sampled character to the target sequence - Repeat until we generate the end-of-sequence character or we hit the character limit. # Data download English to French sentence pairs. http://www.manythings.org/anki/fra-eng.zip Lots of neat sentence pairs datasets can be found at: http://www.manythings.org/anki/ # References - Sequence to Sequence Learning with Neural Networks https://arxiv.org/abs/1409.3215 - Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation https://arxiv.org/abs/1406.1078 ''' from __future__ import print_function from keras.models import Model from keras.layers import Input, LSTM, Dense import numpy as np from keras.callbacks import ModelCheckpoint, EarlyStopping batch_size = 64 # Batch size for training. epochs = 100 # Number of epochs to train for. latent_dim = 256 # Latent dimensionality of the encoding space. num_samples = 100000 # Number of samples to train on. # Path to the data txt file on disk. data_path = 'tripletTranslate' # Vectorize the data. input_texts = [] target_texts = [] input_characters = set() target_characters = set() with open(data_path, 'r', encoding='utf-8') as f: lines = f.read().split('\n') for line in lines[: min(num_samples, len(lines) - 1)]: arr = line.split('|') input_text = arr[0] target_text = arr[1] arr = input_text.split() # data quality issues. In the canonical name, the first name had better not be just a letter or empty if arr[0].strip()=='' or len(arr[0].strip()) <= 1: continue # We use "tab" as the "start sequence" character # for the targets, and "\n" as "end sequence" character. target_text = '\t' + target_text + '\n' input_texts.append(input_text) target_texts.append(target_text) for char in input_text: if char not in input_characters: input_characters.add(char) for char in target_text: if char not in target_characters: target_characters.add(char) input_characters = sorted(list(input_characters)) target_characters = sorted(list(target_characters)) num_encoder_tokens = len(input_characters) num_decoder_tokens = len(target_characters) max_encoder_seq_length = max([len(txt) for txt in input_texts]) max_decoder_seq_length = max([len(txt) for txt in target_texts]) print('Number of samples:', len(input_texts)) print('Number of unique input tokens:', num_encoder_tokens) print('Number of unique output tokens:', num_decoder_tokens) print('Max sequence length for inputs:', max_encoder_seq_length) print('Max sequence length for outputs:', max_decoder_seq_length) input_token_index = dict( [(char, i) for i, char in enumerate(input_characters)]) target_token_index = dict( [(char, i) for i, char in enumerate(target_characters)]) encoder_input_data = np.zeros( (len(input_texts), max_encoder_seq_length, num_encoder_tokens), dtype='float32') decoder_input_data = np.zeros( (len(input_texts), max_decoder_seq_length, num_decoder_tokens), dtype='float32') decoder_target_data = np.zeros( (len(input_texts), max_decoder_seq_length, num_decoder_tokens), dtype='float32') for i, (input_text, target_text) in enumerate(zip(input_texts, target_texts)): for t, char in enumerate(input_text): encoder_input_data[i, t, input_token_index[char]] = 1. for t, char in enumerate(target_text): # decoder_target_data is ahead of decoder_input_data by one timestep decoder_input_data[i, t, target_token_index[char]] = 1. if t > 0: # decoder_target_data will be ahead by one timestep # and will not include the start character. decoder_target_data[i, t - 1, target_token_index[char]] = 1. # Define an input sequence and process it. encoder_inputs = Input(shape=(None, num_encoder_tokens)) encoder = LSTM(latent_dim, return_state=True) encoder_outputs, state_h, state_c = encoder(encoder_inputs) # We discard `encoder_outputs` and only keep the states. encoder_states = [state_h, state_c] # Set up the decoder, using `encoder_states` as initial state. decoder_inputs = Input(shape=(None, num_decoder_tokens)) # We set up our decoder to return full output sequences, # and to return internal states as well. We don't use the # return states in the training model, but we will use them in inference. decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True) decoder_outputs, _, _ = decoder_lstm(decoder_inputs, initial_state=encoder_states) decoder_dense = Dense(num_decoder_tokens, activation='softmax') decoder_outputs = decoder_dense(decoder_outputs) # Define the model that will turn # `encoder_input_data` & `decoder_input_data` into `decoder_target_data` model = Model([encoder_inputs, decoder_inputs], decoder_outputs) filepath = 's2s.h5' checkpoint = ModelCheckpoint(filepath, monitor='val_accuracy', verbose=1, save_best_only=True, mode='max') # check 5 epochs early_stop = EarlyStopping(monitor='val_accuracy', patience=1, mode='max') callbacks_list = [checkpoint, early_stop] # Run training model.compile(optimizer='rmsprop', loss='categorical_crossentropy') model.fit([encoder_input_data, decoder_input_data], decoder_target_data, batch_size=batch_size, epochs=epochs, callbacks=callbacks_list, validation_split=0.2) # Next: inference mode (sampling). # Here's the drill: # 1) encode input and retrieve initial decoder state # 2) run one step of decoder with this initial state # and a "start of sequence" token as target. # Output will be the next target token # 3) Repeat with the current target token and current states # Define sampling models encoder_model = Model(encoder_inputs, encoder_states) decoder_state_input_h = Input(shape=(latent_dim,)) decoder_state_input_c = Input(shape=(latent_dim,)) decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c] decoder_outputs, state_h, state_c = decoder_lstm( decoder_inputs, initial_state=decoder_states_inputs) decoder_states = [state_h, state_c] decoder_outputs = decoder_dense(decoder_outputs) decoder_model = Model( [decoder_inputs] + decoder_states_inputs, [decoder_outputs] + decoder_states) # Reverse-lookup token index to decode sequences back to # something readable. reverse_input_char_index = dict( (i, char) for char, i in input_token_index.items()) reverse_target_char_index = dict( (i, char) for char, i in target_token_index.items()) def decode_sequence(input_seq): # Encode the input as state vectors. states_value = encoder_model.predict(input_seq) # Generate empty target sequence of length 1. target_seq = np.zeros((1, 1, num_decoder_tokens)) # Populate the first character of target sequence with the start character. target_seq[0, 0, target_token_index['\t']] = 1. # Sampling loop for a batch of sequences # (to simplify, here we assume a batch of size 1). stop_condition = False decoded_sentence = '' while not stop_condition: output_tokens, h, c = decoder_model.predict( [target_seq] + states_value) # Sample a token sampled_token_index = np.argmax(output_tokens[0, -1, :]) sampled_char = reverse_target_char_index[sampled_token_index] decoded_sentence += sampled_char # Exit condition: either hit max length # or find stop character. if (sampled_char == '\n' or len(decoded_sentence) > max_decoder_seq_length): stop_condition = True # Update the target sequence (of length 1). target_seq = np.zeros((1, 1, num_decoder_tokens)) target_seq[0, 0, sampled_token_index] = 1. # Update states states_value = [h, c] return decoded_sentence for seq_index in range(100): # Take one sequence (part of the training set) # for trying out decoding. input_seq = encoder_input_data[seq_index: seq_index + 1] decoded_sentence = decode_sequence(input_seq) print('-') print('Input sentence:', input_texts[seq_index]) print('Decoded sentence:', decoded_sentence)
9,670
39.634454
106
py
fuzzyJoiner
fuzzyJoiner-master/old/TripletLossFacenetLSTM_hpo.py
import numpy as np import pandas import tensorflow as tf import random as random import json from keras import backend as K from keras.preprocessing.text import Tokenizer from keras.preprocessing.sequence import pad_sequences from keras.layers import Dense, Input, Flatten, Dropout, Lambda, GRU, Activation from keras.layers.wrappers import Bidirectional from keras.layers import Conv1D, MaxPooling1D, Embedding from keras.models import Model, model_from_json, Sequential from embeddings import KazumaCharEmbedding from annoy import AnnoyIndex from keras.callbacks import ModelCheckpoint, EarlyStopping from names_cleanser import NameDataCleanser, CompanyDataCleanser import sys import statistics import argparse #must fix MAX_NB_WORDS = 140000 EMBEDDING_DIM = 100 MAX_SEQUENCE_LENGTH = 10 MARGIN=10 ALPHA=45 DEBUG = False DEBUG_DATA_LENGTH = 100 DEBUG_ANN = False USE_ANGULAR_LOSS=False LOSS_FUNCTION=None TRAIN_NEIGHBOR_LEN=20 TEST_NEIGHBOR_LEN=20 EMBEDDING_TYPE = 'Kazuma' NUM_LAYERS = 3 USE_L2_NORM = False output_file_name_for_hpo = "val_dict_list.json" def f1score(positive, negative): #labels[predictions.ravel() < 0.5].sum() fsocre = 0.0 true_positive = 0.0 false_positive = 0 false_negitive = 0 for i in range(len(positive)): if positive[i] <= negative[i]: true_positive += 1 else: false_negitive += 1 false_positive += 1 print('tp' + str(true_positive)) print('fp' + str(false_positive)) print('fn' + str(false_negitive)) fscore = (2 * true_positive) / ((2 * true_positive) + false_negitive + false_positive) return fscore def get_embedding_layer(tokenizer): word_index = tokenizer.word_index num_words = len(word_index) + 1 embedding_matrix = np.zeros((num_words, EMBEDDING_DIM)) print('about to get kz') kz = KazumaCharEmbedding() print('got kz') for word, i in word_index.items(): if i >= MAX_NB_WORDS: continue embedding_vector = kz.emb(word) if embedding_vector is not None: if sum(embedding_vector) == 0: print("failed to find embedding for:" + word) # words not found in embedding index will be all-zeros. embedding_matrix[i] = embedding_vector print("Number of words:" + str(num_words)) embedding_layer = Embedding(num_words, EMBEDDING_DIM, weights=[embedding_matrix], input_length=MAX_SEQUENCE_LENGTH, trainable=False) return embedding_layer def get_sequences(texts, tokenizer): sequences = {} sequences['anchor'] = tokenizer.texts_to_sequences(texts['anchor']) sequences['anchor'] = pad_sequences(sequences['anchor'], maxlen=MAX_SEQUENCE_LENGTH) sequences['negative'] = tokenizer.texts_to_sequences(texts['negative']) sequences['negative'] = pad_sequences(sequences['negative'], maxlen=MAX_SEQUENCE_LENGTH) sequences['positive'] = tokenizer.texts_to_sequences(texts['positive']) sequences['positive'] = pad_sequences(sequences['positive'], maxlen=MAX_SEQUENCE_LENGTH) return sequences def read_entities(filepath): entities = [] with open(filepath) as fl: for line in fl: entities.append(line) return entities def read_file(file_path): texts = {'anchor':[], 'negative':[], 'positive':[]} fl = open(file_path, 'r') i = 0 for line in fl: line_array = line.split("|") texts['anchor'].append(line_array[0]) texts['positive'].append(line_array[1]) texts['negative'].append(line_array[2]) i += 1 if i > DEBUG_DATA_LENGTH and DEBUG: break return texts def split(entities, test_split = 0.2): if DEBUG: ents = entities[0:DEBUG_DATA_LENGTH] else: random.shuffle(entities) ents = entities num_validation_samples = int(test_split * len(ents)) return ents[:-num_validation_samples], ents[-num_validation_samples:] """ define a single objective function based on angular loss instead of triplet loss """ def angular_loss(y_true, y_pred): alpha = K.constant(ALPHA) a_p = y_pred[:,0,0] n_c = y_pred[:,1,0] return K.mean(K.maximum(K.constant(0), K.square(a_p) - K.constant(4) * K.square(tf.tan(alpha)) * K.square(n_c))) """ Facenet triplet loss function: https://arxiv.org/pdf/1503.03832.pdf """ def schroff_triplet_loss(y_true, y_pred): margin = K.constant(0.2) return K.mean(K.maximum(K.constant(0), K.square(y_pred[:,0,0]) - K.square(y_pred[:,1,0]) + margin)) def triplet_loss(y_true, y_pred): margin = K.constant(MARGIN) return K.mean(K.square(y_pred[:,0,0]) + K.square(margin - y_pred[:,1,0])) def triplet_tanh_loss(y_true, y_pred): return K.mean(K.tanh(y_pred[:,0,0]) + (K.constant(1) - K.tanh(y_pred[:,1,0]))) def triplet_tanh_pn_loss(y_true, y_pred): return K.mean(K.tanh(y_pred[:,0,0]) + ((K.constant(1) - K.tanh(y_pred[:,1,0])) + (K.constant(1) - K.tanh(y_pred[:,2,0]))) / K.constant(2)); # the following triplet loss function is from: Deep Metric Learning with Improved Triplet Loss for # Face clustering in Videos def improved_loss(y_true, y_pred): margin = K.constant(1) lambda_p = K.constant(0.02) threshold = K.constant(0.1) a_p_distance = y_pred[:,0,0] a_n_distance = y_pred[:,1,0] p_n_distance = y_pred[:,2,0] phi = a_p_distance - ((a_n_distance + p_n_distance) / K.constant(2)) + margin psi = a_p_distance - threshold return K.maximum(K.constant(0), phi) + lambda_p * K.maximum(K.constant(0), psi) def accuracy(y_true, y_pred): return K.mean(y_pred[:,0,0] < y_pred[:,1,0]) def l2Norm(x): return K.l2_normalize(x, axis=-1) def euclidean_distance(vects): x, y = vects return K.sqrt(K.maximum(K.sum(K.square(x - y), axis=1, keepdims=True), K.epsilon())) def n_c_angular_distance(vects): x_a, x_p, x_n = vects return K.sqrt(K.maximum(K.sum(K.square(x_n - ((x_a + x_p) / K.constant(2))), axis=1, keepdims=True), K.epsilon())) def a_p_angular_distance(vects): x_a, x_p, x_n = vects return K.sqrt(K.maximum(K.sum(K.square(x_a - x_p), axis=1, keepdims=True), K.epsilon())) def build_unique_entities(entity2same): unique_text = [] entity2index = {} for key in entity2same: entity2index[key] = len(unique_text) unique_text.append(key) vals = entity2same[key] for v in vals: entity2index[v] = len(unique_text) unique_text.append(v) return unique_text, entity2index def generate_triplets_from_ANN(model, sequences, entity2unique, entity2same, unique_text, test): predictions = model.predict(sequences) t = AnnoyIndex(len(predictions[0]), metric='euclidean') # Length of item vector that will be indexed t.set_seed(123) for i in range(len(predictions)): # print(predictions[i]) v = predictions[i] t.add_item(i, v) t.build(100) # 100 trees match = 0 no_match = 0 accuracy = 0 total = 0 triplets = {} pos_distances = [] neg_distances = [] triplets['anchor'] = [] triplets['positive'] = [] triplets['negative'] = [] if test: NNlen = TEST_NEIGHBOR_LEN else: NNlen = TRAIN_NEIGHBOR_LEN for key in entity2same: index = entity2unique[key] nearest = t.get_nns_by_vector(predictions[index], NNlen) nearest_text = set([unique_text[i] for i in nearest]) expected_text = set(entity2same[key]) # annoy has this annoying habit of returning the queried item back as a nearest neighbor. Remove it. if key in nearest_text: nearest_text.remove(key) # print("query={} names = {} true_match = {}".format(unique_text[index], nearest_text, expected_text)) overlap = expected_text.intersection(nearest_text) # collect up some statistics on how well we did on the match m = len(overlap) match += m # since we asked for only x nearest neighbors, and we get at most x-1 neighbors that are not the same as key (!) # make sure we adjust our estimate of no match appropriately no_match += min(len(expected_text), NNlen - 1) - m # sample only the negatives that are true negatives # that is, they are not in the expected set - sampling only 'semi-hard negatives is not defined here' # positives = expected_text - nearest_text positives = expected_text negatives = nearest_text - expected_text # print(key + str(expected_text) + str(nearest_text)) for i in negatives: for j in positives: dist_pos = t.get_distance(index, entity2unique[j]) pos_distances.append(dist_pos) dist_neg = t.get_distance(index, entity2unique[i]) neg_distances.append(dist_neg) if dist_pos < dist_neg: accuracy += 1 total += 1 # print(key + "|" + j + "|" + i) # print(dist_pos) # print(dist_neg) triplets['anchor'].append(key) triplets['positive'].append(j) triplets['negative'].append(i) print("mean positive distance:" + str(statistics.mean(pos_distances))) print("stdev positive distance:" + str(statistics.stdev(pos_distances))) print("max positive distance:" + str(max(pos_distances))) print("mean neg distance:" + str(statistics.mean(neg_distances))) print("stdev neg distance:" + str(statistics.stdev(neg_distances))) print("max neg distance:" + str(max(neg_distances))) print("Accuracy in the ANN for triplets that obey the distance func:" + str(accuracy / total)) obj = {} obj['accuracy'] = accuracy / total obj['steps'] = 1 with open(output_file_name_for_hpo, 'w') as out: json.dump(obj, out) if test: return match/(match + no_match) else: return triplets, match/(match + no_match) def generate_names(entities, people, limit_pairs=False): if people: num_names = 4 generator = NameDataCleanser(0, num_names, limit_pairs=limit_pairs) else: generator = CompanyDataCleanser(limit_pairs) num_names = 2 entity2same = {} for entity in entities: ret = generator.cleanse_data(entity) if ret and len(ret) >= num_names: entity2same[ret[0]] = ret[1:] return entity2same def embedded_representation_model(embedding_layer): seq = Sequential() seq.add(embedding_layer) seq.add(Flatten()) return seq def build_model(embedder): main_input = Input(shape=(MAX_SEQUENCE_LENGTH,)) net = embedder(main_input) for i in range(0, NUM_LAYERS): net = GRU(128, return_sequences=True, activation='relu', name='embed' + str(i))(net) net = GRU(128, activation='relu', name='embed' + str(i+1))(net) if USE_L2_NORM: net = Lambda(l2Norm, output_shape=[128])(net) base_model = Model(embedder.input, net, name='triplet_model') base_model.summary() input_shape=(MAX_SEQUENCE_LENGTH,) input_anchor = Input(shape=input_shape, name='input_anchor') input_positive = Input(shape=input_shape, name='input_pos') input_negative = Input(shape=input_shape, name='input_neg') net_anchor = base_model(input_anchor) net_positive = base_model(input_positive) net_negative = base_model(input_negative) positive_dist = Lambda(euclidean_distance, name='pos_dist', output_shape=(1,))([net_anchor, net_positive]) negative_dist = Lambda(euclidean_distance, name='neg_dist', output_shape=(1,))([net_anchor, net_negative]) if USE_ANGULAR_LOSS: n_c = Lambda(n_c_angular_distance, name='nc_angular_dist')([net_anchor, net_positive, net_negative]) a_p = Lambda(a_p_angular_distance, name='ap_angular_dist')([net_anchor, net_positive, net_negative]) stacked_dists = Lambda( lambda vects: K.stack(vects, axis=1), name='stacked_dists', output_shape=(3, 1) )([a_p, n_c]) model = Model([input_anchor, input_positive, input_negative], stacked_dists, name='triple_siamese') model.compile(optimizer="rmsprop", loss=angular_loss, metrics=[accuracy]) else: exemplar_negative_dist = Lambda(euclidean_distance, name='exemplar_neg_dist', output_shape=(1,))([net_positive, net_negative]) stacked_dists = Lambda( # lambda vects: C.splice(*vects, axis=C.Axis.new_leading_axis()).eval(vects), lambda vects: K.stack(vects, axis=1), name='stacked_dists', output_shape=(3, 1) )([positive_dist, negative_dist, exemplar_negative_dist]) model = Model([input_anchor, input_positive, input_negative], stacked_dists, name='triple_siamese') model.compile(optimizer="rmsprop", loss=LOSS_FUNCTION, metrics=[accuracy]) test_positive_model = Model([input_anchor, input_positive, input_negative], positive_dist) test_negative_model = Model([input_anchor, input_positive, input_negative], negative_dist) inter_model = Model(input_anchor, net_anchor) print("output_shapes") model.summary() # print(positive_dist.output_shape) # print(negative_dist.output_shape) # print(exemplar_negative_dist) # print(neg_dist.output_shape) return model, test_positive_model, test_negative_model, inter_model parser = argparse.ArgumentParser(description='Run fuzzy join algorithm') parser.add_argument('--debug_sample_size', type=int, help='sample size for debug run') parser.add_argument('--margin', type=int, help='margin') parser.add_argument('--loss_function', type=str, help='triplet loss function type: schroff-loss, improved-loss, angular-loss, tanh-loss, improved-tanh-loss') parser.add_argument('--use_l2_norm', type=str, help='whether to add a l2 norm') parser.add_argument('--num_layers', type=int, help='num_layers to use. Minimum is 2') parser.add_argument('--input', type=str, help='Input file') parser.add_argument('--entity_type', type=str, help='people or companies') args = parser.parse_args() LOSS_FUNCTION = None if args.loss_function == 'schroff-loss': LOSS_FUNCTION=schroff_triplet_loss elif args.loss_function == 'improved-loss': LOSS_FUNCTION=improved_loss elif args.loss_function == 'our-loss': LOSS_FUNCTION=triplet_loss elif args.loss_function == 'tanh-loss': LOSS_FUNCTION=triplet_tanh_loss elif args.loss_function == 'improved-tanh-loss': LOSS_FUNCTION=triplet_tanh_pn_loss elif args.loss_function == 'angular-loss': USE_ANGULAR_LOSS = True LOSS_FUNCTION = angular_loss print('Loss function: ' + args.loss_function) if args.debug_sample_size: DEBUG=True DEBUG_DATA_LENGTH=args.debug_sample_size print('Debug data length:' + str(DEBUG_DATA_LENGTH)) print('Margin:' + str(MARGIN)) USE_L2_NORM = args.use_l2_norm.lower() in ("yes", "true", "t", "1") print('Use L2Norm: ' + str(USE_L2_NORM)) print('Use L2Norm: ' + str(args.use_l2_norm)) NUM_LAYERS = args.num_layers - 1 print('Num layers: ' + str(NUM_LAYERS)) people = 'people' in args.entity_type # read all entities and create positive parts of a triplet entities = read_entities(args.input) train, test = split(entities, test_split = .20) print("TRAIN") print(train) print("TEST") print(test) entity2same_train = generate_names(train, people) entity2same_test = generate_names(test, people, limit_pairs=True) print(entity2same_train) print(entity2same_test) # change the default behavior of the tokenizer to ignore all punctuation except , - and . which are important # clues for entity names tokenizer = Tokenizer(num_words=MAX_NB_WORDS, lower=True, filters='!"#$%&()*+/:;<=>?@[\]^_`{|}~', split=" ") # build a set of data structures useful for annoy, the set of unique entities (unique_text), # a mapping of entities in texts to an index in unique_text, a mapping of entities to other same entities, and the actual # vectorized representation of the text. These structures will be used iteratively as we build up the model # so we need to create them once for re-use unique_text, entity2unique = build_unique_entities(entity2same_train) unique_text_test, entity2unique_test = build_unique_entities(entity2same_test) print("train text len:" + str(len(unique_text))) print("test text len:" + str(len(unique_text_test))) tokenizer.fit_on_texts(unique_text + unique_text_test) sequences = tokenizer.texts_to_sequences(unique_text) sequences = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH) sequences_test = tokenizer.texts_to_sequences(unique_text_test) sequences_test = pad_sequences(sequences_test, maxlen=MAX_SEQUENCE_LENGTH) # build models embedder = get_embedding_layer(tokenizer) model, test_positive_model, test_negative_model, inter_model = build_model(embedder) embedder_model = embedded_representation_model(embedder) if DEBUG_ANN: generate_triplets_from_ANN(embedder_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, True) sys.exit() test_data, test_match_stats = generate_triplets_from_ANN(embedder_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, False) test_seq = get_sequences(test_data, tokenizer) print("Test stats:" + str(test_match_stats)) counter = 0 current_model = embedder_model prev_match_stats = 0 train_data, match_stats = generate_triplets_from_ANN(current_model, sequences, entity2unique, entity2same_train, unique_text, False) print("Match stats:" + str(match_stats)) number_of_names = len(train_data['anchor']) # print(train_data['anchor']) print("number of names" + str(number_of_names)) Y_train = np.random.randint(2, size=(1,2,number_of_names)).T filepath="weights.best.hdf5" checkpoint = ModelCheckpoint(filepath, monitor='val_accuracy', verbose=1, save_best_only=True, mode='max') early_stop = EarlyStopping(monitor='val_accuracy', patience=1, mode='max') callbacks_list = [checkpoint, early_stop] train_seq = get_sequences(train_data, tokenizer) # check just for 5 epochs because this gets called many times model.fit([train_seq['anchor'], train_seq['positive'], train_seq['negative']], Y_train, epochs=100, batch_size=40, callbacks=callbacks_list, validation_split=0.2) current_model = inter_model # print some statistics on this epoch print("training data predictions") positives = test_positive_model.predict([train_seq['anchor'], train_seq['positive'], train_seq['negative']]) negatives = test_negative_model.predict([train_seq['anchor'], train_seq['positive'], train_seq['negative']]) print("f1score for train is: {}".format(f1score(positives, negatives))) print("test data predictions") positives = test_positive_model.predict([test_seq['anchor'], test_seq['positive'], test_seq['negative']]) negatives = test_negative_model.predict([test_seq['anchor'], test_seq['positive'], test_seq['negative']]) print("f1score for test is: {}".format(f1score(positives, negatives))) test_match_stats = generate_triplets_from_ANN(current_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, True) print("Test stats:" + str(test_match_stats))
19,474
35.88447
163
py
fuzzyJoiner
fuzzyJoiner-master/old/TripletLossFacenetLSTM-schroffloss.py
import numpy as np import tensorflow as tf import random as random # import cntk as C # """ # The below is necessary in Python 3.2.3 onwards to # have reproducible behavior for certain hash-based operations. # See these references for further details: # https://docs.python.org/3.4/using/cmdline.html#envvar-PYTHONHASHSEED # https://github.com/keras-team/keras/issues/2280#issuecomment-306959926 import os os.environ['PYTHONHASHSEED'] = '0' # The below is necessary for starting Numpy generated random numbers # in a well-defined initial state. np.random.seed(42) # The below is necessary for starting core Python generated random numbers # in a well-defined state. random.seed(12345) # Force TensorFlow to use single thread. # Multiple threads are a potential source of # non-reproducible results. # For further details, see: https://stackoverflow.com/questions/42022950/which-seeds-have-to-be-set-where-to-realize-100-reproducibility-of-training-res session_conf = tf.ConfigProto(intra_op_parallelism_threads=1, inter_op_parallelism_threads=1) # import theano as T from keras import backend as K # The below tf.set_random_seed() will make random number generation # in the TensorFlow backend have a well-defined initial state. # For further details, see: https://www.tensorflow.org/api_docs/python/tf/set_random_seed tf.set_random_seed(1234) sess = tf.Session(graph=tf.get_default_graph(), config=session_conf) K.set_session(sess) # from keras import backend as K from keras.preprocessing.text import Tokenizer from keras.preprocessing.sequence import pad_sequences from keras.layers import Dense, Input, Flatten, Dropout, Lambda, GRU, Activation from keras.layers.wrappers import Bidirectional from keras.layers import Conv1D, MaxPooling1D, Embedding from keras.models import Model, model_from_json, Sequential from embeddings import KazumaCharEmbedding from annoy import AnnoyIndex from keras.callbacks import ModelCheckpoint, EarlyStopping from names_cleanser import NameDataCleanser import sys import statistics import argparse #must fix MAX_NB_WORDS = 140000 EMBEDDING_DIM = 100 MAX_SEQUENCE_LENGTH = 10 MARGIN=1 ALPHA=30 USE_GRU=True DEBUG = False DEBUG_DATA_LENGTH = 20000 DEBUG_ANN = False USE_ANGULAR_LOSS=False LOSS_FUNCTION=None TRAIN_NEIGHBOR_LEN=20 TEST_NEIGHBOR_LEN=20 EMBEDDING_TYPE = 'Kazuma' NUM_LAYERS = 3 USE_L2_NORM = True def f1score(positive, negative): #labels[predictions.ravel() < 0.5].sum() fsocre = 0.0 true_positive = 0.0 false_positive = 0 false_negitive = 0 for i in range(len(positive)): if positive[i] <= negative[i]: true_positive += 1 else: false_negitive += 1 false_positive += 1 print('tp' + str(true_positive)) print('fp' + str(false_positive)) print('fn' + str(false_negitive)) fscore = (2 * true_positive) / ((2 * true_positive) + false_negitive + false_positive) return fscore def get_embedding_layer(tokenizer): word_index = tokenizer.word_index num_words = len(word_index) + 1 embedding_matrix = np.zeros((num_words, EMBEDDING_DIM)) print('about to get kz') kz = KazumaCharEmbedding() print('got kz') for word, i in word_index.items(): if i >= MAX_NB_WORDS: continue embedding_vector = kz.emb(word) if embedding_vector is not None: if sum(embedding_vector) == 0: print("failed to find embedding for:" + word) # words not found in embedding index will be all-zeros. embedding_matrix[i] = embedding_vector print("Number of words:" + str(num_words)) embedding_layer = Embedding(num_words, EMBEDDING_DIM, weights=[embedding_matrix], input_length=MAX_SEQUENCE_LENGTH, trainable=False) return embedding_layer def get_sequences(texts, tokenizer): sequences = {} sequences['anchor'] = tokenizer.texts_to_sequences(texts['anchor']) sequences['anchor'] = pad_sequences(sequences['anchor'], maxlen=MAX_SEQUENCE_LENGTH) sequences['negative'] = tokenizer.texts_to_sequences(texts['negative']) sequences['negative'] = pad_sequences(sequences['negative'], maxlen=MAX_SEQUENCE_LENGTH) sequences['positive'] = tokenizer.texts_to_sequences(texts['positive']) sequences['positive'] = pad_sequences(sequences['positive'], maxlen=MAX_SEQUENCE_LENGTH) return sequences def read_entities(filepath): entities = [] with open(filepath) as fl: for line in fl: entities.append(line) return entities def read_file(file_path): texts = {'anchor':[], 'negative':[], 'positive':[]} fl = open(file_path, 'r') i = 0 for line in fl: line_array = line.split("|") texts['anchor'].append(line_array[0]) texts['positive'].append(line_array[1]) texts['negative'].append(line_array[2]) i += 1 if i > DEBUG_DATA_LENGTH and DEBUG: break return texts def split(entities, test_split = 0.2): if DEBUG: ents = entities[0:DEBUG_DATA_LENGTH] else: random.shuffle(entities) ents = entities num_validation_samples = int(test_split * len(ents)) return ents[:-num_validation_samples], ents[-num_validation_samples:] """ define a single objective function based on angular loss instead of triplet loss """ def angular_loss(y_true, y_pred): alpha = K.constant(ALPHA) a_p = y_pred[:,0,0] n_c = y_pred[:,1,0] return K.mean(K.maximum(K.constant(0), K.square(a_p) - K.constant(4) * K.square(T.tensor.tan(alpha)) * K.square(n_c))) """ Facenet triplet loss function: https://arxiv.org/pdf/1503.03832.pdf """ def schroff_triplet_loss(y_true, y_pred): margin = K.constant(MARGIN) return K.mean(K.maximum(K.constant(0), K.square(y_pred[:,0,0]) - K.square(y_pred[:,1,0]) + margin)) def triplet_loss(y_true, y_pred): # margin = K.constant(MARGIN) # return K.mean(K.square(y_pred[:,0,0]) + K.square(margin - y_pred[:,1,0])) margin = K.constant(MARGIN) return K.mean(K.square(y_pred[:,0,0]) + K.square(margin - y_pred[:,1,0])) # return K.mean(K.square(y_pred[:,0,0]) + K.square(margin - y_pred[:,1,0]) + K.square(margin - y_pred[:,2,0])) # the following triplet loss function is from: Deep Metric Learning with Improved Triplet Loss for # Face clustering in Videos def improved_loss(y_true, y_pred): margin = K.constant(MARGIN) lambda_p = 0.02 threshold = 0.1 a_p_distance = y_pred[:,0,0] a_n_distance = y_pred[:,1,0] p_n_distance = y_pred[:,2,0] phi = a_p_distance - ((a_n_distance + p_n_distance) / 2) + margin psi = a_p_distance - threshold return K.maximum(K.constant(0), phi) + lambda_p * K.maximum(K.constant(0), psi) def accuracy(y_true, y_pred): return K.mean(y_pred[:,0,0] < y_pred[:,1,0]) def l2Norm(x): return K.l2_normalize(x, axis=-1) def tanhNorm(x): square_sum = K.sum(K.square(x), axis=-1, keepdims=True) dist = K.sqrt(K.maximum(square_sum, K.epsilon())) tanh = K.tanh(dist) scale = tanh / dist return x * scale def euclidean_distance(vects): x, y = vects return K.sqrt(K.maximum(K.sum(K.square(x - y), axis=1, keepdims=True), K.epsilon())) def n_c_angular_distance(vects): x_a, x_p, x_n = vects return K.sqrt(K.maximum(K.sum(K.square(x_n - ((x_a + x_p) / K.constant(2))), axis=1, keepdims=True), K.epsilon())) def a_p_angular_distance(vects): x_a, x_p, x_n = vects return K.sqrt(K.maximum(K.sum(K.square(x_a - x_p), axis=1, keepdims=True), K.epsilon())) def build_unique_entities(entity2same): unique_text = [] entity2index = {} for key in entity2same: entity2index[key] = len(unique_text) unique_text.append(key) vals = entity2same[key] for v in vals: entity2index[v] = len(unique_text) unique_text.append(v) return unique_text, entity2index def generate_triplets_from_ANN(model, sequences, entity2unique, entity2same, unique_text, test): predictions = model.predict(sequences) t = AnnoyIndex(len(predictions[0]), metric='euclidean') # Length of item vector that will be indexed t.set_seed(123) for i in range(len(predictions)): # print(predictions[i]) v = predictions[i] t.add_item(i, v) t.build(100) # 100 trees match = 0 no_match = 0 accuracy = 0 total = 0 triplets = {} pos_distances = [] neg_distances = [] triplets['anchor'] = [] triplets['positive'] = [] triplets['negative'] = [] if test: NNlen = TEST_NEIGHBOR_LEN else: NNlen = TRAIN_NEIGHBOR_LEN for key in entity2same: index = entity2unique[key] nearest = t.get_nns_by_vector(predictions[index], NNlen) nearest_text = set([unique_text[i] for i in nearest]) expected_text = set(entity2same[key]) # annoy has this annoying habit of returning the queried item back as a nearest neighbor. Remove it. if key in nearest_text: nearest_text.remove(key) # print("query={} names = {} true_match = {}".format(unique_text[index], nearest_text, expected_text)) overlap = expected_text.intersection(nearest_text) # collect up some statistics on how well we did on the match m = len(overlap) match += m # since we asked for only x nearest neighbors, and we get at most x-1 neighbors that are not the same as key (!) # make sure we adjust our estimate of no match appropriately no_match += min(len(expected_text), NNlen - 1) - m # sample only the negatives that are true negatives # that is, they are not in the expected set - sampling only 'semi-hard negatives is not defined here' # positives = expected_text - nearest_text positives = expected_text negatives = nearest_text - expected_text # print(key + str(expected_text) + str(nearest_text)) for i in negatives: for j in positives: dist_pos = t.get_distance(index, entity2unique[j]) pos_distances.append(dist_pos) dist_neg = t.get_distance(index, entity2unique[i]) neg_distances.append(dist_neg) if dist_pos < dist_neg: accuracy += 1 total += 1 # print(key + "|" + j + "|" + i) # print(dist_pos) # print(dist_neg) triplets['anchor'].append(key) triplets['positive'].append(j) triplets['negative'].append(i) print("mean positive distance:" + str(statistics.mean(pos_distances))) print("stdev positive distance:" + str(statistics.stdev(pos_distances))) print("max positive distance:" + str(max(pos_distances))) print("mean neg distance:" + str(statistics.mean(neg_distances))) print("stdev neg distance:" + str(statistics.stdev(neg_distances))) print("max neg distance:" + str(max(neg_distances))) print("Accuracy in the ANN for triplets that obey the distance func:" + str(accuracy / total)) if test: return match/(match + no_match) else: return triplets, match/(match + no_match) def generate_names(entities, limit_pairs=False): num_names = 4 names_generator = NameDataCleanser(0, num_names, limit_pairs=limit_pairs) entity2same = {} for entity in entities: ret = names_generator.cleanse_data(entity) if ret and len(ret) >= num_names: entity2same[ret[0]] = ret[1:] return entity2same def embedded_representation_model(embedding_layer): seq = Sequential() seq.add(embedding_layer) seq.add(Flatten()) return seq def get_hidden_layer(name, net, is_last): if USE_GRU: if is_last: return GRU(128, activation='relu', name=name)(net) else: return GRU(128, return_sequences=True, activation='relu', name=name)(net) else: return Dense(128, activation='relu', name=name)(net) def build_model(embedder): main_input = Input(shape=(MAX_SEQUENCE_LENGTH,)) net = embedder(main_input) net = GRU(128, return_sequences=True, activation='relu', name='embed')(net) net = GRU(128, return_sequences=True, activation='relu', name='embed2')(net) net = GRU(128, return_sequences=True, activation='relu', name='embed2a')(net) net = GRU(128, activation='relu', name='embed3')(net) """ for i in range(0, NUM_LAYERS): net = get_hidden_layer('embed' + str(i), net, False) net = get_hidden_layer('embed_last', net, True) """ # if USE_L2_NORM: net = Lambda(l2Norm, output_shape=[128])(net) base_model = Model(embedder.input, net, name='triplet_model') base_model.summary() input_shape=(MAX_SEQUENCE_LENGTH,) input_anchor = Input(shape=input_shape, name='input_anchor') input_positive = Input(shape=input_shape, name='input_pos') input_negative = Input(shape=input_shape, name='input_neg') net_anchor = base_model(input_anchor) net_positive = base_model(input_positive) net_negative = base_model(input_negative) positive_dist = Lambda(euclidean_distance, name='pos_dist', output_shape=(1,))([net_anchor, net_positive]) negative_dist = Lambda(euclidean_distance, name='neg_dist', output_shape=(1,))([net_anchor, net_negative]) if USE_ANGULAR_LOSS: n_c = Lambda(n_c_angular_distance, name='nc_angular_dist')([net_anchor, net_positive, net_negative]) a_p = Lambda(a_p_angular_distance, name='ap_angular_dist')([net_anchor, net_positive, net_negative]) stacked_dists = Lambda( lambda vects: K.stack(vects, axis=1), name='stacked_dists', output_shape=(3, 1) )([a_p, n_c]) model = Model([input_anchor, input_positive, input_negative], stacked_dists, name='triple_siamese') model.compile(optimizer="rmsprop", loss=angular_loss, metrics=[accuracy]) else: exemplar_negative_dist = Lambda(euclidean_distance, name='exemplar_neg_dist', output_shape=(1,))([net_positive, net_negative]) stacked_dists = Lambda( # lambda vects: C.splice(*vects, axis=C.Axis.new_leading_axis()).eval(vects), lambda vects: K.stack(vects, axis=1), name='stacked_dists', output_shape=(3, 1) )([positive_dist, negative_dist, exemplar_negative_dist]) model = Model([input_anchor, input_positive, input_negative], stacked_dists, name='triple_siamese') model.compile(optimizer="rmsprop", loss=schroff_triplet_loss, metrics=[accuracy]) test_positive_model = Model([input_anchor, input_positive, input_negative], positive_dist) test_negative_model = Model([input_anchor, input_positive, input_negative], negative_dist) inter_model = Model(input_anchor, net_anchor) print("output_shapes") model.summary() # print(positive_dist.output_shape) # print(negative_dist.output_shape) # print(exemplar_negative_dist) # print(neg_dist.output_shape) return model, test_positive_model, test_negative_model, inter_model parser = argparse.ArgumentParser(description='Run fuzzy join algorithm') parser.add_argument('--debug_sample_size', type=int, help='sample size for debug run') parser.add_argument('--margin', type=int, help='margin') parser.add_argument('--loss_function', type=str, help='loss function type: triplet-loss, improved-triplet-loss, modified-loss, or angular-loss') parser.add_argument('--use_l2_norm', type=bool, help='whether to add a l2 norm') parser.add_argument('--test_neighbor_len', type=int, help='size of the neighborhood for testing') parser.add_argument('--train_neighbor_len', type=int, help='size of the neighborhood for training') parser.add_argument('--embedding_type', type=str, help='encoding type to use for input: Kazuma (for Kazuma character embedding) or one-hot') parser.add_argument('--use_GRU', type=bool, help='use GRU or default to MLP') parser.add_argument('--num_layers', type=int, help='num_layers to use. Minimum is 2') parser.add_argument('--input', type=str, help='Input file') args = parser.parse_args() """ LOSS_FUNCTION = None if args.loss_function == 'triplet-loss': LOSS_FUNCTION=schroff_triplet_loss elif args.loss_function == 'improved_triplet_loss': LOSS_FUNCTION=improved_triplet_loss elif args.loss_function == 'modified_loss': LOSS_FUNCTION=triplet_loss elif args.loss_function == 'angular-loss': USE_ANGULAR_LOSS = true LOSS_FUNCTION = angular_loss print('Loss function: ' + args.loss_function) if args.debug_sample_size: DEBUG=True DEBUG_DATA_LENGTH=args.debug_sample_size print('Debug data length:' + str(DEBUG_DATA_LENGTH)) MARGIN = args.margin print('Margin:' + str(MARGIN)) TRAIN_NEIGHBOR_LEN = args.train_neighbor_len TEST_NEIGHBOR_LEN = args.test_neighbor_len print('Train neighbor length: ' + str(TRAIN_NEIGHBOR_LEN)) print('Test neighbor length: ' + str(TEST_NEIGHBOR_LEN)) USE_L2_NORM = args.use_l2_norm print('Use L2Norm: ' + str(USE_L2_NORM)) EMBEDDING_TYPE = args.embedding_type print('Embedding type: ' + EMBEDDING_TYPE) USE_GRU = args.use_GRU print('Use GRU: ' + str(args.use_GRU)) NUM_LAYERS = args.num_layers - 1 print('Num layers: ' + str(NUM_LAYERS)) """ # read all entities and create positive parts of a triplet entities = read_entities(args.input) train, test = split(entities, test_split = .20) print("TRAIN") print(train) print("TEST") print(test) entity2same_train = generate_names(train) entity2same_test = generate_names(test, limit_pairs=True) print(entity2same_train) print(entity2same_test) # change the default behavior of the tokenizer to ignore all punctuation except , - and . which are important # clues for entity names tokenizer = Tokenizer(num_words=MAX_NB_WORDS, lower=True, filters='!"#$%&()*+/:;<=>?@[\]^_`{|}~', split=" ") # build a set of data structures useful for annoy, the set of unique entities (unique_text), # a mapping of entities in texts to an index in unique_text, a mapping of entities to other same entities, and the actual # vectorized representation of the text. These structures will be used iteratively as we build up the model # so we need to create them once for re-use unique_text, entity2unique = build_unique_entities(entity2same_train) unique_text_test, entity2unique_test = build_unique_entities(entity2same_test) print("train text len:" + str(len(unique_text))) print("test text len:" + str(len(unique_text_test))) tokenizer.fit_on_texts(unique_text + unique_text_test) sequences = tokenizer.texts_to_sequences(unique_text) sequences = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH) sequences_test = tokenizer.texts_to_sequences(unique_text_test) sequences_test = pad_sequences(sequences_test, maxlen=MAX_SEQUENCE_LENGTH) # build models embedder = get_embedding_layer(tokenizer) model, test_positive_model, test_negative_model, inter_model = build_model(embedder) embedder_model = embedded_representation_model(embedder) if DEBUG_ANN: generate_triplets_from_ANN(embedder_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, True) sys.exit() test_data, test_match_stats = generate_triplets_from_ANN(embedder_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, False) test_seq = get_sequences(test_data, tokenizer) print("Test stats:" + str(test_match_stats)) match_stats = 0 # num_iter = 100 num_iter = 1 counter = 0 current_model = embedder_model prev_match_stats = 0 while test_match_stats < .9 and counter < num_iter: counter += 1 train_data, match_stats = generate_triplets_from_ANN(current_model, sequences, entity2unique, entity2same_train, unique_text, False) print("Match stats:" + str(match_stats)) number_of_names = len(train_data['anchor']) # print(train_data['anchor']) print("number of names" + str(number_of_names)) Y_train = np.random.randint(2, size=(1,2,number_of_names)).T filepath="weights.best.hdf5" checkpoint = ModelCheckpoint(filepath, monitor='val_accuracy', verbose=1, save_best_only=True, mode='max') early_stop = EarlyStopping(monitor='val_accuracy', patience=1, mode='max') callbacks_list = [checkpoint, early_stop] train_seq = get_sequences(train_data, tokenizer) # check just for 5 epochs because this gets called many times model.fit([train_seq['anchor'], train_seq['positive'], train_seq['negative']], Y_train, epochs=100, batch_size=40, callbacks=callbacks_list, validation_split=0.2) current_model = inter_model # print some statistics on this epoch print("training data predictions") positives = test_positive_model.predict([train_seq['anchor'], train_seq['positive'], train_seq['negative']]) negatives = test_negative_model.predict([train_seq['anchor'], train_seq['positive'], train_seq['negative']]) print("f1score for train is: {}".format(f1score(positives, negatives))) print("test data predictions") positives = test_positive_model.predict([test_seq['anchor'], test_seq['positive'], test_seq['negative']]) negatives = test_negative_model.predict([test_seq['anchor'], test_seq['positive'], test_seq['negative']]) print("f1score for test is: {}".format(f1score(positives, negatives))) test_match_stats = generate_triplets_from_ANN(current_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, True) print("Test stats:" + str(test_match_stats))
21,846
37.463028
167
py
fuzzyJoiner
fuzzyJoiner-master/old/TripletLossFacenetLSTM-8.29.18.py
import numpy as np import pandas import tensorflow as tf import random as random import json from keras import backend as K from keras.preprocessing.text import Tokenizer from keras.preprocessing.sequence import pad_sequences from keras.layers import Dense, Input, Flatten, Dropout, Lambda, GRU, Activation from keras.layers.wrappers import Bidirectional from keras.layers import Conv1D, MaxPooling1D, Embedding from keras.models import Model, model_from_json, Sequential from embeddings import KazumaCharEmbedding from annoy import AnnoyIndex from keras.callbacks import ModelCheckpoint, EarlyStopping from names_cleanser import NameDataCleanser, CompanyDataCleanser import sys import statistics from scipy.stats.mstats import gmean import argparse #must fix MAX_NB_WORDS = 140000 EMBEDDING_DIM = 100 MAX_SEQUENCE_LENGTH = 10 MARGIN=10 ALPHA=45 DEBUG = False DEBUG_DATA_LENGTH = 100 DEBUG_ANN = False USE_ANGULAR_LOSS=False LOSS_FUNCTION=None TRAIN_NEIGHBOR_LEN=20 TEST_NEIGHBOR_LEN=20 EMBEDDING_TYPE = 'Kazuma' NUM_LAYERS = 3 USE_L2_NORM = False filepath="weights.best.hdf5" output_file_name_for_hpo = "val_dict_list.json" def f1score(positive, negative): #labels[predictions.ravel() < 0.5].sum() fsocre = 0.0 true_positive = 0.0 false_positive = 0 false_negitive = 0 for i in range(len(positive)): if positive[i] <= negative[i]: true_positive += 1 else: false_negitive += 1 false_positive += 1 print('tp' + str(true_positive)) print('fp' + str(false_positive)) print('fn' + str(false_negitive)) fscore = (2 * true_positive) / ((2 * true_positive) + false_negitive + false_positive) return fscore def get_embedding_layer(tokenizer): word_index = tokenizer.word_index num_words = len(word_index) + 1 embedding_matrix = np.zeros((num_words, EMBEDDING_DIM)) print('about to get kz') kz = KazumaCharEmbedding() print('got kz') for word, i in word_index.items(): if i >= MAX_NB_WORDS: continue embedding_vector = kz.emb(word) if embedding_vector is not None: if sum(embedding_vector) == 0: print("failed to find embedding for:" + word) # words not found in embedding index will be all-zeros. embedding_matrix[i] = embedding_vector print("Number of words:" + str(num_words)) embedding_layer = Embedding(num_words, EMBEDDING_DIM, weights=[embedding_matrix], input_length=MAX_SEQUENCE_LENGTH, trainable=False) return embedding_layer def get_sequences(texts, tokenizer): sequences = {} sequences['anchor'] = tokenizer.texts_to_sequences(texts['anchor']) sequences['anchor'] = pad_sequences(sequences['anchor'], maxlen=MAX_SEQUENCE_LENGTH) sequences['negative'] = tokenizer.texts_to_sequences(texts['negative']) sequences['negative'] = pad_sequences(sequences['negative'], maxlen=MAX_SEQUENCE_LENGTH) sequences['positive'] = tokenizer.texts_to_sequences(texts['positive']) sequences['positive'] = pad_sequences(sequences['positive'], maxlen=MAX_SEQUENCE_LENGTH) return sequences def read_entities(filepath): entities = [] with open(filepath) as fl: for line in fl: entities.append(line) return entities def read_file(file_path): texts = {'anchor':[], 'negative':[], 'positive':[]} fl = open(file_path, 'r') i = 0 for line in fl: line_array = line.split("|") texts['anchor'].append(line_array[0]) texts['positive'].append(line_array[1]) texts['negative'].append(line_array[2]) i += 1 if i > DEBUG_DATA_LENGTH and DEBUG: break return texts def split(entities, test_split = 0.2): if DEBUG: ents = entities[0:DEBUG_DATA_LENGTH] else: random.shuffle(entities) ents = entities num_validation_samples = int(test_split * len(ents)) return ents[:-num_validation_samples], ents[-num_validation_samples:] """ define a single objective function based on angular loss instead of triplet loss """ def angular_loss(y_true, y_pred): alpha = K.constant(ALPHA) a_p = y_pred[:,0,0] n_c = y_pred[:,1,0] return K.mean(K.maximum(K.constant(0), K.square(a_p) - K.constant(4) * K.square(tf.tan(alpha)) * K.square(n_c))) """ Facenet triplet loss function: https://arxiv.org/pdf/1503.03832.pdf """ def schroff_triplet_loss(y_true, y_pred): margin = K.constant(0.2) return K.mean(K.maximum(K.constant(0), K.square(y_pred[:,0,0]) - K.square(y_pred[:,1,0]) + margin)) def triplet_loss(y_true, y_pred): margin = K.constant(MARGIN) return K.mean(K.square(y_pred[:,0,0]) + K.square(K.maximum(margin - y_pred[:,1,0], K.constant(0)))) def triplet_tanh_loss(y_true, y_pred): return K.mean(K.tanh(y_pred[:,0,0]) + (K.constant(1) - K.tanh(y_pred[:,1,0]))) def triplet_tanh_pn_loss(y_true, y_pred): return K.mean(K.tanh(y_pred[:,0,0]) + ((K.constant(1) - K.tanh(y_pred[:,1,0])) + (K.constant(1) - K.tanh(y_pred[:,2,0]))) / K.constant(2)); # the following triplet loss function is from: Deep Metric Learning with Improved Triplet Loss for # Face clustering in Videos def improved_loss(y_true, y_pred): margin = K.constant(1) lambda_p = K.constant(0.02) threshold = K.constant(0.1) a_p_distance = y_pred[:,0,0] a_n_distance = y_pred[:,1,0] p_n_distance = y_pred[:,2,0] phi = a_p_distance - ((a_n_distance + p_n_distance) / K.constant(2)) + margin psi = a_p_distance - threshold return K.maximum(K.constant(0), phi) + lambda_p * K.maximum(K.constant(0), psi) def accuracy(y_true, y_pred): return K.mean(y_pred[:,0,0] < y_pred[:,1,0]) def l2Norm(x): return K.l2_normalize(x, axis=-1) def euclidean_distance(vects): x, y = vects return K.sqrt(K.maximum(K.sum(K.square(x - y), axis=1, keepdims=True), K.epsilon())) def n_c_angular_distance(vects): x_a, x_p, x_n = vects return K.sqrt(K.maximum(K.sum(K.square(x_n - ((x_a + x_p) / K.constant(2))), axis=1, keepdims=True), K.epsilon())) def a_p_angular_distance(vects): x_a, x_p, x_n = vects return K.sqrt(K.maximum(K.sum(K.square(x_a - x_p), axis=1, keepdims=True), K.epsilon())) def build_unique_entities(entity2same): unique_text = [] entity2index = {} for key in entity2same: entity2index[key] = len(unique_text) unique_text.append(key) vals = entity2same[key] for v in vals: entity2index[v] = len(unique_text) unique_text.append(v) return unique_text, entity2index def generate_triplets_from_ANN(model, sequences, entity2unique, entity2same, unique_text, test): predictions = model.predict(sequences) t = AnnoyIndex(len(predictions[0]), metric='euclidean') # Length of item vector that will be indexed t.set_seed(123) for i in range(len(predictions)): # print(predictions[i]) v = predictions[i] t.add_item(i, v) t.build(100) # 100 trees match = 0 no_match = 0 ann_accuracy = 0 total = 0 precise = 0 triplets = {} closest_positive_counts = [] pos_distances = [] neg_distances = [] all_pos_distances = [] all_neg_distances = [] triplets['anchor'] = [] triplets['positive'] = [] triplets['negative'] = [] if test: NNlen = TEST_NEIGHBOR_LEN else: NNlen = TRAIN_NEIGHBOR_LEN for key in entity2same: index = entity2unique[key] nearest = t.get_nns_by_vector(predictions[index], NNlen) nearest_text = set([unique_text[i] for i in nearest]) expected_text = set(entity2same[key]) # annoy has this annoying habit of returning the queried item back as a nearest neighbor. Remove it. if key in nearest_text: nearest_text.remove(key) # print("query={} names = {} true_match = {}".format(unique_text[index], nearest_text, expected_text)) overlap = expected_text.intersection(nearest_text) # collect up some statistics on how well we did on the match m = len(overlap) match += m # since we asked for only x nearest neighbors, and we get at most x-1 neighbors that are not the same as key (!) # make sure we adjust our estimate of no match appropriately no_match += min(len(expected_text), NNlen - 1) - m # sample only the negatives that are true negatives # that is, they are not in the expected set - sampling only 'semi-hard negatives is not defined here' # positives = expected_text - nearest_text positives = overlap negatives = nearest_text - expected_text # print(key + str(expected_text) + str(nearest_text)) for i in negatives: for j in positives: dist_pos = t.get_distance(index, entity2unique[j]) pos_distances.append(dist_pos) dist_neg = t.get_distance(index, entity2unique[i]) neg_distances.append(dist_neg) if dist_pos < dist_neg: ann_accuracy += 1 total += 1 # print(key + "|" + j + "|" + i) # print(dist_pos) # print(dist_neg) min_neg_distance = 1000000 for i in negatives: dist_neg = t.get_distance(index, entity2unique[i]) all_neg_distances.append(dist_neg) if dist_neg < min_neg_distance: min_neg_distance = dist_neg for j in expected_text: dist_pos = t.get_distance(index, entity2unique[j]) all_pos_distances.append(dist_pos) closest_pos_count = 0 for p in overlap: dist_pos = t.get_distance(index, entity2unique[p]) if dist_pos < min_neg_distance: closest_pos_count+=1 if closest_pos_count > 0: precise+=1 closest_positive_counts.append(closest_pos_count / min(len(expected_text), NNlen - 1)) for i in negatives: for j in expected_text: triplets['anchor'].append(key) triplets['positive'].append(j) triplets['negative'].append(i) print("mean closest positive count:" + str(statistics.mean(closest_positive_counts))) print("mean positive distance:" + str(statistics.mean(pos_distances))) print("stdev positive distance:" + str(statistics.stdev(pos_distances))) print("max positive distance:" + str(max(pos_distances))) print("mean neg distance:" + str(statistics.mean(neg_distances))) print("stdev neg distance:" + str(statistics.stdev(neg_distances))) print("max neg distance:" + str(max(neg_distances))) print("mean all positive distance:" + str(statistics.mean(all_pos_distances))) print("stdev all positive distance:" + str(statistics.stdev(all_pos_distances))) print("max all positive distance:" + str(max(all_pos_distances))) print("mean all neg distance:" + str(statistics.mean(all_neg_distances))) print("stdev all neg distance:" + str(statistics.stdev(all_neg_distances))) print("max all neg distance:" + str(max(all_neg_distances))) print("Accuracy in the ANN for triplets that obey the distance func:" + str(ann_accuracy / total)) print("Precision at 1: " + str(precise / len(entity2same))) obj = {} obj['accuracy'] = ann_accuracy / total obj['steps'] = 1 with open(output_file_name_for_hpo, 'w') as out: json.dump(obj, out) if test: return match/(match + no_match) else: return triplets, match/(match + no_match) def generate_names(entities, people, limit_pairs=False): if people: num_names = 4 generator = NameDataCleanser(0, num_names, limit_pairs=limit_pairs) else: generator = CompanyDataCleanser(limit_pairs) num_names = 2 entity2same = {} for entity in entities: ret = generator.cleanse_data(entity) if ret and len(ret) >= num_names: entity2same[ret[0]] = ret[1:] return entity2same def embedded_representation_model(embedding_layer): seq = Sequential() seq.add(embedding_layer) seq.add(Flatten()) return seq def build_model(embedder): main_input = Input(shape=(MAX_SEQUENCE_LENGTH,)) net = embedder(main_input) for i in range(0, NUM_LAYERS): net = GRU(128, return_sequences=True, activation='relu', name='embed' + str(i))(net) net = GRU(128, activation='relu', name='embed' + str(i+1))(net) if USE_L2_NORM: net = Lambda(l2Norm, output_shape=[128])(net) base_model = Model(embedder.input, net, name='triplet_model') base_model.summary() input_shape=(MAX_SEQUENCE_LENGTH,) input_anchor = Input(shape=input_shape, name='input_anchor') input_positive = Input(shape=input_shape, name='input_pos') input_negative = Input(shape=input_shape, name='input_neg') net_anchor = base_model(input_anchor) net_positive = base_model(input_positive) net_negative = base_model(input_negative) positive_dist = Lambda(euclidean_distance, name='pos_dist', output_shape=(1,))([net_anchor, net_positive]) negative_dist = Lambda(euclidean_distance, name='neg_dist', output_shape=(1,))([net_anchor, net_negative]) if USE_ANGULAR_LOSS: n_c = Lambda(n_c_angular_distance, name='nc_angular_dist')([net_anchor, net_positive, net_negative]) a_p = Lambda(a_p_angular_distance, name='ap_angular_dist')([net_anchor, net_positive, net_negative]) stacked_dists = Lambda( lambda vects: K.stack(vects, axis=1), name='stacked_dists', output_shape=(3, 1) )([a_p, n_c]) model = Model([input_anchor, input_positive, input_negative], stacked_dists, name='triple_siamese') model.compile(optimizer="rmsprop", loss=angular_loss, metrics=[accuracy]) else: exemplar_negative_dist = Lambda(euclidean_distance, name='exemplar_neg_dist', output_shape=(1,))([net_positive, net_negative]) stacked_dists = Lambda( # lambda vects: C.splice(*vects, axis=C.Axis.new_leading_axis()).eval(vects), lambda vects: K.stack(vects, axis=1), name='stacked_dists', output_shape=(3, 1) )([positive_dist, negative_dist, exemplar_negative_dist]) model = Model([input_anchor, input_positive, input_negative], stacked_dists, name='triple_siamese') model.compile(optimizer="rmsprop", loss=LOSS_FUNCTION, metrics=[accuracy]) test_positive_model = Model([input_anchor, input_positive, input_negative], positive_dist) test_negative_model = Model([input_anchor, input_positive, input_negative], negative_dist) inter_model = Model(input_anchor, net_anchor) print("output_shapes") model.summary() # print(positive_dist.output_shape) # print(negative_dist.output_shape) # print(exemplar_negative_dist) # print(neg_dist.output_shape) return model, test_positive_model, test_negative_model, inter_model parser = argparse.ArgumentParser(description='Run fuzzy join algorithm') parser.add_argument('--debug_sample_size', type=int, help='sample size for debug run') parser.add_argument('--margin', type=int, help='margin') parser.add_argument('--loss_function', type=str, help='triplet loss function type: schroff-loss, improved-loss, angular-loss, tanh-loss, improved-tanh-loss') parser.add_argument('--use_l2_norm', type=str, help='whether to add a l2 norm') parser.add_argument('--num_layers', type=int, help='num_layers to use. Minimum is 2') parser.add_argument('--input', type=str, help='Input file') parser.add_argument('--entity_type', type=str, help='people or companies') parser.add_argument('--model', type=str, help='name for model file') args = parser.parse_args() filepath = args.model LOSS_FUNCTION = None if args.loss_function == 'schroff-loss': LOSS_FUNCTION=schroff_triplet_loss elif args.loss_function == 'improved-loss': LOSS_FUNCTION=improved_loss elif args.loss_function == 'our-loss': LOSS_FUNCTION=triplet_loss elif args.loss_function == 'tanh-loss': LOSS_FUNCTION=triplet_tanh_loss elif args.loss_function == 'improved-tanh-loss': LOSS_FUNCTION=triplet_tanh_pn_loss elif args.loss_function == 'angular-loss': USE_ANGULAR_LOSS = True LOSS_FUNCTION = angular_loss print('Loss function: ' + args.loss_function) if args.debug_sample_size: DEBUG=True DEBUG_DATA_LENGTH=args.debug_sample_size print('Debug data length:' + str(DEBUG_DATA_LENGTH)) print('Margin:' + str(MARGIN)) USE_L2_NORM = args.use_l2_norm.lower() in ("yes", "true", "t", "1") print('Use L2Norm: ' + str(USE_L2_NORM)) print('Use L2Norm: ' + str(args.use_l2_norm)) NUM_LAYERS = args.num_layers - 1 print('Num layers: ' + str(NUM_LAYERS)) people = 'people' in args.entity_type # read all entities and create positive parts of a triplet entities = read_entities(args.input) train, test = split(entities, test_split = .20) print("TRAIN") print(train) print("TEST") print(test) entity2same_train = generate_names(train, people) entity2same_test = generate_names(test, people, limit_pairs=True) print(entity2same_train) print(entity2same_test) # change the default behavior of the tokenizer to ignore all punctuation except , - and . which are important # clues for entity names tokenizer = Tokenizer(num_words=MAX_NB_WORDS, lower=True, filters='!"#$%&()*+/:;<=>?@[\]^_`{|}~', split=" ") # build a set of data structures useful for annoy, the set of unique entities (unique_text), # a mapping of entities in texts to an index in unique_text, a mapping of entities to other same entities, and the actual # vectorized representation of the text. These structures will be used iteratively as we build up the model # so we need to create them once for re-use unique_text, entity2unique = build_unique_entities(entity2same_train) unique_text_test, entity2unique_test = build_unique_entities(entity2same_test) print("train text len:" + str(len(unique_text))) print("test text len:" + str(len(unique_text_test))) tokenizer.fit_on_texts(unique_text + unique_text_test) sequences = tokenizer.texts_to_sequences(unique_text) sequences = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH) sequences_test = tokenizer.texts_to_sequences(unique_text_test) sequences_test = pad_sequences(sequences_test, maxlen=MAX_SEQUENCE_LENGTH) # build models embedder = get_embedding_layer(tokenizer) model, test_positive_model, test_negative_model, inter_model = build_model(embedder) embedder_model = embedded_representation_model(embedder) if DEBUG_ANN: generate_triplets_from_ANN(embedder_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, True) sys.exit() test_data, test_match_stats = generate_triplets_from_ANN(embedder_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, False) test_seq = get_sequences(test_data, tokenizer) print("Test stats:" + str(test_match_stats)) counter = 0 current_model = embedder_model prev_match_stats = 0 train_data, match_stats = generate_triplets_from_ANN(current_model, sequences, entity2unique, entity2same_train, unique_text, False) print("Match stats:" + str(match_stats)) number_of_names = len(train_data['anchor']) # print(train_data['anchor']) print("number of names" + str(number_of_names)) Y_train = np.random.randint(2, size=(1,2,number_of_names)).T checkpoint = ModelCheckpoint(filepath, monitor='val_accuracy', verbose=1, save_best_only=True, mode='max') early_stop = EarlyStopping(monitor='val_accuracy', patience=1, mode='max') callbacks_list = [checkpoint, early_stop] train_seq = get_sequences(train_data, tokenizer) # check just for 5 epochs because this gets called many times model.fit([train_seq['anchor'], train_seq['positive'], train_seq['negative']], Y_train, epochs=100, batch_size=40, callbacks=callbacks_list, validation_split=0.2) current_model = inter_model # print some statistics on this epoch print("training data predictions") positives = test_positive_model.predict([train_seq['anchor'], train_seq['positive'], train_seq['negative']]) negatives = test_negative_model.predict([train_seq['anchor'], train_seq['positive'], train_seq['negative']]) print("f1score for train is: {}".format(f1score(positives, negatives))) print("test data predictions") positives = test_positive_model.predict([test_seq['anchor'], test_seq['positive'], test_seq['negative']]) negatives = test_negative_model.predict([test_seq['anchor'], test_seq['positive'], test_seq['negative']]) print("f1score for test is: {}".format(f1score(positives, negatives))) test_match_stats = generate_triplets_from_ANN(current_model, sequences_test, entity2unique_test, entity2same_test, unique_text_test, True) print("Test stats:" + str(test_match_stats))
21,235
36.061082
163
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
pyzor
pyzor-master/docs/conf.py
# -*- coding: utf-8 -*- # # Pyzor documentation build configuration file, created by # sphinx-quickstart on Sat Jun 7 15:20:07 2014. # # 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. import sys import os # 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. # sys.path.insert(0, os.path.abspath('.')) # -- 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.todo', 'sphinx.ext.coverage', 'sphinx.ext.mathjax', 'sphinx.ext.ifconfig', 'sphinx.ext.viewcode', ] # Add any paths that contain templates here, relative to this directory. templates_path = ['.templates'] # The suffix of source filenames. source_suffix = '.rst' # The encoding of source files. # source_encoding = 'utf-8-sig' # The master toctree document. master_doc = 'index' # General information about the project. project = u'Pyzor' copyright = u'2014, Frank Tobin' # 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. version = '1.0' # The full version, including alpha/beta/rc tags. release = '1.0' # The language for content autogenerated by Sphinx. Refer to documentation # for a list of supported languages. # language = None # There are two options for replacing |today|: either, you set today to some # non-false value, then it is used: #today = '' # Else, today_fmt is used as the format for a strftime call. #today_fmt = '%B %d, %Y' # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. exclude_patterns = ['.build'] # The reST default role (used for this markup: `text`) to use for all # documents. #default_role = None # If true, '()' will be appended to :func: etc. cross-reference text. #add_function_parentheses = True # If true, the current module name will be prepended to all description # unit titles (such as .. function::). #add_module_names = True # If true, sectionauthor and moduleauthor directives will be shown in the # output. They are ignored by default. #show_authors = False # The name of the Pygments (syntax highlighting) style to use. pygments_style = 'sphinx' # A list of ignored prefixes for module index sorting. #modindex_common_prefix = [] # If true, keep warnings as "system message" paragraphs in the built documents. #keep_warnings = 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 = 'default' # Theme options are theme-specific and customize the look and feel of a theme # further. For a list of options available for each theme, see the # documentation. #html_theme_options = {} # Add any paths that contain custom themes here, relative to this directory. #html_theme_path = [] # The name for this set of Sphinx documents. If None, it defaults to # "<project> v<release> documentation". #html_title = None # A shorter title for the navigation bar. Default is the same as html_title. #html_short_title = None # The name of an image file (relative to this directory) to place at the top # of the sidebar. #html_logo = None # The name of an image file (within the static path) to use as favicon of the # docs. This file should be a Windows icon file (.ico) being 16x16 or 32x32 # pixels large. #html_favicon = None # 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'] # Add any extra paths that contain custom files (such as robots.txt or # .htaccess) here, relative to this directory. These files are copied # directly to the root of the documentation. #html_extra_path = [] # If not '', a 'Last updated on:' timestamp is inserted at every page bottom, # using the given strftime format. #html_last_updated_fmt = '%b %d, %Y' # If true, SmartyPants will be used to convert quotes and dashes to # typographically correct entities. #html_use_smartypants = True # Custom sidebar templates, maps document names to template names. #html_sidebars = {} # Additional templates that should be rendered to pages, maps page names to # template names. #html_additional_pages = {} # If false, no module index is generated. #html_domain_indices = True # If false, no index is generated. #html_use_index = True # If true, the index is split into individual pages for each letter. #html_split_index = False # If true, links to the reST sources are added to the pages. #html_show_sourcelink = True # If true, "Created using Sphinx" is shown in the HTML footer. Default is True. #html_show_sphinx = True # If true, "(C) Copyright ..." is shown in the HTML footer. Default is True. html_show_copyright = False # If true, an OpenSearch description file will be output, and all pages will # contain a <link> tag referring to it. The value of this option must be the # base URL from which the finished HTML is served. #html_use_opensearch = '' # This is the file name suffix for HTML files (e.g. ".xhtml"). #html_file_suffix = None # Output file base name for HTML help builder. htmlhelp_basename = 'Pyzordoc' # -- 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': '', } # 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 = [ ('index', 'Pyzor.tex', u'Pyzor Documentation', u'Frank Tobin', 'manual'), ] # The name of an image file (relative to this directory) to place at the top of # the title page. #latex_logo = None # For "manual" documents, if this is true, then toplevel headings are parts, # not chapters. #latex_use_parts = False # If true, show page references after internal links. #latex_show_pagerefs = False # If true, show URL addresses after external links. #latex_show_urls = False # Documents to append as an appendix to all manuals. #latex_appendices = [] # If false, no module index is generated. #latex_domain_indices = True # -- Options for manual page output --------------------------------------- # One entry per manual page. List of tuples # (source start file, name, description, authors, manual section). man_pages = [ ('index', 'pyzor', u'Pyzor Documentation', [u'Frank Tobin'], 1) ] # If true, show URL addresses after external links. #man_show_urls = False # -- 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 = [ ('index', 'Pyzor', u'Pyzor Documentation', u'Frank Tobin', 'Pyzor', 'One line description of project.', 'Miscellaneous'), ] # Documents to append as an appendix to all manuals. #texinfo_appendices = [] # If false, no module index is generated. #texinfo_domain_indices = True # How to display URL addresses: 'footnote', 'no', or 'inline'. #texinfo_show_urls = 'footnote' # If true, do not generate a @detailmenu in the "Top" node's menu. #texinfo_no_detailmenu = False
8,310
30.244361
79
py
MINDER
MINDER-main/scripts/build_fm_index.py
# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import argparse import csv import logging import multiprocessing import re import ftfy import torch import tqdm import pickle from seal.index import FMIndex from datasets import load_dataset import json logger = logging.getLogger() logger.setLevel(logging.ERROR) def process(line): tokens = tokenize(line) return tokens def preprocess_file(input_path, labels, format="kilt", lowercase=False, tokenize=False, pid2query=None): if pid2query: with open(args.pid2query, 'rb') as f: pid2query = pickle.load(f) if args.id2code: with open(args.id2code, 'rb') as f: id2code = pickle.load(f) if format =="msmarco": f = load_dataset(input_path, split="train") pieces_it = ((pp['docid'], pp["title"], pp["text"]) for pp in f) pieces_it = tqdm.tqdm(pieces_it) for idx, title, text in pieces_it: idx = idx.strip() title = title.strip() text = re.sub(r"\s+", " ", text) text = ftfy.fix_text(text) text = text.replace("BULLET::::", "") text = text.replace("SECTION::::", "") text = text.strip() if args.include_query: if args.query_format == 'free': queries = pid2query[idx] query = '' for s in queries: query =query + " " + s if args.query_format == 'stable': queries = pid2query[idx] query = '' for s in queries: query =query + " || " + s + ' @@' if args.include_code: code = id2code[idx] code = " || " + ' '.join([*code.strip()]) + ' @@' if not text: continue if tokenize: title = " ".join(word_tokenize(title)) text = " ".join(word_tokenize(text)) title = f"{title} {args.delim}" if args.include_title and title: text = f"{title} {text}" if args.include_query and query: text = f"{text} {query}" if lowercase: text = text.lower() labels.append(idx) yield text else: with open(input_path, "r", 2**16) as f: if format == "dpr": next(f) pieces_it = csv.reader(f, delimiter="\t", quotechar='"') pieces_it = ((pp[0], pp[2], pp[1]) for pp in pieces_it if len(pp) == 3) elif format == "kilt": pieces_it = (line.strip() for line in f) pieces_it = (line.split("\t", 2) for line in pieces_it) pieces_it = ((pp[0], pp[1], pp[2]) for pp in pieces_it if len(pp) == 3) pieces_it = tqdm.tqdm(pieces_it) for idx, title, text in pieces_it: idx = idx.strip() title = title.strip() text = re.sub(r"\s+", " ", text) text = ftfy.fix_text(text) text = text.replace("BULLET::::", "") text = text.replace("SECTION::::", "") text = text.strip() if args.include_query: if args.query_format == 'free': queries = pid2query[idx] query = '' for s in queries: query =query + " " + s if args.query_format == 'stable': queries = pid2query[idx] query = '' for s in queries: query =query + " || " + s + ' @@' if args.include_code: code = id2code[idx] code = " || " + ' '.join([*code.strip()]) + ' @@' if not text: continue if tokenize: title = " ".join(word_tokenize(title)) text = " ".join(word_tokenize(text)) title = f"{title} {args.delim}" if args.include_title and title: text = f"{title} {text}" if args.include_query and query: text = f"{text} {query}" if args.include_code and code: text = f"{text} {code}" if lowercase: text = text.lower() labels.append(idx) yield text def build_index(input_path): labels = [] index = FMIndex() lines = preprocess_file(input_path, labels, args.format, lowercase=args.lowercase, tokenize=args.tokenize, pid2query=args.pid2query) print('start build index') with multiprocessing.Pool(args.jobs) as p: sequences = p.imap(process, lines) index.initialize(sequences) print('start build index 2') index.labels = labels return index def parse_args(): parser = argparse.ArgumentParser() parser.add_argument("input") parser.add_argument("output") parser.add_argument("--jobs", type=int, default=1) parser.add_argument("--include_title", action="store_true") parser.add_argument("--delim", default="@@") parser.add_argument("--format", choices=["kilt", "dpr", "msmarco"], default="kilt") parser.add_argument("--hf_model", default=None, type=str) parser.add_argument("--lowercase", action="store_true") parser.add_argument("--tokenize", action="store_true") parser.add_argument('--pid2query', default=None, type=str) parser.add_argument('--id2code', default=None, type=str) parser.add_argument("--include_query", action="store_true") parser.add_argument("--include_code", action="store_true") parser.add_argument("--query_format", choices=["free", "stable"], default="free") return parser.parse_args() if __name__ == "__main__": args = parse_args() print(args) if args.tokenize: from spacy.lang.en import English nlp = English() _tokenizer = nlp.tokenizer def word_tokenize(text): return [t.text.strip() for t in _tokenizer(text)] if args.hf_model is not None: from transformers import AutoTokenizer tokenizer = AutoTokenizer.from_pretrained(args.hf_model, use_fast=False) is_bart = "bart" in args.hf_model def tokenize(text): text = text.strip() if is_bart: text = " " + text with tokenizer.as_target_tokenizer(): return tokenizer(text, add_special_tokens=False)["input_ids"] + [tokenizer.eos_token_id] else: bart = torch.hub.load("pytorch/fairseq", "bart.large").eval() def tokenize(text): return bart.encode(" " + text.strip()).tolist()[1:] delim = tokenize(args.delim)[:-1] index = build_index(args.input) print('start build index 3') index.save(args.output)
7,216
29.974249
136
py
MINDER
MINDER-main/seal/keys.py
# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import math from collections import defaultdict from heapq import heappop, heappush from itertools import chain, islice, product from typing import * import torch from more_itertools import chunked from tqdm import tqdm from seal import FMIndex def deduplicate(list_of_lists): present = set() result = [] for el in list_of_lists: x = el if isinstance(el[0], float): el = el[1] if isinstance(el, torch.Tensor): t_el = tuple(el.tolist()) else: t_el = tuple(el) if t_el in present: continue else: present.add(t_el) result.append(x) return result def decompose_query_into_keys(query, word_tokenizer, length=3): strings = set() query = query.strip() tokens = word_tokenizer(query) tokens = [t.text for t in tokens] for i in range(len(tokens)): for j in range(i + 1, min(1 + len(tokens), i + length + 1)): span = tokens[i:j] for upper in product(*([[True, False]] * (j - i))): ss = [s[0].upper() + s[1:] if u else s for u, s in zip(upper, span)] ss = " " + " ".join(ss) strings.add(ss) strings = list(strings) return strings def strip(seq, symbols_start, symbols_end): i = 0 while i < len(seq) and seq[i] in symbols_start: i += 1 j = len(seq) while j > i and seq[j - 1] in symbols_end: j -= 1 return seq[i:j] @torch.inference_mode() def rescore_keys(model, inputs, list_of_decoded, batch_size=100, length_penalty=0.0, progress_bar=False, prefix=[], strip_from_bos=[], strip_from_eos=[]): device = next(model.parameters()).device if inputs is None: batch_in = [[model.config.bos_token_id, model.config.eos_token_id]] * len(list_of_decoded) else: batch_in = list(inputs) list_of_decoded = [[x[1] if isinstance(x[0], float) else x for x in xx] for xx in list_of_decoded] maxlen = max([len(i) for i in batch_in]) input_ids = [i + ([model.config.pad_token_id] * (maxlen - len(i))) for i in batch_in] input_ids = [torch.LongTensor(i).to(device) for i in input_ids] input_ids = torch.stack(input_ids, 0) attention_mask = input_ids != model.config.pad_token_id attention_mask = attention_mask.byte() encoder_outputs = model._prepare_encoder_decoder_kwargs_for_generation( input_ids, {'attention_mask': attention_mask})['encoder_outputs'].last_hidden_state decoder_inputs = enumerate(list_of_decoded) decoder_inputs = [(idx, di) for idx, ddi in decoder_inputs for di in ddi] all_out = {i: [] for i, _ in enumerate(list_of_decoded)} for batch in chunked(tqdm(decoder_inputs) if progress_bar else decoder_inputs, batch_size): idxs = [] batch_in_decoder_orig = [] batch_in_decoder = [] for i, di in batch: stripped = [model.config.decoder_start_token_id] + prefix + strip(di, strip_from_bos, strip_from_eos) if stripped: idxs.append(i) batch_in_decoder_orig.append(di) batch_in_decoder.append(stripped) batch_in_decoder = [torch.LongTensor(di) for di in batch_in_decoder] batch_in_decoder = [ torch.cat( [torch.LongTensor([model.config.decoder_start_token_id]), di] ) if di[0] != model.config.decoder_start_token_id else di for di in batch_in_decoder] maxlen = max([len(di) for di in batch_in_decoder]) batch_decoder_input_ids = [ torch.cat( [di, torch.LongTensor([model.config.pad_token_id] * (maxlen - len(di)))]) for di in batch_in_decoder] batch_decoder_input_ids = [di for di in batch_decoder_input_ids] batch_decoder_input_ids = torch.stack(batch_decoder_input_ids, 0).to(device) batch_input_ids = torch.stack([input_ids[idx] for idx in idxs], 0) batch_attention_mask = torch.stack([attention_mask[idx] for idx in idxs], 0) batch_encoder_outputs = torch.stack([encoder_outputs[idx] for idx in idxs], 0) logits = model( input_ids=batch_input_ids, attention_mask=batch_attention_mask, encoder_outputs=(batch_encoder_outputs, None, None), decoder_input_ids=batch_decoder_input_ids[:, :-1], ).logits logprobs = logits.log_softmax(-1) logprobs = torch.gather(logprobs, -1, batch_decoder_input_ids[:, 1:].unsqueeze(-1)) logprobs[batch_decoder_input_ids[:, 1:] < 2] = 0.0 logprobs = logprobs[:, len(prefix):] logprobs = logprobs.squeeze(-1).sum(-1) logprobs = logprobs.tolist() for i, di, bdi, ll in zip(idxs, batch_in_decoder_orig, batch_decoder_input_ids, logprobs): sco = ll / (len(di) ** length_penalty) all_out[i].append((sco, di)) return [v for k, v in sorted(all_out.items())] # @torch.inference_mode() @torch.no_grad() def compute_unigram_scores(model, inputs, index: FMIndex, tokenizer=None, tolist=True, temperature=1.0, prefix=[]): device = next(model.parameters()).device if isinstance(inputs[0], str): batch = tokenizer(inputs, padding=True, return_tensors='pt') else: batch_in = list(inputs) maxlen = max([len(i) for i in batch_in]) input_ids = [i + ([model.config.pad_token_id] * (maxlen - len(i))) for i in batch_in] input_ids = [torch.LongTensor(i).to(device) for i in input_ids] input_ids = torch.stack(input_ids, 0) attention_mask = input_ids != model.config.pad_token_id attention_mask = attention_mask.byte() batch = dict(input_ids=input_ids, attention_mask=attention_mask) batch = {k: v.to(device) for k, v in batch.items()} decoder_input_ids = torch.full_like(batch['input_ids'][:, :1 + len(prefix)], model.config.decoder_start_token_id) for i, idx in enumerate(prefix, start=1): decoder_input_ids[:, i] = idx logits = model(**batch, decoder_input_ids=decoder_input_ids).logits[:, 0 + len(prefix)] if temperature != 1.0: logits /= temperature logprobs = logits.log_softmax(-1) if tolist: return logprobs.tolist() else: return logprobs def aggregate_evidence(ngrams_and_scores: List[Tuple[List[int], float]], unigram_scores: Optional[List[float]] = None, index: Optional[FMIndex] = None, max_occurrences_1: int = 1500, max_occurrences_2: int = 10_000_000, n_docs_complete_score: int = 500, alpha: float = 2.0, beta: float = 0.8, length_penalty: float = 0.0, use_fm_index_frequency: bool = True, add_best_unigrams_to_ngrams: bool = False, use_top_k_unigrams=1000, sort_by_length=False, sort_by_freq=False, smoothing=5.0, allow_overlaps=False, single_key=0.0, single_key_add_unigrams=False, unigrams_ignore_free_places=False, tokenizer=None) -> Tuple[List[int], List[float]]: def repetition(ngram, score, coverage): if not coverage: return score ngram = set(ngram) coeff = 1.0 - beta + (beta * len(ngram.difference(coverage)) / len(ngram)) return coeff * score ntokens = float(index.beginnings[-1]) ngrams_and_scores = [(ngram.tolist() if isinstance(ngram, torch.Tensor) else ngram, sr) for ngram, sr in ngrams_and_scores] counts = {tuple(): len(index)} if not use_fm_index_frequency: try: cutoff = sorted(ngrams_and_scores, key=lambda x: x[1])[0][1] - 0.1 except IndexError as e: print(ngrams_and_scores) raise e else: cutoff = None unigrams = {0, 1, 2} for i in range(len(ngrams_and_scores)): ngram, sr = ngrams_and_scores[i] if ngram[0] == 45056: sr = min(sr+ 14, -1.0) # sr *= 0.1 if len(ngram) == 1: unigrams.add(ngram[0]) count = index.get_count(ngram) # print(tokenizer.decode(ngram),sr,count) # if count < 2000: # count = index.get_doc_count(ngram) # print(ngram, count) counts[tuple(ngram)] = count if count == 0: sco = 0.0 elif use_fm_index_frequency: sr -= 1e-10 sr *= (1.0 - length_penalty) ** (len(ngram) - 1.0) snr = math.log((count + smoothing) / (ntokens + smoothing)) sco = \ (sr + math.log(1 - math.exp(snr))) - \ (snr + math.log(1 - math.exp(sr))) sco = max(sco, 0.0) sco **= alpha else: sco = sr - cutoff sco = max(sco, 0.0) sco *= (1.0 - length_penalty) ** (len(ngram) - 1.0) sco **= alpha ngrams_and_scores[i] = (ngram, sco) # print(tokenizer.decode(ngram),'**', sco) if unigram_scores is not None: unigram_scores = unigram_scores[:] best = sorted(range(len(unigram_scores)), reverse=True, key=lambda i: unigram_scores[i]) best = best[:use_top_k_unigrams] best = set(best) unigram_scores = [s if i in best else float('-inf') for i, s in enumerate(unigram_scores)] for i in range(len(unigram_scores)): if i in unigrams: unigram_scores[i] = 0.0 continue sr = unigram_scores[i] ngram = [i] count = index.get_count(ngram) if count == 0: sco = 0.0 elif use_fm_index_frequency: snr = math.log((count + smoothing) / (ntokens + smoothing)) sco = \ (sr + math.log(1 - math.exp(snr))) - \ (snr + math.log(1 - math.exp(sr))) sco = max(sco, 0.0) else: sco = sr - cutoff sco = max(sco, 0.0) sco **= alpha if sco == 0.0: unigram_scores[i] = 0.0 continue unigram_scores[i] = sco if add_best_unigrams_to_ngrams: best_unigrams = sorted(list(range(len(unigram_scores))), key=lambda x: -unigram_scores[x])[:len(ngrams_and_scores)] for i in best_unigrams: counts[tuple([i])] = index.get_count([i]) ngrams_and_scores.append(([i], unigram_scores[i])) # rare ngrams (occurring less than max_hits) --> used for the first stage and full scoring rare_ngrams = defaultdict(float) # frequent ngrams --> used just for full scoring freq_ngrams = defaultdict(float) # computing scores for all ngrams for ngram, sco in ngrams_and_scores: count = index.get_count(ngram) if count > max_occurrences_2: continue elif sco == 0.0: continue elif count > max_occurrences_1 or sco < 0.0: ngrams = freq_ngrams # ngrams = rare_ngrams else: ngrams = rare_ngrams ngram = tuple(ngram) ngrams[ngram] = sco # else: rare_ngrams = {k: v for k, v in sorted(rare_ngrams.items(), key=lambda x: x[1], reverse=True)} # rare_ngrams = remove_redundant_ngrams(rare_ngrams) freq_ngrams = {k: v for k, v in sorted(freq_ngrams.items(), key=lambda x: x[1], reverse=True)} # freq_ngrams = remove_redundant_ngrams(freq_ngrams) all_ngrams = {k: v for k, v in \ sorted( chain(rare_ngrams.items(), freq_ngrams.items()), key=lambda x: x[1], reverse=True)} covered_points = set() first_stage = defaultdict(lambda: [0.0, [], [[], 0.0]]) for ngram, sco in rare_ngrams.items(): # idfs[ngram] = idf(ngram, index) # each ngram only considered once for doc doc_done = defaultdict(set) for row in islice(range(*index.get_range(list(ngram))), max_occurrences_1): tok_end = index.locate(row) tok_start = tok_end - len(ngram) doc = index.get_doc_index(tok_end) new = all([i not in covered_points for i in range(tok_start, tok_end)]) if sort_by_length: order = (len(ngram), sco) max_order = (len(first_stage[doc][2][0]), first_stage[doc][2][1]) elif sort_by_freq: order = (-counts[tuple(ngram)], sco) max_order = (-counts[tuple(first_stage[doc][2][0])], first_stage[doc][2][1]) else: order = sco max_order = first_stage[doc][2][1] if order > max_order: first_stage[doc][2] = [ngram, sco] if new: for tok in range(tok_start, tok_end): covered_points.add(tok) if new or allow_overlaps: if ngram not in doc_done[doc]: doc_done[doc].add(ngram) first_stage[doc][0] += sco first_stage[doc][1].append((ngram, sco)) for doc, doc_info in first_stage.items(): current_coverage = set() current_score = 0.0 for i in range(len(doc_info[1])): tt, sco = doc_info[1][i] tts = set(tt) new_sco = repetition(tts, sco, current_coverage) current_score += new_sco doc_info[1][i] = [tt, new_sco] current_coverage |= tts doc_info[0] = current_score to_fully_score = sorted(first_stage.items(), key=lambda x: (1.0 - single_key) * (-x[1][0]) + single_key * (-x[1][2][1]))[:n_docs_complete_score] results = defaultdict(lambda: [ 0.0, # score [], # ngrams found None, # places filled None, # full doc tokens [[], 0.0] # max ngram ]) trie = {} for ngram, score in all_ngrams.items(): if len(ngram) < 1 or score <= 0.0: continue current = trie for t in ngram: current = current.setdefault(t, {}) current[-1] = score for doc, _ in to_fully_score: doc_tokens = [2] + index.get_doc(doc)[:-1] results[doc][3] = doc_tokens if unigram_scores is not None: type_scores = {t: unigram_scores[t] for t in doc_tokens} else: type_scores = {t: 0.0 for t in doc_tokens} matches = {} open_matches = [] for i in range(len(doc_tokens)): open_matches = [(m.get(doc_tokens[i]), l + 1, n) for (m, l, n) in open_matches] + [ (trie.get(doc_tokens[i]), 1, [])] for _, _, n in open_matches: n.append(doc_tokens[i]) new_open_matches = [] while open_matches: m, l, n = open_matches.pop() if m is None: continue new_open_matches.append((m, l, n)) if -1 in m: start = i - l + 1 end = i + 1 matches.setdefault(tuple(n), [m[-1], []])[1].append((start, end)) open_matches = new_open_matches greedy_matches = [] for n, (s, d) in matches.items(): if sort_by_length: order = (-len(n), -s) max_order = (-len(results[doc][4][0]), -results[doc][4][1]) elif sort_by_freq: order = (counts[tuple(n)], -s) max_order = (counts[tuple(results[doc][4][0])], -results[doc][4][1]) else: order = -s max_order = -results[doc][4][1] for (i, j) in d: heappush(greedy_matches, (-s, n, s, i, j)) if order < max_order: results[doc][4] = [n, s] current_coverage = set() ngrams = [] prev = None f = 0 free = [True] * len(doc_tokens) while greedy_matches: order, n, s, i, j = heappop(greedy_matches) n_set = set(n) if prev == n: new_s = ngrams[-1][1] elif not n_set: new_s = 0.0 else: new_s = repetition(n_set, s, current_coverage) if new_s <= 0.0: continue if allow_overlaps or all(free[i:j]): pass else: continue if prev == n: f += 1 ngrams[-1] = (n, new_s) else: f = 1 prev = n current_coverage |= n_set ngrams.append((n, new_s)) free[i:j] = [False] * (j - i) if unigrams_ignore_free_places: free = [True for _ in free] single_key_score = results[doc][4][1] multi_key_score = sum([s for n, s in ngrams]) unigram_score = 0.0 for t, f in Counter([t for t, b in zip(doc_tokens, free) if b]).items(): s = type_scores[t] if s > 0.0: n = (t,) s = repetition(n, s, current_coverage) if s != 0.0: unigram_score += s ngrams.append((n, s)) if single_key_add_unigrams: single_key_score += unigram_score multi_key_score += unigram_score results[doc][0] = (1.0 - single_key) * multi_key_score + single_key * single_key_score results[doc][1] = ngrams results = {k: v for k, v in sorted(results.items(), key=lambda x: -x[1][0])} return results, all_ngrams ###################test for remove intersestive # if unigrams_ignore_free_places: # free = [True for _ in free] # single_key_score = results[doc][4][1] # multi_key_score = sum([s for n, s in ngrams]) # unigram_score = 0.0 # for t, f in Counter([t for t, b in zip(doc_tokens, free) if b]).items(): # s = type_scores[t] # if s > 0.0: # n = (t,) # s = repetition(n, s, current_coverage) # if s != 0.0: # unigram_score += s # ngrams.append((n, s)) # max_s = 0.0 # max_n = '' # for n, s in ngrams: # if s > max_s: # max_s = s # max_n = n # results[doc][0] = max_s # results[doc][1] = [(max_n,max_s)] # results = {k: v for k, v in sorted(results.items(), key=lambda x: -x[1][0])} # return results, all_ngrams
18,977
34.079482
138
py
MINDER
MINDER-main/seal/utils.py
# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import torch from torch import nn def _remove_ignore_keys_(state_dict): ignore_keys = [ "encoder.version", "decoder.version", "model.encoder.version", "model.decoder.version", "_float_tensor", "decoder.output_projection.weight", ] for k in ignore_keys: state_dict.pop(k, None) def _make_linear_from_emb(emb): vocab_size, emb_size = emb.weight.shape lin_layer = nn.Linear(vocab_size, emb_size, bias=False) lin_layer.weight.data = emb.weight.data return lin_layer def load_state_dict_from_lightning_checkpoint(model, path): state_dict = torch.load(path, map_location="cpu") # state_dict["shared.weight"] = state_dict["decoder.embed_tokens.weight"] # for key in ['shared.weight', 'encoder.embed_tokens.weight', 'decoder.embed_tokens.weight']: # state_dict[key] = torch.cat([state_dict[key], torch.zeros_like(state_dict[key][:1])], 0) # _remove_ignore_keys_(state_dict) # if hasattr(model, "lm_head"): # model.lm_head = _make_linear_from_emb(model.model.shared) model.load_state_dict(state_dict) def load_state_dict_from_fairseq_checkpoint(model, path): state_dict = torch.load(path, map_location="cpu")["model"] state_dict["shared.weight"] = state_dict["decoder.embed_tokens.weight"] for key in ["shared.weight", "encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]: state_dict[key] = torch.cat([state_dict[key], torch.zeros_like(state_dict[key][:1])], 0) _remove_ignore_keys_(state_dict) if hasattr(model, "lm_head"): model.lm_head = _make_linear_from_emb(model.model.shared) new_state_dict = {} for key in model.model.state_dict(): new_state_dict[key] = state_dict[key] model.model.load_state_dict(new_state_dict)
2,021
35.107143
102
py
MINDER
MINDER-main/seal/beam_search.py
# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. from collections import UserDict from typing import * import warnings from more_itertools import chunked import torch from torch import nn import torch.distributed as dist from transformers import LogitsProcessor, BeamScorer, BeamSearchScorer, LogitsProcessorList, StoppingCriteriaList, HammingDiversityLogitsProcessor from transformers.generation_utils import BeamSearchOutput, validate_stopping_criteria, BeamSearchEncoderDecoderOutput, BeamSearchDecoderOnlyOutput from transformers.generation_logits_process import TopKLogitsWarper from seal.index import FMIndex stopword_token_ids = [ 10, # a 41, # an 660, # An 5, # the 1941, # THE 20, # The 7, # to 6, # and ] class IndexBasedLogitsProcessor(LogitsProcessor): """ Class that masks logit, meant to be used during decoding. The logit mask is determined by finding the range of rows in the FM-index that correspond to the previously decoded token ( $O(n log V)$ ), then finding all tokens in that interval ( $O(V log V)$ ). """ def __init__( self, index: FMIndex, num_beams: int, pad_token_id: int = 0, eos_token_id: int = 2, force_decoding_from: Optional[List[int]] = None, stop_at_count: int = 0, always_allow_eos: bool = False, forced_bos_token_id: Optional[int] = None, ): self.index = index self.pad_token_id = pad_token_id self.eos_token_id = eos_token_id self._num_beams = num_beams self.log_odds_weight = 0.0 self.force_decoding_from = force_decoding_from self.force_decoding_second_token = None self.block_initial_stopwords = False self.stop_at_count = stop_at_count self.always_allow_eos = always_allow_eos self.forced_bos_token_id = forced_bos_token_id def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor: mask = torch.full_like(scores, float('-inf')) if self.forced_bos_token_id is not None: if input_ids.size(1) == 1: mask[:, self.forced_bos_token_id] = 0.0 return scores + mask # else: # input_ids = input_ids[:, 1:] if input_ids.size(1) == 1: distinct = self.index.occurring_distinct distinct = torch.LongTensor(distinct).to(scores.device) mask[:, distinct] = 0.0 else: input_ids_list = input_ids.view(-1, self._num_beams, input_ids.shape[-1]).tolist() lows = [] highs = [] fm_index_counts = [] for batch_id, beam_sent in enumerate(input_ids_list): for beam_id, sent in enumerate(beam_sent): if sent[-1] in (self.eos_token_id, self.pad_token_id): low = 0 high = 0 count = 0 elif self.force_decoding_from is not None: low, high = self.index.get_range(self.force_decoding_from + sent[1:]) count = self.index.get_count(self.force_decoding_from + sent[1:-1]) else: low, high = self.index.get_range(sent[1:]) count = self.index.get_count(sent[1:-1]) lows.append(low) highs.append(high) fm_index_counts.append(count) fm_index_result = self.index.get_distinct_count_multi(lows, highs) fm_index_result = fm_index_result[::-1] fm_index_counts = fm_index_counts[::-1] for batch_id, beam_sent in enumerate(input_ids_list): for beam_id, sent in enumerate(beam_sent): if self.stop_at_count > 0 and fm_index_counts[-1] <= self.stop_at_count: fm_index_result.pop() fm_index_counts.pop() distinct = [self.eos_token_id] elif sent[-1] == self.eos_token_id: fm_index_result.pop() fm_index_counts.pop() distinct = [self.pad_token_id] elif sent[-1] == self.pad_token_id: fm_index_result.pop() fm_index_counts.pop() distinct = [self.pad_token_id] else: fm_index_counts.pop() distinct, _ = fm_index_result.pop() distinct = torch.LongTensor(distinct).to(scores.device) mask[batch_id * self._num_beams + beam_id, distinct] = 0 if self.always_allow_eos: mask[:, self.eos_token_id] = 0.0 return scores + mask def constrained_beam_search( model, input_ids: torch.LongTensor, beam_scorer: BeamScorer, logits_processor: Optional[LogitsProcessorList] = None, constrained_decoding_processor: Optional[IndexBasedLogitsProcessor] = None, stopping_criteria: Optional[StoppingCriteriaList] = None, max_length: Optional[int] = None, pad_token_id: Optional[int] = None, eos_token_id: Optional[int] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_scores: Optional[bool] = None, return_dict_in_generate: Optional[bool] = None, synced_gpus: Optional[bool] = None, sample: bool = False, topk: int = 0, **model_kwargs, ) -> Union[BeamSearchOutput, torch.LongTensor]: if topk > 0: topk_warper = TopKLogitsWarper(topk) # init values logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList() stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList() if max_length is not None: warnings.warn( "`max_length` is deprecated in this function, use `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.", UserWarning, ) stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length) if len(stopping_criteria) == 0: warnings.warn("You don't have defined any stopping_criteria, this will likely loop forever", UserWarning) pad_token_id = pad_token_id if pad_token_id is not None else model.config.pad_token_id eos_token_id = eos_token_id if eos_token_id is not None else model.config.eos_token_id output_scores = output_scores if output_scores is not None else model.config.output_scores output_attentions = output_attentions if output_attentions is not None else model.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else model.config.output_hidden_states ) return_dict_in_generate = ( return_dict_in_generate if return_dict_in_generate is not None else model.config.return_dict_in_generate ) batch_size = len(beam_scorer._beam_hyps) num_beams = beam_scorer.num_beams batch_beam_size, cur_len = input_ids.shape if num_beams * batch_size != batch_beam_size: raise ValueError( f"Batch dimension of `input_ids` should be {num_beams * batch_size}, but is {batch_beam_size}." ) # init attention / hidden states / scores tuples scores = () if (return_dict_in_generate and output_scores) else None beam_indices = ( tuple(() for _ in range(batch_beam_size)) if (return_dict_in_generate and output_scores) else None ) decoder_attentions = () if (return_dict_in_generate and output_attentions) else None cross_attentions = () if (return_dict_in_generate and output_attentions) else None decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None # if model is an encoder-decoder, retrieve encoder attention weights and hidden states if return_dict_in_generate and model.config.is_encoder_decoder: encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None encoder_hidden_states = ( model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None ) beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device) beam_scores[:, 1:] = -1e9 beam_scores = beam_scores.view((batch_size * num_beams,)) this_peer_finished = False # used by synced_gpus only while True: if synced_gpus: # Under synced_gpus the `forward` call must continue until all gpus complete their sequence. # The following logic allows an early break if all peers finished generating their sequence this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device) # send 0.0 if we finished, 1.0 otherwise dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM) # did all peers finish? the reduced sum will be 0.0 then if this_peer_finished_flag.item() == 0.0: break model_inputs = model.prepare_inputs_for_generation(input_ids, **model_kwargs) outputs = model( **model_inputs, return_dict=True, output_attentions=output_attentions, output_hidden_states=output_hidden_states, ) if synced_gpus and this_peer_finished: cur_len = cur_len + 1 continue # don't waste resources running the code we don't need next_token_logits = outputs.logits[:, -1, :] # hack: adjust tokens for Marian. For Marian we have to make sure that the `pad_token_id` # cannot be generated both before and after the `nn.functional.log_softmax` operation. next_token_logits = model.adjust_logits_during_generation(next_token_logits, cur_len=cur_len) if topk: next_token_logits = topk_warper(input_ids, next_token_logits) next_token_scores = nn.functional.log_softmax( next_token_logits, dim=-1 ) # (batch_size * num_beams, vocab_size) next_token_scores_processed = logits_processor(input_ids, next_token_scores) next_token_scores_no_prev = next_token_scores_processed next_token_scores = next_token_scores_no_prev + beam_scores[:, None].expand_as(next_token_scores) if constrained_decoding_processor is not None: next_token_scores_constrained_no_prev = constrained_decoding_processor(input_ids, next_token_scores_processed) next_token_scores_constrained = next_token_scores_constrained_no_prev + beam_scores[:, None].expand_as(next_token_scores) # if return_masked_scores: # next_token_scores = next_token_scores_constrained else: next_token_scores_constrained_no_prev = next_token_scores_no_prev next_token_scores_constrained = next_token_scores # Store scores, attentions and hidden_states when required if return_dict_in_generate: if output_scores: scores += (next_token_scores_processed,) if output_attentions: decoder_attentions += ( (outputs.decoder_attentions,) if model.config.is_encoder_decoder else (outputs.attentions,) ) if model.config.is_encoder_decoder: cross_attentions += (outputs.cross_attentions,) if output_hidden_states: decoder_hidden_states += ( (outputs.decoder_hidden_states,) if model.config.is_encoder_decoder else (outputs.hidden_states,) ) # reshape for beam search vocab_size = next_token_scores.shape[-1] if sample: next_token_scores = next_token_scores.view(batch_size, num_beams * vocab_size) weights = next_token_scores_constrained_no_prev.view(batch_size * num_beams, vocab_size).log_softmax(-1).exp() nans = torch.isnan(weights.sum(-1)) weights[nans, :] = 0.0 weights[nans, eos_token_id] = 1.0 next_tokens = torch.multinomial(weights, 1, replacement=True).view(batch_size, 1 * num_beams) next_token_scores = next_token_scores.gather(-1, next_tokens) # next_token_scores = next_token_scores.reshape(batch_size, num_beams, 1) # next_token_scores[:, :, :] = 0.0 # next_token_scores = next_token_scores.reshape(batch_size, 1 * num_beams) else: next_token_scores = next_token_scores.view(batch_size, num_beams * vocab_size) next_token_scores_constrained = next_token_scores_constrained.view(batch_size, num_beams * vocab_size) next_token_scores_constrained, next_tokens = torch.topk( next_token_scores_constrained, 2 * num_beams, dim=1, largest=True, sorted=True ) next_token_scores = next_token_scores.gather(-1, next_tokens) next_indices = (next_tokens / vocab_size).long() next_tokens = next_tokens % vocab_size # stateless beam_outputs = beam_scorer.process( input_ids, next_token_scores, next_tokens, next_indices, pad_token_id=pad_token_id, eos_token_id=eos_token_id, ) beam_scores = beam_outputs["next_beam_scores"] beam_next_tokens = beam_outputs["next_beam_tokens"] beam_idx = beam_outputs["next_beam_indices"] input_ids = torch.cat([input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1) model_kwargs = model._update_model_kwargs_for_generation( outputs, model_kwargs, is_encoder_decoder=model.config.is_encoder_decoder ) if model_kwargs["past"] is not None: model_kwargs["past"] = model._reorder_cache(model_kwargs["past"], beam_idx) if return_dict_in_generate and output_scores: beam_indices = tuple((beam_indices[beam_idx[i]] + (beam_idx[i],) for i in range(len(beam_indices)))) # increase cur_len cur_len = cur_len + 1 if beam_scorer.is_done or stopping_criteria(input_ids, scores): if not synced_gpus: break else: this_peer_finished = True sequence_outputs = beam_scorer.finalize( input_ids, beam_scores, next_tokens, next_indices, pad_token_id=pad_token_id, eos_token_id=eos_token_id, max_length=stopping_criteria.max_length, ) if return_dict_in_generate: if not output_scores: sequence_outputs["sequence_scores"] = None else: num_return_sequences = beam_scorer.num_beam_hyps_to_keep # return only as many indices as sequences beam_indices = tuple( (beam_indices[i * num_beams : i * num_beams + num_return_sequences] for i in range(batch_size)) ) beam_indices = sum(beam_indices, ()) if model.config.is_encoder_decoder: return BeamSearchEncoderDecoderOutput( sequences=sequence_outputs["sequences"], sequences_scores=sequence_outputs["sequence_scores"], scores=scores, beam_indices=beam_indices, encoder_attentions=encoder_attentions, encoder_hidden_states=encoder_hidden_states, decoder_attentions=decoder_attentions, cross_attentions=cross_attentions, decoder_hidden_states=decoder_hidden_states, ) else: return BeamSearchDecoderOnlyOutput( sequences=sequence_outputs["sequences"], sequences_scores=sequence_outputs["sequence_scores"], scores=scores, beam_indices=beam_indices, attentions=decoder_attentions, hidden_states=decoder_hidden_states, ) else: return sequence_outputs["sequences"] @torch.inference_mode() def fm_index_generate( model, index: FMIndex, input_ids: torch.LongTensor, attention_mask: torch.LongTensor, min_length: int = 3, max_length: int = 25, length_penalty: float = 1.0, num_beams: int = 3, diverse_bs_groups: int = 1, diverse_bs_penalty: float = 0.0, eos_token_id: Optional[int] = None, force_decoding_from: Optional[List[int]] = None, always_allow_eos: bool = False, keep_history: bool = False, disable_fm_index: bool = False, sample: bool = False, stop_at_count: int = 0, topk: int = 0, transformers_output: bool = False, **kwargs, ): if 'forced_bos_token_id' in kwargs: forced_bos_token_id = kwargs.pop('forced_bos_token_id') else: forced_bos_token_id = model.config.forced_bos_token_id if sample: orig_num_beams = num_beams input_ids = input_ids.repeat(num_beams, 1) attention_mask = attention_mask.repeat(num_beams, 1) num_beams = 1 device = input_ids.device if eos_token_id is None: eos_token_id = model.config.eos_token_id logits_processor = model._get_logits_processor( encoder_input_ids=input_ids, repetition_penalty=None, no_repeat_ngram_size=0, encoder_no_repeat_ngram_size=0, bad_words_ids=None, min_length=min_length, max_length=max_length, eos_token_id=None, prefix_allowed_tokens_fn=None, forced_bos_token_id=forced_bos_token_id, forced_eos_token_id=None, num_beams=num_beams, num_beam_groups=1, diversity_penalty=0.0, remove_invalid_values=True) if diverse_bs_groups > 1 and diverse_bs_penalty > 0.0: logits_processor.append( HammingDiversityLogitsProcessor( diversity_penalty=diverse_bs_penalty, num_beams=num_beams, num_beam_groups=diverse_bs_groups, ) ) if not disable_fm_index: constrained_decoding_processor = IndexBasedLogitsProcessor( num_beams=num_beams // diverse_bs_groups, index=index, pad_token_id=model.config.pad_token_id, eos_token_id=eos_token_id or model.config.eos_token_id, force_decoding_from=force_decoding_from, stop_at_count=stop_at_count, always_allow_eos=always_allow_eos, forced_bos_token_id=forced_bos_token_id, ) if diverse_bs_groups > 1: logits_processor.append(constrained_decoding_processor) else: constrained_decoding_processor = None stopping_criteria = model._get_stopping_criteria( max_length=max_length, max_time=None, #max_new_tokens=None, #start_length=None ) model_kwargs = model._prepare_encoder_decoder_kwargs_for_generation( input_ids, {'attention_mask': attention_mask}) model_kwargs['use_cache'] = True decoder_input_ids = model._prepare_decoder_input_ids_for_generation( batch_size=input_ids.size(0), decoder_start_token_id=model.config.decoder_start_token_id, bos_token_id=model.config.bos_token_id, ) if keep_history: beam_scorer = BeamSearchScorerWithMemory( batch_size=decoder_input_ids.shape[0], num_beams=num_beams, device=device, length_penalty=length_penalty, do_early_stopping=False, num_beam_hyps_to_keep=num_beams, min_length=min_length, max_length=max_length, num_beam_groups=diverse_bs_groups, ) else: beam_scorer = BeamSearchScorer( batch_size=decoder_input_ids.shape[0], num_beams=num_beams, device=device, length_penalty=length_penalty, do_early_stopping=False, num_beam_hyps_to_keep=num_beams, num_beam_groups=diverse_bs_groups, ) decoder_input_ids, model_kwargs = model._expand_inputs_for_generation( decoder_input_ids, expand_size=num_beams, is_encoder_decoder=True, **model_kwargs) if diverse_bs_groups > 1: out = model.group_beam_search( input_ids=decoder_input_ids, beam_scorer=beam_scorer, logits_processor=logits_processor, stopping_criteria=stopping_criteria, output_scores=True, pad_token_id=model.config.pad_token_id, eos_token_id=eos_token_id, **model_kwargs) else: out = constrained_beam_search( model, input_ids=decoder_input_ids, beam_scorer=beam_scorer, logits_processor=logits_processor, constrained_decoding_processor=constrained_decoding_processor, stopping_criteria=stopping_criteria, output_scores=True, pad_token_id=model.config.pad_token_id, eos_token_id=eos_token_id, sample=sample, topk=topk, **model_kwargs) if transformers_output: return out if sample: out = [[(h[0] * h[1].size(0) ** length_penalty, h[1].tolist()) for b in bb for h in b.beams if h[0] > float('-inf')] for bb in chunked(beam_scorer._beam_hyps, orig_num_beams)] else: out = [[(h[0] * h[1].size(0) ** length_penalty, h[1].tolist()) for h in b.beams if h[0] > float('-inf')] for b in beam_scorer._beam_hyps] return out class BeamSearchScorerWithMemory(BeamScorer): def __init__( self, batch_size: int, num_beams: int, device: torch.device, length_penalty: Optional[float] = 1.0, do_early_stopping: Optional[bool] = False, num_beam_hyps_to_keep: Optional[int] = 1, num_beam_groups: Optional[int] = 1, min_length: Optional[int] = 15, max_length: Optional[int] = 25, **kwargs, ): self.num_beams = num_beams self.device = device self.length_penalty = length_penalty self.do_early_stopping = do_early_stopping self.num_beam_hyps_to_keep = num_beam_hyps_to_keep self.num_beam_groups = num_beam_groups self.group_size = self.num_beams // self.num_beam_groups self.min_length = min_length self.max_length = max_length self._is_init = False self._beam_hyps = [ BeamHypothesesWithMemory( num_beams=self.num_beams, length_penalty=self.length_penalty, early_stopping=self.do_early_stopping, min_length=self.min_length, max_length=self.max_length) for _ in range(batch_size) ] self._done = torch.tensor([False for _ in range(batch_size)], dtype=torch.bool, device=self.device) if not isinstance(num_beam_groups, int) or (num_beam_groups > num_beams) or (num_beams % num_beam_groups != 0): raise ValueError( f"`num_beam_groups` has to be an integer smaller or equal than `num_beams` and `num_beams` " f"has to be divisible by `num_beam_groups`, but is {num_beam_groups} with `num_beams` being {num_beams}." ) if "max_length" in kwargs: warnings.warn( "Passing `max_length` to BeamSearchScorer is deprecated and has no effect." "`max_length` should be passed directly to `beam_search(...)`, `beam_sample(...)`" ",or `group_beam_search(...)`." ) @property def is_done(self) -> bool: return self._done.all() def process( self, input_ids: torch.LongTensor, next_scores: torch.FloatTensor, next_tokens: torch.LongTensor, next_indices: torch.LongTensor, pad_token_id: Optional[int] = None, eos_token_id: Optional[int] = None, ) -> Tuple[torch.Tensor]: cur_len = input_ids.shape[-1] batch_size = len(self._beam_hyps) if not (batch_size == (input_ids.shape[0] // self.group_size)): if self.num_beam_groups > 1: raise ValueError( f"A group beam size of {input_ids.shape[0]} is used as the input, but a group beam " f"size of {self.group_size} is expected by the beam scorer." ) else: raise ValueError( f"A beam size of {input_ids.shape[0]} is used as the input, but a beam size of " f"{self.group_size} is expected by the beam scorer." ) device = input_ids.device next_beam_scores = torch.zeros((batch_size, self.group_size), dtype=next_scores.dtype, device=device) next_beam_tokens = torch.zeros((batch_size, self.group_size), dtype=next_tokens.dtype, device=device) next_beam_indices = torch.zeros((batch_size, self.group_size), dtype=next_indices.dtype, device=device) for batch_idx, beam_hyp in enumerate(self._beam_hyps): if self._done[batch_idx]: if self.num_beams < len(beam_hyp): raise ValueError(f"Batch can only be done if at least {self.num_beams} beams have been generated") if eos_token_id is None or pad_token_id is None: raise ValueError("Generated beams >= num_beams -> eos_token_id and pad_token have to be defined") # pad the batch next_beam_scores[batch_idx, :] = 0 next_beam_tokens[batch_idx, :] = pad_token_id next_beam_indices[batch_idx, :] = 0 continue # next tokens for this sentence beam_idx = 0 broken = False for beam_token_rank, (next_token, next_score, next_index) in enumerate( zip(next_tokens[batch_idx], next_scores[batch_idx], next_indices[batch_idx]) ): batch_beam_idx = batch_idx * self.group_size + next_index beam_hyp.add( torch.cat([ input_ids[batch_beam_idx], next_token.view(1), ]), next_score.item(), ) # add to generated hypotheses if end of sentence if broken: pass elif (eos_token_id is not None) and (next_token.item() == eos_token_id): pass else: # add next predicted token since it is not eos_token next_beam_scores[batch_idx, beam_idx] = next_score next_beam_tokens[batch_idx, beam_idx] = next_token next_beam_indices[batch_idx, beam_idx] = batch_beam_idx beam_idx += 1 # once the beam for next step is full, don't add more tokens to it. if beam_idx == self.group_size: broken = True if beam_idx < self.group_size: raise ValueError( f"At most {self.group_size} tokens in {next_tokens[batch_idx]} can be equal to `eos_token_id: {eos_token_id}`. Make sure {next_tokens[batch_idx]} are corrected." ) # Check if we are done so that we can save a pad step if all(done) self._done[batch_idx] = self._done[batch_idx] or beam_hyp.is_done( next_scores[batch_idx].max().item(), cur_len ) return UserDict( { "next_beam_scores": next_beam_scores.view(-1), "next_beam_tokens": next_beam_tokens.view(-1), "next_beam_indices": next_beam_indices.view(-1), } ) def finalize( self, input_ids: torch.LongTensor, final_beam_scores: torch.FloatTensor, final_beam_tokens: torch.LongTensor, final_beam_indices: torch.LongTensor, max_length: int, pad_token_id: Optional[int] = None, eos_token_id: Optional[int] = None, ) -> Tuple[torch.LongTensor]: batch_size = len(self._beam_hyps) for batch_idx, beam_hyp in enumerate(self._beam_hyps): for beam_id in range(self.num_beams): batch_beam_idx = batch_idx * self.num_beams + beam_id final_score = final_beam_scores[batch_beam_idx].item() final_tokens = input_ids[batch_beam_idx] beam_hyp.add( final_tokens.clone(), final_score) decoded: torch.LongTensor = input_ids.new(batch_size * self.num_beam_hyps_to_keep, 3) best_scores = torch.zeros(batch_size * self.num_beam_hyps_to_keep, device=self.device, dtype=torch.float32) return UserDict( { "sequences": decoded, "sequence_scores": best_scores, } ) class BeamHypothesesWithMemory: def __init__(self, num_beams: int, length_penalty: float, early_stopping: bool, min_length: int, max_length: int): self.length_penalty = length_penalty self.early_stopping = early_stopping self.num_beams = num_beams self.beams = [] self.worst_score = 1e9 self.min_length = min_length self.max_length = max_length self._best = None def __len__(self): return len(self.beams) def add(self, hyp: torch.LongTensor, sum_logprobs: float): size = hyp.size(0) score = sum_logprobs / (size ** self.length_penalty) self.beams.append((score, hyp)) def is_done(self, best_sum_logprobs: float, cur_len: int) -> bool: return cur_len >= self.max_length
31,228
40.090789
183
py
SRU_for_GCI
SRU_for_GCI-master/main.py
#!/usr/bin/env python # coding: utf-8 # Import header files import math import argparse import torch from torch import autograd import torch.nn as nn import torch.nn.functional as F import matplotlib import sys import numpy as np import pylab from matplotlib import pyplot as plt import time import sys from models.sru import SRU, trainSRU from models.eSRU_1LF import eSRU_1LF, train_eSRU_1LF from models.eSRU_2LF import eSRU_2LF, train_eSRU_2LF from utils.utilFuncs import env_config, loadTrainingData, loadTrueNetwork, getCausalNodes, count_parameters, getGeneTrainingData # Read input command line arguments parser = argparse.ArgumentParser() parser.add_argument('--device', type=str, default='cuda:3', help='device, default: cuda:3') parser.add_argument('--dataset', type=str, default='VAR', help='dataset type, default: VAR') parser.add_argument('--dsid', type=int, default=1, help='dataset id, default: 1') parser.add_argument('--T', type=int, default=10, help='training size, default: 10') parser.add_argument('--F', type=int, default=10, help='chaos, default: 10') parser.add_argument('--n', type=int, default=10, help='num of timeseries, default: 10') parser.add_argument('--model', type=str, default='sru', help='[sru, gru, lstm]: select your model') parser.add_argument('--nepochs', type=int, default=500, help='sets max_iter, default: 500') parser.add_argument('--mu1', type=float, default=1, help='sets mu1 parameter, default: 1') parser.add_argument('--mu2', type=float, default=1, help='sets mu2 parameter, default: 1') parser.add_argument('--mu3', type=float, default=1, help='sets mu3 parameter, default: 1') parser.add_argument('--lr', type=float, default=0.005, help='sets learning rate, default: 0.005') parser.add_argument('--joblog', type=str, default="", help='name of job logfile, default=""') args = parser.parse_args() deviceName = args.device model_name = args.model max_iter = args.nepochs mu1 = args.mu1 mu2 = args.mu2 mu3 = args.mu3 dataset = args.dataset dataset_id = args.dsid T = args.T F = args.F n = args.n lr = args.lr jobLogFilename = args.joblog ############################### # Global simulation settings ############################### verbose = 0 # Verbosity level ################################# # Pytorch environment ################################# device, seed = env_config(True, deviceName) # true --> use GPU print("Computational Resource: %s" % (device)) ###################################### # Create input data in batch format ###################################### if(dataset == 'gene'): Xtrain, Gref = getGeneTrainingData(dataset_id, device) n1 = Xtrain.shape[0] if(n != n1): print("Error::Dimension mismatch for input training data..") numTotalSamples = Xtrain.shape[1] Xtrain = Xtrain.float().to(device) # Make input signal zero mean and appropriately scaled Xtrain = Xtrain - Xtrain.mean() inputSignalMultiplier = 50 Xtrain = inputSignalMultiplier * Xtrain elif(dataset == 'var'): fileName = "data/var/S_%s_T_%s_dataset_%s.npz" % (F, T, dataset_id) ld = np.load(fileName) X_np = ld['X_np'] Gref = ld['Gref'] numTotalSamples = T Xtrain = torch.from_numpy(X_np) Xtrain = Xtrain.float().to(device) inputSignalMultiplier = 1 Xtrain = inputSignalMultiplier * Xtrain elif(dataset == 'lorenz'): fileName = "data/lorenz96/F_%s_T_%s_dataset_%s.npz" % (F, T, dataset_id) ld = np.load(fileName) X_np = ld['X_np'] Gref = ld['Gref'] numTotalSamples = T Xtrain = torch.from_numpy(X_np) Xtrain = Xtrain.float().to(device) inputSignalMultiplier = 1 Xtrain = inputSignalMultiplier * Xtrain elif(dataset == 'netsim'): fileName = "data/netsim/sim3_subject_%s.npz" % (dataset_id) ld = np.load(fileName) X_np = ld['X_np'] Gref = ld['Gref'] numTotalSamples = T Xtrain = torch.from_numpy(X_np) Xtrain = Xtrain.float().to(device) inputSignalMultiplier = 1 Xtrain = inputSignalMultiplier * Xtrain else: print("Dataset is not supported") if(verbose >= 1): plt.figure(1) plt.xlabel("t") plt.ylabel("x0(t)") plt.plot(range(numTotalSamples),Xtrain.cpu().numpy()[0][:]) plt.show(block=False) plt.pause(0.1) ###################################### # SRU Cell parameters ###################################### ####################################### # Model training parameters ###################################### if(model_name == 'sru'): lr_gamma = 0.99 lr_update_gap = 4 staggerTrainWin = 1 stoppingThresh = 1e-5; trainVerboseLvl = 2 lr = lr lambda1 = mu1 lambda2 = mu2 n_inp_channels = n n_out_channels = 1 if(dataset == 'gene'): A = [0.0, 0.01, 0.1, 0.5, 0.99]; #0.75 dim_iid_stats = 10 #math.ceil(n) #1.5n dim_rec_stats = 10 #math.ceil(n) #1.5n dim_final_stats = 10 #d * len(A) #math.ceil(n/2) dim_rec_stats_feedback = 10 #d * len(A) batchSize = 21 blk_size = batchSize numBatches = int(numTotalSamples/batchSize) elif(dataset == 'var'): A = [0.0, 0.01, 0.1, 0.99]; dim_iid_stats = 10 #math.ceil(n) #1.5n dim_rec_stats = 10 #math.ceil(n) #1.5n dim_final_stats = 10 #d * len(A) #math.ceil(n/2) #n dim_rec_stats_feedback = 10 #d * len(A) #math.ceil(n/2) #n batchSize = 250 blk_size = int(batchSize/2) numBatches = int(numTotalSamples/batchSize) elif(dataset == 'lorenz'): A = [0.0, 0.01, 0.1, 0.99]; dim_iid_stats = 10 dim_rec_stats = 10 dim_final_stats = 10 dim_rec_stats_feedback = 10 batchSize = 250 blk_size = int(batchSize/2) numBatches = int(numTotalSamples/batchSize) elif(dataset == 'netsim'): A = [0.0, 0.01, 0.05, 0.1, 0.99]; dim_iid_stats = 10 dim_rec_stats = 10 dim_final_stats = 10 dim_rec_stats_feedback = 10 batchSize = 10 #100 blk_size = int(batchSize/2) numBatches = int(numTotalSamples/batchSize) else: print("Unsupported dataset encountered") elif(model_name == 'eSRU_1LF' or model_name == 'eSRU_2LF'): lr_gamma = 0.99 lr_update_gap = 4 staggerTrainWin = 1 stoppingThresh = 1e-5; trainVerboseLvl = 2 lr = lr lambda1 = mu1 lambda2 = mu2 lambda3 = mu3 n_inp_channels = n n_out_channels = 1 if(dataset == 'gene'): A = [0.05, 0.1, 0.2, 0.99]; dim_iid_stats = 10 dim_rec_stats = 10 dim_final_stats = 10 dim_rec_stats_feedback = 10 batchSize = 21 blk_size = int(batchSize) numBatches = int(numTotalSamples/batchSize) elif(dataset == 'var'): A = [0.0, 0.01, 0.1, 0.99]; dim_iid_stats = 10 #math.ceil(n) #1.5n dim_rec_stats = 10 #math.ceil(n) #1.5n dim_final_stats = 10 #d * len(A) #math.ceil(n/2) #n dim_rec_stats_feedback = 10 #d * len(A) #math.ceil(n/2) #n batchSize = 250 blk_size = int(batchSize/2) numBatches = int(numTotalSamples/batchSize) elif(dataset == 'lorenz'): #lr = 0.01 A = [0.0, 0.01, 0.1, 0.99]; dim_iid_stats = 10 dim_rec_stats = 10 dim_final_stats = 10 #d*len(A) dim_rec_stats_feedback = 10 #d*len(A) batchSize = 250 blk_size = int(batchSize/2) numBatches = int(numTotalSamples/batchSize) elif(dataset == 'netsim'): A = [0.0, 0.01, 0.1, 0.99]; dim_iid_stats = 10 dim_rec_stats = 10 dim_final_stats = 10 #d*len(A) dim_rec_stats_feedback = 10 #d*len(A) batchSize = 10 #10 #100 blk_size = int(batchSize/2) numBatches = int(numTotalSamples/batchSize) else: print("Unsupported dataset encountered") else: print("Unsupported model encountered") ############################################ # Evaluate ROC plots (regress mu2) ############################################ if 1: Gest = torch.zeros(n, n, requires_grad = False) if(model_name == 'sru'): for predictedNode in range(n): start = time.time() print("node = %d" % (predictedNode)) model = SRU(n_inp_channels, n_out_channels, dim_iid_stats, dim_rec_stats, dim_rec_stats_feedback, dim_final_stats,A, device) model.to(device) # shift to CPU/GPU memory print(count_parameters(model)) model, lossVec = trainSRU(model, Xtrain, device, numBatches, batchSize, blk_size, predictedNode, max_iter, lambda1, lambda2, lr, lr_gamma, lr_update_gap, staggerTrainWin, stoppingThresh, trainVerboseLvl) Gest.data[predictedNode, :] = torch.norm(model.lin_xr2phi.weight.data[:,:n], p=2, dim=0) print("Elapsed time (1) = % s seconds" % (time.time() - start)) elif(model_name == 'eSRU_1LF'): for predictedNode in range(n): start = time.time() print("node = %d" % (predictedNode)) model = eSRU_1LF(n_inp_channels, n_out_channels, dim_iid_stats, dim_rec_stats, dim_rec_stats_feedback, dim_final_stats,A, device) model.to(device) # shift to CPU/GPU memory print(count_parameters(model)) model, lossVec = train_eSRU_1LF(model, Xtrain, device, numBatches, batchSize, blk_size, predictedNode, max_iter, lambda1, lambda2, lambda3, lr, lr_gamma, lr_update_gap, staggerTrainWin, stoppingThresh, trainVerboseLvl) Gest.data[predictedNode, :] = torch.norm(model.lin_xr2phi.weight.data[:,:n], p=2, dim=0) print("Elapsed time (1) = % s seconds" % (time.time() - start)) elif(model_name == 'eSRU_2LF'): for predictedNode in range(n): start = time.time() print("node = %d" % (predictedNode)) model = eSRU_2LF(n_inp_channels, n_out_channels, dim_iid_stats, dim_rec_stats, dim_rec_stats_feedback, dim_final_stats,A, device) model.to(device) # shift to CPU/GPU memory print(count_parameters(model)) model, lossVec = train_eSRU_2LF(model, Xtrain, device, numBatches, batchSize, blk_size, predictedNode, max_iter, lambda1, lambda2, lambda3, lr, lr_gamma, lr_update_gap, staggerTrainWin, stoppingThresh, trainVerboseLvl) Gest.data[predictedNode, :] = torch.norm(model.lin_xr2phi.weight.data[:,:n], p=2, dim=0) print("Elapsed time (1) = % s seconds" % (time.time() - start)) else: print("Unsupported model encountered") print(Gref) print(Gest) if(jobLogFilename != ""): if(model_name == 'eSRU_1LF' or model_name == 'eSRU_2LF'): np.savez(jobLogFilename, Gref=Gref, Gest=Gest.detach().cpu().numpy(), model=model_name, dataset=dataset, dsid=dataset_id, T=T, F=F, nepochs=max_iter, mu1=mu1, mu2=mu2, mu3=mu3, lr=lr, batchSize=batchSize, blk_size=blk_size, numBatches=numBatches, dim_iid_stats=dim_iid_stats, dim_rec_stats=dim_rec_stats, dim_final_stats=dim_final_stats, dim_rec_stats_feedback=dim_rec_stats_feedback) else: np.savez(jobLogFilename, Gref=Gref, Gest=Gest.detach().cpu().numpy(), model=model_name, dataset=dataset, dsid=dataset_id, T=T, F=F, nepochs=max_iter, mu1=mu1, mu2=mu2, lr=lr) # sleep for one seconds followed by printing # the exit key for tmux consumption time.sleep(1) print("#RUN_COMPLETE #RUN_COMPLETE #RUN_COMPLETE #RUN_COMPLETE")
12,432
33.72905
186
py
SRU_for_GCI
SRU_for_GCI-master/models/esru_2LF.py
import time import torch from torch import autograd import torch.nn as nn import torch.nn.functional as F import numpy as np import math from copy import deepcopy # Statistical Recurrent Unit class (based on paper by Junier B. Oliva, arXiv:1703.00381v1) class eSRU_2LF(torch.nn.Module): def __init__(self, n_inp_channels, # dimension of input sequence n_out_channels, # dimension of output (predicted) sequence dim_iid_stats, # dimension of iid statistics \phi dim_rec_stats, # dimension of recurrent stats u dim_rec_stats_feedback, # dimension of recurrent starts fed back as 'r' to generate iid stats dim_final_stats, # dimension of final stats u A, # Set of scales for exponentially weighted moving averages device # CPU/GPU memory for storing tensors ): # inherit the default attributes of Module class super(eSRU_2LF, self).__init__() # initialization of SRU parameters self.type = 'eSRU_2LF' self.n_inp_channels = n_inp_channels # dimension of input data self.n_out_channels = n_out_channels # dimension of predicted output self.dim_iid_stats = dim_iid_stats # dimension of 'phi_t' self.dim_rec_stats = dim_rec_stats # dimension of 'u_t' self.dim_final_stats = dim_final_stats # dimension of 'o_t' self.dim_rec_stats_feedback = dim_rec_stats_feedback # dimension of 'r_t' self.numScales = len(A) # Take kroneck product: A \otimes 1_{dim_iid_stats} self.A_mask = torch.Tensor([x for x in(A) for i in range(dim_iid_stats)]).view(1, -1) self.A_mask.requires_grad = False self.A_mask = self.A_mask.to(device) # shift to GPU memory # Initialization of SRU cell's tensors self.phi_t = torch.zeros(dim_iid_stats,1, requires_grad=True, device=device) self.phi_tile = torch.zeros(dim_iid_stats*self.numScales,1, requires_grad=True, device=device) self.r_t = torch.zeros(dim_rec_stats_feedback,1, requires_grad=True, device=device) self.o_t = torch.zeros(dim_final_stats,1, requires_grad=True, device=device) self.y_t = torch.zeros(n_out_channels,1, requires_grad=True, device=device) self.u_t = torch.zeros(1, dim_rec_stats * self.numScales, requires_grad=True, device=device) self.u_t_prev = torch.zeros(1, dim_rec_stats * self.numScales, device=device) # MLPs in SRU cell self.lin_xr2phi = nn.Linear(n_inp_channels + dim_rec_stats_feedback, dim_iid_stats, bias=True) self.lin_r1 = nn.Linear(dim_rec_stats_feedback, dim_rec_stats_feedback, bias=True) self.lin_r2 = nn.Linear(dim_rec_stats_feedback, dim_rec_stats_feedback, bias=True) self.lin_o = nn.Linear(self.numScales*dim_rec_stats, dim_final_stats, bias=True) self.lin_y = nn.Linear(dim_final_stats, n_out_channels, bias=True) # Fixed random matrices for sketching hidden state to lower dimensions self.intrMat_h2r_transpose = (1/math.sqrt(dim_rec_stats_feedback)) * torch.randn(self.numScales*dim_rec_stats, dim_rec_stats_feedback, requires_grad=False, device=device) # SRU forward pass def forward(self, x_t): # Generate feedback statistics self.r_t = torch.matmul(self.u_t_prev, self.intrMat_h2r_transpose) # sketch of hidden state self.r_t = F.elu(self.lin_r1(self.r_t)) # layer 1 self.r_t = F.elu(self.lin_r2(self.r_t)) # layer 2 # Generate iid statistics: phi_t self.phi_t = F.elu(self.lin_xr2phi(torch.cat((x_t, torch.flatten(self.r_t))))) # Generate multiscale recurrent statistics: u_t self.phi_tile = self.phi_t.repeat(1, self.numScales) self.u_t = torch.mul(self.A_mask, self.u_t_prev) + torch.mul((1-self.A_mask), self.phi_tile) self.u_t_prev.data = self.u_t.data # Generate final statistics: o_t self.o_t = F.elu(self.lin_o(self.u_t)) # Generate predicted output: y_t self.y_t = self.lin_y(self.o_t) return self.y_t def reset_recurrent_stats(self): self.u_t_prev.fill_(0) ############################################ # trainSRU_eSRU_2LF ############################################ def train_eSRU_2LF(model, trainingData, device, numBatches, batchSize, blk_size, predictedIdx, max_iter, lambda1, lambda2, lambda3, lr, lr_gamma, lr_update_gap, staggerTrainWin, stoppingThresh, verbose): stoppingCntr = 0 stoppingCntrThr = 10 n = trainingData.shape[0] numTotalSamples = trainingData.shape[1] wtMtxRow = torch.zeros(model.numScales * model.dim_final_stats, 1, requires_grad = False, device=device) lin_xr2phi_weight = deepcopy(model.lin_xr2phi.weight.data) lin_xr2phi_bias = deepcopy(model.lin_xr2phi.bias.data) lin_r1_weight = deepcopy(model.lin_r1.weight.data) lin_r1_bias = deepcopy(model.lin_r1.bias.data) lin_r2_weight = deepcopy(model.lin_r2.weight.data) lin_r2_bias = deepcopy(model.lin_r2.bias.data) lin_o_weight = deepcopy(model.lin_o.weight.data) lin_o_bias = deepcopy(model.lin_o.bias.data) lin_y_weight = deepcopy(model.lin_y.weight.data) lin_y_bias = deepcopy(model.lin_y.bias.data) ##################################### # Initialize miscellaneous tensors ##################################### IdxArr = torch.unsqueeze(torch.arange(1,n+1, dtype=torch.float),1) # 1 to n array for plotting purposes estWeights = torch.zeros(n, 1, requires_grad = False) prevWeights = torch.zeros(model.dim_iid_stats, n, requires_grad = False, device=device) lossVec = torch.zeros(max_iter,2) lossVec.to(device) mseLoss = nn.MSELoss(reduction = 'sum') L1Loss = nn.L1Loss(reduction = 'sum') softshrink1 = torch.nn.Softshrink(lambda1) optimizer = torch.optim.Adam(model.parameters(), lr=lr) scheduler = torch.optim.lr_scheduler.StepLR(optimizer, lr_update_gap, lr_gamma) batchCntr = 0 trainingLoss = 0 fitErr = 0 start_time = 0 stop_time = start_time + blk_size -1 optimizer.zero_grad() for epoch in range(max_iter): start1 = time.time() # Make deep copy of trainable model parameters to use later in checking the stopping criterion with torch.no_grad(): lin_xr2phi_weight[:,:] = model.lin_xr2phi.weight.data[:,:] lin_xr2phi_bias[:] = model.lin_xr2phi.bias.data[:] lin_r1_weight[:,:] = model.lin_r1.weight.data[:,:] lin_r1_bias[:] = model.lin_r1.bias.data[:] lin_r2_weight[:,:] = model.lin_r2.weight.data[:,:] lin_r2_bias[:] = model.lin_r2.bias.data[:] lin_o_weight[:,:] = model.lin_o.weight.data[:,:] lin_o_bias[:] = model.lin_o.bias.data[:] lin_y_weight[:,:] = model.lin_y.weight.data[:,:] lin_y_bias[:] = model.lin_y.bias.data[:] # Update start and stop times for next training batch printEpoch = 0 batchCntr = batchCntr + 1 if(batchCntr == numBatches+1): batchCntr = 1 trainingLoss = 0 fitErr = 0 # print epoch summary if(verbose > 0): printEpoch = 1 if(staggerTrainWin == 0): offset = 0 else: offset = math.floor(np.random.uniform()*(batchSize-blk_size)) start_time = (batchCntr-1)*batchSize + offset stop_time = start_time + blk_size - 1 # Reset recurrent stats u_t optimizer.zero_grad() model.reset_recurrent_stats() # Forward pass smooth_loss_list = [] for tt in range(start_time,stop_time,1): model.forward(trainingData[:,tt]) smooth_loss = (1/(blk_size-1))*mseLoss(torch.flatten(model.y_t), torch.unsqueeze(trainingData[predictedIdx,tt+1], 0)) smooth_loss_list.append(smooth_loss) #lossVec[epoch][0] = smooth_loss.item() # Use autograd to compute the backward pass (accumulate gradients on each pass). model.lin_xr2phi.weight.retain_grad() sum([smooth_loss_list[i] for i in range(blk_size-1)]).backward() lossVec[epoch][0] = sum([smooth_loss_list[i].item() for i in range(blk_size-1)]) #print("111: %s" % torch.cuda.memory_allocated(device)) optimizer.step() optimizer.zero_grad() #Adjust for regularization lr_current = optimizer.param_groups[0]['lr'] softshrink1 = nn.Softshrink(lambda1*lr) softshrink2 = nn.Softshrink(lambda2*lr) softshrink3 = nn.Softshrink(lambda3*lr) with torch.no_grad(): # Update all network parameters except for input layer weight matrix model.lin_xr2phi.weight[:,n:].data = softshrink1(model.lin_xr2phi.weight[:,n:]).data model.lin_xr2phi.bias.data = softshrink1(model.lin_xr2phi.bias).data model.lin_r1.weight.data = softshrink1(model.lin_r1.weight).data model.lin_r1.bias.data = softshrink1(model.lin_r1.bias).data model.lin_r2.weight.data = softshrink1(model.lin_r2.weight).data model.lin_r2.bias.data = softshrink1(model.lin_r2.bias).data #model.lin_o.weight.data = softshrink1(model.lin_o.weight).data model.lin_o.bias.data = softshrink1(model.lin_o.bias).data model.lin_y.weight.data = softshrink1(model.lin_y.weight).data model.lin_y.bias.data = softshrink1(model.lin_y.bias).data # Update input layer weight matrix inpWgtMtx = model.lin_xr2phi.weight[:,:n] l2normTensor = torch.norm(inpWgtMtx, p=2, dim=0, keepdim=True) # 1 x n row tensor #model.lin_xr2phi.weight.data[:,:n] = ((inpWgtMtx / torch.clamp(l2normTensor, min=(lambda2 * lr_current * 0.1))) # * torch.clamp(l2normTensor - (lr_current * lambda2), min=0.0)) model.lin_xr2phi.weight.data[:,:n] = inpWgtMtx*(softshrink2(l2normTensor)/torch.clamp(l2normTensor, min=lambda2*lr_current*0.1)) # Update the weight matrix mapping multi-time scale hidden state to # the lag sensitive features for prediction purpose for rr in range(model.dim_final_stats): wtMtxRow.data = model.lin_o.weight.data[rr,:] #reshape wtMtxRow as (numScales x dim_rec_stats) matrix wtMtxRowReshaped = wtMtxRow.view(model.numScales, model.dim_rec_stats) l2normTensor1 = torch.norm(wtMtxRowReshaped, p=2, dim=0, keepdim=True) # 1 x dim_final_stats row tensor model.lin_o.weight.data[rr,:] = (wtMtxRowReshaped*(softshrink3(l2normTensor1)/torch.clamp(l2normTensor1, min=lambda3*lr_current*0.1))).flatten().data[:] # Compute and log regularization loss without updating gradients loss1 = lambda1*((torch.norm(model.lin_y.weight.data, 1)+ torch.norm(model.lin_y.bias.data, 1) + torch.norm(model.lin_xr2phi.weight[:,n:].data, 1)) + torch.norm(model.lin_xr2phi.bias.data, 1) + torch.norm(model.lin_o.weight.data, 1) + torch.norm(model.lin_o.bias.data, 1) + torch.norm(model.lin_r1.weight.data, 1) + torch.norm(model.lin_r1.bias.data, 1) + torch.norm(model.lin_r2.weight.data, 1) + torch.norm(model.lin_r2.bias.data, 1)) lossVec[epoch][1] = lossVec[epoch][1] + loss1.item() loss2 = lambda2*torch.sum(torch.norm(model.lin_xr2phi.weight.data, p=2, dim=0)[:n]) lossVec[epoch][1] = lossVec[epoch][1] + loss2.item() # Again force gradient to be zero (just to be extra safe) optimizer.zero_grad() scheduler.step() # Record total-loss for current epoch lossVec[epoch][1] = lossVec[epoch][1] + lossVec[epoch][0] trainingLoss = trainingLoss + lossVec[epoch][1] fitErr = fitErr + lossVec[epoch][0] with torch.no_grad(): paramDelta = (mseLoss(model.lin_y.weight, lin_y_weight) + mseLoss(model.lin_y.bias, lin_y_bias) + mseLoss(model.lin_xr2phi.weight, lin_xr2phi_weight) + mseLoss(model.lin_xr2phi.bias, lin_xr2phi_bias) + mseLoss(model.lin_o.weight, lin_o_weight) + mseLoss(model.lin_o.bias, lin_o_bias) + mseLoss(model.lin_r1.weight, lin_r1_weight) + mseLoss(model.lin_r1.bias, lin_r1_bias) + mseLoss(model.lin_r2.weight, lin_r2_weight) + mseLoss(model.lin_r2.bias, lin_r2_bias)).data if(printEpoch == 1): print('Predicted Node = %d \t epoch = %s \t lr = %.4f \t Training loss = %.4f \t Fit error = %.4f \t Delta = %f' % (predictedIdx, epoch, optimizer.param_groups[0]['lr'], trainingLoss, fitErr, paramDelta)) #for col in range(n): # estWeights.data[col] = torch.norm(model.lin_xr2phi.weight.data[:,col], 2) #print(torch.cat((IdxArr, estWeights), 1)[:10]) #print(sruCell.lin_xr2phi.weight.grad.data[:,:n_inp_channels]) #print(optimizer.param_groups[0]['lr']*sruCell.lin_o.weight.grad.data[0,:]) #print(model.lin_o.weight.grad.data) #print(model.lin_xr2phi.weight.data[0,:]) #print(optimizer.param_groups[0]['lr']*model.lin_xr2phi.weight.grad.data[0,:]) #print(model.o_t.data) #print(model.lin_r.weight.data[0,:]) #print(model.lin_y.weight.data) #print("-------") # Stopping criterion if(paramDelta < stoppingThresh): stoppingCntr = stoppingCntr + 1 if(stoppingCntr == stoppingCntrThr): break else: stoppingCntr = 0 # run your code if(printEpoch == 1): print("Elapsed time (1) = % s seconds" % (time.time() - start1)) return model, lossVec
14,315
47.040268
216
py
SRU_for_GCI
SRU_for_GCI-master/models/esru_1LF.py
import time import torch from torch import autograd import torch.nn as nn import torch.nn.functional as F import numpy as np import math from copy import deepcopy # Statistical Recurrent Unit class (based on paper by Junier B. Oliva, arXiv:1703.00381v1) class eSRU_1LF(torch.nn.Module): def __init__(self, n_inp_channels, # dimension of input sequence n_out_channels, # dimension of output (predicted) sequence dim_iid_stats, # dimension of iid statistics \phi dim_rec_stats, # dimension of recurrent stats u dim_rec_stats_feedback, # dimension of recurrent starts fed back as 'r' to generate iid stats dim_final_stats, # dimension of final stats u A, # Set of scales for exponentially weighted moving averages device # CPU/GPU memory for storing tensors ): # inherit the default attributes of Module class super(eSRU_1LF, self).__init__() # initialization of SRU parameters self.type = 'eSRU_1LF' self.n_inp_channels = n_inp_channels # dimension of input data self.n_out_channels = n_out_channels # dimension of predicted output self.dim_iid_stats = dim_iid_stats # dimension of 'phi_t' self.dim_rec_stats = dim_rec_stats # dimension of 'u_t' self.dim_final_stats = dim_final_stats # dimension of 'o_t' self.dim_rec_stats_feedback = dim_rec_stats_feedback # dimension of 'r_t' self.numScales = len(A) # Take kroneck product: A \otimes 1_{dim_iid_stats} self.A_mask = torch.Tensor([x for x in(A) for i in range(dim_iid_stats)]).view(1, -1) self.A_mask.requires_grad = False self.A_mask = self.A_mask.to(device) # shift to GPU memory # Initialization of SRU cell's tensors self.phi_t = torch.zeros(dim_iid_stats,1, requires_grad=True, device=device) self.phi_tile = torch.zeros(dim_iid_stats*self.numScales,1, requires_grad=True, device=device) self.r_t = torch.zeros(dim_rec_stats_feedback,1, requires_grad=True, device=device) self.o_t = torch.zeros(dim_final_stats,1, requires_grad=True, device=device) self.y_t = torch.zeros(n_out_channels,1, requires_grad=True, device=device) self.u_t = torch.zeros(1, dim_rec_stats * self.numScales, requires_grad=True, device=device) self.u_t_prev = torch.zeros(1, dim_rec_stats * self.numScales, device=device) # MLPs in SRU cell self.lin_xr2phi = nn.Linear(n_inp_channels + dim_rec_stats_feedback, dim_iid_stats, bias=True) self.lin_r1 = nn.Linear(dim_rec_stats_feedback, dim_rec_stats_feedback, bias=True) self.lin_o = nn.Linear(self.numScales*dim_rec_stats, dim_final_stats, bias=True) self.lin_y = nn.Linear(dim_final_stats, n_out_channels, bias=True) # Fixed random matrices for sketching hidden state to lower dimensions self.intrMat_h2r_transpose = (1/math.sqrt(dim_rec_stats_feedback)) * torch.randn(self.numScales*dim_rec_stats, dim_rec_stats_feedback, requires_grad=False, device=device) # SRU forward pass def forward(self, x_t): # Generate feedback statistics self.r_t = torch.matmul(self.u_t_prev, self.intrMat_h2r_transpose) # sketch of hidden state self.r_t = F.elu(self.lin_r1(self.r_t)) # Generate iid statistics: phi_t self.phi_t = F.elu(self.lin_xr2phi(torch.cat((x_t, torch.flatten(self.r_t))))) # Generate multiscale recurrent statistics: u_t self.phi_tile = self.phi_t.repeat(1, self.numScales) self.u_t = torch.mul(self.A_mask, self.u_t_prev) + torch.mul((1-self.A_mask), self.phi_tile) self.u_t_prev.data = self.u_t.data # Generate final statistics: o_t self.o_t = F.elu(self.lin_o(self.u_t)) # Generate predicted output: y_t self.y_t = self.lin_y(self.o_t) return self.y_t def reset_recurrent_stats(self): self.u_t_prev.fill_(0) ############################################ # trainSRU ############################################ def train_eSRU_1LF(model, trainingData, device, numBatches, batchSize, blk_size, predictedIdx, max_iter, lambda1, lambda2, lambda3, lr, lr_gamma, lr_update_gap, staggerTrainWin, stoppingThresh, verbose): stoppingCntr = 0 stoppingCntrThr = 10 n = trainingData.shape[0] numTotalSamples = trainingData.shape[1] wtMtxRow = torch.zeros(model.numScales * model.dim_final_stats, 1, requires_grad = False, device=device) lin_xr2phi_weight = deepcopy(model.lin_xr2phi.weight.data) lin_xr2phi_bias = deepcopy(model.lin_xr2phi.bias.data) lin_r1_weight = deepcopy(model.lin_r1.weight.data) lin_r1_bias = deepcopy(model.lin_r1.bias.data) lin_o_weight = deepcopy(model.lin_o.weight.data) lin_o_bias = deepcopy(model.lin_o.bias.data) lin_y_weight = deepcopy(model.lin_y.weight.data) lin_y_bias = deepcopy(model.lin_y.bias.data) ##################################### # Initialize miscellaneous tensors ##################################### IdxArr = torch.unsqueeze(torch.arange(1,n+1, dtype=torch.float),1) # 1 to n array for plotting purposes estWeights = torch.zeros(n, 1, requires_grad = False) prevWeights = torch.zeros(model.dim_iid_stats, n, requires_grad = False, device=device) lossVec = torch.zeros(max_iter,2) lossVec.to(device) mseLoss = nn.MSELoss(reduction = 'sum') L1Loss = nn.L1Loss(reduction = 'sum') softshrink1 = torch.nn.Softshrink(lambda1) optimizer = torch.optim.Adam(model.parameters(), lr=lr) scheduler = torch.optim.lr_scheduler.StepLR(optimizer, lr_update_gap, lr_gamma) batchCntr = 0 trainingLoss = 0 fitErr = 0 start_time = 0 stop_time = start_time + blk_size -1 optimizer.zero_grad() for epoch in range(max_iter): start1 = time.time() # Make deep copy of trainable model parameters with torch.no_grad(): lin_xr2phi_weight[:,:] = model.lin_xr2phi.weight.data[:,:] lin_xr2phi_bias[:] = model.lin_xr2phi.bias.data[:] lin_r1_weight[:,:] = model.lin_r1.weight.data[:,:] lin_r1_bias[:] = model.lin_r1.bias.data[:] lin_o_weight[:,:] = deepcopy(model.lin_o.weight.data[:,:]) lin_o_bias[:] = deepcopy(model.lin_o.bias.data[:]) lin_y_weight[:,:] = deepcopy(model.lin_y.weight.data[:,:]) lin_y_bias[:] = deepcopy(model.lin_y.bias.data[:]) # Update start and stop times for next training batch printEpoch = 0 batchCntr = batchCntr + 1 if(batchCntr == numBatches+1): batchCntr = 1 trainingLoss = 0 fitErr = 0 # print epoch summary if(verbose > 0): printEpoch = 1 if(staggerTrainWin == 0): offset = 0 else: offset = math.floor(np.random.uniform()*(batchSize-blk_size)) start_time = (batchCntr-1)*batchSize + offset stop_time = start_time + blk_size - 1 # Reset recurrent stats u_t optimizer.zero_grad() model.reset_recurrent_stats() # Forward pass smooth_loss_list = [] for tt in range(start_time,stop_time,1): model.forward(trainingData[:,tt]) smooth_loss = (1/(blk_size-1))*mseLoss(torch.flatten(model.y_t), torch.unsqueeze(trainingData[predictedIdx,tt+1], 0)) smooth_loss_list.append(smooth_loss) #lossVec[epoch][0] = smooth_loss.item() # Use autograd to compute the backward pass (accumulate gradients on each pass). model.lin_xr2phi.weight.retain_grad() sum([smooth_loss_list[i] for i in range(blk_size-1)]).backward() lossVec[epoch][0] = sum([smooth_loss_list[i].item() for i in range(blk_size-1)]) #print("111: %s" % torch.cuda.memory_allocated(device)) optimizer.step() optimizer.zero_grad() #Adjust for regularization lr_current = optimizer.param_groups[0]['lr'] softshrink1 = nn.Softshrink(lambda1*lr) softshrink2 = nn.Softshrink(lambda2*lr) softshrink3 = nn.Softshrink(lambda3*lr) with torch.no_grad(): # Update all network parameters except for input layer weight matrix model.lin_xr2phi.weight[:,n:].data = softshrink1(model.lin_xr2phi.weight[:,n:]).data model.lin_xr2phi.bias.data = softshrink1(model.lin_xr2phi.bias).data model.lin_r1.weight.data = softshrink1(model.lin_r1.weight).data model.lin_r1.bias.data = softshrink1(model.lin_r1.bias).data #model.lin_o.weight.data = softshrink1(model.lin_o.weight).data model.lin_o.bias.data = softshrink1(model.lin_o.bias).data model.lin_y.weight.data = softshrink1(model.lin_y.weight).data model.lin_y.bias.data = softshrink1(model.lin_y.bias).data # Update input layer weight matrix inpWgtMtx = model.lin_xr2phi.weight[:,:n] l2normTensor = torch.norm(inpWgtMtx, p=2, dim=0, keepdim=True) # 1 x n row tensor #model.lin_xr2phi.weight.data[:,:n] = ((inpWgtMtx / torch.clamp(l2normTensor, min=(lambda2 * lr_current * 0.1))) # * torch.clamp(l2normTensor - (lr_current * lambda2), min=0.0)) model.lin_xr2phi.weight.data[:,:n] = inpWgtMtx*(softshrink2(l2normTensor)/torch.clamp(l2normTensor, min=lambda2*lr_current*0.1)) # Update the weight matrix mapping multi-time scale hidden state to # the lag sensitive features for prediction purpose for rr in range(model.dim_final_stats): wtMtxRow.data = model.lin_o.weight.data[rr,:] #reshape wtMtxRow as (numScales x dim_rec_stats) matrix wtMtxRowReshaped = wtMtxRow.view(model.numScales, model.dim_rec_stats) l2normTensor1 = torch.norm(wtMtxRowReshaped, p=2, dim=0, keepdim=True) # 1 x dim_final_stats row tensor model.lin_o.weight.data[rr,:] = (wtMtxRowReshaped*(softshrink3(l2normTensor1)/torch.clamp(l2normTensor1, min=lambda3*lr_current*0.1))).flatten().data[:] # Compute and log regularization loss without updating gradients loss1 = lambda1*((torch.norm(model.lin_y.weight.data, 1)+ torch.norm(model.lin_y.bias.data, 1) + torch.norm(model.lin_xr2phi.weight[:,n:].data, 1)) + torch.norm(model.lin_xr2phi.bias.data, 1) + torch.norm(model.lin_o.weight.data, 1) + torch.norm(model.lin_o.bias.data, 1) + torch.norm(model.lin_r1.weight.data, 1) + torch.norm(model.lin_r1.bias.data, 1)) lossVec[epoch][1] = lossVec[epoch][1] + loss1.item() loss2 = lambda2*torch.sum(torch.norm(model.lin_xr2phi.weight.data, p=2, dim=0)[:n]) lossVec[epoch][1] = lossVec[epoch][1] + loss2.item() # Again force gradient to be zero (just to be extra safe) optimizer.zero_grad() scheduler.step() # Record total-loss for current epoch lossVec[epoch][1] = lossVec[epoch][1] + lossVec[epoch][0] trainingLoss = trainingLoss + lossVec[epoch][1] fitErr = fitErr + lossVec[epoch][0] with torch.no_grad(): paramDelta = (mseLoss(model.lin_y.weight, lin_y_weight) + mseLoss(model.lin_y.bias, lin_y_bias) + mseLoss(model.lin_xr2phi.weight, lin_xr2phi_weight) + mseLoss(model.lin_xr2phi.bias, lin_xr2phi_bias) + mseLoss(model.lin_o.weight, lin_o_weight) + mseLoss(model.lin_o.bias, lin_o_bias) + mseLoss(model.lin_r1.weight, lin_r1_weight) + mseLoss(model.lin_r1.bias, lin_r1_bias)).data if(printEpoch == 1): print('Predicted Node = %d \t epoch = %s \t lr = %.4f \t Training loss = %.4f \t Fit error = %.4f \t Delta = %f' % (predictedIdx, epoch, optimizer.param_groups[0]['lr'], trainingLoss, fitErr, paramDelta)) #for col in range(n): # estWeights.data[col] = torch.norm(model.lin_xr2phi.weight.data[:,col], 2) #print(torch.cat((IdxArr, estWeights), 1)[:10]) #print(sruCell.lin_xr2phi.weight.grad.data[:,:n_inp_channels]) #print(optimizer.param_groups[0]['lr']*sruCell.lin_o.weight.grad.data[0,:]) #print(model.lin_o.weight.grad.data) #print(model.lin_xr2phi.weight.data[0,:]) #print(optimizer.param_groups[0]['lr']*model.lin_xr2phi.weight.grad.data[0,:]) #print(model.o_t.data) #print(model.lin_r.weight.data[0,:]) #print(model.lin_y.weight.data) #print("-------") # Stopping criterion if(paramDelta < stoppingThresh): stoppingCntr = stoppingCntr + 1 if(stoppingCntr == stoppingCntrThr): break else: stoppingCntr = 0 # run your code if(printEpoch == 1): print("Elapsed time (1) = % s seconds" % (time.time() - start1)) return model, lossVec
13,531
45.501718
216
py
SRU_for_GCI
SRU_for_GCI-master/models/sru.py
import time import torch from torch import autograd import torch.nn as nn import torch.nn.functional as F import numpy as np import math from copy import deepcopy # Statistical Recurrent Unit class (based on paper by Junier B. Oliva, arXiv:1703.00381v1) class SRU(torch.nn.Module): def __init__(self, n_inp_channels, # dimension of input sequence n_out_channels, # dimension of output (predicted) sequence dim_iid_stats, # dimension of iid statistics \phi dim_rec_stats, # dimension of recurrent stats u dim_rec_stats_feedback, # dimension of recurrent starts fed back as 'r' to generate iid stats dim_final_stats, # dimension of final stats u A, # Set of scales for exponentially weighted moving averages device # CPU/GPU memory for storing tensors ): # inherit the default attributes of Module class super(SRU, self).__init__() # initialization of SRU parameters self.type = 'sru' self.n_inp_channels = n_inp_channels # dimension of input data self.n_out_channels = n_out_channels # dimension of predicted output self.dim_iid_stats = dim_iid_stats # dimension of 'phi_t' self.dim_rec_stats = dim_rec_stats # dimension of 'u_t' self.dim_final_stats = dim_final_stats # dimension of 'o_t' self.dim_rec_stats_feedback = dim_rec_stats_feedback # dimension of 'r_t' self.numScales = len(A) # Take kroneck product: A \otimes 1_{dim_iid_stats} self.A_mask = torch.Tensor([x for x in(A) for i in range(dim_iid_stats)]).view(1, -1) self.A_mask.requires_grad = False self.A_mask = self.A_mask.to(device) # shift to GPU memory # Initialization of SRU cell's tensors self.phi_t = torch.zeros(dim_iid_stats,1, requires_grad=True, device=device) self.phi_tile = torch.zeros(dim_iid_stats*self.numScales,1, requires_grad=True, device=device) self.r_t = torch.zeros(dim_rec_stats_feedback,1, requires_grad=True, device=device) self.o_t = torch.zeros(dim_final_stats,1, requires_grad=True, device=device) self.y_t = torch.zeros(n_out_channels,1, requires_grad=True, device=device) self.u_t = torch.zeros(1, dim_rec_stats * self.numScales, requires_grad=True, device=device) self.u_t_prev = torch.zeros(1, dim_rec_stats * self.numScales, device=device) # MLPs in SRU cell self.lin_xr2phi = nn.Linear(n_inp_channels + dim_rec_stats_feedback, dim_iid_stats, bias=True) self.lin_r = nn.Linear(self.numScales*dim_rec_stats, dim_rec_stats_feedback, bias=True) self.lin_o = nn.Linear(self.numScales*dim_rec_stats, dim_final_stats, bias=True) self.lin_y = nn.Linear(dim_final_stats, n_out_channels, bias=True) #self.lin_xr2phi.weight.data.uniform_(-0.1,0.1) #self.lin_r.weight.data.uniform_(-0.1,0.1) #self.lin_o.weight.data.uniform_(-0.1,0.1) #self.lin_y.weight.data.uniform_(-0.1,0.1) # total number of parameteres self.numParams = ( (n_inp_channels + dim_rec_stats_feedback)*dim_iid_stats + self.numScales*dim_rec_stats*dim_rec_stats_feedback + self.numScales*dim_rec_stats*dim_final_stats + dim_final_stats*n_out_channels + dim_iid_stats + dim_rec_stats_feedback + dim_final_stats + n_out_channels) # SRU forward pass def forward(self, x_t): # Update r_t self.r_t = F.elu(self.lin_r(self.u_t_prev)) # Generate iid statistics: phi_t self.phi_t = F.elu(self.lin_xr2phi(torch.cat((x_t, torch.flatten(self.r_t))))) # Generate multiscale recurrent statistics: u_t self.phi_tile = self.phi_t.repeat(1, self.numScales) self.u_t = torch.mul(self.A_mask, self.u_t_prev) + torch.mul((1-self.A_mask), self.phi_tile) self.u_t_prev.data = self.u_t.data # Generate final statistics: o_t self.o_t = F.elu(self.lin_o(self.u_t)) # Generate predicted output: y_t self.y_t = self.lin_y(self.o_t) return self.y_t def reset_recurrent_stats(self): self.u_t_prev.fill_(0) ############################################ # trainSRU ############################################ def trainSRU(model, trainingData, device, numBatches, batchSize, blk_size, predictedIdx, max_iter, lambda1, lambda2, lr, lr_gamma, lr_update_gap, staggerTrainWin, stoppingThresh, verbose): stoppingCntr = 0 stoppingCntrThr = 10 proxUpdate = True n = trainingData.shape[0] numTotalSamples = trainingData.shape[1] lin_xr2phi_weight = deepcopy(model.lin_xr2phi.weight.data) lin_xr2phi_bias = deepcopy(model.lin_xr2phi.bias.data) lin_r_weight = deepcopy(model.lin_r.weight.data) lin_r_bias = deepcopy(model.lin_r.bias.data) lin_o_weight = deepcopy(model.lin_o.weight.data) lin_o_bias = deepcopy(model.lin_o.bias.data) lin_y_weight = deepcopy(model.lin_y.weight.data) lin_y_bias = deepcopy(model.lin_y.bias.data) ##################################### # Initialize miscellaneous tensors ##################################### IdxArr = torch.unsqueeze(torch.arange(1,n+1, dtype=torch.float),1) # 1 to n array for plotting purposes estWeights = torch.zeros(n, 1, requires_grad = False) prevWeights = torch.zeros(model.dim_iid_stats, n, requires_grad = False, device=device) lossVec = torch.zeros(max_iter,2) lossVec.to(device) mseLoss = nn.MSELoss(reduction = 'sum') L1Loss = nn.L1Loss(reduction = 'sum') if(proxUpdate): softshrink1 = torch.nn.Softshrink(lambda1) #hardshrink2 = nn.Hardshrink(hs2) optimizer = torch.optim.Adam(model.parameters(), lr=lr) scheduler = torch.optim.lr_scheduler.StepLR(optimizer, lr_update_gap, lr_gamma) batchCntr = 0 trainingLoss = 0 fitErr = 0 start_time = 0 stop_time = start_time + blk_size -1 optimizer.zero_grad() for epoch in range(max_iter): start1 = time.time() # Make deep copy of trainable model parameters with torch.no_grad(): lin_xr2phi_weight[:,:] = model.lin_xr2phi.weight.data[:,:] lin_xr2phi_bias[:] = model.lin_xr2phi.bias.data[:] lin_r_weight[:,:] = model.lin_r.weight.data[:,:] lin_r_bias[:] = model.lin_r.bias.data[:] lin_o_weight[:,:] = deepcopy(model.lin_o.weight.data[:,:]) lin_o_bias[:] = deepcopy(model.lin_o.bias.data[:]) lin_y_weight[:,:] = deepcopy(model.lin_y.weight.data[:,:]) lin_y_bias[:] = deepcopy(model.lin_y.bias.data[:]) # Update start and stop times for next training batch printEpoch = 0 batchCntr = batchCntr + 1 if(batchCntr == numBatches+1): batchCntr = 1 trainingLoss = 0 fitErr = 0 # print epoch summary if(verbose > 0): printEpoch = 1 if(staggerTrainWin == 0): offset = 0 else: offset = math.floor(np.random.uniform()*(batchSize-blk_size)) start_time = (batchCntr-1)*batchSize + offset stop_time = start_time + blk_size - 1 # Reset recurrent stats u_t optimizer.zero_grad() model.reset_recurrent_stats() # Forward pass smooth_loss_list = [] for tt in range(start_time,stop_time,1): model.forward(trainingData[:,tt]) smooth_loss = (1/(blk_size-1))*mseLoss(torch.flatten(model.y_t), torch.unsqueeze(trainingData[predictedIdx,tt+1], 0)) smooth_loss_list.append(smooth_loss) #lossVec[epoch][0] = smooth_loss.item() # Use autograd to compute the backward pass (accumulate gradients on each pass). model.lin_xr2phi.weight.retain_grad() sum([smooth_loss_list[i] for i in range(blk_size-1)]).backward() lossVec[epoch][0] = sum([smooth_loss_list[i].item() for i in range(blk_size-1)]) #smooth_loss.backward(retain_graph = True) #print("111: %s" % torch.cuda.memory_allocated(device)) # Compute gradient energy (without accounting for regularization terms) #total_grad_norm = 0; #for p in list(filter(lambda p: p.grad is not None, model.parameters())): # total_grad_norm = total_grad_norm + p.grad.data.norm(2).item() optimizer.step() optimizer.zero_grad() #Adjust for regularization if(proxUpdate): lr_current = optimizer.param_groups[0]['lr'] softshrink1 = nn.Softshrink(lambda1*lr) softshrink2 = nn.Softshrink(lambda2*lr) with torch.no_grad(): # Update all network parameters except for input layer weight matrix model.lin_xr2phi.weight[:,n:].data = softshrink1(model.lin_xr2phi.weight[:,n:]).data model.lin_xr2phi.bias.data = softshrink1(model.lin_xr2phi.bias).data model.lin_r.weight.data = softshrink1(model.lin_r.weight).data model.lin_r.bias.data = softshrink1(model.lin_r.bias).data model.lin_o.weight.data = softshrink1(model.lin_o.weight).data model.lin_o.bias.data = softshrink1(model.lin_o.bias).data model.lin_y.weight.data = softshrink1(model.lin_y.weight).data model.lin_y.bias.data = softshrink1(model.lin_y.bias).data # Update input layer weight matrix inpWgtMtx = model.lin_xr2phi.weight[:,:n] l2normTensor = torch.norm(inpWgtMtx, p=2, dim=0, keepdim=True) # 1 x n row tensor #model.lin_xr2phi.weight.data[:,:n] = ((inpWgtMtx / torch.clamp(l2normTensor, min=(lambda2 * lr_current * 0.1))) # * torch.clamp(l2normTensor - (lr_current * lambda2), min=0.0)) model.lin_xr2phi.weight.data[:,:n] = inpWgtMtx*(softshrink2(l2normTensor)/torch.clamp(l2normTensor, min=1e-8)) # Compute and log regularization loss without updating gadients loss1 = lambda1*((torch.norm(model.lin_y.weight.data, 1)+ torch.norm(model.lin_y.bias.data, 1) + torch.norm(model.lin_xr2phi.weight[:,n:].data, 1)) + torch.norm(model.lin_xr2phi.bias.data, 1) + torch.norm(model.lin_o.weight.data, 1) + torch.norm(model.lin_o.bias.data, 1) + torch.norm(model.lin_r.weight.data, 1) + torch.norm(model.lin_r.bias.data, 1)) lossVec[epoch][1] = lossVec[epoch][1] + loss1.item() loss2 = lambda2*torch.sum(torch.norm(model.lin_xr2phi.weight.data, p=2, dim=0)[:n]) lossVec[epoch][1] = lossVec[epoch][1] + loss2.item() # Again force gradient to be zero (just to be extra safe) optimizer.zero_grad() scheduler.step() else: loss1 = lambda1*((torch.norm(model.lin_y.weight, 1)+ torch.norm(model.lin_y.bias, 1) + torch.norm(model.lin_xr2phi.weight[:,n:], 1)) + torch.norm(model.lin_xr2phi.bias, 1) + torch.norm(model.lin_o.weight, 1) + torch.norm(model.lin_o.bias, 1) + torch.norm(model.lin_r.weight, 1) + torch.norm(model.lin_r.bias, 1)) lossVec[epoch][1] = lossVec[epoch][1] + loss1.item() model.lin_xr2phi.weight.retain_grad() loss1.backward(retain_graph = True) optimizer.step() optimizer.zero_grad() #for col in range(n): # loss2 = lambda2*torch.norm(model.lin_xr2phi.weight[:,col], 2) # model.lin_xr2phi.weight.retain_grad() # loss2.backward(retain_graph = True) # lossVec[epoch][1] = lossVec[epoch][1] + loss2.item() loss2 = lambda2*torch.sum(torch.norm(model.lin_xr2phi.weight, p=2, dim=0)[:n]) lossVec[epoch][1] = lossVec[epoch][1] + loss2.item() model.lin_xr2phi.weight.retain_grad() loss2.backward(retain_graph = True) optimizer.step() scheduler.step() # prune small weights #model.lin_xr2phi.weight.data[:,predictedIdx] = 0 #model.lin_xr2phi.weight.grad.data[:,predictedIdx] = 0 #model.lin_xr2phi.weight.data = hardshrink1(model.lin_xr2phi.weight.data) #model.lin_r.weight.data = hardshrink2(model.lin_r.weight.data) #model.lin_o.weight.data = hardshrink2(model.lin_o.weight.data) # Record total-loss for current epoch lossVec[epoch][1] = lossVec[epoch][1] + lossVec[epoch][0] trainingLoss = trainingLoss + lossVec[epoch][1] fitErr = fitErr + lossVec[epoch][0] with torch.no_grad(): paramDelta = (mseLoss(model.lin_y.weight, lin_y_weight) + mseLoss(model.lin_y.bias, lin_y_bias) + mseLoss(model.lin_xr2phi.weight, lin_xr2phi_weight) + mseLoss(model.lin_xr2phi.bias, lin_xr2phi_bias) + mseLoss(model.lin_o.weight, lin_o_weight) + mseLoss(model.lin_o.bias, lin_o_bias) + mseLoss(model.lin_r.weight, lin_r_weight) + mseLoss(model.lin_r.bias, lin_r_bias)).data if(printEpoch == 1): print('Predicted Node = %d \t epoch = %s \t lr = %.4f \t Training loss = %.4f \t Fit error = %.4f \t Delta = %f' % (predictedIdx, epoch, optimizer.param_groups[0]['lr'], trainingLoss, fitErr, paramDelta)) #for col in range(n): # estWeights.data[col] = torch.norm(model.lin_xr2phi.weight.data[:,col], 2) #print(torch.cat((IdxArr, estWeights), 1)[:10]) #print(sruCell.lin_xr2phi.weight.grad.data[:,:n_inp_channels]) #print(optimizer.param_groups[0]['lr']*sruCell.lin_o.weight.grad.data[0,:]) #print(model.lin_o.weight.grad.data) #print(model.lin_xr2phi.weight.data[0,:]) #print(optimizer.param_groups[0]['lr']*model.lin_xr2phi.weight.grad.data[0,:]) #print(model.o_t.data) #print(model.lin_r.weight.data[0,:]) #print(model.lin_y.weight.data) #print("-------") # Stopping criterion if(paramDelta < stoppingThresh): stoppingCntr = stoppingCntr + 1 if(stoppingCntr == stoppingCntrThr): break else: stoppingCntr = 0 # run your code if(printEpoch == 1): print("Elapsed time (1) = % s seconds" % (time.time() - start1)) return model, lossVec
15,047
45.018349
216
py
SRU_for_GCI
SRU_for_GCI-master/utils/utilFuncs.py
# Import header files import math import torch from torch import autograd import torch.nn as nn import torch.nn.functional as F import matplotlib import sys import numpy as np import pylab from matplotlib import pyplot as plt import time import sys import csv ########################################### # Python/numpy/pytorch environment config ########################################### def env_config(GPUTrue, deviceName): global_seed = 2 # Disable debug mode #torch.backends.cudnn.enabled=False torch.autograd.set_detect_anomaly(False) # Shrink very small values to zero in tensors for computational speedup torch.set_flush_denormal(True) # Set seed for random number generation (for reproducibility of results) torch.manual_seed(global_seed) torch.cuda.manual_seed(global_seed) np.random.seed(global_seed) # Set device as GPU if available, otherwise default to CPU if(GPUTrue): device = torch.device(deviceName if torch.cuda.is_available() else "cpu") else: device = torch.device("cpu") return device, global_seed ###################################### # Function for loading input data ###################################### def loadTrainingData(inputDataFilePath, device): # Load and parse input data (create batch data) inpData = torch.load(inputDataFilePath) Xtrain = torch.zeros(inpData['TsData'].shape[1], inpData['TsData'].shape[0], requires_grad = False, device=device) Xtrain1 = inpData['TsData'].t() Xtrain.data[:,:] = Xtrain1.data[:,:] return Xtrain ####################################################### # Function for reading ground truth network from file ####################################################### def loadTrueNetwork(inputFilePath, networkSize): with open(inputFilePath) as tsvin: reader = csv.reader(tsvin, delimiter='\t') numrows = 0 for row in reader: numrows = numrows + 1 network = np.zeros((numrows,2),dtype=np.int16) with open(inputFilePath) as tsvin: reader = csv.reader(tsvin, delimiter='\t') rowcounter = 0 for row in reader: network[rowcounter][0] = int(row[0][1:]) network[rowcounter][1] = int(row[1][1:]) rowcounter = rowcounter + 1 Gtrue = np.zeros((networkSize,networkSize), dtype=np.int16) for row in range(0,len(network),1): Gtrue[network[row][1]-1][network[row][0]-1] = 1 return Gtrue ############################################# # getCausalNodes ###################################### def getCausalNodes(model, threshold): n = model.n_inp_channels causalNodeMask = torch.zeros(n, 1, requires_grad = False, dtype=torch.int16) for col in range(n): #print(torch.norm(model.lin_xr2phi.weight.data[:,col],2)) if(torch.norm(model.lin_xr2phi.weight.data[:,col], 2) > threshold): causalNodeMask.data[col] = 1 return causalNodeMask ####################################################################### # Calculates false positive negatives and true positives negatives ##################################################################### def calcPerfMetrics(Gtrue, Gest): TP = 0 # True positive FP = 0 # False positive TN = 0 # True negative FN = 0 # False negative #n = Gest.shape[0] GTGE = (Gtrue * Gest) GestComplement = -1*(Gest-1) GtrueComplement = -1*(Gtrue-1) GTCGEC = (GtrueComplement * GestComplement) TP = np.sum(GTGE) FP = np.sum(Gest) - TP TN = np.sum(GTCGEC) FN = np.sum(GestComplement) - np.sum(GTCGEC) TPR = float(TP)/float(TP+FN) FPR = float(FP)/float(FP+TN) Recall = float(TP)/float(TP+FN) if(TP > 0 and FP > 0): Precision = float(TP)/float(TP+FP) else: Precision = 0 return TPR, FPR, Precision, Recall #################################################### # Calculates area under ROC curve # # (In) xin: numpy float array of false positive entries # (In) yin: numpy float array of true positive entries # (Out) auroc: calculated area under ROC curve # # Notes: xin and yin should sorted and be of same dimension # and contain bounded entries in (0,1) #################################################### def calcAUROC(xin, yin, verbose): xin, yin = parallel_sort(xin, yin) if(verbose > 0): for ii in range(len(xin)): print("%d\t %.6f \t %.6f" %(ii, xin[ii], yin[ii])) # Update input arrays to include extreme points (0,0) and (1,1) to the ROC plot xin = np.insert(xin,0,0) yin = np.insert(yin,0,0) xin = np.append(xin,1) yin = np.append(yin,1) n = len(xin) auroc = 0 for ii in range(n-1): h = xin[ii+1]-xin[ii] b1 = yin[ii] b2 = yin[ii+1] trapezoid_area = 0.5*h*(b1 + b2) auroc = auroc + trapezoid_area return auroc, xin, yin #################################################### # Calculates area under Precision-Recall curve # # (In) xin: numpy float array of precision values # (In) yin: numpy float array of recall values # (Out) aupr: calculated area under precision-recall curve # # Notes: xin and yin should sorted and be of same dimension # and contain bounded entries in (0,1) #################################################### def calcAUPR(xin, yin): ll = len(xin) # Update input arrays to include extreme points (0,1) and (1,0) to the precision-recall plot if(xin[0] > 0): xin = np.insert(xin,0,0) yin = np.insert(yin,0,1) if(xin[ll-1] < 1): xin = np.append(xin,1) yin = np.append(yin,0) n = len(xin) aupr = 0 for ii in range(n-1): h = xin[ii+1]-xin[ii] b1 = yin[ii] b2 = yin[ii+1] trapezoid_area = 0.5*h*(b1 + b2) aupr = aupr + trapezoid_area return aupr ########################### # Count the number of tunable parameters in the model ########################## def count_parameters(model): return sum(p.numel() for p in model.parameters() if p.requires_grad) ################################## # Calc metrics ################################## def calcMetrics(jobLogFilename, model, dataset, verbose): ld = np.load(jobLogFilename) Gest = ld['Gest'] Gref = ld['Gref'] model_name = ld['model'] dataset = ld['dataset'] dsid = ld['dsid'] nepochs = ld['nepochs'] T = ld['T'] F = ld['F'] mu1 = ld['mu1'] mu2 = ld['mu2'] lr = ld['lr'] # if esru2 model, then register esru2 specific parameters, namely mu3 if(model_name == 'esru2' or model_name == 'esru3'): mu3 = ld['mu3'] else: mu3 = 0 n = Gest.shape[0] Gest1 = np.ones((n,n), dtype=np.int16) thresh = 0 thresh_idx = (Gest <= thresh) Gest1.fill(1) Gest1[thresh_idx] = 0 # remove self loops for gene causal network estimate if(dataset == 'gene'): for ii in range(n): Gest1[ii][ii] = 0 #print(Gref) #print(Gest1) TPR, FPR, Precision, Recall = calcPerfMetrics(Gref, Gest1) if(verbose > 0): print("thresh = %1.4f, \t TPR = %1.4f, \t FPR = %1.4f, \t Precision = %.4f, \t Recall = %.4f" % (thresh, TPR, FPR, Precision, Recall)) return model_name, dataset, dsid, T, F, nepochs, lr, mu1, mu2, mu3, TPR, FPR, Precision, Recall ################################## # Calc metrics ################################## def calcMetricsTCDF(jobLogFilename, model, dataset, threshold, verbose): ld = np.load(jobLogFilename) Gest = ld['Gest'] Gref = ld['Gref'] model_name = ld['model'] dataset = ld['dataset'] dsid = ld['dsid'] nepochs = ld['nepochs'] T = ld['T'] F = ld['F'] nepochs = ld['nepochs'] kernel = ld['kernel_size'] level = ld['levels'] lr = ld['lr'] dilation = ld['dilation_c'] n = Gest.shape[0] Gest1 = np.ones((n,n), dtype=np.int16) thresh_idx = (Gest <= threshold) Gest1.fill(1) Gest1[thresh_idx] = 0 # remove self loops for gene causal network estimate if(dataset == 'gene'): for ii in range(n): Gest1[ii][ii] = 0 #print(Gref) #print(Gest1) TPR, FPR, Precision, Recall = calcPerfMetrics(Gref, Gest1) if(verbose > 0): print("thresh = %1.4f, \t TPR = %1.4f, \t FPR = %1.4f, \t Precision = %.4f, \t Recall = %.4f" % (thresh, TPR, FPR, Precision, Recall)) return model_name, dataset, dsid, T, F, nepochs, lr, kernel, level, dilation, TPR, FPR, Precision, Recall ################################################### # parallel sort in ascending order ################################################### def parallel_sort(xin, yin): n = len(xin) xin_sorted_idx = np.argsort(xin) yin_sorted_idx = np.argsort(yin) xout = xin[xin_sorted_idx] ysorted_by_x = yin[xin_sorted_idx] yout = yin #for ii in range(n): # print("%d\t %.4f \t %.4f" %(ii, xout[ii], ysorted_by_x[ii])) # for fixed xin[.], further sort yin[...] x_prev = xout[0] same_x_start_idx = 0 yout=[] for ii in range(0, n, 1): x = xout[ii] if((x > x_prev) or (ii == n-1)): if(ii == n-1): same_x_stop_idx = n-1 else: same_x_stop_idx = ii-1 if(same_x_start_idx == same_x_stop_idx): y_arr_for_same_x = ysorted_by_x[same_x_start_idx] else: y_arr_for_same_x = np.sort(ysorted_by_x[same_x_start_idx:same_x_stop_idx+1:1]) #print("%d, %d, %.4f" %(same_x_start_idx, same_x_stop_idx, x_prev)) #print(ysorted_by_x[same_x_start_idx:same_x_stop_idx+1:1]) #print(y_arr_for_same_x) yout = np.append(yout, y_arr_for_same_x) #print("%d, %d, %.4f" %(same_x_start_idx, same_x_stop_idx, x_prev)) same_x_start_idx = ii x_prev = xout[ii] return xout, yout def getGeneTrainingData(dataset_id, device): if(dataset_id == 1): InputDataFilePath = "data/dream3/Dream3TensorData/Size100Ecoli1.pt" RefNetworkFilePath = "data/dream3/TrueGeneNetworks/InSilicoSize100-Ecoli1.tsv" elif(dataset_id == 2): InputDataFilePath = "data/dream3/Dream3TensorData/Size100Ecoli2.pt" RefNetworkFilePath = "data/dream3/TrueGeneNetworks/InSilicoSize100-Ecoli2.tsv" elif(dataset_id == 3): InputDataFilePath = "data/dream3/Dream3TensorData/Size100Yeast1.pt" RefNetworkFilePath = "data/dream3/TrueGeneNetworks/InSilicoSize100-Yeast1.tsv" elif(dataset_id == 4): InputDataFilePath = "data/dream3/Dream3TensorData/Size100Yeast2.pt" RefNetworkFilePath = "data/dream3/TrueGeneNetworks/InSilicoSize100-Yeast2.tsv" elif(dataset_id == 5): InputDataFilePath = "data/dream3/Dream3TensorData/Size100Yeast3.pt" RefNetworkFilePath = "data/dream3/TrueGeneNetworks/InSilicoSize100-Yeast3.tsv" else: print("Error while loading gene training data") Xtrain = loadTrainingData(InputDataFilePath, device) n = Xtrain.shape[0] Gref = loadTrueNetwork(RefNetworkFilePath, n) return Xtrain, Gref
11,267
28.730871
142
py
SRU_for_GCI
SRU_for_GCI-master/utils/lorenz96Checker.py
import numpy as np import torch from utilFuncs import calcPerfMetrics, calcAUROC, calcAUPR # lorenz96 params T = 1000 F = 40.0 model_name = 'lstm' mu = 6.6 # F = 10, mu = 0.2| F = 40, mu = 4.0 n = 10 numDatasets = 5 max_iter = 500 verbose = 0 thresholdVec = np.arange(0, 1, 0.05) #thresholdVec = np.arange(0, 0.1, 0.001) TPRVec = np.zeros(len(thresholdVec), dtype=np.float32) FPRVec = np.zeros(len(thresholdVec), dtype=np.float32) RecallVec = np.zeros(len(thresholdVec), dtype=np.float32) PrecisionVec = np.zeros(len(thresholdVec), dtype=np.float32) Gest1 = np.ones((n,n), dtype=np.int16) Gest2 = np.ones((n,n), dtype=np.int16) Gref1 = np.zeros((n,n), dtype=np.int16) AUROCList = np.zeros(numDatasets) AUPRList = np.zeros(numDatasets) for dsid in range(numDatasets): filename = "../logs/lorenz96/LORENZ%s_T%s_F%s_%s_niter%s_mu_%s.pt" % (dsid+1, T, F, model_name, max_iter, mu) savedTensors = torch.load(filename) Gest = savedTensors['Gest'] Gref = savedTensors['Gref'] Gest.requires_grad = False Gest1 = Gest.cpu().numpy() Gest2.fill(1) for ii in range(len(thresholdVec)): thresh = thresholdVec[ii] thresh_idx = (Gest1 < thresh) Gest2[thresh_idx] = 0 TPR, FPR, Precision, Recall = calcPerfMetrics(Gref, Gest2) if(verbose > 0): print("thresh = %1.4f, \t TPR = %1.4f, \t FPR = %1.4f, \t Precision = %.4f, \t Recall = %.4f" % (thresh, TPR, FPR, Precision, Recall)) TPRVec[ii] = TPR FPRVec[ii] = FPR PrecisionVec[ii] = Precision RecallVec[ii] = Recall AUROCList[dsid] = calcAUROC(np.flip(FPRVec), np.flip(TPRVec)) AUPRList[dsid] = calcAUPR(RecallVec, PrecisionVec) print("%s_LORENZ%d_T%s_F%s: AUROC = %.4f, \t AUPR = %.4f" % (model_name, dsid, T, F, AUROCList[dsid], AUPRList[dsid])) print("Mean AUROC = %.4f, \t Mean AUPR = %.4f" % (AUROCList.mean(), AUPRList.mean()))
1,909
32.508772
146
py
SRU_for_GCI
SRU_for_GCI-master/utils/perfChk.py
import math import torch import matplotlib #import sys import numpy as np import pylab from matplotlib import pyplot as plt #import time from utilFuncs import loadTrueNetwork, getCausalNodes, calcPerfMetrics, calcAUROC, calcAUPR dataset = 'LORENZ' #dataset = 'VAR' #dataset = 'GENE' if(dataset == 'LORENZ'): dataset_id = 1 T = 1000 F = 40.0 model_name = 'sru' max_iter = 500 n = 10 thresh = 0.05 muVec = np.arange(1.0, 11.0, 1.0) #muVec = np.arange(18.0, 40.0, 1.0) TPRVec = np.zeros(len(muVec), dtype=np.float32) FPRVec = np.zeros(len(muVec), dtype=np.float32) RecallVec = np.zeros(len(muVec), dtype=np.float32) PrecisionVec = np.zeros(len(muVec), dtype=np.float32) Gest1 = np.zeros((n,n), dtype=np.int16) for muidx in range(len(muVec)): mu = muVec[muidx] LogPath = "logs/lorenz96/%s/%s%s_T%s_F%s_%s_niter%s_mu_%s.pt" % (model_name, dataset, dataset_id, T, F, model_name, max_iter, mu) savedTensors = torch.load(LogPath) Gref = savedTensors['Gref'] Gest = savedTensors['Gest'] Gest.requires_grad = False Gest1 = Gest.cpu().numpy() #print(Gest1) Gest1[Gest1 <= thresh] = 0 Gest1[Gest1 > 0] = 1 TPR, FPR, Precision, Recall = calcPerfMetrics(Gref, Gest1) print("mu = %.4f, \t thresh = %1.4f, \t TPR = %1.4f, \t FPR = %1.4f, \t Precision = %.4f, \t Recall = %.4f" % (mu, thresh, TPR, FPR, Precision, Recall)) TPRVec[muidx] = TPR FPRVec[muidx] = FPR PrecisionVec[muidx] = Precision RecallVec[muidx] = Recall AUROC, FPRVec, TPRVec = calcAUROC(np.flip(FPRVec), np.flip(TPRVec)) AUPR = calcAUPR(RecallVec, PrecisionVec) print("AUROC = %.4f, \t AUPR = %.4f" % (AUROC, AUPR)) plt.figure(1) plt.title("ROC (TPR vs FPR)") plt.xlabel("False positive rate") plt.ylabel("True positive rate") plt.xlim(0,1) plt.ylim(0,1) plt.plot(FPRVec, TPRVec) plt.show() plt.figure(2) plt.title("ROC (Precision vs Recall)") plt.xlabel("Recall") plt.ylabel("Precision") plt.xlim(0,1) plt.ylim(0,1) plt.plot(RecallVec, PrecisionVec) plt.show() else: print("Dataset is not supported") ########################## # old stuff ########################### if 0: #LogPath = "logs/sru_niter1000_mu_2.0.pt" #LogPath = "logs/sru_mod_niter1000_mu_1.0.pt" #LogPath = "logs/lstm_niter200_mu_1.0.pt" #LogPath = "logs/lstm_niter200_mu_1.0.pt" #LogPath = "LORENZ_mlp_niter50000_mu_1.pt" InputDataFilePath = "Dream3TensorData/Size100Ecoli1.pt" RefNetworkFilePath = "Dream3TrueGeneNetworks/InSilicoSize100-Ecoli1.tsv" n = 100; Gref = loadTrueNetwork(RefNetworkFilePath, n) savedTensors = torch.load(LogPath) Gest = savedTensors['Gest'] print(Gest) Gest1 = torch.zeros(n,n, requires_grad = False, dtype=torch.int16) #thresholdVec = np.arange(0, 0.2, 0.001) thresholdVec = np.arange(0, 2, 0.05) TPRVec = np.zeros(len(thresholdVec), dtype=np.float32) FPRVec = np.zeros(len(thresholdVec), dtype=np.float32) RecallVec = np.zeros(len(thresholdVec), dtype=np.float32) PrecisionVec = np.zeros(len(thresholdVec), dtype=np.float32) #ignore self loops for ii in range(n): Gest.data[ii][ii] = 0 for ii in range(len(thresholdVec)): thresh = thresholdVec[ii] for rr in range(n): for cc in range(n): Gest1.data[rr][cc] = Gest.data[rr][cc] > thresh TPR, FPR, Precision, Recall = calcPerfMetrics(Gref, Gest1) print("thresh = %1.4f, \t TPR = %1.4f, \t FPR = %1.4f, \t Precision = %.4f, \t Recall = %.4f" % (thresh, TPR, FPR, Precision, Recall)) TPRVec[ii] = TPR FPRVec[ii] = FPR PrecisionVec[ii] = Precision RecallVec[ii] = Recall
3,856
30.357724
160
py
Quantized-GBDT
Quantized-GBDT-master/experiments/generate_script.py
import os import argparse arg_parser = argparse.ArgumentParser() arg_parser.add_argument("data_path", type=str) arg_parser.add_argument("--use-discretized-grad", action='store_true') arg_parser.add_argument("--discretized-grad-renew", action='store_true') arg_parser.add_argument("--stochastic-rounding", action='store_true') arg_parser.add_argument("--for-speed", action='store_true') arg_parser.add_argument("--device", type=str, default='cpu') arg_parser.add_argument("--algorithm", type=str, default='lgb') arg_parser.add_argument("--log-path", type=str, default='logs') script_fname = 'run.sh' running = open(script_fname, 'w') os.system(f"chmod +x {script_fname}") data = [ 'higgs', 'epsilon', 'criteo', 'bosch', 'kitsune', 'yahoo', 'msltr', 'year' ] task = [ 'binary', 'binary', 'binary', 'binary', 'binary', 'ranking', 'ranking', 'regression' ] col_wise_data = ['epsilon', 'year', 'yahoo', 'bosch'] bins = [2, 3, 4, 5] def generate_script(data_path, use_discretized_grad, discretized_grad_renew, stochastic_rounding, for_speed, device, algorithm, log_dir): data_path = data_path.rstrip('/') dataset = [ f'data={data_path}/higgs.train', f'data={data_path}/epsilon.train', f'data={data_path}/criteo.train', f'data={data_path}/bosch.train', f'data={data_path}/kitsune.train', f'data={data_path}/yahoo.train', f'data={data_path}/msltr.train', f'data={data_path}/year.train', ] validset = [ f'valid={data_path}/higgs.test', f'valid={data_path}/epsilon.test', f'valid={data_path}/criteo.test', f'valid={data_path}/bosch.test', f'valid={data_path}/kitsune.test', f'valid={data_path}/yahoo.test', f'valid={data_path}/msltr.test', f'valid={data_path}/year.test' ] os.system(f"mkdir -p {log_dir}") if algorithm == 'lgb': use_discretized_grad_str = str(use_discretized_grad).lower() discretized_grad_renew_str = str(discretized_grad_renew).lower() stochastic_rounding_str = str(stochastic_rounding).lower() num_k = 4 if use_discretized_grad else 1 for i in range(8): for j in range(5): for k in range(num_k): base_conf_fname = 'train_model.conf' if task[i] == 'binary' else ('train_rank_model.conf' if task[i] == 'ranking' else 'train_reg_model.conf') args = '' args += dataset[i] if not for_speed: args += ' ' + validset[i] args += ' seed=' + str(j) if use_discretized_grad: args += ' grad_discretize_bins='+str(2**bins[k]-2) log_name = f'./{log_dir}/train_' + data[i] + '_seed'+str(j) + '_bins' + str(bins[k])+'.log' else: log_name = f'./{log_dir}/train_' + data[i] + '_seed'+str(j)+ '_fp32' + '.log' args += f' use_discretized_grad={use_discretized_grad_str} discretized_grad_renew={discretized_grad_renew_str} stochastic_rounding={stochastic_rounding_str}' if data[i] == 'bosch': args += ' learning_rate=0.015 num_leaves=45' if data[i] in col_wise_data: args += ' force_row_wise=false force_col_wise=true' if device != 'cpu': args += f' device_type=cuda gpu_device_id=0 num_threads=24' if not use_discretized_grad: running.write(f'../LightGBM-master/lightgbm config={base_conf_fname} {args} > {log_name} 2>&1\n') else: running.write(f'../LightGBM/lightgbm config={base_conf_fname} {args} > {log_name} 2>&1\n') elif algorithm == 'xgb': for i in range(8): for j in range(5): log_name = f'./{log_dir}/train_' + data[i] + '_seed'+str(j)+ '_xgb' + '.log' base_conf_fname = 'xgboost.conf' args = '' args += dataset[i] if task[i] == 'ranking': args += '.xgb?format=libsvm' args += ' seed=' + str(j) if not for_speed: args += ' ' + validset[i].replace('valid=', 'eval[test]=') if task[i] == 'ranking': args += '.xgb?format=libsvm' metric = 'auc' if task[i] == 'binary' else ('rmse' if task[i] == 'regression' else 'ndcg@10') args += f' eval_metric={metric}' objective = 'binary:logistic' if task[i] == 'binary' else ('reg:linear' if task[i] == 'regression' else 'rank:pairwise') args += f' objective={objective}' if data[i] == 'bosch': args += ' eta=0.015 max_leaves=45' # max_leaves=45 for xgboost to reduce time cost for post pruning if device != 'cpu': args += ' tree_method=gpu_hist nthread=24' running.write(f'../xgboost/xgboost {base_conf_fname} {args} > {log_name} 2>&1\n') elif algorithm == 'cat': for i in range(8): for j in range(5): log_name = f'./{log_dir}/train_' + data[i] + '_seed'+str(j)+ '_cat' + '.log' base_conf_fname = 'catboost.json' args = '' args += f"--params-file {base_conf_fname}" if data[i] == 'bosch': args += " --learning-rate 0.015 --max-leaves 45" data_path_prefix = 'libsvm://' if task[i] != 'ranking' else '' data_path_suffix = '' if task[i] != 'ranking' else '.cat' data_path_for_catboost = dataset[i].split('=')[-1] args += f" --learn-set {data_path_prefix}{data_path_for_catboost}{data_path_suffix}" if not for_speed: valid_path_for_catboost = validset[i].split('=')[-1] args += f" --test-set {data_path_prefix}{valid_path_for_catboost}{data_path_suffix}" args += f" --column-description {data_path_for_catboost.split('.')[0]}.cd" loss_function = "Logloss" if task[i] == 'binary' else ("RMSE" if task[i] == 'regression' else "YetiRank") args += f" --loss-function {loss_function}" if not for_speed: eval_metric = "AUC" if task[i] == 'binary' else ("RMSE" if task[i] == 'regression' else "NDCG:top=10\\;type=Exp") args += f" --eval-metric {eval_metric}" args += " --metric-period 1" task_type = "CPU" if device == 'cpu' else "GPU --devices 0 --thread-count 24" args += f" --task-type {task_type}" args += f" --random-seed {j}" args += f" --bootstrap-type No --random-strength 0.0 --rsm 1.0" # remove known randomness running.write(f"../catboost/catboost/app/catboost fit {args} > {log_name} 2>&1\n") if __name__ == '__main__': args = arg_parser.parse_args() generate_script(args.data_path, args.use_discretized_grad, args.discretized_grad_renew, args.stochastic_rounding, args.for_speed, args.device, args.algorithm, args.log_path)
7,377
46.294872
177
py
neuron-merging
neuron-merging-main/main.py
from __future__ import print_function import warnings warnings.simplefilter("ignore", UserWarning) import argparse import numpy as np import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim import os import sys import pickle import copy cwd = os.getcwd() sys.path.append(cwd+'/../') import models from torchvision import datasets, transforms from torch.autograd import Variable from decompose import Decompose def save_state(model, acc): print('==> Saving model ...') state = { 'acc': acc, 'state_dict': model.state_dict(), } for key in state['state_dict'].keys(): if 'module' in key: print(key) state['state_dict'][key.replace('module.', '')] = \ state['state_dict'].pop(key) # save if args.model_type == 'original': if args.arch == 'WideResNet' : model_filename = '.'.join([args.arch, args.dataset, args.model_type, '_'.join(map(str, args.depth_wide)), 'pth.tar']) elif args.arch == 'ResNet' : model_filename = '.'.join([args.arch, args.dataset, args.model_type, str(args.depth_wide), 'pth.tar']) else: model_filename = '.'.join([args.arch, args.dataset, args.model_type, 'pth.tar']) else: # retrain if args.arch == 'WideResNet' : model_filename = '.'.join([args.arch, '_'.join(map(str, args.depth_wide)), args.dataset, args.model_type, args.criterion, str(args.pruning_ratio), 'pth.tar']) elif args.arch == 'ResNet' : model_filename = '.'.join([args.arch, str(args.depth_wide), args.dataset, args.model_type, args.criterion, str(args.pruning_ratio), 'pth.tar']) else : model_filename = '.'.join([args.arch, args.dataset, args.model_type, args.criterion, str(args.pruning_ratio), 'pth.tar']) torch.save(state, os.path.join('saved_models/', model_filename)) def train(epoch): model.train() for batch_idx, (data, target) in enumerate(train_loader): if args.cuda: data, target = data.cuda(), target.cuda() data, target = Variable(data), Variable(target) optimizer.zero_grad() output = model(data) loss = criterion(output, target) loss.backward() optimizer.step() if batch_idx % args.log_interval == 0: print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format( epoch, batch_idx * len(data), len(train_loader.dataset), 100. * batch_idx / len(train_loader), loss.data)) return def test(epoch, evaluate=False): global best_acc global best_epoch model.eval() test_loss = 0 correct = 0 for data, target in test_loader: if args.cuda: data, target = data.cuda(), target.cuda() data, target = Variable(data, volatile=True), Variable(target) output = model(data) test_loss += criterion(output, target).data pred = output.data.max(1, keepdim=True)[1] correct += pred.eq(target.data.view_as(pred)).cpu().sum() acc = 100. * float(correct) / len(test_loader.dataset) if (acc > best_acc): best_acc = acc best_epoch = epoch if not evaluate: save_state(model, best_acc) test_loss /= len(test_loader.dataset) print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.2f}%)'.format( test_loss * args.batch_size, correct, len(test_loader.dataset), 100. * float(correct) / len(test_loader.dataset))) print('Best Accuracy: {:.2f}%, Best Epoch: {}\n'.format(best_acc, best_epoch)) return def adjust_learning_rate(optimizer, epoch, gammas, schedule): lr = args.lr for (gamma, step) in zip (gammas, schedule): if(epoch>= step) and (args.epochs * 3 //4 >= epoch): lr = lr * gamma elif(epoch>= step) and (args.epochs * 3 //4 < epoch): lr = lr * gamma * gamma else: break print('learning rate : ', lr) for param_group in optimizer.param_groups: param_group['lr'] = lr return def weight_init(model, decomposed_weight_list, target): for layer in model.state_dict(): decomposed_weight = decomposed_weight_list.pop(0) model.state_dict()[layer].copy_(decomposed_weight) return model if __name__=='__main__': # settings parser = argparse.ArgumentParser(description='Neuron Merging Example') parser.add_argument('--batch-size', type=int, default=128, metavar='N', help='input batch size for training (default: 128)') parser.add_argument('--test-batch-size', type=int, default=256, metavar='N', help='input batch size for testing (default: 256)') parser.add_argument('--epochs', type=int, default=200, metavar='N', help='number of epochs to train (default: 200)') parser.add_argument('--lr', type=float, default=0.1, metavar='LR', help='learning rate (default: 0.1)') parser.add_argument('--momentum', type=float, default=0.9, metavar='M', help='SGD momentum (default: 0.9)') parser.add_argument('--weight-decay', '--wd', default=5e-4, type=float, metavar='W', help='weight decay (default: 5e-4)') parser.add_argument('--no-cuda', action='store_true', default=False, help='disables CUDA training') parser.add_argument('--seed', type=int, default=1, metavar='S', help='random seed (default: 1)') parser.add_argument('--log-interval', type=int, default=100, metavar='N', help='how many batches to wait before logging training status') parser.add_argument('--arch', action='store', default='VGG', help='network structure: VGG | ResNet | WideResNet | LeNet_300_100') parser.add_argument('--pretrained', action='store', default=None, help='pretrained model') parser.add_argument('--evaluate', action='store_true', default=False, help='whether to run evaluation') parser.add_argument('--retrain', action='store_true', default=False, help='whether to retrain') parser.add_argument('--model-type', action='store', default='original', help='model type: original | prune | merge') parser.add_argument('--target', action='store', default='conv', help='decomposing target: default=None | conv | ip') parser.add_argument('--dataset', action='store', default='cifar10', help='dataset: cifar10 | cifar100 | FashionMNIST') parser.add_argument('--criterion', action='store', default='l1-norm', help='criterion : l1-norm | l2-norm | l2-GM') parser.add_argument('--threshold', type=float, default=1, help='threshold (default: 1)') parser.add_argument('--lamda', type=float, default=0.8, help='lamda (default: 0.8)') parser.add_argument('--pruning-ratio', type=float, default=0.7, help='pruning ratio : (default: 0.7)') parser.add_argument('--gammas', type=float, nargs='+', default=[0.1,0.1], help='gammas : (default: [0.1,0.1])') parser.add_argument('--schedule', type=int, nargs='+', default=[100,200], help='schedule : (default: [100,200])') parser.add_argument('--depth-wide', action='store', default=None, help='depth and wide (default: None)') args = parser.parse_args() # check options if not (args.model_type in [None, 'original', 'merge', 'prune']): print('ERROR: Please choose the correct model type') exit() if not (args.target in [None, 'conv', 'ip']): print('ERROR: Please choose the correct decompose target') exit() if not (args.arch in ['VGG','ResNet','WideResNet','LeNet_300_100']): print('ERROR: specified arch is not suppported') exit() torch.manual_seed(args.seed) args.cuda = not args.no_cuda and torch.cuda.is_available() if args.cuda: torch.cuda.manual_seed(args.seed) torch.backends.cudnn.deterministic=True # load data num_classes = 10 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), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) train_data = datasets.CIFAR10(root='./data', train=True, download=True, transform=transform_train) test_data = datasets.CIFAR10(root='./data', train=False, download=True, transform=transform_test) train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=True, num_workers=2) test_loader = torch.utils.data.DataLoader(test_data, batch_size=args.test_batch_size, shuffle=False, num_workers=2) num_classes = 10 elif args.dataset == 'cifar100': transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) train_data = datasets.CIFAR100(root='./data', train=True, download=True, transform=transform_train) test_data = datasets.CIFAR100(root='./data', train=False, download=True, transform=transform_test) train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=True, num_workers=2) test_loader = torch.utils.data.DataLoader(test_data, batch_size=args.test_batch_size, shuffle=False, num_workers=2) num_classes = 100 elif args.dataset == 'FashionMNIST': transform = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,)) ]) train_data = datasets.FashionMNIST('data', train=True, download=True, transform=transform) test_data = datasets.FashionMNIST('data', train=False, download=True, transform=transform) kwargs = {'num_workers': 1, 'pin_memory': True} if args.cuda else {} train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=True, **kwargs) test_loader = torch.utils.data.DataLoader(test_data, batch_size=args.test_batch_size, shuffle=False, **kwargs) num_classes = 10 else : pass if args.depth_wide: args.depth_wide = eval(args.depth_wide) cfg = None # make cfg if args.retrain: if args.target == 'conv' : if args.arch == 'VGG': if args.dataset == 'cifar10': cfg = [32, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 256, 256, 256, 'M', 256, 256, 256] elif args.dataset == 'cifar100': cfg = [32, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 256, 'M', 256, 256, 256] temp_cfg = list(filter(('M').__ne__, cfg)) elif args.arch == 'ResNet': cfg = [16, 32, 64] for i in range(len(cfg)): cfg[i] = int(cfg[i] * (1 - args.pruning_ratio)) temp_cfg = cfg elif args.arch == 'WideResNet': cfg = [16, 32, 64] temp_cfg = [16, 32, 32] for i in range(len(cfg)): cfg[i] = int(cfg[i] * (1 - args.pruning_ratio)) temp_cfg[i] = cfg[i] * args.depth_wide[1] elif args.target == 'ip' : if args.arch == 'LeNet_300_100': cfg = [300,100] for i in range(len(cfg)): cfg[i] = round(cfg[i] * (1 - args.pruning_ratio)) temp_cfg = cfg pass # generate the model if args.arch == 'VGG': model = models.VGG(num_classes, cfg=cfg) elif args.arch == 'LeNet_300_100': model = models.LeNet_300_100(bias_flag=True, cfg=cfg) elif args.arch == 'ResNet': model = models.ResNet(int(args.depth_wide) ,num_classes,cfg=cfg) elif args.arch == 'WideResNet': model = models.WideResNet(args.depth_wide[0], num_classes, widen_factor=args.depth_wide[1], cfg=cfg) else: pass if args.cuda: model.cuda() # pretrain best_acc = 0.0 best_epoch = 0 if args.pretrained: pretrained_model = torch.load(args.pretrained) best_epoch = 0 if args.model_type == 'original': best_acc = pretrained_model['acc'] model.load_state_dict(pretrained_model['state_dict']) # weight initialization if args.retrain: decomposed_list = Decompose(args.arch, pretrained_model['state_dict'], args.criterion, args.threshold, args.lamda, args.model_type, temp_cfg, args.cuda).main() model = weight_init(model, decomposed_list, args.target) # print the number of model parameters model_parameters = filter(lambda p: p.requires_grad, model.parameters()) params = sum([np.prod(p.size()) for p in model_parameters]) print('Total parameter number:', params, '\n') optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay) criterion = nn.CrossEntropyLoss() if args.evaluate: test(0, evaluate=True) exit() for epoch in range(1, args.epochs + 1): adjust_learning_rate(optimizer, epoch, args.gammas, args.schedule) train(epoch) test(epoch)
14,964
38.485488
167
py
neuron-merging
neuron-merging-main/decompose.py
from __future__ import print_function import argparse import pickle import numpy as np from sklearn.utils.extmath import randomized_svd from sklearn.metrics import pairwise_distances from sklearn.metrics.pairwise import cosine_similarity from scipy.spatial.distance import cosine import torch import torch.nn as nn import torch.nn.functional as F from scipy.spatial import distance import sys import os import scipy import random cwd = os.getcwd() sys.path.append(cwd+'/../') def create_scaling_mat_ip_thres_bias(weight, ind, threshold, model_type): ''' weight - 2D matrix (n_{i+1}, n_i), np.ndarray ind - chosen indices to remain, np.ndarray threshold - cosine similarity threshold ''' assert(type(weight) == np.ndarray) assert(type(ind) == np.ndarray) cosine_sim = 1-pairwise_distances(weight, metric="cosine") weight_chosen = weight[ind, :] scaling_mat = np.zeros([weight.shape[0], weight_chosen.shape[0]]) for i in range(weight.shape[0]): if i in ind: # chosen ind_i, = np.where(ind == i) assert(len(ind_i) == 1) # check if only one index is found scaling_mat[i, ind_i] = 1 else: # not chosen if model_type == 'prune': continue max_cos_value = np.max(cosine_sim[i][ind]) max_cos_value_index = np.argpartition(cosine_sim[i][ind], -1)[-1] if threshold and max_cos_value < threshold: continue baseline_weight = weight_chosen[max_cos_value_index] current_weight = weight[i] baseline_norm = np.linalg.norm(baseline_weight) current_norm = np.linalg.norm(current_weight) scaling_factor = current_norm / baseline_norm scaling_mat[i, max_cos_value_index] = scaling_factor return scaling_mat def create_scaling_mat_conv_thres_bn(weight, ind, threshold, bn_weight, bn_bias, bn_mean, bn_var, lam, model_type): ''' weight - 4D tensor(n, c, h, w), np.ndarray ind - chosen indices to remain threshold - cosine similarity threshold bn_weight, bn_bias - parameters of batch norm layer right after the conv layer bn_mean, bn_var - running_mean, running_var of BN (for inference) lam - how much to consider cosine sim over bias, float value between 0 and 1 ''' assert(type(weight) == np.ndarray) assert(type(ind) == np.ndarray) assert(type(bn_weight) == np.ndarray) assert(type(bn_bias) == np.ndarray) assert(type(bn_mean) == np.ndarray) assert(type(bn_var) == np.ndarray) assert(bn_weight.shape[0] == weight.shape[0]) assert(bn_bias.shape[0] == weight.shape[0]) assert(bn_mean.shape[0] == weight.shape[0]) assert(bn_var.shape[0] == weight.shape[0]) weight = weight.reshape(weight.shape[0], -1) cosine_dist = pairwise_distances(weight, metric="cosine") weight_chosen = weight[ind, :] scaling_mat = np.zeros([weight.shape[0], weight_chosen.shape[0]]) for i in range(weight.shape[0]): if i in ind: # chosen ind_i, = np.where(ind == i) assert(len(ind_i) == 1) # check if only one index is found scaling_mat[i, ind_i] = 1 else: # not chosen if model_type == 'prune': continue current_weight = weight[i] current_norm = np.linalg.norm(current_weight) current_cos = cosine_dist[i] gamma_1 = bn_weight[i] beta_1 = bn_bias[i] mu_1 = bn_mean[i] sigma_1 = bn_var[i] # choose one cos_list = [] scale_list = [] bias_list = [] for chosen_i in ind: chosen_weight = weight[chosen_i] chosen_norm = np.linalg.norm(chosen_weight, ord = 2) chosen_cos = current_cos[chosen_i] gamma_2 = bn_weight[chosen_i] beta_2 = bn_bias[chosen_i] mu_2 = bn_mean[chosen_i] sigma_2 = bn_var[chosen_i] # compute cosine sim cos_list.append(chosen_cos) # compute s s = current_norm/chosen_norm # compute scale term scale_term_inference = s * (gamma_2 / gamma_1) * (sigma_1 / sigma_2) scale_list.append(scale_term_inference) # compute bias term bias_term_inference = abs((gamma_2/sigma_2) * (s * (-(sigma_1*beta_1/gamma_1) + mu_1) - mu_2) + beta_2) bias_term_inference = bias_term_inference/scale_term_inference bias_list.append(bias_term_inference) assert(len(cos_list) == len(ind)) assert(len(scale_list) == len(ind)) assert(len(bias_list) == len(ind)) # merge cosine distance and bias distance bias_list = (bias_list - np.min(bias_list)) / (np.max(bias_list)-np.min(bias_list)) score_list = lam * np.array(cos_list) + (1-lam) * np.array(bias_list) # find index and scale with minimum distance min_ind = np.argmin(score_list) min_scale = scale_list[min_ind] min_cosine_sim = 1-cos_list[min_ind] # check threshold - second if threshold and min_cosine_sim < threshold: continue scaling_mat[i, min_ind] = min_scale return scaling_mat class Decompose: def __init__(self, arch, param_dict, criterion, threshold, lamda, model_type, cfg, cuda): self.param_dict = param_dict self.arch = arch self.criterion = criterion self.threshold = threshold self.lamda = lamda self.model_type = model_type self.cfg = cfg self.cuda = cuda self.output_channel_index = {} self.decompose_weight = [] def get_output_channel_index(self, value, layer_id): output_channel_index = [] if len(value.size()) : weight_vec = value.view(value.size()[0], -1) weight_vec = weight_vec.cuda() # l1-norm if self.criterion == 'l1-norm': norm = torch.norm(weight_vec, 1, 1) norm_np = norm.cpu().detach().numpy() arg_max = np.argsort(norm_np) arg_max_rev = arg_max[::-1][:self.cfg[layer_id]] output_channel_index = sorted(arg_max_rev.tolist()) # l2-norm elif self.criterion == 'l2-norm': norm = torch.norm(weight_vec, 2, 1) norm_np = norm.cpu().detach().numpy() arg_max = np.argsort(norm_np) arg_max_rev = arg_max[::-1][:self.cfg[layer_id]] output_channel_index = sorted(arg_max_rev.tolist()) # l2-GM elif self.criterion == 'l2-GM': weight_vec = weight_vec.cpu().detach().numpy() matrix = distance.cdist(weight_vec, weight_vec, 'euclidean') similar_sum = np.sum(np.abs(matrix), axis=0) output_channel_index = np.argpartition(similar_sum, -self.cfg[layer_id])[-self.cfg[layer_id]:] return output_channel_index def get_decompose_weight(self): # scale matrix z = None # copy original weight self.decompose_weight = list(self.param_dict.values()) # cfg index layer_id = -1 for index, layer in enumerate(self.param_dict): original = self.param_dict[layer] # VGG if self.arch == 'VGG': # feature if 'feature' in layer : # conv if len(self.param_dict[layer].shape) == 4: layer_id += 1 # get index self.output_channel_index[index] = self.get_output_channel_index(self.param_dict[layer], layer_id) # Merge scale matrix if z != None: original = original[:,input_channel_index,:,:] for i, f in enumerate(self.param_dict[layer]): o = f.view(f.shape[0],-1) o = torch.mm(z,o) o = o.view(z.shape[0],f.shape[1],f.shape[2]) original[i,:,:,:] = o # make scale matrix with batchNorm bn = list(self.param_dict.values()) bn_weight = bn[index+1].cpu().detach().numpy() bn_bias = bn[index+2].cpu().detach().numpy() bn_mean = bn[index+3].cpu().detach().numpy() bn_var = bn[index+4].cpu().detach().numpy() x = create_scaling_mat_conv_thres_bn(self.param_dict[layer].cpu().detach().numpy(), np.array(self.output_channel_index[index]), self.threshold, bn_weight, bn_bias, bn_mean, bn_var, self.lamda, self.model_type) z = torch.from_numpy(x).type(dtype=torch.float) if self.cuda: z = z.cuda() z = z.t() # pruned pruned = original[self.output_channel_index[index],:,:,:] # update next input channel input_channel_index = self.output_channel_index[index] # update decompose weight self.decompose_weight[index] = pruned # batchNorm elif len(self.param_dict[layer].shape): # pruned pruned = self.param_dict[layer][input_channel_index] # update decompose weight self.decompose_weight[index] = pruned # first classifier else: pruned = torch.zeros(original.shape[0],z.shape[0]) if self.cuda: pruned = pruned.cuda() for i, f in enumerate(original): o_old = f.view(z.shape[1],-1) o = torch.mm(z,o_old).view(-1) pruned[i,:] = o self.decompose_weight[index] = pruned break # ResNet elif self.arch == 'ResNet': # block if 'layer' in layer : # last layer each block if '0.conv1.weight' in layer : layer_id += 1 # Pruning if 'conv1' in layer : # get index self.output_channel_index[index] = self.get_output_channel_index(self.param_dict[layer], layer_id) # make scale matrix with batchNorm bn = list(self.param_dict.values()) bn_weight = bn[index+1].cpu().detach().numpy() bn_bias = bn[index+2].cpu().detach().numpy() bn_mean = bn[index+3].cpu().detach().numpy() bn_var = bn[index+4].cpu().detach().numpy() x = create_scaling_mat_conv_thres_bn(self.param_dict[layer].cpu().detach().numpy(), np.array(self.output_channel_index[index]), self.threshold, bn_weight, bn_bias, bn_mean, bn_var, self.lamda, self.model_type) z = torch.from_numpy(x).type(dtype=torch.float) if self.cuda: z = z.cuda() z = z.t() # pruned pruned = original[self.output_channel_index[index],:,:,:] # update next input channel input_channel_index = self.output_channel_index[index] # update decompose weight self.decompose_weight[index] = pruned # batchNorm elif 'bn1' in layer : if len(self.param_dict[layer].shape): # pruned pruned = self.param_dict[layer][input_channel_index] # update decompose weight self.decompose_weight[index] = pruned # Merge scale matrix elif 'conv2' in layer : if z != None: original = original[:,input_channel_index,:,:] for i, f in enumerate(self.param_dict[layer]): o = f.view(f.shape[0],-1) o = torch.mm(z,o) o = o.view(z.shape[0],f.shape[1],f.shape[2]) original[i,:,:,:] = o scaled = original # update decompose weight self.decompose_weight[index] = scaled # WideResNet elif self.arch == 'WideResNet': # block if 'block' in layer : # last layer each block if '0.conv1.weight' in layer : layer_id += 1 # Pruning if 'conv1' in layer : # get index self.output_channel_index[index] = self.get_output_channel_index(self.param_dict[layer], layer_id) # make scale matrix with batchNorm bn = list(self.param_dict.values()) bn_weight = bn[index+1].cpu().detach().numpy() bn_bias = bn[index+2].cpu().detach().numpy() bn_mean = bn[index+3].cpu().detach().numpy() bn_var = bn[index+4].cpu().detach().numpy() x = create_scaling_mat_conv_thres_bn(self.param_dict[layer].cpu().detach().numpy(), np.array(self.output_channel_index[index]), self.threshold, bn_weight, bn_bias, bn_mean, bn_var, self.lamda, self.model_type) z = torch.from_numpy(x).type(dtype=torch.float) if self.cuda: z = z.cuda() z = z.t() # pruned pruned = original[self.output_channel_index[index],:,:,:] # update next input channel input_channel_index = self.output_channel_index[index] # update decompose weight self.decompose_weight[index] = pruned # BatchNorm elif 'bn2' in layer : if len(self.param_dict[layer].shape): # pruned pruned = self.param_dict[layer][input_channel_index] # update decompose weight self.decompose_weight[index] = pruned # Merge scale matrix elif 'conv2' in layer : # scale if z != None: original = original[:,input_channel_index,:,:] for i, f in enumerate(self.param_dict[layer]): o = f.view(f.shape[0],-1) o = torch.mm(z,o) o = o.view(z.shape[0],f.shape[1],f.shape[2]) original[i,:,:,:] = o scaled = original # update decompose weight self.decompose_weight[index] = scaled # LeNet_300_100 elif self.arch == 'LeNet_300_100': # ip if layer in ['ip1.weight','ip2.weight'] : # Merge scale matrix if z != None: original = torch.mm(original,z) layer_id += 1 # concatenate weight and bias if layer in 'ip1.weight' : weight = self.param_dict['ip1.weight'].cpu().detach().numpy() bias = self.param_dict['ip1.bias'].cpu().detach().numpy() elif layer in 'ip2.weight' : weight = self.param_dict['ip2.weight'].cpu().detach().numpy() bias = self.param_dict['ip2.bias'].cpu().detach().numpy() bias_reshaped = bias.reshape(bias.shape[0],-1) concat_weight = np.concatenate([weight, bias_reshaped], axis = 1) # get index self.output_channel_index[index] = self.get_output_channel_index(torch.from_numpy(concat_weight), layer_id) # make scale matrix with bias x = create_scaling_mat_ip_thres_bias(concat_weight, np.array(self.output_channel_index[index]), self.threshold, self.model_type) z = torch.from_numpy(x).type(dtype=torch.float) if self.cuda: z = z.cuda() # pruned pruned = original[self.output_channel_index[index],:] # update next input channel input_channel_index = self.output_channel_index[index] # update decompose weight self.decompose_weight[index] = pruned elif layer in 'ip3.weight': original = torch.mm(original,z) # update decompose weight self.decompose_weight[index] = original # update bias elif layer in ['ip1.bias','ip2.bias']: self.decompose_weight[index] = original[input_channel_index] else : pass def main(self): if self.cuda == False: for layer in self.param_dict: self.param_dict[layer] = self.param_dict[layer].cpu() self.get_decompose_weight() return self.decompose_weight
19,186
35.616412
168
py
neuron-merging
neuron-merging-main/models/ResNet.py
import torch import torch.nn as nn import math import torch.utils.model_zoo as model_zoo import torch.nn.functional as F def conv3x3(in_planes, out_planes, stride=1): "3x3 convolution with padding" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, cfg, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, cfg, stride) self.bn1 = nn.BatchNorm2d(cfg) self.conv2 = conv3x3(cfg, planes) self.bn2 = nn.BatchNorm2d(planes) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): x = F.relu(x) residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual # out = self.relu(out) return out def bn_feature(self, x): x = F.relu(x) residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) bn_feature = out if self.downsample is not None: residual = self.downsample(x) out += residual # out = self.relu(out) return out, bn_feature class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, cfg, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, planes * Bottleneck.expansion, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * Bottleneck.expansion) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): x = F.relu(x) residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual # out = self.relu(out) return out class ResNet(nn.Module): def __init__(self, depth, num_classes, cfg=None, bottleneck=False): super(ResNet, self).__init__() if cfg == None: cfg = [16, 32, 64] self.inplanes = 16 #print(bottleneck) if bottleneck == True: n = int((depth - 2) / 9) block = Bottleneck else: n = int((depth - 2) / 6) block = BasicBlock self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(self.inplanes) self.relu = nn.ReLU(inplace=True) self.layer1 = self._make_layer(block, 16, n, cfg[0]) self.layer2 = self._make_layer(block, 32, n, cfg[1], stride=2) self.layer3 = self._make_layer(block, 64, n, cfg[2], stride=2) self.avgpool = nn.AvgPool2d(8) self.fc = nn.Linear(64 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, cfg, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, cfg, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes, cfg)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = F.relu(x) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x def get_bn_before_relu(self): if isinstance(self.layer1[0], Bottleneck): bn1 = self.layer1[-1].bn3 bn2 = self.layer2[-1].bn3 bn3 = self.layer3[-1].bn3 elif isinstance(self.layer1[0], BasicBlock): bn1 = self.layer1[-1].bn2 bn2 = self.layer2[-1].bn2 bn3 = self.layer3[-1].bn2 else: print('ResNet unknown block error !!!') return [bn1, bn2, bn3] def get_channel_num(self): return [16, 32, 64] def extract_feature(self, x, preReLU=False): x = x.cuda() x = self.conv1(x) x = self.bn1(x) feat1 = self.layer1(x) feat2 = self.layer2(feat1) feat3 = self.layer3(feat2) x = F.relu(feat3) x = self.avgpool(x) x = x.view(x.size(0), -1) #out = self.fc(x) if not preReLU: feat1 = F.relu(feat1) feat2 = F.relu(feat2) feat3 = F.relu(feat3) return [feat1, feat2, feat3]#, out def bn_feature(self,x): bn_feature_list = [] x = self.conv1(x) x = self.bn1(x) for block in self.layer1: if isinstance(block,BasicBlock): x, bn_feature = block.bn_feature(x) temp = bn_feature.cpu().detach().numpy() bn_feature_list.append(temp) for block in self.layer2: if isinstance(block,BasicBlock): x, bn_feature = block.bn_feature(x) temp = bn_feature.cpu().detach().numpy() bn_feature_list.append(temp) for block in self.layer3: if isinstance(block,BasicBlock): x, bn_feature = block.bn_feature(x) temp = bn_feature.cpu().detach().numpy() bn_feature_list.append(temp) return bn_feature_list def ware(self, x): x = self.conv1(x) x = self.bn1(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = F.relu(x) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x.cpu().detach().numpy()
7,173
26.381679
99
py
neuron-merging
neuron-merging-main/models/LeNet_300_100.py
from __future__ import print_function import torch import torch.nn as nn import os class LeNet_300_100(nn.Module): def __init__(self, bias_flag, cfg): if cfg == None: cfg = [300,100] super(LeNet_300_100, self).__init__() self.ip1 = nn.Linear(28*28, cfg[0], bias=bias_flag) self.relu_ip1 = nn.ReLU(inplace=True) self.ip2 = nn.Linear(cfg[0], cfg[1], bias=bias_flag) self.relu_ip2 = nn.ReLU(inplace=True) self.ip3 = nn.Linear(cfg[1], 10, bias=bias_flag) return def forward(self, x): x = x.view(x.size(0), 28*28) x = self.ip1(x) x = self.relu_ip1(x) x = self.ip2(x) x = self.relu_ip2(x) x = self.ip3(x) return x
749
29
60
py
neuron-merging
neuron-merging-main/models/VGG.py
from __future__ import print_function import math import torch import torch.nn as nn from torch.autograd import Variable import torch.nn.functional as F __all__ = ['VGG'] defaultcfg = { 11: [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512], 13: [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512], 16: [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512], 19: [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512], } class VGG(nn.Module): def __init__(self, out_classes=10, depth=16, init_weights=True, cfg=None): super(VGG, self).__init__() if cfg is None: cfg = defaultcfg[depth] self.cfg = cfg self.feature = self.make_layers(cfg, True) self.classifier = nn.Sequential( nn.Linear(cfg[-1], 512), nn.BatchNorm1d(512), nn.ReLU(inplace=True), nn.Linear(512, out_classes) ) if init_weights: self._initialize_weights() def make_layers(self, cfg, batch_norm=False): layers = [] in_channels = 3 for v in cfg: if v == 'M': layers += [nn.MaxPool2d(kernel_size=2, stride=2)] else: conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1, bias=False) 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) def forward(self, x): x = self.feature(x) x = nn.AvgPool2d(2)(x) x = x.view(x.size(0), -1) y = self.classifier(x) return y def _initialize_weights(self): for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(0.5) m.bias.data.zero_() elif isinstance(m, nn.Linear): m.weight.data.normal_(0, 0.01) m.bias.data.zero_() def extract_feature(self, x, preReLU=False): x = x.cuda() feat1 = self.feature(x) if not preReLU: feat1 = F.relu(feat1) return [feat1] def bn_feature(self,x): bn_feature = [] for layer in self.feature: x = layer(x) if isinstance(layer, nn.BatchNorm2d) : # temp = torch.sum(x,0).cpu().detach().numpy() temp = x.cpu().detach().numpy() bn_feature.append(temp) return bn_feature def ware(self, x): x = self.feature(x) x = nn.AvgPool2d(2)(x) x = x.view(x.size(0), -1) y = self.classifier(x) return y.cpu().detach().numpy()
3,154
26.920354
107
py
neuron-merging
neuron-merging-main/models/WideResNet.py
import math import torch import torch.nn as nn import torch.nn.functional as F class BasicBlock(nn.Module): def __init__(self, in_planes, out_planes, stride, cfg, dropRate=0.0): super(BasicBlock, self).__init__() self.bn1 = nn.BatchNorm2d(in_planes) self.relu1 = nn.ReLU(inplace=True) self.conv1 = nn.Conv2d(in_planes, cfg, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(cfg) self.relu2 = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(cfg, out_planes, kernel_size=3, stride=1, padding=1, bias=False) self.droprate = dropRate self.equalInOut = (in_planes == out_planes) self.convShortcut = (not self.equalInOut) and nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, padding=0, bias=False) or None def forward(self, x): if not self.equalInOut: x = self.relu1(self.bn1(x)) else: out = self.relu1(self.bn1(x)) out = self.relu2(self.bn2(self.conv1(out if self.equalInOut else x))) if self.droprate > 0: out = F.dropout(out, p=self.droprate, training=self.training) out = self.conv2(out) return torch.add(x if self.equalInOut else self.convShortcut(x), out) class NetworkBlock(nn.Module): def __init__(self, nb_layers, in_planes, out_planes, block, stride, cfg, dropRate=0.0): super(NetworkBlock, self).__init__() self.layer = self._make_layer(block, in_planes, out_planes, nb_layers, stride, cfg, dropRate) def _make_layer(self, block, in_planes, out_planes, nb_layers, stride, cfg, dropRate): layers = [] for i in range(int(nb_layers)): layers.append(block(i == 0 and in_planes or out_planes, out_planes, i == 0 and stride or 1, cfg, dropRate)) return nn.Sequential(*layers) def forward(self, x): return self.layer(x) class WideResNet(nn.Module): def __init__(self, depth, num_classes, widen_factor=1, dropRate=0.0, cfg=None): super(WideResNet, self).__init__() if cfg == None: cfg = [16, 32, 64] nChannels = [16, 16*widen_factor, 32*widen_factor, 64*widen_factor] assert((depth - 4) % 6 == 0) n = (depth - 4) / 6 block = BasicBlock # 1st conv before any network block self.conv1 = nn.Conv2d(3, nChannels[0], kernel_size=3, stride=1, padding=1, bias=False) # 1st block self.block1 = NetworkBlock(n, nChannels[0], nChannels[1], block, 1, cfg[0]*widen_factor, dropRate) # 2nd block self.block2 = NetworkBlock(n, nChannels[1], nChannels[2], block, 2, cfg[1]*widen_factor, dropRate) # 3rd block self.block3 = NetworkBlock(n, nChannels[2], nChannels[3], block, 2, cfg[2]*widen_factor, dropRate) # global average pooling and classifier self.bn1 = nn.BatchNorm2d(nChannels[3]) self.relu = nn.ReLU(inplace=True) self.fc = nn.Linear(nChannels[3], num_classes) self.nChannels = nChannels for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.Linear): m.bias.data.zero_() def forward(self, x): out = self.conv1(x) out = self.block1(out) out = self.block2(out) out = self.block3(out) out = self.relu(self.bn1(out)) out = F.avg_pool2d(out, 8) out = out.view(-1, self.nChannels[3]) return self.fc(out) def get_bn_before_relu(self): bn1 = self.block2.layer[0].bn1 bn2 = self.block3.layer[0].bn1 bn3 = self.bn1 return [bn1, bn2, bn3] def get_channel_num(self): return self.nChannels[1:] def extract_feature(self, x, preReLU=False): x = x.cuda() out = self.conv1(x) feat1 = self.block1(out) feat2 = self.block2(feat1) feat3 = self.block3(feat2) out = self.relu(self.bn1(feat3)) out = F.avg_pool2d(out, 8) out = out.view(-1, self.nChannels[3]) out = self.fc(out) if preReLU: feat1 = self.block2.layer[0].bn1(feat1) feat2 = self.block3.layer[0].bn1(feat2) feat3 = self.bn1(feat3) return [feat1, feat2, feat3] def ware(self, x): out = self.conv1(x) out = self.block1(out) out = self.block2(out) out = self.block3(out) out = self.relu(self.bn1(out)) out = F.avg_pool2d(out, 8) out = out.view(-1, self.nChannels[3]) return self.fc(out).cpu().detach().numpy()
4,955
35.711111
119
py
LAP-PAL
LAP-PAL-master/continuous/main.py
import numpy as np import torch import gym import argparse import os import time import utils import TD3 import LAP_TD3 import PAL_TD3 import PER_TD3 # Runs policy for X episodes and returns average reward def eval_policy(policy, env, seed, eval_episodes=10): eval_env = gym.make(env) eval_env.seed(seed + 100) avg_reward = 0. for _ in range(eval_episodes): state, done = eval_env.reset(), False while not done: action = policy.select_action(np.array(state), test=True) state, reward, done, _ = eval_env.step(action) avg_reward += reward avg_reward /= eval_episodes print("---------------------------------------") print(f"Evaluation over {eval_episodes} episodes: {avg_reward:.3f}") print("---------------------------------------") return avg_reward if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--algorithm", default="LAP_TD3") # Algorithm nameu parser.add_argument("--env", default="HalfCheetah-v3") # OpenAI gym environment name parser.add_argument("--seed", default=0, type=int) # Sets Gym, PyTorch and Numpy seeds parser.add_argument("--start_timesteps", default=25e3, type=int)# Time steps initial random policy is used parser.add_argument("--eval_freq", default=5e3, type=int) # How often (time steps) we evaluate parser.add_argument("--max_timesteps", default=3e6, type=int) # Max time steps to run environment parser.add_argument("--expl_noise", default=0.1) # Std of Gaussian exploration noise parser.add_argument("--batch_size", default=256, type=int) # Batch size for both actor and critic parser.add_argument("--discount", default=0.99) # Discount factor parser.add_argument("--tau", default=0.005) # Target network update rate parser.add_argument("--policy_noise", default=0.2) # Noise added to target policy during critic update parser.add_argument("--noise_clip", default=0.5) # Range to clip target policy noise parser.add_argument("--policy_freq", default=2, type=int) # Frequency of delayed policy updates parser.add_argument("--alpha", default=0.4) # Priority = TD^alpha (only used by LAP/PAL) parser.add_argument("--min_priority", default=1, type=int) # Minimum priority (set to 1 in paper, only used by LAP/PAL) args = parser.parse_args() file_name = "%s_%s_%s" % (args.algorithm, args.env, str(args.seed)) print("---------------------------------------") print(f"Settings: {file_name}") print("---------------------------------------") if not os.path.exists("./results"): os.makedirs("./results") env = gym.make(args.env) # Set seeds env.seed(args.seed) torch.manual_seed(args.seed) np.random.seed(args.seed) state_dim = env.observation_space.shape[0] action_dim = env.action_space.shape[0] max_action = float(env.action_space.high[0]) kwargs = { "state_dim": state_dim, "action_dim": action_dim, "max_action": max_action, "discount": args.discount, "tau": args.tau, "policy_noise": args.policy_noise * max_action, "noise_clip": args.noise_clip * max_action, "policy_freq": args.policy_freq } # Initialize policy and replay buffer if args.algorithm == "TD3": policy = TD3.TD3(**kwargs) replay_buffer = utils.ReplayBuffer(state_dim, action_dim) elif args.algorithm == "PER_TD3": policy = PER_TD3.PER_TD3(**kwargs) replay_buffer = utils.PrioritizedReplayBuffer(state_dim, action_dim) kwargs["alpha"] = args.alpha kwargs["min_priority"] = args.min_priority if args.algorithm == "LAP_TD3": policy = LAP_TD3.LAP_TD3(**kwargs) replay_buffer = utils.PrioritizedReplayBuffer(state_dim, action_dim) elif args.algorithm == "PAL_TD3": policy = PAL_TD3.PAL_TD3(**kwargs) replay_buffer = utils.ReplayBuffer(state_dim, action_dim) # Evaluate untrained policy evaluations = [eval_policy(policy, args.env, args.seed)] state, done = env.reset(), False episode_reward = 0 episode_timesteps = 0 episode_num = 0 for t in range(int(args.max_timesteps)): episode_timesteps += 1 # Select action randomly or according to policy if t < args.start_timesteps: action = env.action_space.sample() else: action = ( policy.select_action(np.array(state)) + np.random.normal(0, max_action * args.expl_noise, size=action_dim) ).clip(-max_action, max_action) # Perform action next_state, reward, done, _ = env.step(action) done_bool = float(done) if episode_timesteps < env._max_episode_steps else 0 # Store data in replay buffer replay_buffer.add(state, action, next_state, reward, done_bool) state = next_state episode_reward += reward # Train agent after collecting sufficient data if t >= args.start_timesteps: #>= policy.train(replay_buffer, args.batch_size) if done: # +1 to account for 0 indexing. +0 on ep_timesteps since it will increment +1 even if done=True print(f"Total T: {t+1} Episode Num: {episode_num+1} Episode T: {episode_timesteps} Reward: {episode_reward:.3f}") state, done = env.reset(), False episode_reward = 0 episode_timesteps = 0 episode_num += 1 # Evaluate episode if (t + 1) % args.eval_freq == 0: evaluations.append(eval_policy(policy, args.env, args.seed)) np.save("./results/%s" % (file_name), evaluations)
5,222
33.361842
121
py
LAP-PAL
LAP-PAL-master/continuous/PAL_TD3.py
import copy import numpy as np import torch import torch.nn as nn import torch.nn.functional as F device = torch.device("cuda" if torch.cuda.is_available() else "cpu") class Actor(nn.Module): def __init__(self, state_dim, action_dim, max_action): super(Actor, self).__init__() self.l1 = nn.Linear(state_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, action_dim) self.max_action = max_action def forward(self, state): a = F.relu(self.l1(state)) a = F.relu(self.l2(a)) return self.max_action * torch.tanh(self.l3(a)) def act(self, state): a = F.relu(self.l1(state)) a = F.relu(self.l2(a)) a = self.l3(a) return self.max_action * torch.tanh(a), a class Critic(nn.Module): def __init__(self, state_dim, action_dim): super(Critic, self).__init__() # Q1 architecture self.l1 = nn.Linear(state_dim + action_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, 1) # Q2 architecture self.l4 = nn.Linear(state_dim + action_dim, 256) self.l5 = nn.Linear(256, 256) self.l6 = nn.Linear(256, 1) def forward(self, state, action): sa = torch.cat([state, action], 1) q1 = F.relu(self.l1(sa)) q1 = F.relu(self.l2(q1)) q1 = self.l3(q1) q2 = F.relu(self.l4(sa)) q2 = F.relu(self.l5(q2)) q2 = self.l6(q2) return q1, q2 def Q1(self, state, action): sa = torch.cat([state, action], 1) q1 = F.relu(self.l1(sa)) q1 = F.relu(self.l2(q1)) q1 = self.l3(q1) return q1 class PAL_TD3(object): def __init__( self, state_dim, action_dim, max_action, discount=0.99, tau=0.005, policy_noise=0.2, noise_clip=0.5, policy_freq=2, alpha=0.4, min_priority=1 ): self.actor = Actor(state_dim, action_dim, max_action).to(device) self.actor_target = copy.deepcopy(self.actor) self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=3e-4) self.critic = Critic(state_dim, action_dim).to(device) self.critic_target = copy.deepcopy(self.critic) self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=3e-4) self.max_action = max_action self.discount = discount self.tau = tau self.policy_noise = policy_noise self.noise_clip = noise_clip self.policy_freq = policy_freq self.alpha = alpha self.min_priority = min_priority self.total_it = 0 def select_action(self, state, test=False): state = torch.FloatTensor(state.reshape(1, -1)).to(device) return self.actor(state).cpu().data.numpy().flatten() def train(self, replay_buffer, batch_size=256): self.total_it += 1 # Sample replay buffer state, action, next_state, reward, not_done = replay_buffer.sample(batch_size) with torch.no_grad(): # Select action according to policy and add clipped noise noise = ( torch.randn_like(action) * self.policy_noise ).clamp(-self.noise_clip, self.noise_clip) next_action = ( self.actor_target(next_state) + noise ).clamp(-self.max_action, self.max_action) # Compute the target Q value target_Q1, target_Q2 = self.critic_target(next_state, next_action) target_Q = torch.min(target_Q1, target_Q2) target_Q = reward + not_done * self.discount * target_Q # Get current Q estimates current_Q1, current_Q2 = self.critic(state, action) td_loss1 = (current_Q1 - target_Q) td_loss2 = (current_Q2 - target_Q) critic_loss = self.PAL(td_loss1) + self.PAL(td_loss2) critic_loss /= torch.max(td_loss1.abs(), td_loss2.abs()).clamp(min=self.min_priority).pow(self.alpha).mean().detach() # Optimize the critic self.critic_optimizer.zero_grad() critic_loss.backward() self.critic_optimizer.step() # Delayed policy updates if self.total_it % self.policy_freq == 0: actor_loss = -self.critic.Q1(state, self.actor(state)).mean() # Optimize the actor self.actor_optimizer.zero_grad() actor_loss.backward() self.actor_optimizer.step() # Update the frozen target models for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) # If min_priority=1, this can be simplified. def PAL(self, x): return torch.where( x.abs() < self.min_priority, (self.min_priority ** self.alpha) * 0.5 * x.pow(2), self.min_priority * x.abs().pow(1. + self.alpha)/(1. + self.alpha) ).mean() def save(self, filename): torch.save(self.critic.state_dict(), filename + "_critic") torch.save(self.critic_optimizer.state_dict(), filename + "_critic_optimizer") torch.save(self.actor.state_dict(), filename + "_actor") torch.save(self.actor_optimizer.state_dict(), filename + "_actor_optimizer") def load(self, filename): self.critic.load_state_dict(torch.load(filename + "_critic")) self.critic_optimizer.load_state_dict(torch.load(filename + "_critic_optimizer")) self.actor.load_state_dict(torch.load(filename + "_actor")) self.actor_optimizer.load_state_dict(torch.load(filename + "_actor_optimizer"))
5,164
26.768817
119
py
LAP-PAL
LAP-PAL-master/continuous/PER_TD3.py
import copy import numpy as np import torch import torch.nn as nn import torch.nn.functional as F device = torch.device("cuda" if torch.cuda.is_available() else "cpu") class Actor(nn.Module): def __init__(self, state_dim, action_dim, max_action): super(Actor, self).__init__() self.l1 = nn.Linear(state_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, action_dim) self.max_action = max_action def forward(self, state): a = F.relu(self.l1(state)) a = F.relu(self.l2(a)) return self.max_action * torch.tanh(self.l3(a)) def act(self, state): a = F.relu(self.l1(state)) a = F.relu(self.l2(a)) a = self.l3(a) return self.max_action * torch.tanh(a), a class Critic(nn.Module): def __init__(self, state_dim, action_dim): super(Critic, self).__init__() # Q1 architecture self.l1 = nn.Linear(state_dim + action_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, 1) # Q2 architecture self.l4 = nn.Linear(state_dim + action_dim, 256) self.l5 = nn.Linear(256, 256) self.l6 = nn.Linear(256, 1) def forward(self, state, action): sa = torch.cat([state, action], 1) q1 = F.relu(self.l1(sa)) q1 = F.relu(self.l2(q1)) q1 = self.l3(q1) q2 = F.relu(self.l4(sa)) q2 = F.relu(self.l5(q2)) q2 = self.l6(q2) return q1, q2 def Q1(self, state, action): sa = torch.cat([state, action], 1) q1 = F.relu(self.l1(sa)) q1 = F.relu(self.l2(q1)) q1 = self.l3(q1) return q1 class PER_TD3(object): def __init__( self, state_dim, action_dim, max_action, discount=0.99, tau=0.005, policy_noise=0.2, noise_clip=0.5, policy_freq=2, alpha=0.6, ): self.actor = Actor(state_dim, action_dim, max_action).to(device) self.actor_target = copy.deepcopy(self.actor) self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=3e-4) self.critic = Critic(state_dim, action_dim).to(device) self.critic_target = copy.deepcopy(self.critic) self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=3e-4) self.max_action = max_action self.discount = discount self.tau = tau self.policy_noise = policy_noise self.noise_clip = noise_clip self.policy_freq = policy_freq self.alpha = alpha self.total_it = 0 def select_action(self, state, test=False): state = torch.FloatTensor(state.reshape(1, -1)).to(device) return self.actor(state).cpu().data.numpy().flatten() def train(self, replay_buffer, batch_size=256): self.total_it += 1 # Sample replay buffer state, action, next_state, reward, not_done, ind, weights = replay_buffer.sample(batch_size) with torch.no_grad(): # Select action according to policy and add clipped noise noise = ( torch.randn_like(action) * self.policy_noise ).clamp(-self.noise_clip, self.noise_clip) next_action = ( self.actor_target(next_state) + noise ).clamp(-self.max_action, self.max_action) # Compute the target Q value target_Q1, target_Q2 = self.critic_target(next_state, next_action) target_Q = torch.min(target_Q1, target_Q2) target_Q = reward + not_done * self.discount * target_Q # Get current Q estimates current_Q1, current_Q2 = self.critic(state, action) td_loss1 = (current_Q1 - target_Q).abs() td_loss2 = (current_Q2 - target_Q).abs() # Compute critic loss critic_loss = ( (weights * F.mse_loss(current_Q1, target_Q, reduction='none')).mean() + (weights * F.mse_loss(current_Q2, target_Q, reduction='none')).mean() ) # Optimize the critic self.critic_optimizer.zero_grad() critic_loss.backward() self.critic_optimizer.step() priority = torch.max(td_loss1, td_loss2).pow(self.alpha).cpu().data.numpy().flatten() replay_buffer.update_priority(ind, priority) # Delayed policy updates if self.total_it % self.policy_freq == 0: actor_loss = -self.critic.Q1(state, self.actor(state)).mean() # Optimize the actor self.actor_optimizer.zero_grad() actor_loss.backward() self.actor_optimizer.step() # Update the frozen target models for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) def save(self, filename): torch.save(self.critic.state_dict(), filename + "_critic") torch.save(self.critic_optimizer.state_dict(), filename + "_critic_optimizer") torch.save(self.actor.state_dict(), filename + "_actor") torch.save(self.actor_optimizer.state_dict(), filename + "_actor_optimizer") def load(self, filename): self.critic.load_state_dict(torch.load(filename + "_critic")) self.critic_optimizer.load_state_dict(torch.load(filename + "_critic_optimizer")) self.actor.load_state_dict(torch.load(filename + "_actor")) self.actor_optimizer.load_state_dict(torch.load(filename + "_actor_optimizer"))
5,032
26.653846
94
py
LAP-PAL
LAP-PAL-master/continuous/utils.py
import numpy as np import torch class ReplayBuffer(object): def __init__(self, state_dim, action_dim, max_size=int(1e6)): self.max_size = max_size self.ptr = 0 self.size = 0 self.state = np.zeros((max_size, state_dim)) self.action = np.zeros((max_size, action_dim)) self.next_state = np.zeros((max_size, state_dim)) self.reward = np.zeros((max_size, 1)) self.not_done = np.zeros((max_size, 1)) self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") def add(self, state, action, next_state, reward, done): self.state[self.ptr] = state self.action[self.ptr] = action self.next_state[self.ptr] = next_state self.reward[self.ptr] = reward self.not_done[self.ptr] = 1. - done self.ptr = (self.ptr + 1) % self.max_size self.size = min(self.size + 1, self.max_size) def sample(self, batch_size): ind = np.random.randint(self.size, size=batch_size) return ( torch.FloatTensor(self.state[ind]).to(self.device), torch.FloatTensor(self.action[ind]).to(self.device), torch.FloatTensor(self.next_state[ind]).to(self.device), torch.FloatTensor(self.reward[ind]).to(self.device), torch.FloatTensor(self.not_done[ind]).to(self.device) ) class PrioritizedReplayBuffer(): def __init__(self, state_dim, action_dim, max_size=int(1e6)): self.max_size = max_size self.ptr = 0 self.size = 0 self.state = np.zeros((max_size, state_dim)) self.action = np.zeros((max_size, action_dim)) self.next_state = np.zeros((max_size, state_dim)) self.reward = np.zeros((max_size, 1)) self.not_done = np.zeros((max_size, 1)) self.tree = SumTree(max_size) self.max_priority = 1.0 self.beta = 0.4 self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") def add(self, state, action, next_state, reward, done): self.state[self.ptr] = state self.action[self.ptr] = action self.next_state[self.ptr] = next_state self.reward[self.ptr] = reward self.not_done[self.ptr] = 1. - done self.tree.set(self.ptr, self.max_priority) self.ptr = (self.ptr + 1) % self.max_size self.size = min(self.size + 1, self.max_size) def sample(self, batch_size): ind = self.tree.sample(batch_size) weights = self.tree.levels[-1][ind] ** -self.beta weights /= weights.max() self.beta = min(self.beta + 2e-7, 1) # Hardcoded: 0.4 + 2e-7 * 3e6 = 1.0. Only used by PER. return ( torch.FloatTensor(self.state[ind]).to(self.device), torch.FloatTensor(self.action[ind]).to(self.device), torch.FloatTensor(self.next_state[ind]).to(self.device), torch.FloatTensor(self.reward[ind]).to(self.device), torch.FloatTensor(self.not_done[ind]).to(self.device), ind, torch.FloatTensor(weights).to(self.device).reshape(-1, 1) ) def update_priority(self, ind, priority): self.max_priority = max(priority.max(), self.max_priority) self.tree.batch_set(ind, priority) class SumTree(object): def __init__(self, max_size): self.levels = [np.zeros(1)] # Tree construction # Double the number of nodes at each level level_size = 1 while level_size < max_size: level_size *= 2 self.levels.append(np.zeros(level_size)) # Batch binary search through sum tree # Sample a priority between 0 and the max priority # and then search the tree for the corresponding index def sample(self, batch_size): value = np.random.uniform(0, self.levels[0][0], size=batch_size) ind = np.zeros(batch_size, dtype=int) for nodes in self.levels[1:]: ind *= 2 left_sum = nodes[ind] is_greater = np.greater(value, left_sum) # If value > left_sum -> go right (+1), else go left (+0) ind += is_greater # If we go right, we only need to consider the values in the right tree # so we subtract the sum of values in the left tree value -= left_sum * is_greater return ind def set(self, ind, new_priority): priority_diff = new_priority - self.levels[-1][ind] for nodes in self.levels[::-1]: np.add.at(nodes, ind, priority_diff) ind //= 2 def batch_set(self, ind, new_priority): # Confirm we don't increment a node twice ind, unique_ind = np.unique(ind, return_index=True) priority_diff = new_priority[unique_ind] - self.levels[-1][ind] for nodes in self.levels[::-1]: np.add.at(nodes, ind, priority_diff) ind //= 2
4,293
28.410959
93
py
LAP-PAL
LAP-PAL-master/continuous/TD3.py
import copy import numpy as np import torch import torch.nn as nn import torch.nn.functional as F device = torch.device("cuda" if torch.cuda.is_available() else "cpu") class Actor(nn.Module): def __init__(self, state_dim, action_dim, max_action): super(Actor, self).__init__() self.l1 = nn.Linear(state_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, action_dim) self.max_action = max_action def forward(self, state): a = F.relu(self.l1(state)) a = F.relu(self.l2(a)) return self.max_action * torch.tanh(self.l3(a)) class Critic(nn.Module): def __init__(self, state_dim, action_dim): super(Critic, self).__init__() # Q1 architecture self.l1 = nn.Linear(state_dim + action_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, 1) # Q2 architecture self.l4 = nn.Linear(state_dim + action_dim, 256) self.l5 = nn.Linear(256, 256) self.l6 = nn.Linear(256, 1) def forward(self, state, action): sa = torch.cat([state, action], 1) q1 = F.relu(self.l1(sa)) q1 = F.relu(self.l2(q1)) q1 = self.l3(q1) q2 = F.relu(self.l4(sa)) q2 = F.relu(self.l5(q2)) q2 = self.l6(q2) return q1, q2 def Q1(self, state, action): sa = torch.cat([state, action], 1) q1 = F.relu(self.l1(sa)) q1 = F.relu(self.l2(q1)) q1 = self.l3(q1) return q1 class TD3(object): def __init__( self, state_dim, action_dim, max_action, discount=0.99, tau=0.005, policy_noise=0.2, noise_clip=0.5, policy_freq=2 ): self.actor = Actor(state_dim, action_dim, max_action).to(device) self.actor_target = copy.deepcopy(self.actor) self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=3e-4) self.critic = Critic(state_dim, action_dim).to(device) self.critic_target = copy.deepcopy(self.critic) self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=3e-4) self.max_action = max_action self.discount = discount self.tau = tau self.policy_noise = policy_noise self.noise_clip = noise_clip self.policy_freq = policy_freq self.total_it = 0 def select_action(self, state, test=False): state = torch.FloatTensor(state.reshape(1, -1)).to(device) return self.actor(state).cpu().data.numpy().flatten() def train(self, replay_buffer, batch_size=256): self.total_it += 1 # Sample replay buffer state, action, next_state, reward, not_done = replay_buffer.sample(batch_size) with torch.no_grad(): # Select action according to policy and add clipped noise noise = ( torch.randn_like(action) * self.policy_noise ).clamp(-self.noise_clip, self.noise_clip) next_action = ( self.actor_target(next_state) + noise ).clamp(-self.max_action, self.max_action) # Compute the target Q value target_Q1, target_Q2 = self.critic_target(next_state, next_action) target_Q = torch.min(target_Q1, target_Q2) target_Q = reward + not_done * self.discount * target_Q # Get current Q estimates current_Q1, current_Q2 = self.critic(state, action) # Compute critic loss critic_loss = F.mse_loss(current_Q1, target_Q) + F.mse_loss(current_Q2, target_Q) # Optimize the critic self.critic_optimizer.zero_grad() critic_loss.backward() self.critic_optimizer.step() # Delayed policy updates if self.total_it % self.policy_freq == 0: # Compute actor loss actor_loss = -self.critic.Q1(state, self.actor(state)).mean() # Optimize the actor self.actor_optimizer.zero_grad() actor_loss.backward() self.actor_optimizer.step() # Update the frozen target models for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) def save(self, filename): torch.save(self.critic.state_dict(), filename + "_critic") torch.save(self.critic_optimizer.state_dict(), filename + "_critic_optimizer") torch.save(self.actor.state_dict(), filename + "_actor") torch.save(self.actor_optimizer.state_dict(), filename + "_actor_optimizer") def load(self, filename): self.critic.load_state_dict(torch.load(filename + "_critic")) self.critic_optimizer.load_state_dict(torch.load(filename + "_critic_optimizer")) self.actor.load_state_dict(torch.load(filename + "_actor")) self.actor_optimizer.load_state_dict(torch.load(filename + "_actor_optimizer"))
4,551
26.756098
93
py
LAP-PAL
LAP-PAL-master/continuous/LAP_TD3.py
import copy import numpy as np import torch import torch.nn as nn import torch.nn.functional as F device = torch.device("cuda" if torch.cuda.is_available() else "cpu") class Actor(nn.Module): def __init__(self, state_dim, action_dim, max_action): super(Actor, self).__init__() self.l1 = nn.Linear(state_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, action_dim) self.max_action = max_action def forward(self, state): a = F.relu(self.l1(state)) a = F.relu(self.l2(a)) return self.max_action * torch.tanh(self.l3(a)) def act(self, state): a = F.relu(self.l1(state)) a = F.relu(self.l2(a)) a = self.l3(a) return self.max_action * torch.tanh(a), a class Critic(nn.Module): def __init__(self, state_dim, action_dim): super(Critic, self).__init__() # Q1 architecture self.l1 = nn.Linear(state_dim + action_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, 1) # Q2 architecture self.l4 = nn.Linear(state_dim + action_dim, 256) self.l5 = nn.Linear(256, 256) self.l6 = nn.Linear(256, 1) def forward(self, state, action): sa = torch.cat([state, action], 1) q1 = F.relu(self.l1(sa)) q1 = F.relu(self.l2(q1)) q1 = self.l3(q1) q2 = F.relu(self.l4(sa)) q2 = F.relu(self.l5(q2)) q2 = self.l6(q2) return q1, q2 def Q1(self, state, action): sa = torch.cat([state, action], 1) q1 = F.relu(self.l1(sa)) q1 = F.relu(self.l2(q1)) q1 = self.l3(q1) return q1 class LAP_TD3(object): def __init__( self, state_dim, action_dim, max_action, discount=0.99, tau=0.005, policy_noise=0.2, noise_clip=0.5, policy_freq=2, alpha=0.4, min_priority=1 ): self.actor = Actor(state_dim, action_dim, max_action).to(device) self.actor_target = copy.deepcopy(self.actor) self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=3e-4) self.critic = Critic(state_dim, action_dim).to(device) self.critic_target = copy.deepcopy(self.critic) self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=3e-4) self.max_action = max_action self.discount = discount self.tau = tau self.policy_noise = policy_noise self.noise_clip = noise_clip self.policy_freq = policy_freq self.alpha = alpha self.min_priority = min_priority self.total_it = 0 def select_action(self, state, test=False): state = torch.FloatTensor(state.reshape(1, -1)).to(device) return self.actor(state).cpu().data.numpy().flatten() def train(self, replay_buffer, batch_size=256): self.total_it += 1 # Sample replay buffer state, action, next_state, reward, not_done, ind, weights = replay_buffer.sample(batch_size) with torch.no_grad(): # Select action according to policy and add clipped noise noise = ( torch.randn_like(action) * self.policy_noise ).clamp(-self.noise_clip, self.noise_clip) next_action = ( self.actor_target(next_state) + noise ).clamp(-self.max_action, self.max_action) # Compute the target Q value target_Q1, target_Q2 = self.critic_target(next_state, next_action) target_Q = torch.min(target_Q1, target_Q2) target_Q = reward + not_done * self.discount * target_Q # Get current Q estimates current_Q1, current_Q2 = self.critic(state, action) td_loss1 = (current_Q1 - target_Q).abs() td_loss2 = (current_Q2 - target_Q).abs() # Compute critic loss critic_loss = self.huber(td_loss1) + self.huber(td_loss2) # Optimize the critic self.critic_optimizer.zero_grad() critic_loss.backward() self.critic_optimizer.step() priority = torch.max(td_loss1, td_loss2).clamp(min=self.min_priority).pow(self.alpha).cpu().data.numpy().flatten() replay_buffer.update_priority(ind, priority) # Delayed policy updates if self.total_it % self.policy_freq == 0: actor_loss = -self.critic.Q1(state, self.actor(state)).mean() # Optimize the actor self.actor_optimizer.zero_grad() actor_loss.backward() self.actor_optimizer.step() # Update the frozen target models for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) def huber(self, x): return torch.where(x < self.min_priority, 0.5 * x.pow(2), self.min_priority * x).mean() def save(self, filename): torch.save(self.critic.state_dict(), filename + "_critic") torch.save(self.critic_optimizer.state_dict(), filename + "_critic_optimizer") torch.save(self.actor.state_dict(), filename + "_actor") torch.save(self.actor_optimizer.state_dict(), filename + "_actor_optimizer") def load(self, filename): self.critic.load_state_dict(torch.load(filename + "_critic")) self.critic_optimizer.load_state_dict(torch.load(filename + "_critic_optimizer")) self.actor.load_state_dict(torch.load(filename + "_actor")) self.actor_optimizer.load_state_dict(torch.load(filename + "_actor_optimizer"))
5,118
26.67027
116
py
LAP-PAL
LAP-PAL-master/discrete/main.py
import argparse import copy import importlib import json import os import numpy as np import torch import DDQN import PER_DDQN import LAP_DDQN import PAL_DDQN import utils def main(env, replay_buffer, is_atari, state_dim, num_actions, args, parameters, device): # Initialize and load policy kwargs = { "is_atari": is_atari, "num_actions": num_actions, "state_dim": state_dim, "device": device, "discount": parameters["discount"], "optimizer": parameters["optimizer"], "optimizer_parameters": parameters["optimizer_parameters"], "polyak_target_update": parameters["polyak_target_update"], "target_update_frequency": parameters["target_update_freq"], "tau": parameters["tau"], "initial_eps": parameters["initial_eps"], "end_eps": parameters["end_eps"], "eps_decay_period": parameters["eps_decay_period"], "eval_eps": parameters["eval_eps"] } if args.algorithm == "DDQN": policy = DDQN.DDQN(**kwargs) elif args.algorithm == "PER_DDQN": policy = PER_DDQN.PER_DDQN(**kwargs) kwargs["alpha"] = parameters["alpha"] kwargs["min_priority"] = parameters["min_priority"] if args.algorithm == "LAP_DDQN": policy = LAP_DDQN.LAP_DDQN(**kwargs) elif args.algorithm == "PAL_DDQN": policy = PAL_DDQN.PAL_DDQN(**kwargs) evaluations = [] state, done = env.reset(), False episode_start = True episode_reward = 0 episode_timesteps = 0 episode_num = 0 # Interact with the environment for max_timesteps for t in range(int(args.max_timesteps)): episode_timesteps += 1 #if args.train_behavioral: if t < parameters["start_timesteps"]: action = env.action_space.sample() else: action = policy.select_action(np.array(state)) # Perform action and log results next_state, reward, done, info = env.step(action) episode_reward += reward # Only consider "done" if episode terminates due to failure condition done_float = float(done) if episode_timesteps < env._max_episode_steps else 0 # For atari, info[0] = clipped reward, info[1] = done_float if is_atari: reward = info[0] done_float = info[1] # Store data in replay buffer replay_buffer.add(state, action, next_state, reward, done_float, done, episode_start) state = copy.copy(next_state) episode_start = False # Train agent after collecting sufficient data if t >= parameters["start_timesteps"] and (t + 1) % parameters["train_freq"] == 0: policy.train(replay_buffer) if done: # +1 to account for 0 indexing. +0 on ep_timesteps since it will increment +1 even if done=True print(f"Total T: {t+1} Episode Num: {episode_num+1} Episode T: {episode_timesteps} Reward: {episode_reward:.3f}") # Reset environment state, done = env.reset(), False episode_start = True episode_reward = 0 episode_timesteps = 0 episode_num += 1 # Evaluate episode if (t + 1) % parameters["eval_freq"] == 0: evaluations.append(eval_policy(policy, args.env, args.seed)) np.save(f"./results/{setting}.npy", evaluations) # Runs policy for X episodes and returns average reward # A fixed seed is used for the eval environment def eval_policy(policy, env_name, seed, eval_episodes=10): eval_env, _, _, _ = utils.make_env(env_name, atari_preprocessing) eval_env.seed(seed + 100) avg_reward = 0. for _ in range(eval_episodes): state, done = eval_env.reset(), False while not done: action = policy.select_action(np.array(state), eval=True) state, reward, done, _ = eval_env.step(action) avg_reward += reward avg_reward /= eval_episodes print("---------------------------------------") print(f"Evaluation over {eval_episodes} episodes: {avg_reward:.3f}") print("---------------------------------------") return avg_reward if __name__ == "__main__": # Atari Specific atari_preprocessing = { "frame_skip": 4, "frame_size": 84, "state_history": 4, "done_on_life_loss": False, "reward_clipping": True, "max_episode_timesteps": 27e3 } atari_parameters = { # LAP/PAL "alpha": 0.6, "min_priority": 1e-2, # Exploration "start_timesteps": 2e4, "initial_eps": 1, "end_eps": 1e-2, "eps_decay_period": 25e4, # Evaluation "eval_freq": 5e4, "eval_eps": 1e-3, # Learning "discount": 0.99, "buffer_size": 1e6, "batch_size": 32, "optimizer": "RMSprop", "optimizer_parameters": { "lr": 0.0000625, "alpha": 0.95, "centered": True, "eps": 0.00001 }, "train_freq": 4, "polyak_target_update": False, "target_update_freq": 8e3, "tau": 1 } regular_parameters = { # LAP/PAL "alpha": 0.4, "min_priority": 1, # Exploration "start_timesteps": 1e3, "initial_eps": 0.1, "end_eps": 0.1, "eps_decay_period": 1, # Evaluation "eval_freq": 5e3, "eval_eps": 0, # Learning "discount": 0.99, "buffer_size": 1e6, "batch_size": 64, "optimizer": "Adam", "optimizer_parameters": { "lr": 3e-4 }, "train_freq": 1, "polyak_target_update": True, "target_update_freq": 1, "tau": 0.005 } # Load parameters parser = argparse.ArgumentParser() parser.add_argument("--algorithm", default="LAP_DDQN") # OpenAI gym environment name parser.add_argument("--env", default="PongNoFrameskip-v0") # OpenAI gym environment name #PongNoFrameskip-v0 parser.add_argument("--seed", default=0, type=int) # Sets Gym, PyTorch and Numpy seeds parser.add_argument("--buffer_name", default="Default") # Prepends name to filename parser.add_argument("--max_timesteps", default=50e6, type=int) # Max time steps to run environment or train for args = parser.parse_args() print("---------------------------------------") print(f"Setting: Algorithm: {args.algorithm}, Env: {args.env}, Seed: {args.seed}") print("---------------------------------------") setting = f"{args.algorithm}_{args.env}_{args.seed}" if not os.path.exists("./results"): os.makedirs("./results") # Make env and determine properties env, is_atari, state_dim, num_actions = utils.make_env(args.env, atari_preprocessing) parameters = atari_parameters if is_atari else regular_parameters env.seed(args.seed) torch.manual_seed(args.seed) np.random.seed(args.seed) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # Initialize buffer prioritized = True if args.algorithm == "PER_DDQN" or args.algorithm == "LAP_DDQN" else False replay_buffer = utils.ReplayBuffer( state_dim, prioritized, is_atari, atari_preprocessing, parameters["batch_size"], parameters["buffer_size"], device ) main(env, replay_buffer, is_atari, state_dim, num_actions, args, parameters, device)
6,543
26.846809
116
py
LAP-PAL
LAP-PAL-master/discrete/utils.py
import cv2 import gym import numpy as np import torch def ReplayBuffer(state_dim, prioritized, is_atari, atari_preprocessing, batch_size, buffer_size, device): if is_atari: return PrioritizedAtariBuffer(state_dim, atari_preprocessing, batch_size, buffer_size, device, prioritized) else: return PrioritizedStandardBuffer(state_dim, batch_size, buffer_size, device, prioritized) class PrioritizedAtariBuffer(object): def __init__(self, state_dim, atari_preprocessing, batch_size, buffer_size, device, prioritized): self.batch_size = batch_size self.max_size = int(buffer_size) self.device = device self.state_history = atari_preprocessing["state_history"] self.ptr = 0 self.size = 0 self.state = np.zeros(( self.max_size + 1, atari_preprocessing["frame_size"], atari_preprocessing["frame_size"] ), dtype=np.uint8) self.action = np.zeros((self.max_size, 1), dtype=np.int64) self.reward = np.zeros((self.max_size, 1)) # not_done only consider "done" if episode terminates due to failure condition # if episode terminates due to timelimit, the transition is not added to the buffer self.not_done = np.zeros((self.max_size, 1)) self.first_timestep = np.zeros(self.max_size, dtype=np.uint8) self.prioritized = prioritized if self.prioritized: self.tree = SumTree(self.max_size) self.max_priority = 1.0 self.beta = 0.4 def add(self, state, action, next_state, reward, done, env_done, first_timestep): # If dones don't match, env has reset due to timelimit # and we don't add the transition to the buffer if done != env_done: return self.state[self.ptr] = state[0] self.action[self.ptr] = action self.reward[self.ptr] = reward self.not_done[self.ptr] = 1. - done self.first_timestep[self.ptr] = first_timestep self.ptr = (self.ptr + 1) % self.max_size self.size = min(self.size + 1, self.max_size) if self.prioritized: self.tree.set(self.ptr, self.max_priority) def sample(self): ind = self.tree.sample(self.batch_size) if self.prioritized \ else np.random.randint(0, self.size, size=self.batch_size) # Note + is concatenate here state = np.zeros(((self.batch_size, self.state_history) + self.state.shape[1:]), dtype=np.uint8) next_state = np.array(state) state_not_done = 1. next_not_done = 1. for i in range(self.state_history): # Wrap around if the buffer is filled if self.size == self.max_size: j = (ind - i) % self.max_size k = (ind - i + 1) % self.max_size else: j = ind - i k = (ind - i + 1).clip(min=0) # If j == -1, then we set state_not_done to 0. state_not_done *= (j + 1).clip(min=0, max=1).reshape(-1, 1, 1) j = j.clip(min=0) # State should be all 0s if the episode terminated previously state[:, i] = self.state[j] * state_not_done next_state[:, i] = self.state[k] * next_not_done # If this was the first timestep, make everything previous = 0 next_not_done *= state_not_done state_not_done *= (1. - self.first_timestep[j]).reshape(-1, 1, 1) batch = ( torch.ByteTensor(state).to(self.device).float(), torch.LongTensor(self.action[ind]).to(self.device), torch.ByteTensor(next_state).to(self.device).float(), torch.FloatTensor(self.reward[ind]).to(self.device), torch.FloatTensor(self.not_done[ind]).to(self.device) ) if self.prioritized: weights = np.array(self.tree.nodes[-1][ind]) ** -self.beta weights /= weights.max() self.beta = min(self.beta + 4.8e-8, 1) # Hardcoded: 0.4 + 4.8e-8 * 12.5e6 = 1.0. Only used by PER. batch += (ind, torch.FloatTensor(weights).to(self.device).reshape(-1, 1)) return batch def update_priority(self, ind, priority): self.max_priority = max(priority.max(), self.max_priority) self.tree.batch_set(ind, priority) # Replay buffer for standard gym tasks class PrioritizedStandardBuffer(): def __init__(self, state_dim, batch_size, buffer_size, device, prioritized): self.batch_size = batch_size self.max_size = int(buffer_size) self.device = device self.ptr = 0 self.size = 0 self.state = np.zeros((self.max_size, state_dim)) self.action = np.zeros((self.max_size, 1)) self.next_state = np.array(self.state) self.reward = np.zeros((self.max_size, 1)) self.not_done = np.zeros((self.max_size, 1)) self.prioritized = prioritized if self.prioritized: self.tree = SumTree(self.max_size) self.max_priority = 1.0 self.beta = 0.4 def add(self, state, action, next_state, reward, done, env_done, first_timestep): self.state[self.ptr] = state self.action[self.ptr] = action self.next_state[self.ptr] = next_state self.reward[self.ptr] = reward self.not_done[self.ptr] = 1. - done if self.prioritized: self.tree.set(self.ptr, self.max_priority) self.ptr = (self.ptr + 1) % self.max_size self.size = min(self.size + 1, self.max_size) def sample(self): ind = self.tree.sample(self.batch_size) if self.prioritized \ else np.random.randint(0, self.size, size=self.batch_size) batch = ( torch.FloatTensor(self.state[ind]).to(self.device), torch.LongTensor(self.action[ind]).to(self.device), torch.FloatTensor(self.next_state[ind]).to(self.device), torch.FloatTensor(self.reward[ind]).to(self.device), torch.FloatTensor(self.not_done[ind]).to(self.device) ) if self.prioritized: weights = np.array(self.tree.nodes[-1][ind]) ** -self.beta weights /= weights.max() self.beta = min(self.beta + 2e-7, 1) # Hardcoded: 0.4 + 2e-7 * 3e6 = 1.0. Only used by PER. batch += (ind, torch.FloatTensor(weights).to(self.device).reshape(-1, 1)) return batch def update_priority(self, ind, priority): self.max_priority = max(priority.max(), self.max_priority) self.tree.batch_set(ind, priority) class SumTree(object): def __init__(self, max_size): self.nodes = [] # Tree construction # Double the number of nodes at each level level_size = 1 for _ in range(int(np.ceil(np.log2(max_size))) + 1): nodes = np.zeros(level_size) self.nodes.append(nodes) level_size *= 2 # Batch binary search through sum tree # Sample a priority between 0 and the max priority # and then search the tree for the corresponding index def sample(self, batch_size): query_value = np.random.uniform(0, self.nodes[0][0], size=batch_size) node_index = np.zeros(batch_size, dtype=int) for nodes in self.nodes[1:]: node_index *= 2 left_sum = nodes[node_index] is_greater = np.greater(query_value, left_sum) # If query_value > left_sum -> go right (+1), else go left (+0) node_index += is_greater # If we go right, we only need to consider the values in the right tree # so we subtract the sum of values in the left tree query_value -= left_sum * is_greater return node_index def set(self, node_index, new_priority): priority_diff = new_priority - self.nodes[-1][node_index] for nodes in self.nodes[::-1]: np.add.at(nodes, node_index, priority_diff) node_index //= 2 def batch_set(self, node_index, new_priority): # Confirm we don't increment a node twice node_index, unique_index = np.unique(node_index, return_index=True) priority_diff = new_priority[unique_index] - self.nodes[-1][node_index] for nodes in self.nodes[::-1]: np.add.at(nodes, node_index, priority_diff) node_index //= 2 # Atari Preprocessing # Code is based on https://github.com/openai/gym/blob/master/gym/wrappers/atari_preprocessing.py class AtariPreprocessing(object): def __init__( self, env, frame_skip=4, frame_size=84, state_history=4, done_on_life_loss=False, reward_clipping=True, # Clips to a range of -1,1 max_episode_timesteps=27000 ): self.env = env.env self.done_on_life_loss = done_on_life_loss self.frame_skip = frame_skip self.frame_size = frame_size self.reward_clipping = reward_clipping self._max_episode_steps = max_episode_timesteps self.observation_space = np.zeros((frame_size, frame_size)) self.action_space = self.env.action_space self.lives = 0 self.episode_length = 0 # Tracks previous 2 frames self.frame_buffer = np.zeros( (2, self.env.observation_space.shape[0], self.env.observation_space.shape[1]), dtype=np.uint8 ) # Tracks previous 4 states self.state_buffer = np.zeros((state_history, frame_size, frame_size), dtype=np.uint8) def reset(self): self.env.reset() self.lives = self.env.ale.lives() self.episode_length = 0 self.env.ale.getScreenGrayscale(self.frame_buffer[0]) self.frame_buffer[1] = 0 self.state_buffer[0] = self.adjust_frame() self.state_buffer[1:] = 0 return self.state_buffer # Takes single action is repeated for frame_skip frames (usually 4) # Reward is accumulated over those frames def step(self, action): total_reward = 0. self.episode_length += 1 for frame in range(self.frame_skip): _, reward, done, _ = self.env.step(action) total_reward += reward if self.done_on_life_loss: crt_lives = self.env.ale.lives() done = True if crt_lives < self.lives else done self.lives = crt_lives if done: break # Second last and last frame f = frame + 2 - self.frame_skip if f >= 0: self.env.ale.getScreenGrayscale(self.frame_buffer[f]) self.state_buffer[1:] = self.state_buffer[:-1] self.state_buffer[0] = self.adjust_frame() done_float = float(done) if self.episode_length >= self._max_episode_steps: done = True return self.state_buffer, total_reward, done, [np.clip(total_reward, -1, 1), done_float] def adjust_frame(self): # Take maximum over last two frames np.maximum( self.frame_buffer[0], self.frame_buffer[1], out=self.frame_buffer[0] ) # Resize image = cv2.resize( self.frame_buffer[0], (self.frame_size, self.frame_size), interpolation=cv2.INTER_AREA ) return np.array(image, dtype=np.uint8) def seed(self, seed): self.env.seed(seed) # Create environment, add wrapper if necessary and create env_properties def make_env(env_name, atari_preprocessing): env = gym.make(env_name) is_atari = gym.envs.registry.spec(env_name).entry_point == 'gym.envs.atari:AtariEnv' env = AtariPreprocessing(env, **atari_preprocessing) if is_atari else env state_dim = ( atari_preprocessing["state_history"], atari_preprocessing["frame_size"], atari_preprocessing["frame_size"] ) if is_atari else env.observation_space.shape[0] return ( env, is_atari, state_dim, env.action_space.n )
10,436
28.483051
109
py
LAP-PAL
LAP-PAL-master/discrete/PER_DDQN.py
import copy import numpy as np import torch import torch.nn as nn import torch.nn.functional as F # Used for Atari class Conv_Q(nn.Module): def __init__(self, frames, num_actions): super(Conv_Q, self).__init__() self.c1 = nn.Conv2d(frames, 32, kernel_size=8, stride=4) self.c2 = nn.Conv2d(32, 64, kernel_size=4, stride=2) self.c3 = nn.Conv2d(64, 64, kernel_size=3, stride=1) self.l1 = nn.Linear(3136, 512) self.l2 = nn.Linear(512, num_actions) def forward(self, state): q = F.relu(self.c1(state)) q = F.relu(self.c2(q)) q = F.relu(self.c3(q)) q = F.relu(self.l1(q.reshape(-1, 3136))) return self.l2(q) # Used for Box2D / Toy problems class FC_Q(nn.Module): def __init__(self, state_dim, num_actions): super(FC_Q, self).__init__() self.l1 = nn.Linear(state_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, num_actions) def forward(self, state): q = F.relu(self.l1(state)) q = F.relu(self.l2(q)) return self.l3(q) class PER_DDQN(object): def __init__( self, is_atari, num_actions, state_dim, device, discount=0.99, optimizer="Adam", optimizer_parameters={}, polyak_target_update=False, target_update_frequency=8e3, tau=0.005, initial_eps = 1, end_eps = 0.001, eps_decay_period = 25e4, eval_eps=0.001, ): self.device = device # Determine network type self.Q = Conv_Q(4, num_actions).to(self.device) if is_atari else FC_Q(state_dim, num_actions).to(self.device) self.Q_target = copy.deepcopy(self.Q) self.Q_optimizer = getattr(torch.optim, optimizer)(self.Q.parameters(), **optimizer_parameters) self.discount = discount # Target update rule self.maybe_update_target = self.polyak_target_update if polyak_target_update else self.copy_target_update self.target_update_frequency = target_update_frequency self.tau = tau # Decay for eps self.initial_eps = initial_eps self.end_eps = end_eps self.slope = (self.end_eps - self.initial_eps) / eps_decay_period # Evaluation hyper-parameters self.state_shape = (-1, 4, 84, 84) if is_atari else (-1, state_dim) ### need to pass framesize self.eval_eps = eval_eps self.num_actions = num_actions # Number of training iterations self.iterations = 0 def select_action(self, state, eval=False): eps = self.eval_eps if eval \ else max(self.slope * self.iterations + self.initial_eps, self.end_eps) # Select action according to policy with probability (1-eps) # otherwise, select random action if np.random.uniform(0,1) > eps: with torch.no_grad(): state = torch.FloatTensor(state).reshape(self.state_shape).to(self.device) return int(self.Q(state).argmax(1)) else: return np.random.randint(self.num_actions) def train(self, replay_buffer): # Sample replay buffer state, action, next_state, reward, done, ind, weights = replay_buffer.sample() # Compute the target Q value with torch.no_grad(): next_action = self.Q(next_state).argmax(1, keepdim=True) target_Q = ( reward + done * self.discount * self.Q_target(next_state).gather(1, next_action).reshape(-1, 1) ) # Get current Q estimate current_Q = self.Q(state).gather(1, action) td_loss = (current_Q - target_Q).abs() Q_loss = self.huber(td_loss) # Optimize the Q network self.Q_optimizer.zero_grad() Q_loss.backward() self.Q_optimizer.step() # Update target network by polyak or full copy every X iterations. self.iterations += 1 self.maybe_update_target() priority = td_loss.pow(0.6).clamp(min=0.06309573444).cpu().data.numpy().flatten() replay_buffer.reinsert(ind, priority) def train(self, replay_buffer): # Sample replay buffer state, action, next_state, reward, done, ind, weights = replay_buffer.sample() # Compute the target Q value with torch.no_grad(): next_action = self.Q(next_state).argmax(1, keepdim=True) target_Q = ( reward + done * self.discount * self.Q_target(next_state).gather(1, next_action).reshape(-1, 1) ) # Get current Q estimate current_Q = self.Q(state).gather(1, action) # Compute Q loss Q_loss = (weights * F.smooth_l1_loss(current_Q, target_Q, reduction='none')).mean() # Optimize the Q network self.Q_optimizer.zero_grad() Q_loss.backward() self.Q_optimizer.step() # Update target network by polyak or full copy every X iterations. self.iterations += 1 self.maybe_update_target() priority = ((current_Q - target_Q).abs() + 1e-10).pow(0.6).cpu().data.numpy().flatten() replay_buffer.update_priority(ind, priority) def polyak_target_update(self): for param, target_param in zip(self.Q.parameters(), self.Q_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) def copy_target_update(self): if self.iterations % self.target_update_frequency == 0: self.Q_target.load_state_dict(self.Q.state_dict()) def save(self, filename): torch.save(self.iterations, filename + "iterations") torch.save(self.Q.state_dict(), f"{filename}Q_{self.iterations}") torch.save(self.Q_optimizer.state_dict(), filename + "optimizer") def load(self, filename): self.iterations = torch.load(filename + "iterations") self.Q.load_state_dict(torch.load(f"{filename}Q_{self.iterations}")) self.Q_target = copy.deepcopy(self.Q) self.Q_optimizer.load_state_dict(torch.load(filename + "optimizer"))
5,366
28.010811
111
py
LAP-PAL
LAP-PAL-master/discrete/PAL_DDQN.py
import copy import numpy as np import torch import torch.nn as nn import torch.nn.functional as F # Used for Atari class Conv_Q(nn.Module): def __init__(self, frames, num_actions): super(Conv_Q, self).__init__() self.c1 = nn.Conv2d(frames, 32, kernel_size=8, stride=4) self.c2 = nn.Conv2d(32, 64, kernel_size=4, stride=2) self.c3 = nn.Conv2d(64, 64, kernel_size=3, stride=1) self.l1 = nn.Linear(3136, 512) self.l2 = nn.Linear(512, num_actions) def forward(self, state): q = F.relu(self.c1(state)) q = F.relu(self.c2(q)) q = F.relu(self.c3(q)) q = F.relu(self.l1(q.reshape(-1, 3136))) return self.l2(q) # Used for Box2D / Toy problems class FC_Q(nn.Module): def __init__(self, state_dim, num_actions): super(FC_Q, self).__init__() self.l1 = nn.Linear(state_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, num_actions) def forward(self, state): q = F.relu(self.l1(state)) q = F.relu(self.l2(q)) return self.l3(q) class PAL_DDQN(object): def __init__( self, is_atari, num_actions, state_dim, device, discount=0.99, optimizer="Adam", optimizer_parameters={}, polyak_target_update=False, target_update_frequency=8e3, tau=0.005, initial_eps = 1, end_eps = 0.001, eps_decay_period = 25e4, eval_eps=0.001, alpha=0.6, min_priority=1e-2 ): self.device = device # Determine network type self.Q = Conv_Q(4, num_actions).to(self.device) if is_atari else FC_Q(state_dim, num_actions).to(self.device) self.Q_target = copy.deepcopy(self.Q) self.Q_optimizer = getattr(torch.optim, optimizer)(self.Q.parameters(), **optimizer_parameters) self.discount = discount # PAL hyper-parameters self.alpha = alpha self.min_priority = min_priority # Target update rule self.maybe_update_target = self.polyak_target_update if polyak_target_update else self.copy_target_update self.target_update_frequency = target_update_frequency self.tau = tau # Decay for eps self.initial_eps = initial_eps self.end_eps = end_eps self.slope = (self.end_eps - self.initial_eps) / eps_decay_period # Evaluation hyper-parameters self.state_shape = (-1,) + state_dim if is_atari else (-1, state_dim) self.eval_eps = eval_eps self.num_actions = num_actions # Number of training iterations self.iterations = 0 def select_action(self, state, eval=False): eps = self.eval_eps if eval \ else max(self.slope * self.iterations + self.initial_eps, self.end_eps) # Select action according to policy with probability (1-eps) # otherwise, select random action if np.random.uniform(0,1) > eps: with torch.no_grad(): state = torch.FloatTensor(state).reshape(self.state_shape).to(self.device) return int(self.Q(state).argmax(1)) else: return np.random.randint(self.num_actions) def train(self, replay_buffer): # Sample replay buffer state, action, next_state, reward, done = replay_buffer.sample() # Compute the target Q value with torch.no_grad(): next_action = self.Q(next_state).argmax(1, keepdim=True) target_Q = ( reward + done * self.discount * self.Q_target(next_state).gather(1, next_action).reshape(-1, 1) ) # Get current Q estimate current_Q = self.Q(state).gather(1, action) td_loss = (current_Q - target_Q).abs() weight = td_loss.clamp(min=self.min_priority).pow(self.alpha).mean().detach() # Compute critic loss Q_loss = self.PAL(td_loss)/weight.detach() # Optimize the Q self.Q_optimizer.zero_grad() Q_loss.backward() self.Q_optimizer.step() # Update target network by polyak or full copy every X iterations. self.iterations += 1 self.maybe_update_target() def PAL(self, x): return torch.where( x.abs() < self.min_priority, (self.min_priority ** self.alpha) * 0.5 * x.pow(2), self.min_priority * x.abs().pow(1. + self.alpha)/(1. + self.alpha) ).mean() def polyak_target_update(self): for param, target_param in zip(self.Q.parameters(), self.Q_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) def copy_target_update(self): if self.iterations % self.target_update_frequency == 0: self.Q_target.load_state_dict(self.Q.state_dict()) def save(self, filename): torch.save(self.iterations, filename + "iterations") torch.save(self.Q.state_dict(), f"{filename}Q_{self.iterations}") torch.save(self.Q_optimizer.state_dict(), filename + "optimizer") def load(self, filename): self.iterations = torch.load(filename + "iterations") self.Q.load_state_dict(torch.load(f"{filename}Q_{self.iterations}")) self.Q_target = copy.deepcopy(self.Q) self.Q_optimizer.load_state_dict(torch.load(filename + "optimizer"))
4,712
27.053571
111
py
LAP-PAL
LAP-PAL-master/discrete/LAP_DDQN.py
import copy import numpy as np import torch import torch.nn as nn import torch.nn.functional as F # Used for Atari class Conv_Q(nn.Module): def __init__(self, frames, num_actions): super(Conv_Q, self).__init__() self.c1 = nn.Conv2d(frames, 32, kernel_size=8, stride=4) self.c2 = nn.Conv2d(32, 64, kernel_size=4, stride=2) self.c3 = nn.Conv2d(64, 64, kernel_size=3, stride=1) self.l1 = nn.Linear(3136, 512) self.l2 = nn.Linear(512, num_actions) def forward(self, state): q = F.relu(self.c1(state)) q = F.relu(self.c2(q)) q = F.relu(self.c3(q)) q = F.relu(self.l1(q.reshape(-1, 3136))) return self.l2(q) # Used for Box2D / Toy problems class FC_Q(nn.Module): def __init__(self, state_dim, num_actions): super(FC_Q, self).__init__() self.l1 = nn.Linear(state_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, num_actions) def forward(self, state): q = F.relu(self.l1(state)) q = F.relu(self.l2(q)) return self.l3(q) class LAP_DDQN(object): def __init__( self, is_atari, num_actions, state_dim, device, discount=0.99, optimizer="Adam", optimizer_parameters={}, polyak_target_update=False, target_update_frequency=8e3, tau=0.005, initial_eps = 1, end_eps = 0.001, eps_decay_period = 25e4, eval_eps=0.001, alpha=0.6, min_priority=1e-2 ): self.device = device # Determine network type self.Q = Conv_Q(4, num_actions).to(self.device) if is_atari else FC_Q(state_dim, num_actions).to(self.device) self.Q_target = copy.deepcopy(self.Q) self.Q_optimizer = getattr(torch.optim, optimizer)(self.Q.parameters(), **optimizer_parameters) self.discount = discount # LAP hyper-parameters self.alpha = alpha self.min_priority = min_priority # Target update rule self.maybe_update_target = self.polyak_target_update if polyak_target_update else self.copy_target_update self.target_update_frequency = target_update_frequency self.tau = tau # Decay for eps self.initial_eps = initial_eps self.end_eps = end_eps self.slope = (self.end_eps - self.initial_eps) / eps_decay_period # Evaluation hyper-parameters self.state_shape = (-1,) + state_dim if is_atari else (-1, state_dim) self.eval_eps = eval_eps self.num_actions = num_actions # Number of training iterations self.iterations = 0 def select_action(self, state, eval=False): eps = self.eval_eps if eval \ else max(self.slope * self.iterations + self.initial_eps, self.end_eps) # Select action according to policy with probability (1-eps) # otherwise, select random action if np.random.uniform(0,1) > eps: with torch.no_grad(): state = torch.FloatTensor(state).reshape(self.state_shape).to(self.device) return int(self.Q(state).argmax(1)) else: return np.random.randint(self.num_actions) def train(self, replay_buffer): # Sample replay buffer state, action, next_state, reward, done, ind, weights = replay_buffer.sample() # Compute the target Q value with torch.no_grad(): next_action = self.Q(next_state).argmax(1, keepdim=True) target_Q = ( reward + done * self.discount * self.Q_target(next_state).gather(1, next_action).reshape(-1, 1) ) # Get current Q estimate current_Q = self.Q(state).gather(1, action) td_loss = (current_Q - target_Q).abs() Q_loss = self.huber(td_loss) # Optimize the Q network self.Q_optimizer.zero_grad() Q_loss.backward() self.Q_optimizer.step() # Update target network by polyak or full copy every X iterations. self.iterations += 1 self.maybe_update_target() priority = td_loss.clamp(min=self.min_priority).pow(self.alpha).cpu().data.numpy().flatten() replay_buffer.update_priority(ind, priority) def huber(self, x): return torch.where(x < self.min_priority, 0.5 * x.pow(2), self.min_priority * x).mean() def polyak_target_update(self): for param, target_param in zip(self.Q.parameters(), self.Q_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) def copy_target_update(self): if self.iterations % self.target_update_frequency == 0: self.Q_target.load_state_dict(self.Q.state_dict()) def save(self, filename): torch.save(self.iterations, filename + "iterations") torch.save(self.Q.state_dict(), f"{filename}Q_{self.iterations}") torch.save(self.Q_optimizer.state_dict(), filename + "optimizer") def load(self, filename): self.iterations = torch.load(filename + "iterations") self.Q.load_state_dict(torch.load(f"{filename}Q_{self.iterations}")) self.Q_target = copy.deepcopy(self.Q) self.Q_optimizer.load_state_dict(torch.load(filename + "optimizer"))
4,658
27.408537
111
py
LAP-PAL
LAP-PAL-master/discrete/DDQN.py
import copy import numpy as np import torch import torch.nn as nn import torch.nn.functional as F # Used for Atari class Conv_Q(nn.Module): def __init__(self, frames, num_actions): super(Conv_Q, self).__init__() self.c1 = nn.Conv2d(frames, 32, kernel_size=8, stride=4) self.c2 = nn.Conv2d(32, 64, kernel_size=4, stride=2) self.c3 = nn.Conv2d(64, 64, kernel_size=3, stride=1) self.l1 = nn.Linear(3136, 512) self.l2 = nn.Linear(512, num_actions) def forward(self, state): q = F.relu(self.c1(state)) q = F.relu(self.c2(q)) q = F.relu(self.c3(q)) q = F.relu(self.l1(q.reshape(-1, 3136))) return self.l2(q) # Used for Box2D / Toy problems class FC_Q(nn.Module): def __init__(self, state_dim, num_actions): super(FC_Q, self).__init__() self.l1 = nn.Linear(state_dim, 256) self.l2 = nn.Linear(256, 256) self.l3 = nn.Linear(256, num_actions) def forward(self, state): q = F.relu(self.l1(state)) q = F.relu(self.l2(q)) return self.l3(q) class DDQN(object): def __init__( self, is_atari, num_actions, state_dim, device, discount=0.99, optimizer="Adam", optimizer_parameters={}, polyak_target_update=False, target_update_frequency=8e3, tau=0.005, initial_eps = 1, end_eps = 0.001, eps_decay_period = 25e4, eval_eps=0.001, ): self.device = device # Determine network type self.Q = Conv_Q(4, num_actions).to(self.device) if is_atari else FC_Q(state_dim, num_actions).to(self.device) self.Q_target = copy.deepcopy(self.Q) self.Q_optimizer = getattr(torch.optim, optimizer)(self.Q.parameters(), **optimizer_parameters) self.discount = discount # Target update rule self.maybe_update_target = self.polyak_target_update if polyak_target_update else self.copy_target_update self.target_update_frequency = target_update_frequency self.tau = tau # Decay for eps self.initial_eps = initial_eps self.end_eps = end_eps self.slope = (self.end_eps - self.initial_eps) / eps_decay_period # Evaluation hyper-parameters self.state_shape = (-1,) + state_dim if is_atari else (-1, state_dim) self.eval_eps = eval_eps self.num_actions = num_actions # Number of training iterations self.iterations = 0 def select_action(self, state, eval=False): eps = self.eval_eps if eval \ else max(self.slope * self.iterations + self.initial_eps, self.end_eps) # Select action according to policy with probability (1-eps) # otherwise, select random action if np.random.uniform(0,1) > eps: with torch.no_grad(): state = torch.FloatTensor(state).reshape(self.state_shape).to(self.device) return int(self.Q(state).argmax(1)) else: return np.random.randint(self.num_actions) def train(self, replay_buffer): # Sample replay buffer state, action, next_state, reward, done = replay_buffer.sample() # Compute the target Q value with torch.no_grad(): next_action = self.Q(next_state).argmax(1, keepdim=True) target_Q = ( reward + done * self.discount * self.Q_target(next_state).gather(1, next_action).reshape(-1, 1) ) # Get current Q estimate current_Q = self.Q(state).gather(1, action) # Compute Q loss Q_loss = F.smooth_l1_loss(current_Q, target_Q) # Optimize the Q network self.Q_optimizer.zero_grad() Q_loss.backward() self.Q_optimizer.step() # Update target network by polyak or full copy every X iterations. self.iterations += 1 self.maybe_update_target() def polyak_target_update(self): for param, target_param in zip(self.Q.parameters(), self.Q_target.parameters()): target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data) def copy_target_update(self): if self.iterations % self.target_update_frequency == 0: self.Q_target.load_state_dict(self.Q.state_dict()) def save(self, filename): torch.save(self.iterations, filename + "iterations") torch.save(self.Q.state_dict(), f"{filename}Q_{self.iterations}") torch.save(self.Q_optimizer.state_dict(), filename + "optimizer") def load(self, filename): self.iterations = torch.load(filename + "iterations") self.Q.load_state_dict(torch.load(f"{filename}Q_{self.iterations}")) self.Q_target = copy.deepcopy(self.Q) self.Q_optimizer.load_state_dict(torch.load(filename + "optimizer"))
4,265
27.251656
111
py
sa-nmt
sa-nmt-master/Loss.py
""" This file handles the details of the loss function during training. This includes: loss criterion, training statistics, and memory optimizations. """ from __future__ import division import time import sys import math import torch import torch.nn as nn def nmt_criterion(vocab_size, pad_id=0): """ Construct the standard NMT Criterion """ weight = torch.ones(vocab_size) weight[pad_id] = 0 crit = nn.NLLLoss(weight, size_average=False) return crit class Statistics: """ Training loss function statistics. """ def __init__(self, loss=0, n_words=0, n_correct=0): self.loss = loss self.n_words = n_words self.n_correct = n_correct self.start_time = time.time() def update(self, stat): self.loss += stat.loss self.n_words += stat.n_words self.n_correct += stat.n_correct def accuracy(self): return 100 * (self.n_correct / self.n_words) def ppl(self): print(self.loss, self.n_words) return math.exp(min(self.loss / self.n_words, 100)) def elapsed_time(self): return time.time() - self.start_time def output(self, epoch, uidx, max_updates, start): t = self.elapsed_time() print(("Epoch %2d, %5d/%5d; acc: %6.2f; ppl: %6.2f; " + "%3.0f tgt tok/s; %6.0f s elapsed") % (epoch, uidx, max_updates, self.accuracy(), self.ppl(), self.n_words / (t + 1e-5), time.time() - start)) sys.stdout.flush() def log(self, prefix, experiment, optim): t = self.elapsed_time() experiment.add_scalar_value(prefix + "_ppl", self.ppl()) experiment.add_scalar_value(prefix + "_accuracy", self.accuracy()) experiment.add_scalar_value(prefix + "_tgtper", self.n_words / t) experiment.add_scalar_value(prefix + "_lr", optim.lr) @staticmethod def score(loss, scores, targ, pad): pred = scores.max(1)[1] non_padding = targ.ne(pad) num_correct = pred.eq(targ) \ .masked_select(non_padding).int() \ .sum().item() return Statistics(loss, non_padding.int().sum().item(), num_correct) def filter_gen_state(state): for k, v in state.items(): if v is not None: yield k, v def new_split(x, size): xs = [] for u in torch.split(x, size): v = u.detach() if u.requires_grad: v.requires_grad_(True) xs += [v] return tuple(xs) def shards(state, shard_size, eval=False): if eval: yield state else: non_none = dict(filter_gen_state(state)) keys, values = zip(*((k, new_split(v, shard_size)) for k, v in non_none.items())) for shard_tensors in zip(*values): yield dict(zip(keys, shard_tensors)) # Assumed backprop'd variables = [] for i, k in enumerate(keys): dv = [v.grad for v in values[i] if v.grad is not None] if dv: dv = torch.cat(dv) variables += [(state[k], dv)] inputs, grads = zip(*variables) torch.autograd.backward(inputs, grads) class LossCompute: def __init__(self, generator, crit): self.generator = generator self.crit = crit def make_loss_batch(self, outputs, targets): """ Create all the variables that need to be sharded. This needs to match compute loss exactly. """ return {"out": outputs, "target": targets} def compute_loss(self, out, target): def bottle(v): return v.view(-1, v.size(2)) target = target.view(-1) # Standard generator. scores = self.generator(bottle(out)) loss = self.crit(scores, target) scores_data = scores.detach() target = target.clone() # Coverage loss term. stats = Statistics.score(loss.item(), scores_data, target, 0) return loss, stats
4,092
27.227586
77
py
sa-nmt
sa-nmt-master/translate.py
import argparse import torch import modelx as models import infer import string # build args parser parser = argparse.ArgumentParser(description='Training NMT') parser.add_argument('-checkpoint', required=True, help='saved checkpoit.') parser.add_argument('-input', required=True, help='Text file to translate.') parser.add_argument('-output', default='trans.bpe', help='output file') parser.add_argument('-beam_size', default=5, type=int, help="Beam size.") parser.add_argument('-gpuid', default=[], nargs='+', type=int, help="Use CUDA") opt = parser.parse_args() use_gpu = len(opt.gpuid) > 0 if torch.cuda.is_available() and not use_gpu: print("so you should probably run with -gpus 0") checkpoint = torch.load(opt.checkpoint) train_opt = checkpoint['opt'] print('| train configuration') train_opt.min_thres = -5.0 train_opt.max_thres = 7.0 #if train_opt.hard is None: #train_opt.hard = False print(train_opt) model = models.make_base_model(train_opt, use_gpu, checkpoint) if train_opt.encoder_type == "sabrnn": punct_idx = set() for p in set(string.punctuation): if p in model.dicts[0]: punct_idx.add(model.dicts[0][p]) model.encoder.punct(punct_idx) # over-write some options train_opt.beam_size = opt.beam_size agent = infer.Beam(train_opt, model) agent.translate(opt.input, opt.output)
1,416
28.520833
71
py
sa-nmt
sa-nmt-master/extract_tree.py
import argparse import torch from torch.autograd import Variable import modelx as models import networkx as nx from networkx.algorithms.tree import maximum_spanning_arborescence import string # build args parser parser = argparse.ArgumentParser(description='Training NMT') parser.add_argument('-checkpoint', required=True, help='saved checkpoit.') parser.add_argument('-input', required=True, help='Text file to translate.') parser.add_argument('-output', default='tree.txt', help='output file') parser.add_argument('-gpuid', default=[], nargs='+', type=int, help="Use CUDA") opt = parser.parse_args() use_gpu = True if torch.cuda.is_available() and not use_gpu: print("so you should probably run with -gpus 0") checkpoint = torch.load(opt.checkpoint) train_opt = checkpoint['opt'] print('| train configuration') print(train_opt) use_gpu = len(train_opt.gpuid) > 0 model = models.make_base_model(train_opt, use_gpu, checkpoint) if train_opt.encoder_type in ["sabrnn", "fabrnn"]: punct_idx = set() for p in set(string.punctuation): if p in model.dicts[0]: punct_idx.add(model.dicts[0][p]) print('Add punctuation constraint') model.encoder.punct(punct_idx) # get the encoder encoder = model.encoder dicts = model.dicts tt = torch.cuda if use_gpu else torch def encode_string(ss): ss = ss.split() ss = [dicts[0].get(w, 1) for w in ss] ss = Variable(tt.LongTensor(ss).view(-1, 1), volatile=True) return ss def collapse_bpe(s, score): """Collapse BPEed tokens Args: s: a bped sentence score: beta in the paper, a 2D tensor of p(z|s), sum over the last dimension should be 1 """ punct = set(string.punctuation) # (1) identify bpe tokens = s.split() bpe = [] indices = [] punct_idx = [] for i, w in enumerate(tokens): if w in punct: punct_idx += [i] if w.endswith("@@"): bpe += [i] else: if len(bpe) == 0: indices += [[i]] else: bpe += [i] # add the last bped token indices += [bpe] bpe = [] # collapsing from here s_ = [] for bpe in indices: # (1) collapse heads s_.append(score[:, bpe].sum(1).view(-1, 1)) s_ = torch.cat(s_, 1) collapsed_score = [] for bpe in indices: # (2) collapse childs collapsed_score += [s_[bpe, :].sum(0).view(1, -1)] collapsed_score = torch.cat(collapsed_score, 0) s = s.replace("@@ ", "") # the original string return s, collapsed_score def build_graph(score): """Build graph from potential score matrix Args: score: FloatTensor (n, n), score.sum(1) = 1 Returns: a graph object """ # return a list of (parent, child, weight) arcs = [] n = score.size(0) # find the root first val, idx = score.diag().max(0) arcs.append((0, idx[0] + 1, val[0])) for i in range(n): for j in range(n): if i == j: # root on the diagonal continue # arcs.append([0, i+1, score[i, j]]) else: arcs.append([j+1, i+1, score[i, j]]) g = nx.DiGraph() g.add_weighted_edges_from(arcs) return g def mst(score): """Get spaning tree from the adjacent matrix""" g = build_graph(score) mst = maximum_spanning_arborescence(g) tree = [] for e in mst.edges(): head, child = e tree.append('%s->%s' % (head, child)) return ' '.join(tree) def renorm(m): x = m.exp() x = x / x.sum(1, keepdim=True) return x.log() def collapse(): fw = open(opt.output, 'w') for line in open(opt.input): x = encode_string(line) model.encoder(x) if train_opt.encoder_type == 'sabrnn': score = model.encoder.tree_attn.edge_score.squeeze(0) else: score = model.encoder.attn.score.squeeze(0).log() s, score = collapse_bpe(line, score.data) try: tree = mst(score) out = '%s ||| %s\n' % (s.strip(), tree) fw.write(out) except: pass fw.close() collapse()
4,276
26.242038
70
py
sa-nmt
sa-nmt-master/models.py
import torch import torch.nn as nn from torch.autograd import Variable from attention import GlobalAttention, SelfAttention from Utils import aeq from torch.nn.utils.rnn import pack_padded_sequence as pack from torch.nn.utils.rnn import pad_packed_sequence as unpack import math class EncoderBase(nn.Module): """ EncoderBase class for sharing code among various encoder. """ def _check_args(self, input, lengths=None, hidden=None): s_len, n_batch = input.size() if lengths is not None: n_batch_ = len(lengths) # lengths.size() aeq(n_batch, n_batch_) def forward(self, input, lengths=None, hidden=None): """ Args: input (LongTensor): len x batch x nfeat. lengths (LongTensor): batch hidden: Initial hidden state. Returns: hidden_t (Variable): Pair of layers x batch x rnn_size - final encoder state outputs (FloatTensor): len x batch x rnn_size - Memory bank """ raise NotImplementedError class Encoder(EncoderBase): """ The standard RNN encoder. """ def __init__(self, word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True): super(Encoder, self).__init__() num_directions = 2 if bidirectional else 1 assert hidden_size % num_directions == 0 hidden_size = hidden_size // num_directions self.embeddings = nn.Embedding(word_vocab_size, embedding_size, padding_idx=word_padding_idx) self.rnn = nn.LSTM(input_size=embedding_size, hidden_size=hidden_size, num_layers=num_layers, dropout=dropout, bidirectional=bidirectional) def forward(self, input, lengths=None, hidden=None): """ See EncoderBase.forward() for description of args and returns.""" self._check_args(input, lengths, hidden) emb = self.embeddings(input) s_len, batch, emb_dim = emb.size() self.word_embs = emb # for accessing later packed_emb = emb if lengths is not None: packed_emb = pack(emb, lengths) outputs, hidden = self.rnn(packed_emb, hidden) if lengths is not None: outputs = unpack(outputs)[0] return outputs, hidden # Structured Attention and our models class MatrixTree(nn.Module): """Implementation of the matrix-tree theorem for computing marginals of non-projective dependency parsing. This attention layer is used in the paper "Learning Structured Text Representations." """ def __init__(self, eps=1e-5, device=torch.device('cpu')): self.eps = eps super(MatrixTree, self).__init__() self.device = device def forward(self, input, lengths=None): laplacian = input.exp() output = input.clone() output.data.fill_(0) for b in range(input.size(0)): lx = lengths[b] if lengths is not None else input.size(1) input_b = input[b, :lx, :lx] lap = laplacian[b, :lx, :lx].masked_fill( torch.eye(lx).to(self.device).ne(0), 0) lap = -lap + torch.diag(lap.sum(0)) # store roots on diagonal lap[0] = input_b.diag().exp() inv_laplacian = lap.inverse() factor = inv_laplacian.diag().unsqueeze(1)\ .expand(lx, lx).transpose(0, 1) term1 = input_b.exp().mul(factor).clone() term2 = input_b.exp().mul(inv_laplacian.transpose(0, 1)).clone() term1[:, 0] = 0 term2[0] = 0 output_b = term1 - term2 roots_output = input_b.diag().exp().mul( inv_laplacian.transpose(0, 1)[0]) output[b, :lx, :lx] = output_b + torch.diag(roots_output) return output class TreeAttention(nn.Module): """Structured attention class""" def __init__(self, dim, min_thres=-5, max_thres=7, hard=False, device=torch.device('cpu')): super(TreeAttention, self).__init__() self.q = nn.Linear(dim, dim, bias=False) self.k = nn.Linear(dim, dim, bias=False) self.v = nn.Linear(dim, dim, bias=False) self.root_query = nn.Parameter(torch.randn(dim)) self.scale = math.sqrt(1 / dim) self.dtree = MatrixTree() self.min_thres = min_thres self.max_thres = max_thres self.hard = hard self.device = device def forward(self, input, punct_mask=None, lengths=None): s_len, batch, dim = input.size() input = input.contiguous().transpose(0, 1) \ .contiguous().view(-1, dim) q = self.q(input).view(batch, s_len, -1) # (batch, s_len, dim) k = self.k(input).view(batch, s_len, -1) # (batch, s_len, dim) v = self.v(input).view(batch, s_len, -1) # (batch, s_len, dim) _score = torch.bmm(q, k.transpose(1, 2)) # (batch, s_len, s_len) # compute root r_ = self.root_query.view(1, -1, 1).expand(batch, dim, 1) root = torch.bmm(k, r_).squeeze(-1) # (batch, s_len) mask = torch.eye(s_len).to(self.device) score = _score.clone() for b in range(batch): score[b] = _score[b] * mask + torch.diag(root[b]) # normalized score *= self.scale if punct_mask is not None: punct_mask = punct_mask.transpose(0, 1) punct_mask = punct_mask[:, None, :].expand(batch, s_len, s_len) \ .transpose(1, 2) score.data.masked_fill_(punct_mask, -math.inf) score = score.clamp(self.min_thres, self.max_thres) self.edge_score = score.transpose(1, 2) edge_score = self.dtree(score, lengths).transpose(1, 2) # edge_score.sum(2) == 1 if self.hard: y = edge_score.data.new(edge_score.size()).fill_(0) _, max_idx = edge_score.data.max(2) y.scatter_(2, max_idx[:, :, None], 1) hard_edge = (Variable(y) - edge_score).detach() + edge_score edge_score = hard_edge return torch.bmm(edge_score, v) class SAEncoder(Encoder): """ The structured attention RNN encoder. """ def __init__(self, word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True, encode_multi_key=True, min_thres=-5, max_thres=7, hard=False): super(SAEncoder, self).__init__(word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True) self.tree_attn = TreeAttention(hidden_size, min_thres, max_thres, hard) self.encode_multi_key = encode_multi_key if not self.encode_multi_key: self.linear = nn.Linear(hidden_size, hidden_size, bias=False) self.punct_idx = None def punct(self, punct_idx): self.punct_idx = punct_idx def forward(self, input, lengths=None): outputs, hidden = Encoder.forward(self, input, lengths) # find puncts punct_mask = None if self.punct_idx is not None: punct = set(input.data.contiguous().view(-1).tolist()) punct &= self.punct_idx if len(punct) > 0: punct_mask = 0 for p in punct: punct_mask += input.data.eq(p) tree_outputs = self.tree_attn(outputs, punct_mask, lengths) # tree_outputs has size (batch_size, s_len, hidden_size) if not self.encode_multi_key: # compute gate syntax z = self.linear(outputs.view(-1, outputs.size(2))).sigmoid() gtree = z.view_as(outputs) * tree_outputs.transpose(0, 1) outputs = outputs + gtree return outputs, hidden return (outputs, tree_outputs), hidden class FAEncoder(Encoder): """ The flat attention RNN encoder. """ def __init__(self, word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True, encode_multi_key=True): super(FAEncoder, self).__init__(word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, bidirectional=True) self.attn = SelfAttention(hidden_size) self.encode_multi_key = encode_multi_key def punct(self, punct_idx): self.punct_idx = punct_idx def forward(self, input, lengths=None): mask = input.data.eq(0).t() outputs, hidden = Encoder.forward(self, input, lengths) punct_mask = None if self.punct_idx is not None: punct = set(input.data.contiguous().view(-1).tolist()) punct &= self.punct_idx if len(punct) > 0: punct_mask = 0 for p in punct: punct_mask += input.data.eq(p) self_attn_outputs = self.attn(outputs, mask, punct_mask) if not self.encode_multi_key: outputs = outputs + self_attn_outputs.transpose(0, 1) return outputs, hidden return (outputs, self_attn_outputs), hidden class StackedLSTM(nn.Module): """ Our own implementation of stacked LSTM. Needed for the decoder, because we do input feeding. """ def __init__(self, num_layers, input_size, rnn_size, dropout): super(StackedLSTM, self).__init__() self.dropout = nn.Dropout(dropout) self.num_layers = num_layers self.layers = nn.ModuleList() for i in range(num_layers): self.layers.append(nn.LSTMCell(input_size, rnn_size)) input_size = rnn_size def forward(self, input, hidden): h_0, c_0 = hidden h_1, c_1 = [], [] for i, layer in enumerate(self.layers): h_1_i, c_1_i = layer(input, (h_0[i], c_0[i])) input = h_1_i if i + 1 != self.num_layers: input = self.dropout(input) h_1 += [h_1_i] c_1 += [c_1_i] h_1 = torch.stack(h_1) c_1 = torch.stack(c_1) return input, (h_1, c_1) class Decoder(nn.Module): def __init__(self, word_vocab_size, embedding_size, word_padding_idx, num_layers, hidden_size, dropout, multi_key=False, shared_attn=False): input_size = embedding_size + hidden_size self.hidden_size = hidden_size super(Decoder, self).__init__() self.embeddings = nn.Embedding(word_vocab_size, embedding_size, padding_idx=word_padding_idx) self.rnn = StackedLSTM(num_layers, input_size, hidden_size, dropout) self.attn = GlobalAttention(hidden_size, multi_key, shared_attn) self.dropout = nn.Dropout(dropout) def forward(self, input, hidden, context, mask=None, init_output=None): emb = self.embeddings(input) batch_size = input.size(1) h_size = (batch_size, self.hidden_size) outputs = [] if init_output is None: output = Variable(emb.data.new(*h_size).zero_(), requires_grad=False) else: output = init_output attns = [] # set mask if mask is not None: self.attn.apply_mask(mask) for i, emb_t in enumerate(emb.split(1)): emb_t = emb_t.squeeze(0) emb_t = torch.cat([emb_t, output], 1) rnn_output, hidden = self.rnn(emb_t, hidden) attn_output, attn = self.attn(rnn_output, context) output = self.dropout(attn_output) outputs += [output] attns.append(attn) attns = torch.stack(attns) outputs = torch.stack(outputs) return outputs, hidden, attns class NMT(nn.Module): def __init__(self, encoder, decoder): super(NMT, self).__init__() self.encoder = encoder self.decoder = decoder def _fix_enc_hidden(self, h): """ The encoder hidden is (layers*directions) x batch x dim. We need to convert it to layers x batch x (directions*dim). """ return torch.cat([h[0:h.size(0):2], h[1:h.size(0):2]], 2) def _init_hidden(self, enc_hidden): hidden = (self._fix_enc_hidden(enc_hidden[0]), self._fix_enc_hidden(enc_hidden[1])) return hidden def forward(self, src, tgt, src_lengths): context, enc_hidden = self.encoder(src, src_lengths) if isinstance(context, tuple): context_ = (context[0].transpose(0, 1), context[1]) else: context_ = context.transpose(0, 1) enc_hidden = self._init_hidden(enc_hidden) src_pad_mask = src.data.eq(0).t() out, dec_hidden, attn = self.decoder(tgt, enc_hidden, context_, src_pad_mask) return out def make_encoder(opt): """ Various encoder dispatcher function. Args: opt: the option in current environment. """ if opt.encoder_type == "sabrnn": return SAEncoder(opt.src_vocab_size, opt.word_vec_size, 0, opt.layers, opt.rnn_size, opt.dropout, bidirectional=True, encode_multi_key=opt.encode_multi_key, min_thres=opt.min_thres, max_thres=opt.max_thres, hard=opt.hard) elif opt.encoder_type == "fabrnn": return FAEncoder(opt.src_vocab_size, opt.word_vec_size, 0, opt.layers, opt.rnn_size, opt.dropout, bidirectional=True, encode_multi_key=opt.encode_multi_key) else: return Encoder(opt.src_vocab_size, opt.word_vec_size, 0, opt.layers, opt.rnn_size, opt.dropout) def make_decoder(opt): """ Various decoder dispatcher function. Args: opt: the option in current environment. embeddings (Embeddings): vocab embeddings for this decoder. """ return Decoder(opt.tgt_vocab_size, opt.word_vec_size, 0, opt.layers, opt.rnn_size, opt.dropout, opt.encode_multi_key, opt.share_attn) def make_base_model(model_opt, checkpoint=None): """ Args: model_opt: the option loaded from checkpoint. fields: `Field` objects for the model. gpu: Boolean: whether to use gpu. checkpoint: the snapshot model. Returns: the NMTModel. """ # Make encoder. encoder = make_encoder(model_opt) decoder = make_decoder(model_opt) # Make NMT (= encoder + decoder). model = NMT(encoder, decoder) # Make Generator. generator = nn.Sequential( nn.Linear(model_opt.rnn_size, model_opt.tgt_vocab_size), nn.LogSoftmax(dim=-1)) if model_opt.share_decoder_embeddings: generator[0].weight = decoder.embeddings.weight model.generator = generator # Load the model states from checkpoint. if checkpoint is not None: print('Loading model') model.load_state_dict(checkpoint['model']) model.dicts = checkpoint['dicts'] return model
15,669
36.488038
79
py
sa-nmt
sa-nmt-master/infer.py
import torch from torch.autograd import Variable import pickle as pkl import math # TODO: documentation of functions class Beam(object): r"""Beam search class for NMT. This is a simple beam search object. It takes model, which can be used to compute the next probable output and dictionaries that will be used to map from word indices to the real strings. opt: opt (argpaser): that contains path to dictionaries model (pytorch network): this should be creat ed before """ def __init__(self, opt, model): self.opt = opt self.tt = torch.cuda if len(opt.gpuid) > 0 else torch self.model = model self.model.eval() self.dicts = model.dicts # create an inverse map from int->word for target side self.idx2w = {} for w, idx in self.dicts[1].items(): self.idx2w[idx] = w self.bos_idx = self.dicts[1]['<bos>'] self.eos_idx = self.dicts[1]['<eos>'] self.pad_idx = self.dicts[1]['<pad>'] def encode_string(self, ss): ss = ss.split() ss = [self.dicts[0].get(w, 1) for w in ss] if self.opt.src_vocab_size > 0: ss = [w if w < self.opt.src_vocab_size else 1 for w in ss] ss = Variable(self.tt.LongTensor(ss).view(-1, 1), volatile=True) return ss def decode_string(self, tidx): ts = [] for i in list(tidx): if i == self.eos_idx: break else: ts += [self.idx2w[i]] return ' '.join(ts) def beam_search(self, input): """ Beam search function. opt: trained optiongs input: Tensor (bptt x 1) """ k = self.opt.beam_size completed_hyps = [] input = input.expand(input.size(0), k) max_len = int(input.size(0) * 1.5) hypos = self.tt.LongTensor(max_len, k).fill_(2) init_target = self.tt.LongTensor(1, k).fill_(2) init_target = Variable(init_target, volatile=True) scores = self.tt.FloatTensor(k).fill_(-math.inf) scores[0] = 0 #lengths = [input.size(0) for i in range(k)] context, enc_hidden = self.model.encoder(input) init_hidden = self.model._init_hidden(enc_hidden) # alias decoder = self.model.decoder generator = self.model.generator init_output = None if isinstance(context, tuple): context = (context[0].transpose(0, 1), context[1]) else: context = context.transpose(0, 1) decoder.attn.mask = None for t in range(max_len): out, dec_hidden, attn = decoder(init_target, init_hidden, context, init_output=init_output) log_probs = generator(out.squeeze(0)).data scores_t, idx_t = log_probs.topk(k, 1) scores_t = scores_t + scores.view(-1, 1).expand_as(scores_t) scores, k_idx = scores_t.view(-1).topk(k) next_hp = k_idx.div(k) next_ys = idx_t.view(-1).index_select(0, k_idx) done_y = next_ys.eq(self.eos_idx) if done_y.sum() > 0 and t > 0: for i in range(k): if next_ys[i] == self.eos_idx: j = next_hp[i] text = self.decode_string(hypos[0:t, j]) completed_hyps.append((text, scores[i] / (t+1))) k -= 1 if k > 0: cont_y = next_ys.ne(self.eos_idx) next_ys = next_ys.masked_select(cont_y) next_hp = next_hp.masked_select(cont_y) if isinstance(context, tuple): context = (context[0][:k], context[1][:k]) else: context = context[:k] scores = scores.masked_select(cont_y) if k == 0: break hypos = hypos.index_select(1, next_hp) hypos[t] = next_ys init_target = Variable(next_ys.view(1, -1), volatile=True) next_hp = Variable(next_hp) init_output = out.squeeze(0).index_select(0, next_hp) init_hidden = [h.index_select(1, next_hp) for h in dec_hidden] if len(completed_hyps) > 0: completed_hyps.sort(key=lambda tup: tup[1]) best_h = completed_hyps.pop() return best_h[0] else: best_s, idx = scores.topk(1) best_h = hypos.index_select(1, idx).view(-1) return self.decode_string(best_h) def translate(self, text_file, out_file='output.txt'): fw = open(out_file, 'w') for line in open(text_file): src_idx = self.encode_string(line) s = self.beam_search(src_idx) fw.write(s + '\n') fw.close()
4,938
35.316176
78
py
sa-nmt
sa-nmt-master/attention.py
import torch import torch.nn as nn from Utils import aeq import math import torch.nn.functional as F class SelfAttention(nn.Module): """Self attention class""" def __init__(self, dim): super(SelfAttention, self).__init__() self.q = nn.Linear(dim, dim, bias=False) self.k = nn.Linear(dim, dim, bias=False) self.v = nn.Linear(dim, dim, bias=False) self.scale = math.sqrt(1 / dim) def forward(self, input, mask=None, punct_mask=None): """ input (FloatTensor): s_len, batch, dim mask (ByteTensor): batch x s_len """ s_len, batch, dim = input.size() input = input.contiguous().transpose(0, 1) \ .contiguous().view(-1, dim) q = self.q(input).view(batch, s_len, -1) k = self.k(input).view(batch, s_len, -1) v = self.v(input).view(batch, s_len, -1) score = torch.bmm(q, k.transpose(1, 2)) * self.scale # now (batch, s_len, s_len) FloatTensor if mask is not None: mask = mask[:, None, :].expand(batch, s_len, s_len) score.data.masked_fill_(mask, -math.inf) if punct_mask is not None: punct_mask = punct_mask.transpose(0, 1) # batch, s_len punct_mask = punct_mask[:, None, :].expand(batch, s_len, s_len) score.data.masked_fill_(punct_mask, -math.inf) # need proper masking attn = F.softmax(score.view(-1, s_len), dim=-1).view(-1, s_len, s_len) self.score = attn return torch.bmm(attn, v) class GlobalAttention(nn.Module): """ Luong Attention. This implement general attention Concrete distribution: The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables """ def __init__(self, dim, multi_key=False, share_attn=False): """ dim (Int): dimension of input vector multi_key (Boolean): using multi keys encoder share_attn (Boolean): sharing attention weights between semantic and syntactic annotations """ super(GlobalAttention, self).__init__() # make a local copy of hyper-parameters self.dim = dim self.share_attn = share_attn self.multi_key = multi_key self.linear_in = nn.Linear(dim, dim, bias=False) self.share_attn = share_attn if multi_key: if not share_attn: # using a separate attention self.linear_sa = nn.Linear(dim, dim, bias=False) self.linear_out = nn.Linear(dim*3, dim, bias=False) self.gate = nn.Linear(dim, dim) else: self.linear_out = nn.Linear(dim*2, dim, bias=False) self.mask = None def apply_mask(self, mask): self.mask = mask def score(self, h_t, h_s, sa_attn=False): """ h_t (FloatTensor): batch x tgt_len x dim h_s (FloatTensor): batch x src_len x dim returns scores (FloatTensor): batch x tgt_len x src_len: raw attention scores for each src index sa_attn (Boolean): using a separate attention for syntax context """ # Check input sizes src_batch, src_len, src_dim = h_s.size() tgt_batch, tgt_len, tgt_dim = h_t.size() aeq(src_batch, tgt_batch) aeq(src_dim, tgt_dim) aeq(self.dim, src_dim) h_t_ = h_t.view(tgt_batch*tgt_len, tgt_dim) if sa_attn: h_t_ = self.linear_sa(h_t_) else: h_t_ = self.linear_in(h_t_) h_t = h_t_.view(tgt_batch, tgt_len, tgt_dim) h_s_ = h_s.transpose(1, 2) # (batch, t_len, d) x (batch, d, s_len) --> (batch, t_len, s_len) return torch.bmm(h_t, h_s_) def forward(self, input, context): """ input (FloatTensor): batch x tgt_len x dim: decoder's rnn's output. context (FloatTensor): batch x src_len x dim: src hidden states """ # one step input if isinstance(context, tuple): context, tree_context = context if input.dim() == 2: one_step = True input = input.unsqueeze(1) else: one_step = False batch, sourceL, dim = context.size() batch_, targetL, dim_ = input.size() aeq(batch, batch_) aeq(dim, dim_) aeq(self.dim, dim) # compute attention scores, as in Luong et al. align = self.score(input, context) if self.mask is not None: mask_ = self.mask[:, None, :] align.data.masked_fill_(mask_, -math.inf) # Softmax to normalize attention weights align_vectors = F.softmax(align, dim=-1) # each context vector c_t is the weighted average # over all the source hidden states c = torch.bmm(align_vectors, context) if self.multi_key: # sharing attention weight if self.share_attn: sc = torch.bmm(align_vectors, tree_context) else: # computing attention scores for syntax tree_align = self.score(input, tree_context, True) if self.mask is not None: tree_align.data.masked_fill_(self.mask[:, None, :], -math.inf) tree_align_vectors = F.softmax(tree_align, dim=-1) sc = torch.bmm(tree_align_vectors, tree_context) z = F.sigmoid(self.gate(input)) # batch x tgt_len x dim self.z = z # for visualization sc = sc * z concat_c = torch.cat([c, input, sc], 2).view(batch*targetL, dim*3) else: concat_c = torch.cat([c, input], 2).view(batch*targetL, dim*2) attn_h = self.linear_out(concat_c).view(batch, targetL, dim) attn_h = F.tanh(attn_h) if one_step: attn_h = attn_h.squeeze(1) align_vectors = align_vectors.squeeze(1) # Check output sizes batch_, dim_ = attn_h.size() aeq(batch, batch_) aeq(dim, dim_) batch_, sourceL_ = align_vectors.size() aeq(batch, batch_) aeq(sourceL, sourceL_) else: attn_h = attn_h.transpose(0, 1).contiguous() align_vectors = align_vectors.transpose(0, 1).contiguous() # Check output sizes targetL_, batch_, dim_ = attn_h.size() aeq(targetL, targetL_) aeq(batch, batch_) aeq(dim, dim_) targetL_, batch_, sourceL_ = align_vectors.size() aeq(targetL, targetL_) aeq(batch, batch_) aeq(sourceL, sourceL_) return attn_h, align_vectors
6,737
35.032086
78
py
sa-nmt
sa-nmt-master/train.py
import argparse import torch from Iterator import TextIterator import models from itertools import zip_longest import random import Loss import opts import os import math import subprocess from infer import Beam import re from torch.optim.lr_scheduler import ReduceLROnPlateau parser = argparse.ArgumentParser(description='train.py') # Data and loading options parser.add_argument('-datasets', required=True, default=[], nargs='+', type=str, help='source_file target_file.') parser.add_argument('-valid_datasets', required=True, default=[], nargs='+', type=str, help='valid_source valid target files.') parser.add_argument('-beam_size', default=12, type=int, help="beam size") # dictionaries parser.add_argument('-dicts', required=True, default=[], nargs='+', help='source_vocab.pkl target_vocab.pkl files.') # opts.py opts.add_md_help_argument(parser) opts.model_opts(parser) opts.train_opts(parser) opts.preprocess_opts(parser) opt = parser.parse_args() device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") # for reproducibility torch.manual_seed(opt.seed) random.seed(opt.seed) if torch.cuda.is_available(): torch.cuda.manual_seed(opt.seed) print(opt) # batch preparation def prepare_data(seqs_x, seqs_y): mb = [(seqs_x[i], seqs_y[i]) for i in range(len(seqs_x))] mb.sort(key=lambda x: len(x[0]), reverse=True) xs = torch.LongTensor( list(zip_longest(*map(lambda x: x[0], mb), fillvalue=0))).to(device) ys = torch.LongTensor( list(zip_longest(*map(lambda x: x[1], mb), fillvalue=0))).to(device) lengths_x = [len(x[0]) for x in mb] return xs, ys, lengths_x def eval(model, criterion, valid_data): stats = Loss.Statistics() model.eval() loss = Loss.LossCompute(model.generator, criterion) for src, tgt in valid_data: src, tgt, src_lengths = prepare_data(src, tgt, True) outputs = model(src, tgt[:-1], src_lengths) gen_state = loss.make_loss_batch(outputs, tgt[1:]) _, batch_stats = loss.compute_loss(**gen_state) stats.update(batch_stats) model.train() return stats def init_uniform(model, init_range=0.04): """Simple uniform initialization of all the weights""" for p in model.parameters(): p.data.uniform_(-init_range, init_range) def tally_parameters(model): n_params = sum([p.nelement() for p in model.parameters()]) print('* number of parameters: %d' % n_params) enc = 0 dec = 0 for name, param in model.named_parameters(): if 'encoder' in name: enc += param.nelement() elif 'decoder' or 'generator' in name: dec += param.nelement() print('encoder: ', enc) print('decoder: ', dec) def check_model_path(): save_model_path = os.path.abspath(opt.save_model) model_dirname = os.path.dirname(save_model_path) if not os.path.exists(model_dirname): os.makedirs(model_dirname) def train(opt): print('| build data iterators') train = TextIterator(*opt.datasets, *opt.dicts, src_vocab_size=opt.src_vocab_size, tgt_vocab_size=opt.tgt_vocab_size, batch_size=opt.batch_size, max_seq_length=opt.max_seq_length) valid = TextIterator(*opt.valid_datasets, *opt.dicts, src_vocab_size=opt.src_vocab_size, tgt_vocab_size=opt.tgt_vocab_size, batch_size=opt.batch_size, max_seq_length=opt.max_seq_length) if opt.src_vocab_size < 0: opt.src_vocab_size = len(train.source_dict) if opt.tgt_vocab_size < 0: opt.tgt_vocab_size = len(train.target_dict) print('| vocabulary size. source = %d; target = %d' % (opt.src_vocab_size, opt.tgt_vocab_size)) dicts = [train.source_dict, train.target_dict] crit = Loss.nmt_criterion(opt.tgt_vocab_size, 0).to(device) if opt.train_from != '': print('| Load trained model!') checkpoint = torch.load(opt.train_from) model = models.make_base_model(opt, checkpoint) else: model = models.make_base_model(opt) init_uniform(model) model.to(device) if opt.encoder_type in ["sabrnn", "fabrnn"]: print('Add punctuation constrain!') model.encoder.punct(train.src_punct) print(model) model.dicts = dicts check_model_path() tally_parameters(model) optimizer = torch.optim.Adam(model.parameters(), lr=opt.learning_rate) scheduler = ReduceLROnPlateau(optimizer, 'min', factor=opt.learning_rate_decay, patience=0) uidx = 0 # number of updates estop = False min_lr = opt.learning_rate * math.pow(opt.learning_rate_decay, 5) best_bleu = -1 for eidx in range(1, opt.epochs + 1): closs = Loss.LossCompute(model.generator, crit) tot_loss = 0 total_stats = Loss.Statistics() report_stats = Loss.Statistics() for x, y in train: model.zero_grad() src, tgt, lengths_x = prepare_data(x, y) out = model(src, tgt[:-1], lengths_x) gen_state = closs.make_loss_batch(out, tgt[1:]) shard_size = opt.max_generator_batches batch_size = len(lengths_x) batch_stats = Loss.Statistics() for shard in Loss.shards(gen_state, shard_size): loss, stats = closs.compute_loss(**shard) loss.div(batch_size).backward() batch_stats.update(stats) tot_loss += loss.item() torch.nn.utils.clip_grad_norm_(model.parameters(), opt.max_grad_norm) optimizer.step() total_stats.update(batch_stats) report_stats.update(batch_stats) uidx += 1 if uidx % opt.report_every == 0: report_stats.output(eidx, uidx, opt.max_updates, total_stats.start_time) report_stats = Loss.Statistics() if uidx % opt.eval_every == 0: valid_stats = eval(model, crit, valid) # maybe adjust learning rate scheduler.step(valid_stats.ppl()) cur_lr = optimizer.param_groups[0]['lr'] print('Validation perplexity %d: %g' % (uidx, valid_stats.ppl())) print('Learning rate: %g' % cur_lr) if cur_lr < min_lr: print('Reaching minimum learning rate. Stop training!') estop = True break model_state_dict = model.state_dict() if eidx >= opt.start_checkpoint_at: checkpoint = { 'model': model_state_dict, 'opt': opt, 'dicts': dicts } # evaluate with BLEU score inference = Beam(opt, model) output_bpe = opt.save_model + '.bpe' output_txt = opt.save_model + '.txt' inference.translate(opt.valid_datasets[0], output_bpe) model.train() subprocess.call("sed 's/@@ //g' {:s} > {:s}" .format(output_bpe, output_txt), shell=True) ref = opt.valid_datasets[1][:-4] subprocess.call("sed 's/@@ //g' {:s} > {:s}" .format(opt.valid_datasets[1], ref), shell=True) cmd = "perl data/multi-bleu.perl {} < {}" \ .format(ref, output_txt) p = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE) \ .stdout.read().decode('utf-8') bleu = re.search("[\d]+.[\d]+", p) bleu = float(bleu.group()) print('Validation BLEU %d: %g' % (uidx, bleu)) if bleu > best_bleu: best_bleu = bleu torch.save(checkpoint, '%s_best.pt' % opt.save_model) print('Saved model: %d | BLEU %.2f' % (uidx, bleu)) if uidx >= opt.max_updates: print('Finishing after {:d} iterations!'.format(uidx)) estop = True break if estop: break train(opt)
8,797
36.598291
77
py
cogcn
cogcn-main/cogcn/utils.py
import pickle as pkl import os import networkx as nx import numpy as np import scipy.sparse as sp import torch import pandas as pd from sklearn.metrics import roc_auc_score, average_precision_score from matplotlib import pyplot as plt def load_data_cma(dataset): adj_file = os.path.join(dataset, "struct.csv") feat_file = os.path.join(dataset, "content.csv") # Load reate adjacency matrix adj = pd.read_csv(adj_file, header=None) adj = adj.values adj = nx.from_numpy_matrix(adj) adj = nx.adjacency_matrix(adj) print("Adjacency matrix shape:", adj.shape) # Load features feat = pd.read_csv(feat_file, header=None) feat = feat.values features = torch.FloatTensor(feat) print("Features shape:", features.shape) return adj, features def parse_index_file(filename): index = [] for line in open(filename): index.append(int(line.strip())) return index def sparse_to_tuple(sparse_mx): if not sp.isspmatrix_coo(sparse_mx): sparse_mx = sparse_mx.tocoo() coords = np.vstack((sparse_mx.row, sparse_mx.col)).transpose() values = sparse_mx.data shape = sparse_mx.shape return coords, values, shape def preprocess_graph(adj): adj = sp.coo_matrix(adj) adj_ = adj + sp.eye(adj.shape[0]) rowsum = np.array(adj_.sum(1)) degree_mat_inv_sqrt = sp.diags(np.power(rowsum, -0.5).flatten()) adj_normalized = adj_.dot(degree_mat_inv_sqrt).transpose().dot(degree_mat_inv_sqrt).tocoo() # return sparse_to_tuple(adj_normalized) return sparse_mx_to_torch_sparse_tensor(adj_normalized) def sparse_mx_to_torch_sparse_tensor(sparse_mx): """Convert a scipy sparse matrix to a torch sparse tensor.""" sparse_mx = sparse_mx.tocoo().astype(np.float32) indices = torch.from_numpy( np.vstack((sparse_mx.row, sparse_mx.col)).astype(np.int64)) values = torch.from_numpy(sparse_mx.data) shape = torch.Size(sparse_mx.shape) return torch.sparse.FloatTensor(indices, values, shape) def plot_losses(losses, epoch_mark): for i in range(4): plt.subplot(2,2,i+1) plt.plot(losses[:,i]) plt.axvline(epoch_mark, color='r') plt.axvline(epoch_mark*2, color='g') plt.show()
2,240
29.69863
95
py
cogcn
cogcn-main/cogcn/model.py
import torch import torch.nn as nn import torch.nn.functional as F from layers import GraphConvolution class GCNAE(nn.Module): def __init__(self, input_feat_dim, hidden_dim1, hidden_dim2, dropout): super(GCNAE, self).__init__() self.encgc1 = GraphConvolution(input_feat_dim, hidden_dim1, dropout, act=F.relu) self.encgc2 = GraphConvolution(hidden_dim1, hidden_dim2, dropout, act=lambda x: x) self.decgc1 = GraphConvolution(hidden_dim2, hidden_dim1, dropout, act=F.relu) self.decgc2 = GraphConvolution(hidden_dim1, input_feat_dim, dropout, act=lambda x: x) def encode(self, x, adj): hidden1 = self.encgc1(x, adj) hidden2 = self.encgc2(hidden1, adj) return hidden2 def decode(self, hidden, adj): hidden1 = self.decgc1(hidden, adj) recon = self.decgc2(hidden1, adj) return recon def forward(self, x, adj): enc = self.encode(x, adj) dec = self.decode(enc, adj) return dec, enc class InnerProductDecoder(nn.Module): """Decoder for using inner product for prediction.""" def __init__(self, dropout, act=torch.sigmoid): super(InnerProductDecoder, self).__init__() self.dropout = dropout self.act = act def forward(self, z): z = F.dropout(z, self.dropout, training=self.training) adj = self.act(torch.mm(z, z.t())) return adj
1,415
31.930233
93
py
cogcn
cogcn-main/cogcn/kmeans.py
import sys import torch import torch.nn as nn from sklearn.cluster import KMeans class Clustering(object): def __init__(self, K, n_init=5, max_iter=250): self.K = K self.n_init = n_init self.max_iter = max_iter self.u = None self.M = None def cluster(self, embed): embed_np = embed.detach().cpu().numpy() clustering = KMeans(n_clusters=self.K, n_init=self.n_init, max_iter=self.max_iter) clustering.fit(embed_np) self.M = clustering.labels_ self.u = self._compute_centers(self.M, embed_np) def get_loss(self, embed): loss = torch.Tensor([0.]) #TODO: This may be slightly inefficient, we can fix it later to use matrix multiplications for i, clusteridx in enumerate(self.M): x = embed[i] c = self.u[clusteridx] difference = x - c err = torch.sum(torch.mul(difference, difference)) loss += err return loss def get_membership(self): return self.M def _compute_centers(self,labels, embed): """ sklearn kmeans may not give accurate cluster centers in some cases (see doc), so we compute ourselves """ clusters = {} for i,lbl in enumerate(labels): if clusters.get(lbl) is None: clusters[lbl] = [] clusters[lbl].append(torch.FloatTensor(embed[i])) centers = {} for k in clusters: all_embed = torch.stack(clusters[k]) center = torch.mean(all_embed, 0) centers[k] = center return centers
1,646
28.410714
109
py
cogcn
cogcn-main/cogcn/layers.py
import torch import torch.nn.functional as F from torch.nn.modules.module import Module from torch.nn.parameter import Parameter class GraphConvolution(Module): """ Simple GCN layer, similar to https://arxiv.org/abs/1609.02907 """ def __init__(self, in_features, out_features, dropout=0., act=F.relu): super(GraphConvolution, self).__init__() self.in_features = in_features self.out_features = out_features self.dropout = dropout self.act = act self.weight = Parameter(torch.FloatTensor(in_features, out_features)) self.reset_parameters() def reset_parameters(self): torch.nn.init.xavier_uniform_(self.weight) def forward(self, input, adj): input = F.dropout(input, self.dropout, self.training) support = torch.mm(input, self.weight) output = torch.spmm(adj, support) output = self.act(output) return output def __repr__(self): return self.__class__.__name__ + ' (' \ + str(self.in_features) + ' -> ' \ + str(self.out_features) + ')'
1,110
30.742857
77
py
cogcn
cogcn-main/cogcn/train.py
from __future__ import division from __future__ import print_function import argparse import time import sys import os import pickle import numpy as np import scipy.sparse as sp import torch import torch.nn as nn from torch import optim from matplotlib import pyplot as plt from model import GCNAE from optimizer import compute_structure_loss, compute_attribute_loss, update_o1, update_o2 from utils import load_data_cma, preprocess_graph, plot_losses from kmeans import Clustering parser = argparse.ArgumentParser() parser.add_argument('--model', type=str, default='gcn_vae', help="models used") parser.add_argument('--seed', type=int, default=42, help='Random seed.') # 5000 epochs is where the pretrain loss starts hovering around 5.0 - 6.0 parser.add_argument('--k', type=int, default=6, help='Number of clusters.') parser.add_argument('--preepochs', type=int, default=350, help='Number of epochs to pre-train.') parser.add_argument('--clusepochs', type=int, default=1, help='Number of epochs to pre-train for clustering.') parser.add_argument('--epochs', type=int, default=300, help='Number of epochs to train.') parser.add_argument('--hidden1', type=int, default=64, help='Number of units in hidden layer 1.') parser.add_argument('--hidden2', type=int, default=32, help='Number of units in hidden layer 2.') parser.add_argument('--lr', type=float, default=0.01, help='Initial learning rate.') parser.add_argument('--lambda1', type=float, default=0.1, help='Structure loss weight.') parser.add_argument('--lambda2', type=float, default=0.1, help='Attribute loss weight.') parser.add_argument('--lambda3', type=float, default=0.8, help='Clustering loss weight.') parser.add_argument('--dropout', type=float, default=0.2, help='Dropout rate (1 - keep probability).') parser.add_argument('--dataset-str', type=str, default=None, help='type of dataset.') parser.add_argument('--outfile', type=str, default='embeddings', help='output embeddings file.') args = parser.parse_args() def gae_for(args): print("Using {} dataset".format(args.dataset_str)) adj, features = load_data_cma(args.dataset_str) n_nodes, feat_dim = features.shape # Some preprocessing adj_norm = preprocess_graph(adj) model = GCNAE(feat_dim, args.hidden1, args.hidden2, args.dropout) optimizer = optim.Adam(model.parameters(), lr=args.lr) gamma = 0.98 schedule_update_interval = 400 scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=gamma) # initialize all outlier scores with the equal values summing to 1 init_value = [1./n_nodes] * n_nodes o_1 = torch.FloatTensor(init_value) # structural outlier o_2 = torch.FloatTensor(init_value) # attribute outlier lossfn = nn.MSELoss(reduction='none') optimizer = optim.Adam(model.parameters(), lr=args.lr) lambda1 = args.lambda1 / (args.lambda1 + args.lambda2) lambda2 = args.lambda2 / (args.lambda1 + args.lambda2) kmeans = Clustering(args.k) # PRETRAIN ON STRUCTURE AND ATTRIBUTE LOSSES, NO OUTLIER LOSS for epoch in range(args.preepochs): model.train() optimizer.zero_grad() recon, embed = model(features, adj_norm) structure_loss = compute_structure_loss(adj_norm, embed, o_1) attribute_loss = compute_attribute_loss(lossfn, features, recon, o_2) loss = lambda1 * structure_loss + lambda2 * attribute_loss # Update the functions F and G (embedding network) loss.backward() cur_loss = loss.item() optimizer.step() if (epoch+1) % 100 == 0: #print("Epoch:", '%04d' % (epoch + 1), "train_loss=", "{:.5f}".format(cur_loss), "time=", "{:.5f}".format(time.time() - t), "Pretrain:",pretrain) print("Epoch:", '%04d' % (epoch + 1), "train_loss=", "{:.5f}".format(cur_loss), "lr=", "{:.5f}".format(scheduler.get_last_lr()[0])) # Initialize clusters recon, embed = model(features, adj_norm) kmeans.cluster(embed) # TRAIN ON ALL THREE LOSES WITH OUTLIER UPDATES for epoch in range(args.epochs): # Update the values of O_i1 and O_i2 o_1 = update_o1(adj_norm, embed) o_2 = update_o2(features, recon) if (epoch+1) % schedule_update_interval == 0: scheduler.step() model.train() optimizer.zero_grad() recon, embed = model(features, adj_norm) kmeans.cluster(embed) structure_loss = compute_structure_loss(adj_norm, embed, o_1) attribute_loss = compute_attribute_loss(lossfn, features, recon, o_2) clustering_loss = kmeans.get_loss(embed) loss = (args.lambda1 * structure_loss) + (args.lambda2 * attribute_loss) + (args.lambda3 * clustering_loss) #loss = (args.lambda1 * structure_loss) + (args.lambda2 * attribute_loss) # Update the functions F and G (embedding network) loss.backward() cur_loss = loss.item() optimizer.step() if (epoch+1) % 100 == 0: #print("Epoch:", '%04d' % (epoch + 1), "train_loss=", "{:.5f}".format(cur_loss), "time=", "{:.5f}".format(time.time() - t), "Pretrain:",pretrain) print("Epoch:", '%04d' % (epoch + 1), "train_loss=", "{:.5f}".format(cur_loss), "lr=", "{:.5f}".format(scheduler.get_last_lr()[0])) # Extract embeddings adj_norm = preprocess_graph(adj) recon, embed = model(features, adj_norm) embed = embed.detach().cpu().numpy() embfile = os.path.join(args.dataset_str, args.outfile+".pkl") with open(embfile,"wb") as f: pickle.dump(embed, f) o_1 = o_1.detach().cpu().numpy() o_2 = o_2.detach().cpu().numpy() outlfile = os.path.join(args.dataset_str, args.outfile+"_outliers.pkl") with open(outlfile,"wb") as f: pickle.dump([o_1, o_2], f) membfile = os.path.join(args.dataset_str, args.outfile+"_membership.pkl") with open(membfile,"wb") as f: pickle.dump(kmeans.get_membership(), f) if __name__ == '__main__': gae_for(args)
5,980
39.412162
157
py
cogcn
cogcn-main/cogcn/optimizer.py
import sys import torch import torch.nn as nn import torch.nn.modules.loss import torch.nn.functional as F from sklearn.cluster import KMeans def compute_attribute_loss(lossfn, features, recon, outlier_wt): loss = lossfn(features, recon) loss = loss.sum(dim=1) outlier_wt = torch.log(1/outlier_wt) attr_loss = torch.sum(torch.mul(outlier_wt, loss)) return attr_loss def compute_structure_loss(adj, embed, outlier_wt): # to compute F(x_i).F(x_j) embeddot = torch.mm(embed, torch.transpose(embed, 0, 1)) adj_tensor = adj.to_dense() # compute A_ij - F(x_i)*F(x_j) difference = adj_tensor - embeddot # square difference and sum loss = torch.sum(torch.mul(difference, difference), dim=1) outlier_wt = torch.log(1/outlier_wt) struct_loss = torch.sum(torch.mul(outlier_wt, loss)) return struct_loss def update_o1(adj, embed): # to compute F(x_i).F(x_j) embed = embed.data embeddot = torch.mm(embed, torch.transpose(embed, 0, 1)) adj_tensor = adj.to_dense() # compute A_ij - F(x_i)*F(x_j) difference = adj_tensor - embeddot # square difference and sum error = torch.sum(torch.mul(difference, difference), dim=1) # compute the denominator normalization_factor = torch.sum(error) # normalize the errors o1 = error/normalization_factor return o1 def update_o2(features, recon): features = features.data recon = recon.data # error = x - F(G(x)) error = features - recon # error now = (x - F(G(x)))^2, summed across dim 1 error = torch.sum(torch.mul(error, error), dim=1) # compute the denominator normalization_factor = torch.sum(error) # normalize the errors o2 = error/normalization_factor return o2
1,768
23.915493
64
py
deepglo
deepglo-master/DeepGLO/DeepGLO.py
from __future__ import print_function import torch, h5py import numpy as np from scipy.io import loadmat from torch.nn.utils import weight_norm import torch.nn as nn import torch.optim as optim import numpy as np # import matplotlib from torch.autograd import Variable import sys import itertools import torch.nn.functional as F import copy import os import gc from DeepGLO.data_loader import * from sklearn.decomposition import NMF use_cuda = True #### Assuming you have a GPU ###### from DeepGLO.utilities import * from DeepGLO.LocalModel import * from DeepGLO.metrics import * import copy import random import pickle np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) def get_model(A, y, lamb=0): """ Regularized least-squares """ n_col = A.shape[1] return np.linalg.lstsq( A.T.dot(A) + lamb * np.identity(n_col), A.T.dot(y), rcond=None ) class DeepGLO(object): def __init__( self, Ymat, vbsize=150, hbsize=256, num_channels_X=[32, 32, 32, 32, 1], num_channels_Y=[32, 32, 32, 32, 1], kernel_size=7, dropout=0.2, rank=64, kernel_size_Y=7, lr=0.0005, val_len=24, end_index=20000, normalize=False, start_date="2016-1-1", freq="H", covariates=None, use_time=True, dti=None, svd=False, period=None, forward_cov=False, ): self.start_date = start_date self.freq = freq self.covariates = covariates self.use_time = use_time self.dti = dti self.dropout = dropout self.period = period self.forward_cov = forward_cov self.Xseq = TemporalConvNet( num_inputs=1, num_channels=num_channels_X, kernel_size=kernel_size, dropout=dropout, init=True, ) if normalize: self.s = np.std(Ymat[:, 0:end_index], axis=1) # self.s[self.s == 0] = 1.0 self.s += 1.0 self.m = np.mean(Ymat[:, 0:end_index], axis=1) self.Ymat = (Ymat - self.m[:, None]) / self.s[:, None] self.mini = np.abs(np.min(self.Ymat)) self.Ymat = self.Ymat + self.mini else: self.Ymat = Ymat self.normalize = normalize n, T = self.Ymat.shape t0 = end_index + 1 if t0 > T: self.Ymat = np.hstack([self.Ymat, self.Ymat[:, -1].reshape(-1, 1)]) if svd: indices = np.random.choice(self.Ymat.shape[0], rank, replace=False) X = self.Ymat[indices, 0:t0] mX = np.std(X, axis=1) mX[mX == 0] = 1.0 X = X / mX[:, None] Ft = get_model(X.transpose(), self.Ymat[:, 0:t0].transpose(), lamb=0.1) F = Ft[0].transpose() self.X = torch.from_numpy(X).float() self.F = torch.from_numpy(F).float() else: R = torch.zeros(rank, t0).float() X = torch.normal(R, 0.1) C = torch.zeros(n, rank).float() F = torch.normal(C, 0.1) self.X = X.float() self.F = F.float() self.vbsize = vbsize self.hbsize = hbsize self.num_channels_X = num_channels_X self.num_channels_Y = num_channels_Y self.kernel_size_Y = kernel_size_Y self.rank = rank self.kernel_size = kernel_size self.lr = lr self.val_len = val_len self.end_index = end_index self.D = data_loader( Ymat=self.Ymat, vbsize=vbsize, hbsize=hbsize, end_index=end_index, val_len=val_len, shuffle=False, ) def tensor2d_to_temporal(self, T): T = T.view(1, T.size(0), T.size(1)) T = T.transpose(0, 1) return T def temporal_to_tensor2d(self, T): T = T.view(T.size(0), T.size(2)) return T def calculate_newX_loss_vanilla(self, Xn, Fn, Yn, Xf, alpha): Yout = torch.mm(Fn, Xn) cr1 = nn.L1Loss() cr2 = nn.MSELoss() l1 = cr2(Yout, Yn) / torch.mean(Yn ** 2) l2 = cr2(Xn, Xf) / torch.mean(Xf ** 2) return (1 - alpha) * l1 + alpha * l2 def recover_future_X( self, last_step, future, cpu=True, num_epochs=50, alpha=0.5, vanilla=True, tol=1e-7, ): rg = max( 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels_X) - 1), 1 + 2 * (self.kernel_size_Y - 1) * 2 ** (len(self.num_channels_Y) - 1), ) X = self.X[:, last_step - rg : last_step] X = self.tensor2d_to_temporal(X) outX = self.predict_future(model=self.Xseq, inp=X, future=future, cpu=cpu) outX = self.temporal_to_tensor2d(outX) Xf = outX[:, -future::] Yn = self.Ymat[:, last_step : last_step + future] Yn = torch.from_numpy(Yn).float() if cpu: self.Xseq = self.Xseq.cpu() else: Yn = Yn.cuda() Xf = Xf.cuda() Fn = self.F Xt = torch.zeros(self.rank, future).float() Xn = torch.normal(Xt, 0.1) if not cpu: Xn = Xn.cuda() lprev = 0 for i in range(num_epochs): Xn = Variable(Xn, requires_grad=True) optim_Xn = optim.Adam(params=[Xn], lr=self.lr) optim_Xn.zero_grad() loss = self.calculate_newX_loss_vanilla( Xn, Fn.detach(), Yn.detach(), Xf.detach(), alpha ) loss.backward() optim_Xn.step() # Xn = torch.clamp(Xn.detach(), min=0) if np.abs(lprev - loss.cpu().item()) <= tol: break if i % 1000 == 0: print("Recovery Loss: " + str(loss.cpu().item())) lprev = loss.cpu().item() self.Xseq = self.Xseq.cuda() return Xn.detach() def step_factX_loss(self, inp, out, last_vindex, last_hindex, reg=0.0): Xout = self.X[:, last_hindex + 1 : last_hindex + 1 + out.size(2)] Fout = self.F[self.D.I[last_vindex : last_vindex + out.size(0)], :] if use_cuda: Xout = Xout.cuda() Fout = Fout.cuda() Xout = Variable(Xout, requires_grad=True) out = self.temporal_to_tensor2d(out) optim_X = optim.Adam(params=[Xout], lr=self.lr) Hout = torch.matmul(Fout, Xout) optim_X.zero_grad() loss = torch.mean(torch.pow(Hout - out.detach(), 2)) l2 = torch.mean(torch.pow(Xout, 2)) r = loss.detach() / l2.detach() loss = loss + r * reg * l2 loss.backward() optim_X.step() # Xout = torch.clamp(Xout, min=0) self.X[:, last_hindex + 1 : last_hindex + 1 + inp.size(2)] = Xout.cpu().detach() return loss def step_factF_loss(self, inp, out, last_vindex, last_hindex, reg=0.0): Xout = self.X[:, last_hindex + 1 : last_hindex + 1 + out.size(2)] Fout = self.F[self.D.I[last_vindex : last_vindex + out.size(0)], :] if use_cuda: Xout = Xout.cuda() Fout = Fout.cuda() Fout = Variable(Fout, requires_grad=True) optim_F = optim.Adam(params=[Fout], lr=self.lr) out = self.temporal_to_tensor2d(out) Hout = torch.matmul(Fout, Xout) optim_F.zero_grad() loss = torch.mean(torch.pow(Hout - out.detach(), 2)) l2 = torch.mean(torch.pow(Fout, 2)) r = loss.detach() / l2.detach() loss = loss + r * reg * l2 loss.backward() optim_F.step() self.F[ self.D.I[last_vindex : last_vindex + inp.size(0)], : ] = Fout.cpu().detach() return loss def step_temporal_loss_X(self, inp, last_vindex, last_hindex): Xin = self.X[:, last_hindex : last_hindex + inp.size(2)] Xout = self.X[:, last_hindex + 1 : last_hindex + 1 + inp.size(2)] for p in self.Xseq.parameters(): p.requires_grad = False if use_cuda: Xin = Xin.cuda() Xout = Xout.cuda() Xin = Variable(Xin, requires_grad=True) Xout = Variable(Xout, requires_grad=True) optim_out = optim.Adam(params=[Xout], lr=self.lr) Xin = self.tensor2d_to_temporal(Xin) Xout = self.tensor2d_to_temporal(Xout) hatX = self.Xseq(Xin) optim_out.zero_grad() loss = torch.mean(torch.pow(Xout - hatX.detach(), 2)) loss.backward() optim_out.step() # Xout = torch.clamp(Xout, min=0) temp = self.temporal_to_tensor2d(Xout.detach()) self.X[:, last_hindex + 1 : last_hindex + 1 + inp.size(2)] = temp return loss def predict_future_batch(self, model, inp, future=10, cpu=True): if cpu: model = model.cpu() inp = inp.cpu() else: inp = inp.cuda() out = model(inp) output = out[:, :, out.size(2) - 1].view(out.size(0), out.size(1), 1) out = torch.cat((inp, output), dim=2) torch.cuda.empty_cache() for i in range(future - 1): inp = out out = model(inp) output = out[:, :, out.size(2) - 1].view(out.size(0), out.size(1), 1) out = torch.cat((inp, output), dim=2) torch.cuda.empty_cache() out = self.temporal_to_tensor2d(out) out = np.array(out.cpu().detach()) return out def predict_future(self, model, inp, future=10, cpu=True, bsize=90): n = inp.size(0) inp = inp.cpu() ids = np.arange(0, n, bsize) ids = list(ids) + [n] out = self.predict_future_batch(model, inp[ids[0] : ids[1], :, :], future, cpu) torch.cuda.empty_cache() for i in range(1, len(ids) - 1): temp = self.predict_future_batch( model, inp[ids[i] : ids[i + 1], :, :], future, cpu ) torch.cuda.empty_cache() out = np.vstack([out, temp]) out = torch.from_numpy(out).float() return self.tensor2d_to_temporal(out) def predict_global( self, ind, last_step=100, future=10, cpu=False, normalize=False, bsize=90 ): if ind is None: ind = np.arange(self.Ymat.shape[0]) if cpu: self.Xseq = self.Xseq.cpu() self.Xseq = self.Xseq.eval() rg = max( 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels_X) - 1), 1 + 2 * (self.kernel_size_Y - 1) * 2 ** (len(self.num_channels_Y) - 1), ) X = self.X[:, last_step - rg : last_step] n = X.size(0) T = X.size(1) X = self.tensor2d_to_temporal(X) outX = self.predict_future( model=self.Xseq, inp=X, future=future, cpu=cpu, bsize=bsize ) outX = self.temporal_to_tensor2d(outX) F = self.F Y = torch.matmul(F, outX) Y = np.array(Y[ind, :].cpu().detach()) self.Xseq = self.Xseq.cuda() del F torch.cuda.empty_cache() for p in self.Xseq.parameters(): p.requires_grad = True if normalize: Y = Y - self.mini Y = Y * self.s[ind, None] + self.m[ind, None] return Y else: return Y def train_Xseq(self, Ymat, num_epochs=20, early_stop=False, tenacity=3): seq = self.Xseq num_channels = self.num_channels_X kernel_size = self.kernel_size vbsize = min(self.vbsize, Ymat.shape[0] / 2) for p in seq.parameters(): p.requires_grad = True TC = LocalModel( Ymat=Ymat, num_inputs=1, num_channels=num_channels, kernel_size=kernel_size, vbsize=vbsize, hbsize=self.hbsize, normalize=False, end_index=self.end_index - self.val_len, val_len=self.val_len, lr=self.lr, num_epochs=num_epochs, ) TC.train_model(early_stop=early_stop, tenacity=tenacity) self.Xseq = TC.seq def train_factors( self, reg_X=0.0, reg_F=0.0, mod=5, early_stop=False, tenacity=3, ind=None, seed=False, ): self.D.epoch = 0 self.D.vindex = 0 self.D.hindex = 0 if use_cuda: self.Xseq = self.Xseq.cuda() for p in self.Xseq.parameters(): p.requires_grad = True l_F = [0.0] l_X = [0.0] l_X_temporal = [0.0] iter_count = 0 vae = float("inf") scount = 0 Xbest = self.X.clone() Fbest = self.F.clone() while self.D.epoch < self.num_epochs: last_epoch = self.D.epoch last_vindex = self.D.vindex last_hindex = self.D.hindex inp, out, vindex, hindex = self.D.next_batch(option=1) if use_cuda: inp = inp.float().cuda() out = out.float().cuda() if iter_count % mod >= 0: l1 = self.step_factF_loss(inp, out, last_vindex, last_hindex, reg=reg_F) l_F = l_F + [l1.cpu().item()] if iter_count % mod >= 0: l1 = self.step_factX_loss(inp, out, last_vindex, last_hindex, reg=reg_X) l_X = l_X + [l1.cpu().item()] if seed == False and iter_count % mod == 1: l2 = self.step_temporal_loss_X(inp, last_vindex, last_hindex) l_X_temporal = l_X_temporal + [l2.cpu().item()] iter_count = iter_count + 1 if self.D.epoch > last_epoch: print("Entering Epoch# ", self.D.epoch) print("Factorization Loss F: ", np.mean(l_F)) print("Factorization Loss X: ", np.mean(l_X)) print("Temporal Loss X: ", np.mean(l_X_temporal)) if ind is None: ind = np.arange(self.Ymat.shape[0]) else: ind = ind inp = self.predict_global( ind, last_step=self.end_index - self.val_len, future=self.val_len, cpu=False, ) R = self.Ymat[ind, self.end_index - self.val_len : self.end_index] S = inp[:, -self.val_len : :] ve = np.abs(R - S).mean() / np.abs(R).mean() print("Validation Loss (Global): ", ve) if ve <= vae: vae = ve scount = 0 Xbest = self.X.clone() Fbest = self.F.clone() # Xseqbest = TemporalConvNet( # num_inputs=1, # num_channels=self.num_channels_X, # kernel_size=self.kernel_size, # dropout=self.dropout, # ) # Xseqbest.load_state_dict(self.Xseq.state_dict()) Xseqbest = pickle.loads(pickle.dumps(self.Xseq)) else: scount += 1 if scount > tenacity and early_stop: print("Early Stopped") self.X = Xbest self.F = Fbest self.Xseq = Xseqbest if use_cuda: self.Xseq = self.Xseq.cuda() break def create_Ycov(self): t0 = self.end_index + 1 self.D.epoch = 0 self.D.vindex = 0 self.D.hindex = 0 Ycov = copy.deepcopy(self.Ymat[:, 0:t0]) Ymat_now = self.Ymat[:, 0:t0] if use_cuda: self.Xseq = self.Xseq.cuda() self.Xseq = self.Xseq.eval() while self.D.epoch < 1: last_epoch = self.D.epoch last_vindex = self.D.vindex last_hindex = self.D.hindex inp, out, vindex, hindex = self.D.next_batch(option=1) if use_cuda: inp = inp.cuda() Xin = self.tensor2d_to_temporal(self.X[:, last_hindex : last_hindex + inp.size(2)]).cuda() Xout = self.temporal_to_tensor2d(self.Xseq(Xin)).cpu() Fout = self.F[self.D.I[last_vindex : last_vindex + out.size(0)], :] output = np.array(torch.matmul(Fout, Xout).detach()) Ycov[ last_vindex : last_vindex + output.shape[0], last_hindex + 1 : last_hindex + 1 + output.shape[1], ] = output for p in self.Xseq.parameters(): p.requires_grad = True if self.period is None: Ycov_wc = np.zeros(shape=[Ycov.shape[0], 1, Ycov.shape[1]]) if self.forward_cov: Ycov_wc[:, 0, 0:-1] = Ycov[:, 1::] else: Ycov_wc[:, 0, :] = Ycov else: Ycov_wc = np.zeros(shape=[Ycov.shape[0], 2, Ycov.shape[1]]) if self.forward_cov: Ycov_wc[:, 0, 0:-1] = Ycov[:, 1::] else: Ycov_wc[:, 0, :] = Ycov Ycov_wc[:, 1, self.period - 1 : :] = Ymat_now[:, 0 : -(self.period - 1)] return Ycov_wc def train_Yseq(self, num_epochs=20, early_stop=False, tenacity=7): Ycov = self.create_Ycov() self.Yseq = LocalModel( self.Ymat, num_inputs=1, num_channels=self.num_channels_Y, kernel_size=self.kernel_size_Y, dropout=self.dropout, vbsize=self.vbsize, hbsize=self.hbsize, num_epochs=num_epochs, lr=self.lr, val_len=self.val_len, test=True, end_index=self.end_index - self.val_len, normalize=False, start_date=self.start_date, freq=self.freq, covariates=self.covariates, use_time=self.use_time, dti=self.dti, Ycov=Ycov, ) self.Yseq.train_model(early_stop=early_stop, tenacity=tenacity) def train_all_models( self, init_epochs=100, alt_iters=10, y_iters=200, tenacity=7, mod=5 ): print("Initializing Factors.....") self.num_epochs = init_epochs self.train_factors() if alt_iters % 2 == 1: alt_iters += 1 print("Starting Alternate Training.....") for i in range(1, alt_iters): if i % 2 == 0: print( "--------------------------------------------Training Factors. Iter#: " + str(i) + "-------------------------------------------------------" ) self.num_epochs = 300 self.train_factors( seed=False, early_stop=True, tenacity=tenacity, mod=mod ) else: print( "--------------------------------------------Training Local Model. Iter#: " + str(i) + "-------------------------------------------------------" ) self.num_epochs = 300 T = np.array(self.X.cpu().detach()) self.train_Xseq( Ymat=T, num_epochs=self.num_epochs, early_stop=True, tenacity=tenacity, ) self.num_epochs = y_iters self.train_Yseq(num_epochs=y_iters, early_stop=True, tenacity=tenacity) def predict( self, ind=None, last_step=100, future=10, cpu=False, normalize=False, bsize=90 ): if ind is None: ind = np.arange(self.Ymat.shape[0]) if cpu: self.Xseq = self.Xseq.cpu() self.Yseq.seq = self.Yseq.seq.eval() self.Xseq = self.Xseq.eval() rg = max( 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels_X) - 1), 1 + 2 * (self.kernel_size_Y - 1) * 2 ** (len(self.num_channels_Y) - 1), ) covs = self.Yseq.covariates[:, last_step - rg : last_step + future] # print(covs.shape) yc = self.predict_global( ind=ind, last_step=last_step, future=future, cpu=cpu, normalize=False, bsize=bsize, ) if self.period is None: ycovs = np.zeros(shape=[yc.shape[0], 1, yc.shape[1]]) if self.forward_cov: ycovs[:, 0, 0:-1] = yc[:, 1::] else: ycovs[:, 0, :] = yc else: ycovs = np.zeros(shape=[yc.shape[0], 2, yc.shape[1]]) if self.forward_cov: ycovs[:, 0, 0:-1] = yc[:, 1::] else: ycovs[:, 0, :] = yc period = self.period while last_step + future - (period - 1) > last_step + 1: period += self.period ycovs[:, 1, period - 1 : :] = self.Ymat[ :, last_step - rg : last_step + future - (period - 1) ] ### this seems like we are looking ahead, but it will not use the last coordinate, which is the only new point added # print(ycovs.shape) Y = self.Yseq.predict_future( data_in=self.Ymat[ind, last_step - rg : last_step], covariates=covs, ycovs=ycovs, future=future, cpu=cpu, bsize=bsize, normalize=False, ) if normalize: Y = Y - self.mini Y = Y * self.s[ind, None] + self.m[ind, None] return Y else: return Y def rolling_validation(self, Ymat, tau=24, n=7, bsize=90, cpu=False, alpha=0.3): prevX = self.X.clone() prev_index = self.end_index out = self.predict( last_step=self.end_index, future=tau, bsize=bsize, cpu=cpu, normalize=self.normalize, ) out_global = self.predict_global( np.arange(self.Ymat.shape[0]), last_step=self.end_index, future=tau, cpu=cpu, normalize=self.normalize, bsize=bsize, ) predicted_values = [] actual_values = [] predicted_values_global = [] S = out[:, -tau::] S_g = out_global[:, -tau::] predicted_values += [S] predicted_values_global += [S_g] R = Ymat[:, self.end_index : self.end_index + tau] actual_values += [R] print("Current window wape: " + str(wape(S, R))) self.Xseq = self.Xseq.eval() self.Yseq.seq = self.Yseq.seq.eval() for i in range(n - 1): Xn = self.recover_future_X( last_step=self.end_index + 1, future=tau, num_epochs=100000, alpha=alpha, vanilla=True, cpu=True, ) self.X = torch.cat([self.X, Xn], dim=1) self.end_index += tau out = self.predict( last_step=self.end_index, future=tau, bsize=bsize, cpu=cpu, normalize=self.normalize, ) out_global = self.predict_global( np.arange(self.Ymat.shape[0]), last_step=self.end_index, future=tau, cpu=cpu, normalize=self.normalize, bsize=bsize, ) S = out[:, -tau::] S_g = out_global[:, -tau::] predicted_values += [S] predicted_values_global += [S_g] R = Ymat[:, self.end_index : self.end_index + tau] actual_values += [R] print("Current window wape: " + str(wape(S, R))) predicted = np.hstack(predicted_values) predicted_global = np.hstack(predicted_values_global) actual = np.hstack(actual_values) dic = {} dic["wape"] = wape(predicted, actual) dic["mape"] = mape(predicted, actual) dic["smape"] = smape(predicted, actual) dic["mae"] = np.abs(predicted - actual).mean() dic["rmse"] = np.sqrt(((predicted - actual) ** 2).mean()) dic["nrmse"] = dic["rmse"] / np.sqrt(((actual) ** 2).mean()) dic["wape_global"] = wape(predicted_global, actual) dic["mape_global"] = mape(predicted_global, actual) dic["smape_global"] = smape(predicted_global, actual) dic["mae_global"] = np.abs(predicted_global - actual).mean() dic["rmse_global"] = np.sqrt(((predicted_global - actual) ** 2).mean()) dic["nrmse_global"] = dic["rmse"] / np.sqrt(((actual) ** 2).mean()) baseline = Ymat[:, Ymat.shape[1] - n * tau - tau : Ymat.shape[1] - tau] dic["baseline_wape"] = wape(baseline, actual) dic["baseline_mape"] = mape(baseline, actual) dic["baseline_smape"] = smape(baseline, actual) self.X = prevX self.end_index = prev_index return dic
25,258
32.235526
131
py
deepglo
deepglo-master/DeepGLO/data_loader.py
import torch, h5py import numpy as np from scipy.io import loadmat import torch.nn as nn import torch.optim as optim import numpy as np # import matplotlib from torch.autograd import Variable import itertools from sklearn.preprocessing import normalize import datetime import json import os, sys import pandas as pd import pyarrow.parquet as pq from DeepGLO.Ftree import * class data_loader(object): """ Data Loader Class for DeepGLO """ def __init__( self, Ymat, covariates=None, Ycov=None, vbsize=200, hbsize=100, end_index=20000, val_len=30, shuffle=False, ): """ Argeuments: Ymat: time-series matrix n*T covariates: global covariates common for all time series r*T, where r is the number of covariates Ycov: per time-series covariates n*l*T, l such covariates per time-series All of the above arguments are numpy arrays vbsize: vertical batch size hbsize: horizontal batch size end_index: training and validation set is only from 0:end_index val_len: validation length. The last 'val_len' time-points for every time-series is the validation set shuffle: data is shuffles if True (this is deprecated and set to False) """ n, T = Ymat.shape self.vindex = 0 self.hindex = 0 self.epoch = 0 self.vbsize = vbsize self.hbsize = hbsize self.Ymat = Ymat self.val_len = val_len self.end_index = end_index self.val_index = np.random.randint(0, n - self.vbsize - 5) self.shuffle = shuffle self.I = np.array(range(n)) self.covariates = covariates self.Ycov = Ycov def next_batch(self, option=1): """ Arguments: option = 1 means data is returned as pytorch tensor of shape nd*cd*td where nd is vbsize, hb is hsize and cd is the number os channels (depends on covariates) option = 0 is deprecated Returns: inp: input batch out: one shifted output batch vindex: strating vertical index of input batch hindex: starting horizontal index of input batch """ n, T = self.Ymat.shape if self.hindex + self.hbsize + 1 >= self.end_index: pr_hindex = self.hindex self.hindex = 0 if self.vindex + self.vbsize >= n: pr_vindex = self.vindex self.vindex = 0 self.epoch = self.epoch + 1 if self.shuffle: I = np.random.choice(n, n, replace=False) self.I = I self.Ymat = self.Ymat[self.I, :] else: pr_vindex = self.vindex self.vindex = self.vindex + self.vbsize else: pr_hindex = self.hindex self.hindex = self.hindex + self.hbsize pr_vindex = self.vindex data = self.Ymat[ int(pr_vindex) : int(pr_vindex + self.vbsize), int(pr_hindex) : int(min(self.end_index, pr_hindex + self.hbsize)), ] out_data = self.Ymat[ int(pr_vindex) : int(pr_vindex + self.vbsize), int(pr_hindex + 1) : int(min(self.end_index, pr_hindex + self.hbsize) + 1), ] nd, Td = data.shape if self.covariates is not None: covs = self.covariates[ :, int(pr_hindex) : int(min(self.end_index, pr_hindex + self.hbsize)) ] rcovs = np.repeat( covs.reshape(1, covs.shape[0], covs.shape[1]), repeats=nd, axis=0 ) if self.Ycov is not None: ycovs = self.Ycov[ int(pr_vindex) : int(pr_vindex + self.vbsize), :, int(pr_hindex) : int(min(self.end_index, pr_hindex + self.hbsize)), ] if option == 1: inp = torch.from_numpy(data).view(1, nd, Td) out = torch.from_numpy(out_data).view(1, nd, Td) if self.covariates is not None: rcovs = torch.from_numpy(rcovs).float() if self.Ycov is not None: ycovs = torch.from_numpy(ycovs).float() inp = inp.transpose(0, 1).float() if self.covariates is not None: inp = torch.cat((inp, rcovs), 1) if self.Ycov is not None: inp = torch.cat((inp, ycovs), 1) out = out.transpose(0, 1).float() else: inp = torch.from_numpy(data).float() out = torch.from_numpy(out_data).float() inp[torch.isnan(inp)] = 0 out[torch.isnan(out)] = 0 return inp, out, self.vindex, self.hindex def supply_test(self, option=1): """ Supplies validation set in the same format as above """ n, T = self.Ymat.shape index = self.val_index in_data = self.Ymat[ int(index) : int(index + self.vbsize), int(self.end_index) : int(self.end_index + self.val_len), ] out_data = self.Ymat[ int(index) : int(index + self.vbsize), int(self.end_index + 1) : int(self.end_index + self.val_len + 1), ] nd, Td = in_data.shape if self.covariates is not None: covs = self.covariates[ :, int(self.end_index) : int(self.end_index + self.val_len) ] rcovs = np.repeat( covs.reshape(1, covs.shape[0], covs.shape[1]), repeats=nd, axis=0 ) if self.Ycov is not None: ycovs = self.Ycov[ int(index) : int(index + self.vbsize), :, int(self.end_index) : int(self.end_index + self.val_len), ] if option == 1: inp = torch.from_numpy(in_data).view(1, nd, Td) inp = inp.transpose(0, 1).float() if self.covariates is not None: rcovs = torch.from_numpy(rcovs).float() if self.Ycov is not None: ycovs = torch.from_numpy(ycovs).float() out = torch.from_numpy(out_data).view(1, nd, Td) if self.covariates is not None: inp = torch.cat((inp, rcovs), 1) if self.Ycov is not None: inp = torch.cat((inp, ycovs), 1) out = out.transpose(0, 1).float() else: inp = torch.from_numpy(in_data).float() out = torch.from_numpy(out_data).float() return inp, out, self.vindex, self.hindex
6,610
34.735135
167
py
deepglo
deepglo-master/DeepGLO/LocalModel.py
import torch, h5py import numpy as np from scipy.io import loadmat from torch.nn.utils import weight_norm import torch.nn as nn import torch.optim as optim import numpy as np # import matplotlib from torch.autograd import Variable import itertools import torch.nn.functional as F from DeepGLO.data_loader import * use_cuda = True #### Assuming you have a GPU ###### from DeepGLO.utilities import * from DeepGLO.time import * from DeepGLO.metrics import * import random import pickle np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) class Chomp1d(nn.Module): def __init__(self, chomp_size): super(Chomp1d, self).__init__() self.chomp_size = chomp_size def forward(self, x): return x[:, :, : -self.chomp_size].contiguous() class TemporalBlock(nn.Module): def __init__( self, n_inputs, n_outputs, kernel_size, stride, dilation, padding, dropout=0.1, init=True, ): super(TemporalBlock, self).__init__() self.kernel_size = kernel_size self.conv1 = weight_norm( nn.Conv1d( n_inputs, n_outputs, kernel_size, stride=stride, padding=padding, dilation=dilation, ) ) self.chomp1 = Chomp1d(padding) self.relu1 = nn.ReLU() self.dropout1 = nn.Dropout(dropout) self.conv2 = weight_norm( nn.Conv1d( n_outputs, n_outputs, kernel_size, stride=stride, padding=padding, dilation=dilation, ) ) self.chomp2 = Chomp1d(padding) self.relu2 = nn.ReLU() self.dropout2 = nn.Dropout(dropout) self.net = nn.Sequential( self.conv1, self.chomp1, self.relu1, self.dropout1, self.conv2, self.chomp2, self.relu2, self.dropout2, ) self.downsample = ( nn.Conv1d(n_inputs, n_outputs, 1) if n_inputs != n_outputs else None ) self.init = init self.relu = nn.ReLU() self.init_weights() def init_weights(self): if self.init: nn.init.normal_(self.conv1.weight, std=1e-3) nn.init.normal_(self.conv2.weight, std=1e-3) self.conv1.weight[:, 0, :] += ( 1.0 / self.kernel_size ) ###new initialization scheme self.conv2.weight += 1.0 / self.kernel_size ###new initialization scheme nn.init.normal_(self.conv1.bias, std=1e-6) nn.init.normal_(self.conv2.bias, std=1e-6) else: nn.init.xavier_uniform_(self.conv1.weight) nn.init.xavier_uniform_(self.conv2.weight) if self.downsample is not None: self.downsample.weight.data.normal_(0, 0.1) def forward(self, x): out = self.net(x) res = x if self.downsample is None else self.downsample(x) return self.relu(out + res) class TemporalBlock_last(nn.Module): def __init__( self, n_inputs, n_outputs, kernel_size, stride, dilation, padding, dropout=0.2, init=True, ): super(TemporalBlock_last, self).__init__() self.kernel_size = kernel_size self.conv1 = weight_norm( nn.Conv1d( n_inputs, n_outputs, kernel_size, stride=stride, padding=padding, dilation=dilation, ) ) self.chomp1 = Chomp1d(padding) self.relu1 = nn.ReLU() self.dropout1 = nn.Dropout(dropout) self.conv2 = weight_norm( nn.Conv1d( n_outputs, n_outputs, kernel_size, stride=stride, padding=padding, dilation=dilation, ) ) self.chomp2 = Chomp1d(padding) self.relu2 = nn.ReLU() self.dropout2 = nn.Dropout(dropout) self.net = nn.Sequential( self.conv1, self.chomp1, self.dropout1, self.conv2, self.chomp2, self.dropout2, ) self.downsample = ( nn.Conv1d(n_inputs, n_outputs, 1) if n_inputs != n_outputs else None ) self.init = init self.relu = nn.ReLU() self.init_weights() def init_weights(self): if self.init: nn.init.normal_(self.conv1.weight, std=1e-3) nn.init.normal_(self.conv2.weight, std=1e-3) self.conv1.weight[:, 0, :] += ( 1.0 / self.kernel_size ) ###new initialization scheme self.conv2.weight += 1.0 / self.kernel_size ###new initialization scheme nn.init.normal_(self.conv1.bias, std=1e-6) nn.init.normal_(self.conv2.bias, std=1e-6) else: nn.init.xavier_uniform_(self.conv1.weight) nn.init.xavier_uniform_(self.conv2.weight) if self.downsample is not None: self.downsample.weight.data.normal_(0, 0.1) def forward(self, x): out = self.net(x) res = x if self.downsample is None else self.downsample(x) return out + res class TemporalConvNet(nn.Module): def __init__(self, num_inputs, num_channels, kernel_size=2, dropout=0.1, init=True): super(TemporalConvNet, self).__init__() layers = [] self.num_channels = num_channels self.num_inputs = num_inputs self.kernel_size = kernel_size self.dropout = dropout num_levels = len(num_channels) for i in range(num_levels): dilation_size = 2 ** i in_channels = num_inputs if i == 0 else num_channels[i - 1] out_channels = num_channels[i] if i == num_levels - 1: layers += [ TemporalBlock_last( in_channels, out_channels, kernel_size, stride=1, dilation=dilation_size, padding=(kernel_size - 1) * dilation_size, dropout=dropout, init=init, ) ] else: layers += [ TemporalBlock( in_channels, out_channels, kernel_size, stride=1, dilation=dilation_size, padding=(kernel_size - 1) * dilation_size, dropout=dropout, init=init, ) ] self.network = nn.Sequential(*layers) def forward(self, x): return self.network(x) class LocalModel(object): def __init__( self, Ymat, num_inputs=1, num_channels=[32, 32, 32, 32, 32, 1], kernel_size=7, dropout=0.2, vbsize=300, hbsize=128, num_epochs=100, lr=0.0005, val_len=10, test=True, end_index=120, normalize=False, start_date="2016-1-1", freq="H", covariates=None, use_time=False, dti=None, Ycov=None, ): """ Arguments: Ymat: input time-series n*T num_inputs: always set to 1 num_channels: list containing channel progression of temporal comvolution network kernel_size: kernel size of temporal convolution filters dropout: dropout rate for each layer vbsize: vertical batch size hbsize: horizontal batch size num_epochs: max. number of epochs lr: learning rate val_len: validation length test: always set to True end_index: no data is touched fro training or validation beyond end_index normalize: normalize dataset before training or not start_data: start data in YYYY-MM-DD format (give a random date if unknown) freq: "H" hourly, "D": daily and for rest see here: https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#timeseries-offset-aliases covariates: global covariates common for all time series r*T, where r is the number of covariates Ycov: per time-series covariates n*l*T, l such covariates per time-series use_time: if false, default trime-covriates are not used dti: date time object can be explicitly supplied here, leave None if default options are to be used """ self.start_date = start_date if use_time: self.time = TimeCovariates( start_date=start_date, freq=freq, normalized=True, num_ts=Ymat.shape[1] ) if dti is not None: self.time.dti = dti time_covariates = self.time.get_covariates() if covariates is None: self.covariates = time_covariates else: self.covariates = np.vstack([time_covariates, covariates]) else: self.covariates = covariates self.Ycov = Ycov self.freq = freq self.vbsize = vbsize self.hbsize = hbsize self.num_inputs = num_inputs self.num_channels = num_channels self.num_epochs = num_epochs self.lr = lr self.val_len = val_len self.Ymat = Ymat self.test = test self.end_index = end_index self.normalize = normalize self.kernel_size = kernel_size if normalize: Y = Ymat m = np.mean(Y[:, 0 : self.end_index], axis=1) s = np.std(Y[:, 0 : self.end_index], axis=1) # s[s == 0] = 1.0 s += 1.0 Y = (Y - m[:, None]) / s[:, None] mini = np.abs(np.min(Y)) self.Ymat = Y + mini self.m = m self.s = s self.mini = mini if self.Ycov is not None: self.num_inputs += self.Ycov.shape[1] if self.covariates is not None: self.num_inputs += self.covariates.shape[0] self.seq = TemporalConvNet( num_inputs=self.num_inputs, num_channels=num_channels, kernel_size=kernel_size, dropout=dropout, init=True, ) self.seq = self.seq.float() self.D = data_loader( Ymat=self.Ymat, vbsize=vbsize, hbsize=hbsize, end_index=end_index, val_len=val_len, covariates=self.covariates, Ycov=self.Ycov, ) self.val_len = val_len if use_cuda: self.seq = self.seq.cuda() def __loss__(self, out, target, dic=None): criterion = nn.L1Loss() return criterion(out, target) / torch.abs(target.data).mean() def __prediction__(self, data): dic = None out = self.seq(data) return out, dic def train_model(self, early_stop=False, tenacity=10): """ early_stop: set true for using early stop tenacity: patience for early_stop """ print("Training Local Model(Tconv)") if use_cuda: self.seq = self.seq.cuda() optimizer = optim.Adam(params=self.seq.parameters(), lr=self.lr) iter_count = 0 loss_all = [] loss_test_all = [] vae = float("inf") scount = 0 while self.D.epoch < self.num_epochs: last_epoch = self.D.epoch inp, out_target, _, _ = self.D.next_batch() if self.test: inp_test, out_target_test, _, _ = self.D.supply_test() current_epoch = self.D.epoch if use_cuda: inp = inp.cuda() out_target = out_target.cuda() inp = Variable(inp) out_target = Variable(out_target) optimizer.zero_grad() out, dic = self.__prediction__(inp) loss = self.__loss__(out, out_target, dic) iter_count = iter_count + 1 for p in self.seq.parameters(): p.requires_grad = True loss.backward() for p in self.seq.parameters(): p.grad.data.clamp_(max=1e5, min=-1e5) optimizer.step() loss_all = loss_all + [loss.cpu().item()] if self.test: if use_cuda: inp_test = inp_test.cuda() out_target_test = out_target_test.cuda() inp_test = Variable(inp_test) out_target_test = Variable(out_target_test) out_test, dic = self.__prediction__(inp_test) losst = self.__loss__(out_test, out_target_test, dic) loss_test_all = loss_test_all + [losst.cpu().item()] if current_epoch > last_epoch: ve = loss_test_all[-1] print("Entering Epoch# ", current_epoch) print("Train Loss:", np.mean(loss_all)) print("Validation Loss:", ve) if ve <= vae: vae = ve scount = 0 # self.saved_seq = TemporalConvNet( # num_inputs=self.seq.num_inputs, # num_channels=self.seq.num_channels, # kernel_size=self.seq.kernel_size, # dropout=self.seq.dropout, # ) # self.saved_seq.load_state_dict(self.seq.state_dict()) self.saved_seq = pickle.loads(pickle.dumps(self.seq)) else: scount += 1 if scount > tenacity and early_stop: self.seq = self.saved_seq if use_cuda: self.seq = self.seq.cuda() break def convert_to_input(self, data, cuda=True): n, m = data.shape inp = torch.from_numpy(data).view(1, n, m) inp = inp.transpose(0, 1).float() if cuda: inp = inp.cuda() return inp def convert_covariates(self, data, covs, cuda=True): nd, td = data.shape rcovs = np.repeat( covs.reshape(1, covs.shape[0], covs.shape[1]), repeats=nd, axis=0 ) rcovs = torch.from_numpy(rcovs).float() if cuda: rcovs = rcovs.cuda() return rcovs def convert_ycovs(self, data, ycovs, cuda=True): nd, td = data.shape ycovs = torch.from_numpy(ycovs).float() if cuda: ycovs = ycovs.cuda() return ycovs def convert_from_output(self, T): out = T.view(T.size(0), T.size(2)) return np.array(out.cpu().detach()) def predict_future_batch( self, data, covariates=None, ycovs=None, future=10, cpu=False ): if cpu: self.seq = self.seq.cpu() else: self.seq = self.seq.cuda() inp = self.convert_to_input(data) if covariates is not None: cov = self.convert_covariates(data, covariates) inp = torch.cat((inp, cov[:, :, 0 : inp.size(2)]), 1) if ycovs is not None: ycovs = self.convert_ycovs(data, ycovs) inp = torch.cat((inp, ycovs[:, :, 0 : inp.size(2)]), 1) if cpu: inp = inp.cpu() cov = cov.cpu() ycovs = ycovs.cpu() out, dic = self.__prediction__(inp) ci = inp.size(2) output = out[:, :, out.size(2) - 1].view(out.size(0), out.size(1), 1) if covariates is not None: output = torch.cat( (output, cov[:, :, ci].view(cov.size(0), cov.size(1), 1)), 1 ) if ycovs is not None: output = torch.cat( (output, ycovs[:, :, ci].view(ycovs.size(0), ycovs.size(1), 1)), 1 ) out = torch.cat((inp, output), dim=2) for i in range(future - 1): inp = out out, dic = self.__prediction__(inp) output = out[:, :, out.size(2) - 1].view(out.size(0), out.size(1), 1) ci += 1 if covariates is not None: output = torch.cat( (output, cov[:, :, ci].view(cov.size(0), cov.size(1), 1)), 1 ) if ycovs is not None: output = torch.cat( (output, ycovs[:, :, ci].view(ycovs.size(0), ycovs.size(1), 1)), 1 ) out = torch.cat((inp, output), dim=2) out = out[:, 0, :].view(out.size(0), 1, out.size(2)) out = out.cuda() y = self.convert_from_output(out) self.seq = self.seq.cuda() return y def predict_future( self, data_in, covariates=None, ycovs=None, future=10, cpu=False, bsize=40, normalize=False, ): """ data_in: input past data in same format of Ymat covariates: input past covariates ycovs: input past individual covariates future: number of time-points to predict cpu: if true then gpu is not used bsize: batch size for processing (determine according to gopu memory limits) normalize: should be set according to the normalization used in the class initialization """ if normalize: data = (data_in - self.m[:, None]) / self.s[:, None] data += self.mini else: data = data_in n, T = data.shape I = list(np.arange(0, n, bsize)) I.append(n) bdata = data[range(I[0], I[1]), :] if ycovs is not None: out = self.predict_future_batch( bdata, covariates, ycovs[range(I[0], I[1]), :, :], future, cpu ) else: out = self.predict_future_batch(bdata, covariates, None, future, cpu) for i in range(1, len(I) - 1): bdata = data[range(I[i], I[i + 1]), :] self.seq = self.seq.cuda() if ycovs is not None: temp = self.predict_future_batch( bdata, covariates, ycovs[range(I[i], I[i + 1]), :, :], future, cpu ) else: temp = self.predict_future_batch(bdata, covariates, None, future, cpu) out = np.vstack([out, temp]) if normalize: temp = (out - self.mini) * self.s[:, None] + self.m[:, None] out = temp return out def rolling_validation(self, Ymat, tau=24, n=7, bsize=90, cpu=False, alpha=0.3): last_step = Ymat.shape[1] - tau * n rg = 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels) - 1) self.seq = self.seq.eval() if self.covariates is not None: covs = self.covariates[:, last_step - rg : last_step + tau] else: covs = None if self.Ycov is not None: ycovs = self.Ycov[:, :, last_step - rg : last_step + tau] else: ycovs = None data_in = Ymat[:, last_step - rg : last_step] out = self.predict_future( data_in, covariates=covs, ycovs=ycovs, future=tau, cpu=cpu, bsize=bsize, normalize=self.normalize, ) predicted_values = [] actual_values = [] S = out[:, -tau::] predicted_values += [S] R = Ymat[:, last_step : last_step + tau] actual_values += [R] print("Current window wape: " + str(wape(S, R))) for i in range(n - 1): last_step += tau rg = 1 + 2 * (self.kernel_size - 1) * 2 ** (len(self.num_channels) - 1) if self.covariates is not None: covs = self.covariates[:, last_step - rg : last_step + tau] else: covs = None if self.Ycov is not None: ycovs = self.Ycov[:, :, last_step - rg : last_step + tau] else: ycovs = None data_in = Ymat[:, last_step - rg : last_step] out = self.predict_future( data_in, covariates=covs, ycovs=ycovs, future=tau, cpu=cpu, bsize=bsize, normalize=self.normalize, ) S = out[:, -tau::] predicted_values += [S] R = Ymat[:, last_step : last_step + tau] actual_values += [R] print("Current window wape: " + str(wape(S, R))) predicted = np.hstack(predicted_values) actual = np.hstack(actual_values) dic = {} dic["wape"] = wape(predicted, actual) dic["mape"] = mape(predicted, actual) dic["smape"] = smape(predicted, actual) dic["mae"] = np.abs(predicted - actual).mean() dic["rmse"] = np.sqrt(((predicted - actual) ** 2).mean()) dic["nrmse"] = dic["rmse"] / np.sqrt(((actual) ** 2).mean()) baseline = Ymat[:, Ymat.shape[1] - n * tau - tau : Ymat.shape[1] - tau] dic["baseline_wape"] = wape(baseline, actual) dic["baseline_mape"] = mape(baseline, actual) dic["baseline_smape"] = smape(baseline, actual) return dic
21,683
31.804841
157
py
deepglo
deepglo-master/run_scripts/run_traffic.py
#### OS and commanline arguments import sys import multiprocessing as mp import gzip import subprocess from pathlib import Path import argparse import logging import os sys.path.append('./') #### DeepGLO model imports from DeepGLO.metrics import * from DeepGLO.DeepGLO import * from DeepGLO.LocalModel import * import pandas as pd import numpy as np import pickle import random np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) import json def str2bool(v): if isinstance(v, bool): return v if v.lower() in ("yes", "true", "t", "y", "1"): return True elif v.lower() in ("no", "false", "f", "n", "0"): return False else: raise argparse.ArgumentTypeError("Boolean value expected.") def bool2str(b): if b: return "true" else: return "false" Ymat = np.load("./datasets/traffic.npy") vbsize = 128 ## vertical batch size hbsize = 256 ## horizontal batch size num_channels_X = [32, 32, 32, 32, 32, 1] ## number of channels for local model num_channels_Y = [32, 32, 32, 32, 32, 1] ## number of channels for hybrid model kernel_size = 7 ## kernel size for local models dropout = 0.2 ## dropout during training rank = 64 ## rank of global model kernel_size_Y = 7 ## kernel size of hybrid model lr = 0.0005 ## learning rate val_len = 24 ## validation length end_index = Ymat.shape[1] - 24 * 7 ## models will not look beyond this during training start_date = "2012-1-1" ## start date time for the time-series freq = "H" ## frequency of data covariates = None ## no covraites specified use_time = True ## us time covariates dti = None ## no spcified time covariates (using default) svd = False ## factor matrices are not initialized by NMF period = None ## periodicity of 24 is expected, leave it out if not known y_iters = 300 ## max. number of iterations while training Tconv models init_epochs = 100 ## max number of iterations while initialiozing factors forward_cov = False logging.basicConfig( stream=sys.stdout, format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger = logging.getLogger(__name__) def main(args): DG = DeepGLO( Ymat, vbsize=vbsize, hbsize=hbsize, num_channels_X=num_channels_X, num_channels_Y=num_channels_Y, kernel_size=kernel_size, dropout=dropout, rank=rank, kernel_size_Y=kernel_size_Y, lr=lr, val_len=val_len, end_index=end_index, normalize=normalize, start_date=start_date, freq=freq, covariates=covariates, use_time=use_time, dti=dti, svd=svd, period=period, forward_cov=forward_cov, ) DG.train_all_models(y_iters=y_iters, init_epochs=init_epochs) result_dic = DG.rolling_validation( Ymat=Ymat, tau=24, n=7, bsize=100, cpu=False, alpha=0.3 ) print(result_dic) out_path = Path( ".", "results", "result_dictionary_traffic_" + bool2str(normalize) + ".pkl", ) pickle.dump(result_dic, open(out_path, "wb")) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "--normalize", type=str2bool, required=True, help="normalize for training or not", ) args = parser.parse_args() global normalize normalize = args.normalize main(args)
3,498
24.727941
87
py
deepglo
deepglo-master/run_scripts/run_wiki.py
#### OS and commanline arguments import sys import multiprocessing as mp import gzip import subprocess from pathlib import Path import argparse import logging import os sys.path.append('./') #### DeepGLO model imports from DeepGLO.metrics import * from DeepGLO.DeepGLO import * from DeepGLO.LocalModel import * import pandas as pd import numpy as np import pickle import random np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) import json def str2bool(v): if isinstance(v, bool): return v if v.lower() in ("yes", "true", "t", "y", "1"): return True elif v.lower() in ("no", "false", "f", "n", "0"): return False else: raise argparse.ArgumentTypeError("Boolean value expected.") def bool2str(b): if b: return "true" else: return "false" Ymat = np.load("./datasets/wiki.npy") vbsize = 2048 ## vertical batch size hbsize = 256 ## horizontal batch size num_channels_X = [32, 32, 32, 32, 1] ## number of channels for local model num_channels_Y = [32, 32, 32, 32, 1] ## number of channels for hybrid model kernel_size = 7 ## kernel size for local models dropout = 0.2 ## dropout during training rank = 128 ## rank of global model kernel_size_Y = 7 ## kernel size of hybrid model lr = 0.0005 ## learning rate val_len = 14 ## validation length end_index = Ymat.shape[1] - 14 * 4 ## models will not look beyond this during training start_date = "2012-1-1" ## start date time for the time-series freq = "D" ## frequency of data covariates = None ## no covraites specified use_time = True ## us time covariates dti = None ## no specified time covariates (using default) svd = True ## factor matrices are initialized by NMF period = 7 ## periodicity of 7 is expected, leave it out if not known y_iters = 300 ## max. number of iterations while training Tconv models init_epochs = 100 ## max number of iterations while initializing factors forward_cov = False logging.basicConfig( stream=sys.stdout, format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger = logging.getLogger(__name__) def main(args): DG = DeepGLO( Ymat, vbsize=vbsize, hbsize=hbsize, num_channels_X=num_channels_X, num_channels_Y=num_channels_Y, kernel_size=kernel_size, dropout=dropout, rank=rank, kernel_size_Y=kernel_size_Y, lr=lr, val_len=val_len, end_index=end_index, normalize=normalize, start_date=start_date, freq=freq, covariates=covariates, use_time=use_time, dti=dti, svd=svd, period=period, forward_cov=forward_cov, ) DG.train_all_models(y_iters=y_iters, init_epochs=init_epochs) result_dic = DG.rolling_validation( Ymat=Ymat, tau=14, n=4, bsize=100, cpu=False, alpha=0.5 ) print(result_dic) out_path = Path( ".", "results", "result_dictionary_wiki_" + bool2str(normalize) + ".pkl", ) pickle.dump(result_dic, open(out_path, "wb")) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "--normalize", type=str2bool, required=True, help="normalize for training or not", ) args = parser.parse_args() global normalize normalize = args.normalize main(args)
3,475
24.940299
87
py
deepglo
deepglo-master/run_scripts/run_pems.py
#### OS and commanline arguments import sys import multiprocessing as mp import gzip import subprocess from pathlib import Path import argparse import logging import os sys.path.append('./') #sys.path.append("/efs/users/rajatse/DeepGLOv2/") #### DeepGLO model imports from DeepGLO.metrics import * from DeepGLO.DeepGLO import * from DeepGLO.LocalModel import * import pandas as pd import numpy as np import pickle import random np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) def str2bool(v): if isinstance(v, bool): return v if v.lower() in ("yes", "true", "t", "y", "1"): return True elif v.lower() in ("no", "false", "f", "n", "0"): return False else: raise argparse.ArgumentTypeError("Boolean value expected.") def bool2str(b): if b: return "true" else: return "false" Ymat = np.load("./datasets/pems.npy") print(Ymat.shape) vbsize = 128 ## vertical batch size hbsize = 256 ## horizontal batch size num_channels_X = [32, 32, 32, 32, 32, 1] ## number of channels for local model num_channels_Y = [16, 16, 16, 16, 16, 1] ## number of channels for hybrid model kernel_size = 7 ## kernel size for local models dropout = 0.1 ## dropout during training rank = 64 ## rank of global model kernel_size_Y = 7 ## kernel size of hybrid model lr = 0.0005 ## learning rate val_len = 24 ## validation length end_index = Ymat.shape[1] - 160 * 9 ## models will not look beyond this during training start_date = "2012-5-1" ## start date time for the time-series freq = "5T" ## frequency of data covariates = None ## no covraites specified use_time = True ## us time covariates dti = None ## no spcified time covariates (using default) svd = True ## factor matrices are initialized by NMF period = None ## periodicity of 24 is expected, leave it out if not known y_iters = 300 ## max. number of iterations while training Tconv models init_epochs = 100 ## max number of iterations while initialiozing factors forward_cov = False logging.basicConfig( stream=sys.stdout, format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger = logging.getLogger(__name__) def main(args): DG = DeepGLO( Ymat, vbsize=vbsize, hbsize=hbsize, num_channels_X=num_channels_X, num_channels_Y=num_channels_Y, kernel_size=kernel_size, dropout=dropout, rank=rank, kernel_size_Y=kernel_size_Y, lr=lr, val_len=val_len, end_index=end_index, normalize=normalize, start_date=start_date, freq=freq, covariates=covariates, use_time=use_time, dti=dti, svd=svd, period=period, forward_cov=forward_cov, ) DG.train_all_models(y_iters=y_iters, init_epochs=init_epochs) result_dic = DG.rolling_validation( Ymat=Ymat, tau=9, n=160, bsize=100, cpu=False, alpha=0.3 ) print(result_dic) out_path = Path( ".", "results", "result_dictionary_pems_" + bool2str(normalize) + ".pkl", ) pickle.dump(result_dic, open(out_path, "wb")) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "--normalize", type=str2bool, required=True, help="normalize for training or not", ) args = parser.parse_args() global normalize normalize = args.normalize main(args)
3,544
25.259259
88
py
deepglo
deepglo-master/run_scripts/run_electricity.py
#### OS and commanline arguments import sys import multiprocessing as mp import gzip import subprocess from pathlib import Path import argparse import logging import os sys.path.append('./') #### DeepGLO model imports from DeepGLO.metrics import * from DeepGLO.DeepGLO import * from DeepGLO.LocalModel import * import pandas as pd import numpy as np import pickle import json import random np.random.seed(111) torch.cuda.manual_seed(111) torch.manual_seed(111) random.seed(111) def str2bool(v): if isinstance(v, bool): return v if v.lower() in ("yes", "true", "t", "y", "1"): return True elif v.lower() in ("no", "false", "f", "n", "0"): return False else: raise argparse.ArgumentTypeError("Boolean value expected.") def bool2str(b): if b: return "true" else: return "false" Ymat = np.load("./datasets/electricity.npy") vbsize = 128 ## vertical batch size hbsize = 256 ## horizontal batch size num_channels_X = [32, 32, 32, 32, 32, 1] ## number of channels for local model num_channels_Y = [32, 32, 32, 32, 32, 1] ## number of channels for hybrid model kernel_size = 7 ## kernel size for local models dropout = 0.2 ## dropout during training rank = 64 ## rank of global model kernel_size_Y = 7 ## kernel size of hybrid model lr = 0.0005 ## learning rate val_len = 24 ## validation length end_index = Ymat.shape[1] - 24 * 7 ## models will not look beyond this during training start_date = "2012-1-1" ## start date time for the time-series freq = "H" ## frequency of data covariates = None ## no covraites specified use_time = True ## us time covariates dti = None ## no spcified time covariates (using default) svd = True ## factor matrices are initialized by NMF period = 24 ## periodicity of 24 is expected, leave it out if not known y_iters = 300 ## max. number of iterations while training Tconv models init_epochs = 100 ## max number of iterations while initialiozing factors forward_cov = False logging.basicConfig( stream=sys.stdout, format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger = logging.getLogger(__name__) def main(args): DG = DeepGLO( Ymat, vbsize=vbsize, hbsize=hbsize, num_channels_X=num_channels_X, num_channels_Y=num_channels_Y, kernel_size=kernel_size, dropout=dropout, rank=rank, kernel_size_Y=kernel_size_Y, lr=lr, val_len=val_len, end_index=end_index, normalize=normalize, start_date=start_date, freq=freq, covariates=covariates, use_time=use_time, dti=dti, svd=svd, period=period, forward_cov=forward_cov, ) DG.train_all_models(y_iters=y_iters, init_epochs=init_epochs) result_dic = DG.rolling_validation( Ymat=Ymat, tau=24, n=7, bsize=100, cpu=False, alpha=0.3 ) print(result_dic) out_path = Path("./results", "result_dictionary_electricity_" + bool2str(normalize) + ".pkl", ) pickle.dump(result_dic, open(out_path, "wb")) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "--normalize", type=str2bool, required=True, help="normalize for training or not", ) args = parser.parse_args() global normalize normalize = args.normalize main(args)
3,478
25.157895
87
py
HyperIMBA
HyperIMBA-main/main.py
import argparse import torch import dataloader as dl import torch.nn.functional as F import numpy as np from models import GatHyper, SageHyper, GcnHyper import test as tt def main(args): if args.dataset == 'all': ds_names = ['Cora','Citeseer','Photo','Actor','chameleon','Squirrel'] else: ds_names = [args.dataset] if args.backbone in ['all','Gcn','Gat','Sage']: if args.backbone == 'all': backbones = [b+'Hyper' for b in ['Gcn','Gat','Sage']] else: backbones = [args.backbone+'Hyper'] else: return for ds in ds_names: for babo in backbones: babotrain_acc={babo:[i for i in range(args.run_times)]} babovalid_acc={babo:[i for i in range(args.run_times)]} babotest_acc={babo:[i for i in range(args.run_times)]} babowf1={babo:[i for i in range(args.run_times)]} f2=open('results/'+ds+babo+'_scores.txt', 'w+') f2.write('{0:7} {1:7}\n'.format(ds,babo)) f2.write('{0:7} {1:7} {2:7} {3:7} {4:7}\n'.format('run','train','valid','m-f1','w-f1')) f2.flush() for run in range(args.run_times): dataset,data,train_mask,val_mask,test_mask = dl.select_dataset(ds, args.split) model,data = globals()[babo].call(data,dataset.name,data.x.size(1),dataset.num_classes,args.hid_dim) optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay) best_val_acc = test_acc = 0.0 best_val_loss = np.inf for epoch in range(1, args.epoch+1): model.train() optimizer.zero_grad() F.nll_loss(model(data,args.loss_hp)[train_mask], data.y[train_mask]).backward() optimizer.step() train_acc,val_acc,tmp_test_acc,val_loss,tmp_w_f1 = tt.test(model, data, train_mask, val_mask, test_mask, args.loss_hp) #print("acc:", train_acc,val_acc,tmp_test_acc,val_loss.item(),tmp_w_f1) if val_acc>=best_val_acc: train_re=train_acc best_val_acc=val_acc test_acc=tmp_test_acc w_f1 = tmp_w_f1 best_val_loss=val_loss wait_step=0 else: wait_step += 1 if wait_step == args.stop_step: #print('Early stop! Validate-- Min loss: ', best_val_loss, ', Max f1-score: ', best_val_acc) break del model del data babotrain_acc[babo][run]=train_re babovalid_acc[babo][run]=best_val_acc babotest_acc[babo][run]=test_acc babowf1[babo][run]=w_f1 log ='Epoch: 200, dataset name: '+ ds + ', Backbone: '+ babo + ', Test: {0:.4f} {1:.4f}\n' print((log.format(babotest_acc[babo][run],babowf1[babo][run]))) f2.write('{0:4d} {1:4f} {2:4f} {3:4f} {4:4f}\n'.format(run,babotrain_acc[babo][run],babovalid_acc[babo][run],babotest_acc[babo][run],babowf1[babo][run])) f2.flush() f2.write('{0:4} {1:4f} {2:4f} {3:4f} {4:4f}\n'.format('std',np.std(babotrain_acc[babo]),np.std(babovalid_acc[babo]),np.std(babotest_acc[babo]),np.std(babowf1[babo]))) f2.write('{0:4} {1:4f} {2:4f} {3:4f} {4:4f}\n'.format('mean',np.mean(babotrain_acc[babo]),np.mean(babovalid_acc[babo]),np.mean(babotest_acc[babo]),np.mean(babowf1[babo]))) f2.close() if __name__ == '__main__': parser = argparse.ArgumentParser(description='Hyperbolic Geometric Hierarchy-IMBAlance Learning') parser.add_argument("--dataset", '-d', type=str, default="Cora", help="all,Cora,Citeseer,Photo,Actor,chameleon,Squirrel") parser.add_argument("--backbone", '-b', type=str, default="Gcn", help="all,Gcn,Gat,Sage") parser.add_argument("--split", '-s', type=str, default=0, help="Way of train-set split: 0~5(random,(0.5,1),(0,0.05),(0.66,1),(0.33,0.66),(0,0.33))") parser.add_argument("--gpu", type=int, default=-1, help="GPU index. Default: -1, using CPU.") parser.add_argument("--hid_dim", type=int, default=256, help="Hidden layer dimension") parser.add_argument("--num_layers", type=int, default=2, help="Number of layers") parser.add_argument("--epoch", type=int, default=200, help="Number of epochs. Default: 200") parser.add_argument("--run_times", type=int, default=10, help="Run times") parser.add_argument("--lr", type=float, default=0.01, help="Learning rate. Default: 0.01") parser.add_argument("--weight_decay", type=float, default=0.0005, help="Weight decay. Default: 0.0005") parser.add_argument("--loss_hp", type=float, default=1, help="Loss hyper-parameters (alpha). Default: 1") #parser.add_argument('--early_stop', action='store_true', default=True, help="Indicates whether to use early stop") parser.add_argument('--stop_step', default=100, help="Step of early stop") args = parser.parse_args() print(args) main(args)
5,254
56.119565
183
py
HyperIMBA
HyperIMBA-main/test.py
import torch from sklearn.metrics import f1_score import torch.nn.functional as F def test(model, data, train_mask, val_mask, test_mask, alpha): with torch.no_grad(): model.eval() logits, accs = model(data, alpha), [] for mask in [train_mask,val_mask,test_mask]: pred = logits[mask].max(1)[1] acc = f1_score(pred.cpu(), data.y[mask].cpu(), average='micro') accs.append(acc) accs.append(F.nll_loss(model(data, alpha)[val_mask], data.y[val_mask])) accs.append(f1_score(pred.cpu(), data.y[mask].cpu(), average='weighted')) return accs
617
37.625
81
py
HyperIMBA
HyperIMBA-main/dataloader.py
import torch_geometric.datasets as dt import torch_geometric.transforms as T import torch import numpy as np from dgl.data.utils import generate_mask_tensor, idx2mask from sklearn.model_selection import train_test_split def select_dataset(ds,spcial): if ds=='Cora' or ds=='Citeseer': ds_loader='Planetoid' elif ds=='Photo': ds_loader='Amazon' elif ds == 'chameleon' or ds == 'Squirrel': ds_loader='WikipediaNetwork' else: ds_loader=ds dataset=load_datas(ds_loader,ds,spcial) if ds == 'Actor': data=dataset.data dataset.name = ds else: data=dataset[0] train_mask=data.train_mask val_mask=data.val_mask test_mask=data.test_mask return dataset,data,train_mask,val_mask,test_mask def load_datas(ds_loader,ds,spcial): if ds_loader=='Planetoid': dataset = dt.Planetoid(root='data/'+ds, name=ds, transform=T.NormalizeFeatures()) else: dataset = getattr(dt, ds_loader)('data/'+ds,ds) if ds_loader == 'Actor': dataset.name = ds data = get_split(dataset, spcial) dataset.data = data return dataset def get_split(dataset, spcial): data = dataset.data values=np.load('hyperemb/'+dataset.name+'_values.npy') sorted, indices = torch.sort(torch.norm(torch.tensor(values),dim=1),descending=True) #train set split ratio 1:1:8 if spcial == 1:#Top 50% in the Poincare weight train_idx, val_idx, test_idx = split_idx1(indices[:data.num_nodes//2],indices[data.num_nodes//2:], 0.2, 0.1, 42) elif spcial == 2:#Bottom 50% train_idx, val_idx, test_idx = split_idx1(indices[data.num_nodes//2:],indices[:data.num_nodes//2], 0.2, 0.1, 42) elif spcial == 3:#Top 33% train_idx, val_idx, test_idx = split_idx1(indices[:data.num_nodes//3],indices[data.num_nodes//3:], 0.3, 0.1, 42) elif spcial == 4:#Middle 33% remaining = torch.cat((indices[:data.num_nodes//3],indices[data.num_nodes//3+data.num_nodes//3:])) train_idx, val_idx, test_idx = split_idx1(indices[data.num_nodes//3:data.num_nodes//3+data.num_nodes//3],remaining, 0.3, 0.1, 42) elif spcial == 5:#Bottom 33% train_idx, val_idx, test_idx = split_idx1(indices[data.num_nodes//3+data.num_nodes//3:],indices[:data.num_nodes//3+data.num_nodes//3], 0.3, 0.1, 42) else:#random train_idx, val_idx, test_idx = split_idx(np.arange(data.num_nodes), 0.1, 0.1, 42) data.train_mask = generate_mask_tensor(idx2mask(train_idx, data.num_nodes)) data.val_mask = generate_mask_tensor(idx2mask(val_idx, data.num_nodes)) data.test_mask = generate_mask_tensor(idx2mask(test_idx, data.num_nodes)) return data def split_idx(samples, train_size, val_size, random_state=None): train, val = train_test_split(samples, train_size=train_size, random_state=random_state) if isinstance(val_size, float): val_size *= len(samples) / len(val) val, test = train_test_split(val, train_size=val_size, random_state=random_state) return train, val, test def split_idx1(samples1, samples2, train_size, val_size, random_state=None): train, val = train_test_split(samples1, train_size=train_size, random_state=random_state) val = torch.cat((val,samples2)) val, test = train_test_split(val, train_size=val_size, random_state=random_state) return train, val, test
3,365
41.607595
156
py
HyperIMBA
HyperIMBA-main/calculator.py
#Calculate Hyperbolic Embedding import argparse import torch import numpy as np from models.Poincare import PoincareModel import dataloader as dl from torch_geometric.utils import degree, to_networkx from GraphRicciCurvature.OllivierRicci import OllivierRicci parser = argparse.ArgumentParser(description='Calculate Hyperbolic Embedding') parser.add_argument('--epochs', type=int, default=1) parser.add_argument('--manifolds', type=str, default='poincare', help="ricci, poincare") parser.add_argument("--dataset", '-d', type=str, default="Cora", help="all,Cora,Citeseer,Photo,Actor,chameleon,Squirrel") parser.add_argument("--split", '-s', type=str, default=0, help="Random split train-set") args = parser.parse_args() print(args) dataset,data,_,_,_ = dl.select_dataset(args.dataset, args.split) if args.manifolds=='ricci': G = to_networkx(data) orc = OllivierRicci(G, alpha=0.5, verbose="TRACE") orc.compute_ricci_curvature() G_orc = orc.G.copy() # save an intermediate result curvature="ricciCurvature" ricci_results = {} ricci = {} for i,(n1,n2) in enumerate(list(G_orc.edges()),0): #ricci_results[i] = G_orc[n1][n2][curvature] ricci[i] = [int(n1),int(n2),G_orc[n1][n2][curvature]] weights = [ricci[i] for i in ricci.keys()] np.savetxt('hyperemb/' + args.dataset + '.edge_list',weights,fmt="%d %d %.16f") else: degrees = np.array(degree(data.edge_index[0],num_nodes=data.num_nodes)+degree(data.edge_index[1],num_nodes=data.num_nodes)) edges_list = list(data.edge_index.t().numpy()) labels = dict(enumerate(data.y.numpy()+1, 0)) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') dim = 2 model = PoincareModel(edges_list,node_weights=degrees*0.2,node_labels=labels, n_components=dim,eta=0.01,n_negative=10, name="hierarchy", device=device) model.init_embeddings() model.train(args.epochs) weights = model.embeddings keys = np.array([item for item in model.emb_dict.keys()]) values = np.array([item for item in model.emb_dict.values()]) np.save('hyperemb/' + args.dataset + '_keys.npy', keys) np.save('hyperemb/' + args.dataset + '_values.npy', values)
2,190
41.134615
155
py
HyperIMBA
HyperIMBA-main/models/GcnHyper.py
from typing import Optional, Tuple import numpy as np import torch from torch import Tensor from torch.nn import Parameter from torch_scatter import scatter_add from torch_sparse import SparseTensor, fill_diag, matmul, mul from torch_sparse import sum as sparsesum import torch.nn.functional as F from torch_geometric.nn.conv import MessagePassing from torch_geometric.nn.dense.linear import Linear from torch_geometric.nn.inits import zeros from torch_geometric.typing import Adj, OptTensor, PairTensor from torch_geometric.utils import add_remaining_self_loops from torch_geometric.utils.num_nodes import maybe_num_nodes from torch.nn import Sequential as seq, Parameter,LeakyReLU,init,Linear from torch_geometric.utils import add_self_loops, remove_self_loops,degree,softmax @torch.jit._overload def gcn_norm(edge_index, edge_weight=None, num_nodes=None, improved=False, add_self_loops=True, flow="source_to_target", dtype=None): # type: (Tensor, OptTensor, Optional[int], bool, bool, str, Optional[int]) -> PairTensor # noqa pass @torch.jit._overload def gcn_norm(edge_index, edge_weight=None, num_nodes=None, improved=False, add_self_loops=True, flow="source_to_target", dtype=None): # type: (SparseTensor, OptTensor, Optional[int], bool, bool, str, Optional[int]) -> SparseTensor # noqa pass def gcn_norm(edge_index, edge_weight=None, num_nodes=None, improved=False, add_self_loops=True, flow="source_to_target", dtype=None): fill_value = 2. if improved else 1. if isinstance(edge_index, SparseTensor): assert flow in ["source_to_target"] adj_t = edge_index if not adj_t.has_value(): adj_t = adj_t.fill_value(1., dtype=dtype) if add_self_loops: adj_t = fill_diag(adj_t, fill_value) deg = sparsesum(adj_t, dim=1) deg_inv_sqrt = deg.pow_(-0.5) deg_inv_sqrt.masked_fill_(deg_inv_sqrt == float('inf'), 0.) adj_t = mul(adj_t, deg_inv_sqrt.view(-1, 1)) adj_t = mul(adj_t, deg_inv_sqrt.view(1, -1)) return adj_t else: assert flow in ["source_to_target", "target_to_source"] num_nodes = maybe_num_nodes(edge_index, num_nodes) if edge_weight is None: edge_weight = torch.ones((edge_index.size(1), ), dtype=dtype, device=edge_index.device) if add_self_loops: edge_index, tmp_edge_weight = add_remaining_self_loops( edge_index, edge_weight, fill_value, num_nodes) assert tmp_edge_weight is not None edge_weight = tmp_edge_weight row, col = edge_index[0], edge_index[1] idx = col if flow == "source_to_target" else row deg = scatter_add(edge_weight, idx, dim=0, dim_size=num_nodes) deg_inv_sqrt = deg.pow_(-0.5) deg_inv_sqrt.masked_fill_(deg_inv_sqrt == float('inf'), 0) return edge_index, deg_inv_sqrt[row] * edge_weight * deg_inv_sqrt[col] class GCNConv(MessagePassing): r"""The graph convolutional operator from the `"Semi-supervised Classification with Graph Convolutional Networks" <https://arxiv.org/abs/1609.02907>`_ paper .. math:: \mathbf{X}^{\prime} = \mathbf{\hat{D}}^{-1/2} \mathbf{\hat{A}} \mathbf{\hat{D}}^{-1/2} \mathbf{X} \mathbf{\Theta}, where :math:`\mathbf{\hat{A}} = \mathbf{A} + \mathbf{I}` denotes the adjacency matrix with inserted self-loops and :math:`\hat{D}_{ii} = \sum_{j=0} \hat{A}_{ij}` its diagonal degree matrix. The adjacency matrix can include other values than :obj:`1` representing edge weights via the optional :obj:`edge_weight` tensor. Its node-wise formulation is given by: .. math:: \mathbf{x}^{\prime}_i = \mathbf{\Theta}^{\top} \sum_{j \in \mathcal{N}(v) \cup \{ i \}} \frac{e_{j,i}}{\sqrt{\hat{d}_j \hat{d}_i}} \mathbf{x}_j with :math:`\hat{d}_i = 1 + \sum_{j \in \mathcal{N}(i)} e_{j,i}`, where :math:`e_{j,i}` denotes the edge weight from source node :obj:`j` to target node :obj:`i` (default: :obj:`1.0`) Args: in_channels (int): Size of each input sample, or :obj:`-1` to derive the size from the first input(s) to the forward method. out_channels (int): Size of each output sample. improved (bool, optional): If set to :obj:`True`, the layer computes :math:`\mathbf{\hat{A}}` as :math:`\mathbf{A} + 2\mathbf{I}`. (default: :obj:`False`) cached (bool, optional): If set to :obj:`True`, the layer will cache the computation of :math:`\mathbf{\hat{D}}^{-1/2} \mathbf{\hat{A}} \mathbf{\hat{D}}^{-1/2}` on first execution, and will use the cached version for further executions. This parameter should only be set to :obj:`True` in transductive learning scenarios. (default: :obj:`False`) add_self_loops (bool, optional): If set to :obj:`False`, will not add self-loops to the input graph. (default: :obj:`True`) normalize (bool, optional): Whether to add self-loops and compute symmetric normalization coefficients on the fly. (default: :obj:`True`) bias (bool, optional): If set to :obj:`False`, the layer will not learn an additive bias. (default: :obj:`True`) **kwargs (optional): Additional arguments of :class:`torch_geometric.nn.conv.MessagePassing`. Shapes: - **input:** node features :math:`(|\mathcal{V}|, F_{in})`, edge indices :math:`(2, |\mathcal{E}|)`, edge weights :math:`(|\mathcal{E}|)` *(optional)* - **output:** node features :math:`(|\mathcal{V}|, F_{out})` """ _cached_edge_index: Optional[Tuple[Tensor, Tensor]] _cached_adj_t: Optional[SparseTensor] def __init__(self, in_channels: int, out_channels: int, k_ricci,e_poinc,n_components,n_components_p, improved: bool = False, cached: bool = False, add_self_loops: bool = True, normalize: bool = True, bias: bool = True, **kwargs): kwargs.setdefault('aggr', 'add') super().__init__(**kwargs) self.in_channels = in_channels self.out_channels = out_channels self.improved = improved self.cached = cached self.add_self_loops = add_self_loops self.normalize = normalize self.k_ricci = k_ricci self.e_poinc = e_poinc self._cached_edge_index = None self._cached_adj_t = None self.lin = Linear(in_channels, out_channels, bias=False) widths=[n_components,out_channels] widths_p=[n_components_p,out_channels] self.hmpnn=create_wmlp(widths,out_channels,1) self.ham=create_wmlp(widths_p,out_channels,1) if bias: self.bias = Parameter(torch.Tensor(out_channels)) else: self.register_parameter('bias', None) self.reset_parameters() def reset_parameters(self): self.lin.reset_parameters() zeros(self.bias) self._cached_edge_index = None self._cached_adj_t = None def forward(self, x: Tensor, edge_index: Adj, alpha_hp: float, edge_weight: OptTensor = None) -> Tensor: """""" if self.normalize: if isinstance(edge_index, Tensor): cache = self._cached_edge_index if cache is None: edge_index, edge_weight = gcn_norm( # yapf: disable edge_index, edge_weight, x.size(self.node_dim), self.improved, self.add_self_loops, self.flow) if self.cached: self._cached_edge_index = (edge_index, edge_weight) else: edge_index, edge_weight = cache[0], cache[1] elif isinstance(edge_index, SparseTensor): cache = self._cached_adj_t if cache is None: edge_index = gcn_norm( # yapf: disable edge_index, edge_weight, x.size(self.node_dim), self.improved, self.add_self_loops, self.flow) if self.cached: self._cached_adj_t = edge_index else: edge_index = cache edge_weight = edge_weight.view(-1, 1) x = self.lin(x) edge_weight=self.hmpnn(self.k_ricci) edge_weight=softmax(edge_weight,edge_index[0]) # propagate_type: (x: Tensor, edge_weight: OptTensor) out = self.propagate(edge_index, x=x, edge_weight=edge_weight, size=None) p_weight=self.ham(self.e_poinc) p_weight=F.leaky_relu(p_weight) if self.bias is not None: out += self.bias return out+alpha_hp*p_weight def message(self, x_j: Tensor, edge_weight: OptTensor) -> Tensor: return x_j if edge_weight is None else edge_weight * x_j def update(self, aggr_out): return aggr_out def message_and_aggregate(self, adj_t: SparseTensor, x: Tensor) -> Tensor: return matmul(adj_t, x, reduce=self.aggr) def create_wmlp(widths,nfeato,lbias): mlp_modules=[] for k in range(len(widths)-1): mlp_modules.append(Linear(widths[k],widths[k+1],bias=False)) mlp_modules.append(LeakyReLU(0.2,True)) mlp_modules.append(Linear(widths[len(widths)-1],nfeato,bias=lbias)) return seq(*mlp_modules) class Net(torch.nn.Module): def __init__(self,data,num_features,num_hidden,num_classes,k_ricci,e_poinc,n_components,n_components_p): super(Net, self).__init__() self.conv1 = GCNConv(num_features, num_hidden,k_ricci,e_poinc,n_components,n_components_p, cached=True) self.conv2 = GCNConv(num_hidden, num_classes, k_ricci,e_poinc,n_components,n_components_p, cached=True) def forward(self,data,alpha): x = F.dropout(data.x,p=0.6,training=self.training) x = self.conv1(x, data.edge_index, alpha) x = F.elu(x) x = F.dropout(x,p=0.6,training=self.training) x = self.conv2(x, data.edge_index, alpha) return F.log_softmax(x, dim=1) def num(strings): try: return int(strings) except ValueError: return float(strings) def call(data,name,num_features,num_classes,num_hidden): #ricci filename='hyperemb/'+name+'.edge_list' f=open(filename) cur_list=list(f) if name=='Cora' or name == 'Actor' or name=='chameleon' or name=='squirrel': ricci_cur=[[] for i in range(len(cur_list))] for i in range(len(cur_list)): ricci_cur[i]=[num(s) for s in cur_list[i].split(' ',2)] else: ricci_cur=[[] for i in range(2*len(cur_list))] for i in range(len(cur_list)): ricci_cur[i]=[num(s) for s in cur_list[i].split(' ',2)] ricci_cur[i+len(cur_list)]=[ricci_cur[i][1],ricci_cur[i][0],ricci_cur[i][2]] ricci_cur=sorted(ricci_cur) k_ricci=[i[2] for i in ricci_cur] k_ricci=k_ricci+[0 for i in range(data.x.size(0))] k_ricci=torch.tensor(k_ricci, dtype=torch.float) data.k_ricci=k_ricci.view(-1,1) data.n_components=1 #poincare data.edge_index, _ = remove_self_loops(data.edge_index) keys=np.load('hyperemb/'+name+'_keys.npy') values=np.load('hyperemb/'+name+'_values.npy') e_poinc = dict(zip(keys, values)) data.n_components_p = values.shape[1] alls = dict(enumerate(np.ones((data.num_nodes,data.n_components_p)), 0)) alls.update(e_poinc) e_poinc = torch.tensor(np.array([alls[i] for i in alls])) data.e_poinc = e_poinc.to(torch.float32) data.edge_index, _ = add_self_loops(data.edge_index,num_nodes=data.x.size(0)) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') data.k_ricci = data.k_ricci.to(device) data.e_poinc = data.e_poinc.to(device) data = data.to(device) model= Net(data,num_features,num_hidden,num_classes,data.k_ricci,data.e_poinc,data.n_components,data.n_components_p).to(device) return model, data
12,196
40.06734
131
py
HyperIMBA
HyperIMBA-main/models/SageHyper.py
import numpy as np import torch from torch.nn import Sequential as seq, Parameter,LeakyReLU,init,Linear from typing import List, Optional, Tuple, Union import torch.nn.functional as F from torch import Tensor from torch.nn import LSTM from torch_sparse import SparseTensor, matmul from torch_geometric.nn.aggr import Aggregation, MultiAggregation from torch_geometric.nn.conv import MessagePassing from torch_geometric.nn.dense.linear import Linear from torch_geometric.typing import Adj, OptPairTensor, Size from torch_geometric.utils import add_self_loops, remove_self_loops,degree,softmax class SAGEConv(MessagePassing): r"""The GraphSAGE operator from the `"Inductive Representation Learning on Large Graphs" <https://arxiv.org/abs/1706.02216>`_ paper .. math:: \mathbf{x}^{\prime}_i = \mathbf{W}_1 \mathbf{x}_i + \mathbf{W}_2 \cdot \mathrm{mean}_{j \in \mathcal{N(i)}} \mathbf{x}_j If :obj:`project = True`, then :math:`\mathbf{x}_j` will first get projected via .. math:: \mathbf{x}_j \leftarrow \sigma ( \mathbf{W}_3 \mathbf{x}_j + \mathbf{b}) as described in Eq. (3) of the paper. Args: in_channels (int or tuple): Size of each input sample, or :obj:`-1` to derive the size from the first input(s) to the forward method. A tuple corresponds to the sizes of source and target dimensionalities. out_channels (int): Size of each output sample. aggr (string or Aggregation, optional): The aggregation scheme to use. Any aggregation of :obj:`torch_geometric.nn.aggr` can be used, *e.g.*, :obj:`"mean"`, :obj:`"max"`, or :obj:`"lstm"`. (default: :obj:`"mean"`) normalize (bool, optional): If set to :obj:`True`, output features will be :math:`\ell_2`-normalized, *i.e.*, :math:`\frac{\mathbf{x}^{\prime}_i} {\| \mathbf{x}^{\prime}_i \|_2}`. (default: :obj:`False`) root_weight (bool, optional): If set to :obj:`False`, the layer will not add transformed root node features to the output. (default: :obj:`True`) project (bool, optional): If set to :obj:`True`, the layer will apply a linear transformation followed by an activation function before aggregation (as described in Eq. (3) of the paper). (default: :obj:`False`) bias (bool, optional): If set to :obj:`False`, the layer will not learn an additive bias. (default: :obj:`True`) **kwargs (optional): Additional arguments of :class:`torch_geometric.nn.conv.MessagePassing`. Shapes: - **inputs:** node features :math:`(|\mathcal{V}|, F_{in})` or :math:`((|\mathcal{V_s}|, F_{s}), (|\mathcal{V_t}|, F_{t}))` if bipartite, edge indices :math:`(2, |\mathcal{E}|)` - **outputs:** node features :math:`(|\mathcal{V}|, F_{out})` or :math:`(|\mathcal{V_t}|, F_{out})` if bipartite """ def __init__( self, in_channels: Union[int, Tuple[int, int]], out_channels: int, k_ricci,e_poinc,n_components,n_components_p, aggr: Optional[Union[str, List[str], Aggregation]] = "mean", normalize: bool = False, root_weight: bool = True, project: bool = False, bias: bool = True, **kwargs, ): self.in_channels = in_channels self.out_channels = out_channels self.normalize = normalize self.root_weight = root_weight self.project = project self.k_ricci = k_ricci self.e_poinc = e_poinc if isinstance(in_channels, int): in_channels = (in_channels, in_channels) if aggr == 'lstm': kwargs.setdefault('aggr_kwargs', {}) kwargs['aggr_kwargs'].setdefault('in_channels', in_channels[0]) kwargs['aggr_kwargs'].setdefault('out_channels', in_channels[0]) super().__init__(aggr, **kwargs) widths=[n_components,out_channels] widths_p=[n_components_p,out_channels] self.hmpnn=create_wmlp(widths,in_channels[0],1) self.ham=create_wmlp(widths_p,out_channels,1) if self.project: self.lin = Linear(in_channels[0], in_channels[0], bias=True) if self.aggr is None: self.fuse = False # No "fused" message_and_aggregate. self.lstm = LSTM(in_channels[0], in_channels[0], batch_first=True) if isinstance(self.aggr_module, MultiAggregation): aggr_out_channels = self.aggr_module.get_out_channels( in_channels[0]) else: aggr_out_channels = in_channels[0] self.lin_l = Linear(aggr_out_channels, out_channels, bias=bias) if self.root_weight: self.lin_r = Linear(in_channels[1], out_channels, bias=False) self.reset_parameters() def reset_parameters(self): if self.project: self.lin.reset_parameters() self.aggr_module.reset_parameters() self.lin_l.reset_parameters() if self.root_weight: self.lin_r.reset_parameters() def forward(self, x: Union[Tensor, OptPairTensor], edge_index: Adj, alpha_hp: float, size: Size = None) -> Tensor: """""" if isinstance(x, Tensor): x: OptPairTensor = (x, x) if self.project and hasattr(self, 'lin'): x = (self.lin(x[0]).relu(), x[1]) # propagate_type: (x: OptPairTensor) out_weight=self.hmpnn(self.k_ricci) out_weight=softmax(out_weight,edge_index[0]) out = self.propagate(x=x,edge_index=edge_index,out_weight=out_weight) out = self.lin_l(out) p_weight=self.ham(self.e_poinc) p_weight=F.leaky_relu(p_weight) out = out+alpha_hp*p_weight x_r = x[1] if self.root_weight and x_r is not None: out += self.lin_r(x_r) if self.normalize: out = F.normalize(out, p=2., dim=-1) return out def message(self, x_j: Tensor, out_weight: Tensor) -> Tensor: return out_weight*x_j def message_and_aggregate(self, adj_t: SparseTensor, x: OptPairTensor) -> Tensor: adj_t = adj_t.set_value(None, layout=None) return matmul(adj_t, x[0], reduce=self.aggr) def __repr__(self) -> str: return (f'{self.__class__.__name__}({self.in_channels}, ' f'{self.out_channels}, aggr={self.aggr})') def create_wmlp(widths,nfeato,lbias): mlp_modules=[] for k in range(len(widths)-1): mlp_modules.append(Linear(widths[k],widths[k+1],bias=False)) mlp_modules.append(LeakyReLU(0.2,True)) mlp_modules.append(Linear(widths[len(widths)-1],nfeato,bias=lbias)) return seq(*mlp_modules) class Net(torch.nn.Module): def __init__(self,data,num_features,num_hidden,num_classes,k_ricci,e_poinc,n_components,n_components_p): super(Net, self).__init__() self.conv1 = SAGEConv(num_features, num_hidden,k_ricci,e_poinc,n_components,n_components_p) self.conv2 = SAGEConv(num_hidden, num_classes,k_ricci,e_poinc,n_components,n_components_p) def forward(self, data, alpha): x, edge_index = data.x, data.edge_index x = F.dropout(x, p=0.6, training=self.training) x = F.relu(self.conv1(x, edge_index, alpha)) x = F.dropout(x, training=self.training) x = self.conv2(x, edge_index, alpha) return F.log_softmax(x, dim=1) def num(strings): try: return int(strings) except ValueError: return float(strings) def call(data,name,num_features,num_classes,num_hidden): #ricci filename='hyperemb/'+name+'.edge_list' f=open(filename) cur_list=list(f) if name=='Cora' or name == 'Actor' or name=='chameleon' or name=='squirrel': ricci_cur=[[] for i in range(len(cur_list))] for i in range(len(cur_list)): ricci_cur[i]=[num(s) for s in cur_list[i].split(' ',2)] else: ricci_cur=[[] for i in range(2*len(cur_list))] for i in range(len(cur_list)): ricci_cur[i]=[num(s) for s in cur_list[i].split(' ',2)] ricci_cur[i+len(cur_list)]=[ricci_cur[i][1],ricci_cur[i][0],ricci_cur[i][2]] ricci_cur=sorted(ricci_cur) k_ricci=[i[2] for i in ricci_cur] k_ricci=k_ricci+[0 for i in range(data.x.size(0))] k_ricci=torch.tensor(k_ricci, dtype=torch.float) data.k_ricci=k_ricci.view(-1,1) data.n_components=1 #poincare data.edge_index, _ = remove_self_loops(data.edge_index) keys=np.load('hyperemb/'+name+'_keys.npy') values=np.load('hyperemb/'+name+'_values.npy') e_poinc = dict(zip(keys, values)) data.n_components_p = values.shape[1] alls = dict(enumerate(np.ones((data.num_nodes,data.n_components_p)), 0)) alls.update(e_poinc) e_poinc = torch.tensor(np.array([alls[i] for i in alls])) data.e_poinc = e_poinc.to(torch.float32) data.edge_index, _ = add_self_loops(data.edge_index,num_nodes=data.x.size(0)) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') data.k_ricci = data.k_ricci.to(device) data.e_poinc = data.e_poinc.to(device) data = data.to(device) model= Net(data,num_features,num_hidden,num_classes,data.k_ricci,data.e_poinc,data.n_components,data.n_components_p).to(device) return model, data
9,477
38.327801
131
py
HyperIMBA
HyperIMBA-main/models/GatHyper.py
from typing import Optional, Tuple, Union import torch import torch.nn.functional as F from torch import Tensor from torch.nn import Parameter from torch_sparse import SparseTensor, set_diag import math import numpy as np from typing import Any from torch.nn import Sequential as seq, Parameter,LeakyReLU,init,Linear from torch_geometric.nn.conv import MessagePassing from torch_geometric.nn.dense.linear import Linear from torch_geometric.typing import ( Adj, NoneType, OptPairTensor, OptTensor, Size, ) from torch_geometric.utils import add_self_loops, remove_self_loops, softmax #from ..inits import glorot, zeros class GATConv(MessagePassing): def __init__( self, in_channels: Union[int, Tuple[int, int]], out_channels: int, heads, k_ricci,e_poinc,n_components,n_components_p, concat: bool = True, negative_slope: float = 0.2, dropout: float = 0.0, add_self_loops: bool = True, edge_dim: Optional[int] = None, fill_value: Union[float, Tensor, str] = 'mean', bias: bool = True, **kwargs, ): kwargs.setdefault('aggr', 'add') super().__init__(node_dim=0, **kwargs) self.in_channels = in_channels self.out_channels = out_channels self.heads = heads self.concat = concat self.negative_slope = negative_slope self.dropout = dropout self.add_self_loops = add_self_loops self.edge_dim = edge_dim self.fill_value = fill_value self.k_ricci = k_ricci self.e_poinc = e_poinc widths=[n_components,out_channels] widths_p=[n_components_p,out_channels*heads] self.hmpnn=create_wmlp(widths,out_channels,1) self.ham=create_wmlp(widths_p,out_channels*heads,1) # In case we are operating in bipartite graphs, we apply separate # transformations 'lin_src' and 'lin_dst' to source and target nodes: if isinstance(in_channels, int): self.lin_src = Linear(in_channels, heads * out_channels, bias=False, weight_initializer='glorot') self.lin_dst = self.lin_src else: self.lin_src = Linear(in_channels[0], heads * out_channels, False, weight_initializer='glorot') self.lin_dst = Linear(in_channels[1], heads * out_channels, False, weight_initializer='glorot') # The learnable parameters to compute attention coefficients: self.att_src = Parameter(torch.Tensor(1, heads, out_channels)) self.att_dst = Parameter(torch.Tensor(1, heads, out_channels)) if edge_dim is not None: self.lin_edge = Linear(edge_dim, heads * out_channels, bias=False, weight_initializer='glorot') self.att_edge = Parameter(torch.Tensor(1, heads, out_channels)) else: self.lin_edge = None self.register_parameter('att_edge', None) if bias and concat: self.bias = Parameter(torch.Tensor(heads * out_channels)) elif bias and not concat: self.bias = Parameter(torch.Tensor(out_channels)) else: self.register_parameter('bias', None) self.reset_parameters() def reset_parameters(self): self.lin_src.reset_parameters() self.lin_dst.reset_parameters() if self.lin_edge is not None: self.lin_edge.reset_parameters() glorot(self.att_src) glorot(self.att_dst) glorot(self.att_edge) zeros(self.bias) def forward(self, x: Union[Tensor, OptPairTensor], edge_index: Adj, alpha_hp: float, edge_attr: OptTensor = None, size: Size = None, return_attention_weights=None): H, C = self.heads, self.out_channels # We first transform the input node features. If a tuple is passed, we # transform source and target node features via separate weights: if isinstance(x, Tensor): assert x.dim() == 2, "Static graphs not supported in 'GATConv'" x_src = x_dst = self.lin_src(x).view(-1, H, C) else: # Tuple of source and target node features: x_src, x_dst = x assert x_src.dim() == 2, "Static graphs not supported in 'GATConv'" x_src = self.lin_src(x_src).view(-1, H, C) if x_dst is not None: x_dst = self.lin_dst(x_dst).view(-1, H, C) x = (x_src, x_dst) # Next, we compute node-level attention coefficients, both for source # and target nodes (if present): alpha_src = (x_src * self.att_src).sum(dim=-1) alpha_dst = None if x_dst is None else (x_dst * self.att_dst).sum(-1) alpha = (alpha_src, alpha_dst) if self.add_self_loops: if isinstance(edge_index, Tensor): # We only want to add self-loops for nodes that appear both as # source and target nodes: num_nodes = x_src.size(0) if x_dst is not None: num_nodes = min(num_nodes, x_dst.size(0)) num_nodes = min(size) if size is not None else num_nodes edge_index, edge_attr = remove_self_loops( edge_index, edge_attr) edge_index, edge_attr = add_self_loops( edge_index, edge_attr, fill_value=self.fill_value, num_nodes=num_nodes) elif isinstance(edge_index, SparseTensor): if self.edge_dim is None: edge_index = set_diag(edge_index) else: raise NotImplementedError( "The usage of 'edge_attr' and 'add_self_loops' " "simultaneously is currently not yet supported for " "'edge_index' in a 'SparseTensor' form") # edge_updater_type: (alpha: OptPairTensor, edge_attr: OptTensor) alpha = self.edge_updater(edge_index, alpha=alpha, edge_attr=edge_attr) # propagate_type: (x: OptPairTensor, alpha: Tensor) #hyperIMBA out_weight=self.hmpnn(self.k_ricci) out_weight=softmax(out_weight,edge_index[0]) alpha = out_weight # alpha = alpha+out_weight out = self.propagate(edge_index, x=x, alpha=alpha, size=size) if self.concat: out = out.view(-1, self.heads * self.out_channels) else: out = out.mean(dim=1) if self.bias is not None: out += self.bias p_weight=self.ham(self.e_poinc) p_weight=F.leaky_relu(p_weight) out = out+alpha_hp*p_weight if isinstance(return_attention_weights, bool): if isinstance(edge_index, Tensor): return out, (edge_index, alpha) elif isinstance(edge_index, SparseTensor): return out, edge_index.set_value(alpha, layout='coo') else: return out def edge_update(self, alpha_j: Tensor, alpha_i: OptTensor, edge_attr: OptTensor, index: Tensor, ptr: OptTensor, size_i: Optional[int]) -> Tensor: # Given edge-level attention coefficients for source and target nodes, # we simply need to sum them up to "emulate" concatenation: alpha = alpha_j if alpha_i is None else alpha_j + alpha_i if edge_attr is not None and self.lin_edge is not None: if edge_attr.dim() == 1: edge_attr = edge_attr.view(-1, 1) edge_attr = self.lin_edge(edge_attr) edge_attr = edge_attr.view(-1, self.heads, self.out_channels) alpha_edge = (edge_attr * self.att_edge).sum(dim=-1) alpha = alpha + alpha_edge alpha = F.leaky_relu(alpha, self.negative_slope) alpha = softmax(alpha, index, ptr, size_i) alpha = F.dropout(alpha, p=self.dropout, training=self.training) return alpha def message(self, x_j: Tensor, alpha: Tensor) -> Tensor: return alpha.unsqueeze(1) * x_j #return alpha.unsqueeze(-1) * x_j def __repr__(self) -> str: return (f'{self.__class__.__name__}({self.in_channels}, ' f'{self.out_channels}, heads={self.heads})') def glorot(value: Any): if isinstance(value, Tensor): stdv = math.sqrt(6.0 / (value.size(-2) + value.size(-1))) value.data.uniform_(-stdv, stdv) else: for v in value.parameters() if hasattr(value, 'parameters') else []: glorot(v) for v in value.buffers() if hasattr(value, 'buffers') else []: glorot(v) def zeros(value: Any): constant(value, 0.) def constant(value: Any, fill_value: float): if isinstance(value, Tensor): value.data.fill_(fill_value) else: for v in value.parameters() if hasattr(value, 'parameters') else []: constant(v, fill_value) for v in value.buffers() if hasattr(value, 'buffers') else []: constant(v, fill_value) def create_wmlp(widths,nfeato,lbias): mlp_modules=[] for k in range(len(widths)-1): mlp_modules.append(Linear(widths[k],widths[k+1],bias=False)) mlp_modules.append(LeakyReLU(0.2,True)) mlp_modules.append(Linear(widths[len(widths)-1],nfeato,bias=lbias)) return seq(*mlp_modules) class Net(torch.nn.Module): def __init__(self,data,num_features,num_hidden,heads,num_classes,k_ricci,e_poinc,n_components,n_components_p): super(Net, self).__init__() self.conv1 = GATConv(num_features, num_hidden, heads,k_ricci,e_poinc,n_components,n_components_p) self.conv2 = GATConv(num_hidden * heads, num_classes, 1,k_ricci,e_poinc,n_components,n_components_p,concat=False) def forward(self, data, alpha): x, edge_index = data.x, data.edge_index x = F.dropout(x, p=0.6, training=self.training) x = F.relu(self.conv1(x, edge_index, alpha)) x = F.dropout(x, training=self.training) x = self.conv2(x, edge_index, alpha) return F.log_softmax(x, dim=1) def num(strings): try: return int(strings) except ValueError: return float(strings) def call(data,name,num_features,num_classes,num_hidden): #ricci filename='hyperemb/'+name+'.edge_list' f=open(filename) cur_list=list(f) if name=='Cora' or name == 'Actor' or name=='chameleon' or name=='squirrel': ricci_cur=[[] for i in range(len(cur_list))] for i in range(len(cur_list)): ricci_cur[i]=[num(s) for s in cur_list[i].split(' ',2)] else: ricci_cur=[[] for i in range(2*len(cur_list))] for i in range(len(cur_list)): ricci_cur[i]=[num(s) for s in cur_list[i].split(' ',2)] ricci_cur[i+len(cur_list)]=[ricci_cur[i][1],ricci_cur[i][0],ricci_cur[i][2]] ricci_cur=sorted(ricci_cur) k_ricci=[i[2] for i in ricci_cur] k_ricci=k_ricci+[0 for i in range(data.x.size(0))] k_ricci=torch.tensor(k_ricci, dtype=torch.float) data.k_ricci=k_ricci.view(-1,1) data.n_components=1 #poincare data.edge_index, _ = remove_self_loops(data.edge_index) keys=np.load('hyperemb/'+name+'_keys.npy') values=np.load('hyperemb/'+name+'_values.npy') e_poinc = dict(zip(keys, values)) data.n_components_p = values.shape[1] alls = dict(enumerate(np.ones((data.num_nodes,data.n_components_p)), 0)) alls.update(e_poinc) e_poinc = torch.tensor(np.array([alls[i] for i in alls])) data.e_poinc = e_poinc.to(torch.float32) data.edge_index, _ = add_self_loops(data.edge_index,num_nodes=data.x.size(0)) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') data.k_ricci = data.k_ricci.to(device) data.e_poinc = data.e_poinc.to(device) data = data.to(device) model= Net(data,num_features,num_hidden//8,8,num_classes,data.k_ricci,data.e_poinc,data.n_components,data.n_components_p).to(device) return model, data
12,092
39.043046
136
py
larq
larq-main/larq/optimizers_test.py
import numpy as np import pytest import tensorflow as tf from packaging import version from tensorflow import keras from tensorflow.python.keras import testing_utils import larq as lq from larq import testing_utils as lq_testing_utils if version.parse(tf.__version__) >= version.parse("2.11"): from tensorflow.keras.optimizers import legacy as optimizers # type: ignore else: from tensorflow.keras import optimizers # type: ignore def _test_optimizer( optimizer, target=0.75, test_kernels_are_binary=True, trainable_bn=True ): np.random.seed(1337) (x_train, y_train), _ = testing_utils.get_test_data( train_samples=1000, test_samples=0, input_shape=(10,), num_classes=2 ) y_train = keras.utils.to_categorical(y_train) model = lq_testing_utils.get_small_bnn_model( x_train.shape[1], 20, y_train.shape[1], trainable_bn=trainable_bn ) model.compile(loss="categorical_crossentropy", optimizer=optimizer, metrics=["acc"]) initial_vars = [tf.keras.backend.get_value(w) for w in model.trainable_weights] history = model.fit(x_train, y_train, epochs=2, batch_size=16, verbose=0) trained_vars = [tf.keras.backend.get_value(w) for w in model.trainable_weights] # check all trainable variables have actually been updated for v0, v1 in zip(initial_vars, trained_vars): assert not np.all(v0 == v1) # Note that when kernels are treated as latent weights they need not be # binary (see https://arxiv.org/abs/1906.02107 for further discussion) if test_kernels_are_binary: for layer in model.layers: if "quant" in layer.name: for weight in layer.trainable_weights: assert np.all(np.isin(tf.keras.backend.get_value(weight), [-1, 1])) assert history.history["acc"][-1] >= target def _test_serialization(optimizer): config = keras.optimizers.serialize(optimizer) optim = keras.optimizers.deserialize(config) new_config = keras.optimizers.serialize(optim) assert config == new_config class TestCaseOptimizer: def test_type_check_predicate(self): with pytest.raises(TypeError): # pytype: disable=wrong-arg-types lq.optimizers.CaseOptimizer((False, lq.optimizers.Bop())) # pytype: enable=wrong-arg-types def test_type_check_optimizer(self): with pytest.raises(TypeError): lq.optimizers.CaseOptimizer((lq.optimizers.Bop.is_binary_variable, False)) def test_type_check_default(self): with pytest.raises(TypeError): lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, lq.optimizers.Bop()), default_optimizer=False, ) def test_overlapping_predicates(self): with pytest.raises(ValueError): naughty_case_opt = lq.optimizers.CaseOptimizer( (lambda var: True, lq.optimizers.Bop()), (lambda var: True, lq.optimizers.Bop()), ) _test_optimizer(naughty_case_opt) def test_missing_default(self): with pytest.warns(Warning): naughty_case_opt = lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, lq.optimizers.Bop()), ) # Simple MNIST model mnist = tf.keras.datasets.mnist (train_images, train_labels), _ = mnist.load_data() model = tf.keras.Sequential( [ tf.keras.layers.Flatten(input_shape=(28, 28)), lq.layers.QuantDense( 64, input_quantizer="ste_sign", kernel_quantizer=lq.quantizers.NoOp(precision=1), activation="relu", ), tf.keras.layers.Dense(10, activation="softmax"), ] ) model.compile( loss="sparse_categorical_crossentropy", optimizer=naughty_case_opt, metrics=["acc"], ) # Should raise on first call to apply_gradients() model.fit(train_images[:1], train_labels[:1], epochs=1) def test_wrong_predicate(self): """Make sure we throw when an optimizer does not claim variables.""" with pytest.raises(ValueError): naughty_case_opt = lq.optimizers.CaseOptimizer( (lambda var: False, lq.optimizers.Bop()), default_optimizer=optimizers.Adam(0.01), ) # Simple MNIST model mnist = tf.keras.datasets.mnist (train_images, train_labels), _ = mnist.load_data() model = tf.keras.Sequential( [ tf.keras.layers.Flatten(input_shape=(28, 28)), tf.keras.layers.Dense(128, activation="relu"), tf.keras.layers.Dense(10, activation="softmax"), ] ) model.compile( loss="sparse_categorical_crossentropy", optimizer=naughty_case_opt, metrics=["acc"], ) # Should raise on first call to apply_gradients() model.fit(train_images[:1], train_labels[:1], epochs=1) def test_weights(self): (train_images, train_labels), _ = tf.keras.datasets.mnist.load_data() model = tf.keras.Sequential( [ tf.keras.layers.Flatten(input_shape=(28, 28)), lq.layers.QuantDense( 64, input_quantizer="ste_sign", kernel_quantizer=lq.quantizers.NoOp(precision=1), activation="relu", ), tf.keras.layers.Dense(10, activation="softmax"), ] ) model.compile( loss="sparse_categorical_crossentropy", optimizer=lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, lq.optimizers.Bop()), default_optimizer=optimizers.SGD(0.1, momentum=0.9), ), ) model.fit(train_images[:1], train_labels[:1], epochs=1) opt_weights = model.optimizer.weights # SGD with momentum and Bop both create a single momentum variable per weight # and one variable each to keep track of iterations assert len(opt_weights) == len(model.weights) + 2 checked_weights = 0 for opt in model.optimizer.optimizers: for weight in opt.weights: assert weight is opt_weights[checked_weights] checked_weights += 1 assert checked_weights == len(opt_weights) @pytest.mark.usefixtures("eager_mode") def test_checkpoint(self, tmp_path): # Build and run a simple model. var = tf.Variable([2.0]) opt = optimizers.SGD(1.0, momentum=1.0) opt = lq.optimizers.CaseOptimizer((lambda var: True, opt)) opt.minimize(lambda: var + 1.0, var_list=[var]) slot_var = opt.optimizers[0].get_slot(var, "momentum") slot_value = slot_var.numpy().item() # Save a checkpoint. checkpoint = tf.train.Checkpoint(optimizer=opt, var=var) save_path = checkpoint.save(tmp_path / "ckpt") # Run model again. opt.minimize(lambda: var + 1.0, var_list=[var]) assert slot_var.numpy().item() != slot_value # Load checkpoint and ensure loss scale is back to its original value. status = checkpoint.restore(save_path) status.assert_consumed() status.run_restore_ops() assert slot_var.numpy().item() == slot_value class TestBopOptimizer: def test_bop_accuracy(self): _test_optimizer( lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, lq.optimizers.Bop()), default_optimizer=optimizers.Adam(0.01), ), test_kernels_are_binary=True, ) # test optimizer on model with only binary trainable vars (low accuracy) _test_optimizer( lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, lq.optimizers.Bop()), default_optimizer=optimizers.Adam(0.01), ), test_kernels_are_binary=True, trainable_bn=False, target=0, ) @pytest.mark.usefixtures("distribute_scope") def test_mixed_precision(self): opt = lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, lq.optimizers.Bop()), default_optimizer=optimizers.Adam(0.01), ) try: opt = tf.keras.mixed_precision.LossScaleOptimizer(opt) except AttributeError: opt = tf.keras.mixed_precision.experimental.LossScaleOptimizer( opt, "dynamic" ) _test_optimizer(opt, test_kernels_are_binary=True) def test_bop_tf_1_14_schedules(self): _test_optimizer( lq.optimizers.CaseOptimizer( ( lq.optimizers.Bop.is_binary_variable, lq.optimizers.Bop( threshold=tf.keras.optimizers.schedules.InverseTimeDecay( 3.0, decay_steps=1.0, decay_rate=0.5 ), gamma=tf.keras.optimizers.schedules.InverseTimeDecay( 3.0, decay_steps=1.0, decay_rate=0.5 ), ), ), default_optimizer=optimizers.Adam(0.01), ), test_kernels_are_binary=True, ) def test_bop_serialization(self): _test_serialization( lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, lq.optimizers.Bop()), default_optimizer=optimizers.Adam(0.01), ), ) @pytest.mark.parametrize( "hyper", [5e-4, tf.keras.optimizers.schedules.PolynomialDecay(5e-4, 100)], ) def test_bop_serialization_schedule(self, hyper): bop = lq.optimizers.Bop( gamma=hyper, threshold=hyper, ) new_bop = lq.optimizers.Bop.from_config(bop.get_config()) assert isinstance(new_bop._get_hyper("gamma"), type(bop._get_hyper("gamma"))) assert isinstance( new_bop._get_hyper("threshold"), type(bop._get_hyper("threshold")) )
10,528
37.01083
88
py
larq
larq-main/larq/callbacks.py
from typing import Any, Callable, MutableMapping, Optional from tensorflow import keras class HyperparameterScheduler(keras.callbacks.Callback): """Generic hyperparameter scheduler. !!! example ```python bop = lq.optimizers.Bop(threshold=1e-6, gamma=1e-3) adam = tf.keras.optimizers.Adam(0.01) optimizer = lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, bop), default_optimizer=adam, ) callbacks = [ HyperparameterScheduler(lambda x: 0.001 * (0.1 ** (x // 30)), "gamma", bop) ] ``` # Arguments schedule: a function that takes an epoch index as input (integer, indexed from 0) and returns a new hyperparameter as output. hyperparameter: str. the name of the hyperparameter to be scheduled. optimizer: the optimizer that contains the hyperparameter that will be scheduled. Defaults to `self.model.optimizer` if `optimizer == None`. update_freq: str (optional), denotes on what update_freq to change the hyperparameter. Can be either "epoch" (default) or "step". verbose: int. 0: quiet, 1: update messages. log_name: str (optional), under which name to log this hyperparameter to Tensorboard. If `None`, defaults to `hyperparameter`. Use this if you have several schedules for the same hyperparameter on different optimizers. """ def __init__( self, schedule: Callable, hyperparameter: str, optimizer: Optional[keras.optimizers.Optimizer] = None, update_freq: str = "epoch", verbose: int = 0, log_name: Optional[str] = None, ): super().__init__() self.optimizer = optimizer self.schedule = schedule self.hyperparameter = hyperparameter self.log_name = log_name or hyperparameter self.verbose = verbose if update_freq not in ["epoch", "step"]: raise ValueError( "HyperparameterScheduler.update_freq can only be 'step' or 'epoch'." f" Received value '{update_freq}'" ) self.update_freq = update_freq def set_model(self, model: keras.models.Model) -> None: super().set_model(model) if self.optimizer is None: # It is not possible for a model to reach this state and not have # an optimizer, so we can safely access it here. self.optimizer = model.optimizer if not hasattr(self.optimizer, self.hyperparameter): raise ValueError( f'Optimizer must have a "{self.hyperparameter}" attribute.' ) def set_hyperparameter(self, t: int) -> Any: hp = getattr(self.optimizer, self.hyperparameter) try: # new API hyperparameter_val = keras.backend.get_value(hp) hyperparameter_val = self.schedule(t, hyperparameter_val) except TypeError: # Support for old API for backward compatibility hyperparameter_val = self.schedule(t) keras.backend.set_value(hp, hyperparameter_val) return hp def on_batch_begin( self, batch: int, logs: Optional[MutableMapping[str, Any]] = None ) -> None: if not self.update_freq == "step": return # We use optimizer.iterations (i.e. global step), since batch only # reflects the batch index in the current epoch. batch = keras.backend.get_value(self.optimizer.iterations) hp = self.set_hyperparameter(batch) if self.verbose > 0: print( f"Batch {batch}: {self.log_name} is now {keras.backend.get_value(hp)}." ) def on_epoch_begin( self, epoch: int, logs: Optional[MutableMapping[str, Any]] = None ) -> None: if not self.update_freq == "epoch": return hp = self.set_hyperparameter(epoch) if self.verbose > 0: print( f"Epoch {epoch}: {self.log_name} is now {keras.backend.get_value(hp)}." ) def on_epoch_end( self, epoch: int, logs: Optional[MutableMapping[str, Any]] = None ) -> None: logs = logs or {} hp = getattr(self.optimizer, self.hyperparameter) logs[self.log_name] = keras.backend.get_value(hp)
4,375
36.724138
89
py
larq
larq-main/larq/quantizers.py
"""A Quantizer defines the way of transforming a full precision input to a quantized output and the pseudo-gradient method used for the backwards pass. Quantizers can either be used through quantizer arguments that are supported for Larq layers, such as `input_quantizer` and `kernel_quantizer`; or they can be used similar to activations, i.e. either through an `Activation` layer, or through the `activation` argument supported by all forward layers: ```python import tensorflow as tf import larq as lq ... x = lq.layers.QuantDense(64, activation=None)(x) x = lq.layers.QuantDense(64, input_quantizer="ste_sign")(x) ``` is equivalent to: ```python x = lq.layers.QuantDense(64)(x) x = tf.keras.layers.Activation("ste_sign")(x) x = lq.layers.QuantDense(64)(x) ``` as well as: ```python x = lq.layers.QuantDense(64, activation="ste_sign")(x) x = lq.layers.QuantDense(64)(x) ``` We highly recommend using the first of these formulations: for the other two formulations, intermediate layers - like batch normalization or average pooling - and shortcut connections may result in non-binary input to the convolutions. Quantizers can either be referenced by string or called directly. The following usages are equivalent: ```python lq.layers.QuantDense(64, kernel_quantizer="ste_sign") ``` ```python lq.layers.QuantDense(64, kernel_quantizer=lq.quantizers.SteSign(clip_value=1.0)) ``` """ from typing import Callable, Union import tensorflow as tf from packaging import version from larq import context, math from larq import metrics as lq_metrics from larq import utils __all__ = [ "ApproxSign", "DoReFa", "DoReFaQuantizer", "MagnitudeAwareSign", "NoOp", "NoOpQuantizer", "Quantizer", "SteHeaviside", "SteSign", "SteTern", "SwishSign", ] def _clipped_gradient(x, dy, clip_value): """Calculate `clipped_gradent * dy`.""" if clip_value is None: return dy zeros = tf.zeros_like(dy) mask = tf.math.less_equal(tf.math.abs(x), clip_value) return tf.where(mask, dy, zeros) def ste_sign(x: tf.Tensor, clip_value: float = 1.0) -> tf.Tensor: @tf.custom_gradient def _call(x): def grad(dy): return _clipped_gradient(x, dy, clip_value) return math.sign(x), grad return _call(x) def _scaled_sign(x): # pragma: no cover return 1.3 * ste_sign(x) @tf.custom_gradient def approx_sign(x: tf.Tensor) -> tf.Tensor: def grad(dy): abs_x = tf.math.abs(x) zeros = tf.zeros_like(dy) mask = tf.math.less_equal(abs_x, 1.0) return tf.where(mask, (1 - abs_x) * 2 * dy, zeros) return math.sign(x), grad def swish_sign(x: tf.Tensor, beta: float = 5.0) -> tf.Tensor: @tf.custom_gradient def _call(x): def grad(dy): b_x = beta * x return dy * beta * (2 - b_x * tf.tanh(b_x * 0.5)) / (1 + tf.cosh(b_x)) return math.sign(x), grad return _call(x) def ste_tern( x: tf.Tensor, threshold_value: float = 0.05, ternary_weight_networks: bool = False, clip_value: float = 1.0, ) -> tf.Tensor: @tf.custom_gradient def _call(x): if ternary_weight_networks: threshold = 0.7 * tf.reduce_sum(tf.abs(x)) / tf.cast(tf.size(x), x.dtype) else: threshold = threshold_value def grad(dy): return _clipped_gradient(x, dy, clip_value) return tf.sign(tf.sign(x + threshold) + tf.sign(x - threshold)), grad return _call(x) def ste_heaviside(x: tf.Tensor, clip_value: float = 1.0) -> tf.Tensor: @tf.custom_gradient def _call(x): def grad(dy): return _clipped_gradient(x, dy, clip_value) return math.heaviside(x), grad return _call(x) class Quantizer(tf.keras.layers.Layer): """Common base class for defining quantizers. # Attributes precision: An integer defining the precision of the output. This value will be used by `lq.models.summary()` for improved logging. """ precision = None def compute_output_shape(self, input_shape): return input_shape class _BaseQuantizer(Quantizer): """Private base class for defining quantizers with Larq metrics.""" def __init__(self, *args, metrics=None, **kwargs): self._custom_metrics = metrics super().__init__(*args, **kwargs) def build(self, input_shape): if self._custom_metrics and "flip_ratio" in self._custom_metrics: self.flip_ratio = lq_metrics.FlipRatio(name=f"flip_ratio/{self.name}") self.flip_ratio.build(input_shape) super().build(input_shape) def call(self, inputs): if hasattr(self, "flip_ratio"): self.add_metric(self.flip_ratio(inputs)) return inputs @property def non_trainable_weights(self): return [] @utils.register_keras_custom_object class NoOp(_BaseQuantizer): r"""Instantiates a serializable no-op quantizer. \\[ q(x) = x \\] !!! warning This quantizer will not change the input variable. It is only intended to mark variables with a desired precision that will be recognized by optimizers like `Bop` and add training metrics to track variable changes. !!! example ```python layer = lq.layers.QuantDense( 16, kernel_quantizer=lq.quantizers.NoOp(precision=1), ) layer.build((32,)) assert layer.kernel.precision == 1 ``` # Arguments precision: Set the desired precision of the variable. This can be used to tag metrics: An array of metrics to add to the layer. If `None` the metrics set in `larq.context.metrics_scope` are used. Currently only the `flip_ratio` metric is available. """ precision = None def __init__(self, precision: int, **kwargs): self.precision = precision super().__init__(**kwargs) def get_config(self): return {**super().get_config(), "precision": self.precision} # `NoOp` used to be called `NoOpQuantizer`; this alias is for # backwards-compatibility. NoOpQuantizer = NoOp @utils.register_alias("ste_sign") @utils.register_keras_custom_object class SteSign(_BaseQuantizer): r"""Instantiates a serializable binary quantizer. \\[ q(x) = \begin{cases} -1 & x < 0 \\\ 1 & x \geq 0 \end{cases} \\] The gradient is estimated using the Straight-Through Estimator (essentially the binarization is replaced by a clipped identity on the backward pass). \\[\frac{\partial q(x)}{\partial x} = \begin{cases} 1 & \left|x\right| \leq \texttt{clip_value} \\\ 0 & \left|x\right| > \texttt{clip_value} \end{cases}\\] ```plot-activation quantizers.SteSign ``` # Arguments clip_value: Threshold for clipping gradients. If `None` gradients are not clipped. metrics: An array of metrics to add to the layer. If `None` the metrics set in `larq.context.metrics_scope` are used. Currently only the `flip_ratio` metric is available. # References - [Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1](https://arxiv.org/abs/1602.02830) """ precision = 1 def __init__(self, clip_value: float = 1.0, **kwargs): self.clip_value = clip_value super().__init__(**kwargs) def call(self, inputs): outputs = ste_sign(inputs, clip_value=self.clip_value) return super().call(outputs) def get_config(self): return {**super().get_config(), "clip_value": self.clip_value} @utils.register_alias("approx_sign") @utils.register_keras_custom_object class ApproxSign(_BaseQuantizer): r"""Instantiates a serializable binary quantizer. \\[ q(x) = \begin{cases} -1 & x < 0 \\\ 1 & x \geq 0 \end{cases} \\] The gradient is estimated using the ApproxSign method. \\[\frac{\partial q(x)}{\partial x} = \begin{cases} (2 - 2 \left|x\right|) & \left|x\right| \leq 1 \\\ 0 & \left|x\right| > 1 \end{cases} \\] ```plot-activation quantizers.ApproxSign ``` # Arguments metrics: An array of metrics to add to the layer. If `None` the metrics set in `larq.context.metrics_scope` are used. Currently only the `flip_ratio` metric is available. # References - [Bi-Real Net: Enhancing the Performance of 1-bit CNNs With Improved Representational Capability and Advanced Training Algorithm](https://arxiv.org/abs/1808.00278) """ precision = 1 def call(self, inputs): outputs = approx_sign(inputs) return super().call(outputs) @utils.register_alias("ste_heaviside") @utils.register_keras_custom_object class SteHeaviside(_BaseQuantizer): r""" Instantiates a binarization quantizer with output values 0 and 1. \\[ q(x) = \begin{cases} +1 & x > 0 \\\ 0 & x \leq 0 \end{cases} \\] The gradient is estimated using the Straight-Through Estimator (essentially the binarization is replaced by a clipped identity on the backward pass). \\[\frac{\partial q(x)}{\partial x} = \begin{cases} 1 & \left|x\right| \leq 1 \\\ 0 & \left|x\right| > 1 \end{cases}\\] ```plot-activation quantizers.SteHeaviside ``` # Arguments clip_value: Threshold for clipping gradients. If `None` gradients are not clipped. metrics: An array of metrics to add to the layer. If `None` the metrics set in `larq.context.metrics_scope` are used. Currently only the `flip_ratio` metric is available. # Returns AND Binarization function """ precision = 1 def __init__(self, clip_value: float = 1.0, **kwargs): self.clip_value = clip_value super().__init__(**kwargs) def call(self, inputs): outputs = ste_heaviside(inputs, clip_value=self.clip_value) return super().call(outputs) def get_config(self): return {**super().get_config(), "clip_value": self.clip_value} @utils.register_alias("swish_sign") @utils.register_keras_custom_object class SwishSign(_BaseQuantizer): r"""Sign binarization function. \\[ q(x) = \begin{cases} -1 & x < 0 \\\ 1 & x \geq 0 \end{cases} \\] The gradient is estimated using the SignSwish method. \\[ \frac{\partial q_{\beta}(x)}{\partial x} = \frac{\beta\left\\{2-\beta x \tanh \left(\frac{\beta x}{2}\right)\right\\}}{1+\cosh (\beta x)} \\] ```plot-activation quantizers.SwishSign ``` # Arguments beta: Larger values result in a closer approximation to the derivative of the sign. metrics: An array of metrics to add to the layer. If `None` the metrics set in `larq.context.metrics_scope` are used. Currently only the `flip_ratio` metric is available. # Returns SwishSign quantization function # References - [BNN+: Improved Binary Network Training](https://arxiv.org/abs/1812.11800) """ precision = 1 def __init__(self, beta: float = 5.0, **kwargs): self.beta = beta super().__init__(**kwargs) def call(self, inputs): outputs = swish_sign(inputs, beta=self.beta) return super().call(outputs) def get_config(self): return {**super().get_config(), "beta": self.beta} @utils.register_alias("magnitude_aware_sign") @utils.register_keras_custom_object class MagnitudeAwareSign(_BaseQuantizer): r"""Instantiates a serializable magnitude-aware sign quantizer for Bi-Real Net. A scaled sign function computed according to Section 3.3 in [Zechun Liu et al](https://arxiv.org/abs/1808.00278). ```plot-activation quantizers._scaled_sign ``` # Arguments clip_value: Threshold for clipping gradients. If `None` gradients are not clipped. metrics: An array of metrics to add to the layer. If `None` the metrics set in `larq.context.metrics_scope` are used. Currently only the `flip_ratio` metric is available. # References - [Bi-Real Net: Enhancing the Performance of 1-bit CNNs With Improved Representational Capability and Advanced Training Algorithm](https://arxiv.org/abs/1808.00278) """ precision = 1 def __init__(self, clip_value: float = 1.0, **kwargs): self.clip_value = clip_value super().__init__(**kwargs) def call(self, inputs): scale_factor = tf.stop_gradient( tf.reduce_mean(tf.abs(inputs), axis=list(range(len(inputs.shape) - 1))) ) outputs = scale_factor * ste_sign(inputs, clip_value=self.clip_value) return super().call(outputs) def get_config(self): return {**super().get_config(), "clip_value": self.clip_value} @utils.register_alias("ste_tern") @utils.register_keras_custom_object class SteTern(_BaseQuantizer): r"""Instantiates a serializable ternarization quantizer. \\[ q(x) = \begin{cases} +1 & x > \Delta \\\ 0 & |x| < \Delta \\\ -1 & x < - \Delta \end{cases} \\] where \\(\Delta\\) is defined as the threshold and can be passed as an argument, or can be calculated as per the Ternary Weight Networks original paper, such that \\[ \Delta = \frac{0.7}{n} \sum_{i=1}^{n} |W_i| \\] where we assume that \\(W_i\\) is generated from a normal distribution. The gradient is estimated using the Straight-Through Estimator (essentially the Ternarization is replaced by a clipped identity on the backward pass). \\[\frac{\partial q(x)}{\partial x} = \begin{cases} 1 & \left|x\right| \leq \texttt{clip_value} \\\ 0 & \left|x\right| > \texttt{clip_value} \end{cases}\\] ```plot-activation quantizers.SteTern ``` # Arguments threshold_value: The value for the threshold, \\(\Delta\\). ternary_weight_networks: Boolean of whether to use the Ternary Weight Networks threshold calculation. clip_value: Threshold for clipping gradients. If `None` gradients are not clipped. metrics: An array of metrics to add to the layer. If `None` the metrics set in `larq.context.metrics_scope` are used. Currently only the `flip_ratio` metric is available. # References - [Ternary Weight Networks](https://arxiv.org/abs/1605.04711) """ precision = 2 def __init__( self, threshold_value: float = 0.05, ternary_weight_networks: bool = False, clip_value: float = 1.0, **kwargs, ): self.threshold_value = threshold_value self.ternary_weight_networks = ternary_weight_networks self.clip_value = clip_value super().__init__(**kwargs) def call(self, inputs): outputs = ste_tern( inputs, threshold_value=self.threshold_value, ternary_weight_networks=self.ternary_weight_networks, clip_value=self.clip_value, ) return super().call(outputs) def get_config(self): return { **super().get_config(), "threshold_value": self.threshold_value, "ternary_weight_networks": self.ternary_weight_networks, "clip_value": self.clip_value, } @utils.register_alias("dorefa_quantizer") @utils.register_keras_custom_object class DoReFa(_BaseQuantizer): r"""Instantiates a serializable k_bit quantizer as in the DoReFa paper. \\[ q(x) = \begin{cases} 0 & x < \frac{1}{2n} \\\ \frac{i}{n} & \frac{2i-1}{2n} < x < \frac{2i+1}{2n} \text{ for } i \in \\{1,n-1\\}\\\ 1 & \frac{2n-1}{2n} < x \end{cases} \\] where \\(n = 2^{\text{k_bit}} - 1\\). The number of bits, k_bit, needs to be passed as an argument. The gradient is estimated using the Straight-Through Estimator (essentially the binarization is replaced by a clipped identity on the backward pass). \\[\frac{\partial q(x)}{\partial x} = \begin{cases} 1 & 0 \leq x \leq 1 \\\ 0 & \text{else} \end{cases}\\] The behavior for quantizing weights should be different in comparison to the quantization of activations: instead of limiting input operands (or in this case: weights) using a hard limiter, a tangens hyperbolicus is applied to achieve a softer limiting with a gradient, which is continuously differentiable itself. \\[ w_{lim}(w) = \tanh(w) \\] Furthermore, the weights of each layer are normed, such that the weight with the largest magnitude gets the largest or smallest (depending on its sign) quantizable value. That way, the full quantizable numeric range is utilized. \\[ w_{norm}(w) = \frac{w}{\max(|w|)} \\] The formulas can be found in the paper in section 2.3. Please note, that the paper refers to weights being quantized on a numeric range of [-1, 1], while activations are quantized on the numeric range [0, 1]. This implementation uses the same ranges as specified in the paper. The activation quantizer defines the function quantizek() from the paper with the correct numeric range of [0, 1]. The weight quantization mode adds pre- and post-processing for numeric range adaptions, soft limiting and norming. The full quantization function including the adaption of numeric ranges is \\[ q(w) = 2 \, quantize_{k}(\frac{w_{norm}\left(w_{lim}\left(w\right)\right)}{2} + \frac{1}{2}) - 1 \\] !!! warning The weight mode works for weights on the range [-1, 1], which matches the default setting of `constraints.weight_clip`. Do not use this quantizer with a different constraint `clip_value` than the default one. __`mode == "activations"`__ ```plot-activation quantizers.DoReFa ``` __`mode == "weights"`__ ```plot-activation quantizers.DoReFa(mode='weights') ``` # Arguments k_bit: number of bits for the quantization. mode: `"activations"` for clipping inputs on [0, 1] range or `"weights"` for soft-clipping and norming weights on [-1, 1] range before applying quantization. metrics: An array of metrics to add to the layer. If `None` the metrics set in `larq.context.metrics_scope` are used. Currently only the `flip_ratio` metric is available. # Returns Quantization function # Raises ValueError for bad value of `mode`. # References - [DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients](https://arxiv.org/abs/1606.06160) """ precision = None def __init__(self, k_bit: int = 2, mode: str = "activations", **kwargs): self.precision = k_bit if mode not in ("activations", "weights"): raise ValueError( f"Invalid DoReFa quantizer mode {mode}. " "Valid values are 'activations' and 'weights'." ) self.mode = mode super().__init__(**kwargs) def weight_preprocess(self, inputs): # Limit inputs to [-1, 1] range limited = tf.math.tanh(inputs) # Divider for max-value norm. dividend = tf.math.reduce_max(tf.math.abs(limited)) # Need to stop the gradient here. Otherwise, for the maximum element, # which gives the dividend, normed is limited/limited (for this one # maximum digit). The derivative of y = x/x, dy/dx is just zero, when # one does the simplification y = x/x = 1. But TF does NOT do this # simplification when computing the gradient for the # normed = limited/dividend operation. As a result, this gradient # becomes complicated, because during the computation, "dividend" is # not just a constant, but depends on "limited" instead. Here, # tf.stop_gradient is used to mark "dividend" as a constant explicitly. dividend = tf.stop_gradient(dividend) # Norm and then scale from value range [-1,1] to [0,1] (the range # expected by the core quantization operation). # If the dividend used for the norm operation is 0, all elements of # the weight tensor are 0 and divide_no_nan returns 0 for all weights. # So if all elements of the weight tensor are zero, nothing is normed. return tf.math.divide_no_nan(limited, 2.0 * dividend) + 0.5 def call(self, inputs): # Depending on quantizer mode (activation or weight) just clip inputs # on [0, 1] range or use weight preprocessing method. if self.mode == "activations": inputs = tf.clip_by_value(inputs, 0.0, 1.0) elif self.mode == "weights": inputs = self.weight_preprocess(inputs) else: raise ValueError( f"Invalid DoReFa quantizer mode {self.mode}. " "Valid values are 'activations' and 'weights'." ) @tf.custom_gradient def _k_bit_with_identity_grad(x): n = 2**self.precision - 1 return tf.round(x * n) / n, lambda dy: dy outputs = _k_bit_with_identity_grad(inputs) # Scale weights from [0, 1] quantization range back to [-1,1] range if self.mode == "weights": outputs = 2.0 * outputs - 1.0 return super().call(outputs) def get_config(self): return {**super().get_config(), "k_bit": self.precision, "mode": self.mode} # `DoReFa` used to be called `DoReFaQuantizer`; this alias is for # backwards-compatibility. DoReFaQuantizer = DoReFa QuantizerType = Union[Quantizer, Callable[[tf.Tensor], tf.Tensor]] def serialize(quantizer: tf.keras.layers.Layer, use_legacy_format=False): if use_legacy_format and version.parse(tf.__version__) >= version.parse("2.13"): return tf.keras.utils.legacy.serialize_keras_object(quantizer) return tf.keras.utils.serialize_keras_object(quantizer) def deserialize(name, custom_objects=None, use_legacy_format=False): if use_legacy_format and version.parse(tf.__version__) >= version.parse("2.13"): return tf.keras.utils.legacy.deserialize_keras_object( name, module_objects=globals(), custom_objects=custom_objects, printable_module_name="quantization function", ) return tf.keras.utils.deserialize_keras_object( name, module_objects=globals(), custom_objects=custom_objects, printable_module_name="quantization function", ) def get(identifier): if identifier is None: return None if isinstance(identifier, dict): use_legacy_format = "module" not in identifier return deserialize(identifier, use_legacy_format=use_legacy_format) if isinstance(identifier, str): config = {"class_name": str(identifier), "config": {}} return get(config) if callable(identifier): return identifier raise ValueError( f"Could not interpret quantization function identifier: {identifier}" ) def get_kernel_quantizer(identifier): """Returns a quantizer from identifier and adds default kernel quantizer metrics. # Arguments identifier: Function or string # Returns `Quantizer` or `None` """ quantizer = get(identifier) if isinstance(quantizer, _BaseQuantizer) and not quantizer._custom_metrics: quantizer._custom_metrics = list(context.get_training_metrics()) return quantizer
23,775
30.449735
141
py
larq
larq-main/larq/context.py
"""Context managers that configure global behaviour of Larq.""" import contextlib import threading __all__ = [ "metrics_scope", "quantized_scope", "get_training_metrics", "should_quantize", ] _quantized_scope = threading.local() _quantized_scope.should_quantize = False @contextlib.contextmanager def quantized_scope(quantize): """A context manager to define the behaviour of `QuantizedVariable`. !!! example ```python model.save("full_precision_model.h5") # save full precision latent weights fp_weights = model.get_weights() # get latent weights with larq.context.quantized_scope(True): model.save("binary_model.h5") # save binarized weights weights = model.get_weights() # get binarized weights ``` # Arguments quantize: If `should_quantize` is `True`, `QuantizedVariable` will return their quantized value in the forward pass. If `False`, `QuantizedVariable` will act as a latent variable. """ backup = should_quantize() _quantized_scope.should_quantize = quantize yield quantize _quantized_scope.should_quantize = backup def should_quantize(): """Returns the current quantized scope.""" return getattr(_quantized_scope, "should_quantize", False) _global_training_metrics = set() _available_metrics = {"flip_ratio"} @contextlib.contextmanager def metrics_scope(metrics=[]): """A context manager to set the training metrics to be used in quantizers. !!! example ```python with larq.context.metrics_scope(["flip_ratio"]): model = tf.keras.models.Sequential( [larq.layers.QuantDense(3, kernel_quantizer="ste_sign", input_shape=(32,))] ) model.compile(loss="mse", optimizer="sgd") ``` # Arguments metrics: Iterable of metrics to add to quantizers defined inside this context. Currently only the `flip_ratio` metric is available. """ for metric in metrics: if metric not in _available_metrics: raise ValueError( f"Unknown training metric '{metric}'. Available metrics: {_available_metrics}." ) backup = _global_training_metrics.copy() _global_training_metrics.update(metrics) yield _global_training_metrics _global_training_metrics.clear() _global_training_metrics.update(backup) def get_training_metrics(): """Retrieves a live reference to the training metrics in the current scope. Updating and clearing training metrics using `larq.context.metrics_scope` is preferred, but `get_training_metrics` can be used to directly access them. !!! example ```python get_training_metrics().clear() get_training_metrics().add("flip_ratio") ``` # Returns A set of training metrics in the current scope. """ return _global_training_metrics
2,953
29.453608
95
py
larq
larq-main/larq/conftest.py
import pytest import tensorflow as tf from packaging import version from tensorflow.python.distribute import strategy_combinations from tensorflow.python.eager import context from larq import context as lq_context if version.parse(tf.__version__) >= version.parse("1.15"): strategy_combinations.set_virtual_cpus_to_at_least(3) distributed_devices = ["/cpu:1", "/cpu:2"] else: distributed_devices = ["/cpu:0"] @pytest.fixture def eager_mode(): """pytest fixture for running test in eager mode""" with context.eager_mode(): yield @pytest.fixture def graph_mode(): """pytest fixture for running test in graph mode""" with context.graph_mode(): with tf.compat.v1.Session().as_default(): yield tf.keras.backend.clear_session() @pytest.fixture(params=["eager", "graph"]) def eager_and_graph_mode(request): """pytest fixture for running test in eager and graph mode""" if request.param == "graph": with context.graph_mode(): with tf.compat.v1.Session().as_default(): yield request.param tf.keras.backend.clear_session() else: with context.eager_mode(): yield request.param @pytest.fixture(params=["graph", "tf_eager", "tf_keras_eager"]) def keras_should_run_eagerly(request): """Fixture to run in graph and two eager modes. The modes are: - Graph mode - TensorFlow eager and Keras eager - TensorFlow eager and Keras not eager The `tf.context` sets graph/eager mode for TensorFlow. The yield is True if Keras should run eagerly. """ if request.param == "graph": if version.parse(tf.__version__) >= version.parse("2"): pytest.skip("Skipping graph mode for TensorFlow 2+.") with context.graph_mode(): yield else: with context.eager_mode(): yield request.param == "tf_keras_eager" @pytest.fixture(params=[False, True]) def distribute_scope(request): if request.param is True: with tf.distribute.MirroredStrategy(distributed_devices).scope(): yield request.param else: yield request.param @pytest.fixture(params=[True, False]) def quantized(request): """pytest fixture for running test quantized and non-quantized""" with lq_context.quantized_scope(request.param): yield request.param @pytest.fixture(params=["channels_last", "channels_first"]) def data_format(request): return request.param
2,508
27.511364
85
py
larq
larq-main/larq/testing_utils.py
import numpy as np import tensorflow as tf import larq as lq def _eval_tensor(tensor): if tensor is None: return None elif callable(tensor): return _eval_helper(tensor()) else: return tensor.numpy() def _eval_helper(tensors): if tensors is None: return None return tf.nest.map_structure(_eval_tensor, tensors) def evaluate(tensors): if tf.executing_eagerly(): return _eval_helper(tensors) else: sess = tf.compat.v1.get_default_session() return sess.run(tensors) def generate_real_values_with_zeros(low=-2, high=2, shape=(4, 10)): real_values = np.random.uniform(low, high, shape) real_values = np.insert(real_values, 1, 0, axis=1) return real_values def get_small_bnn_model(input_dim, num_hidden, output_dim, trainable_bn=True): model = tf.keras.models.Sequential() model.add( lq.layers.QuantDense( units=num_hidden, kernel_quantizer="ste_sign", kernel_constraint="weight_clip", activation="relu", input_shape=(input_dim,), use_bias=False, ) ) model.add(tf.keras.layers.BatchNormalization(trainable=trainable_bn)) model.add( lq.layers.QuantDense( units=output_dim, kernel_quantizer="ste_sign", kernel_constraint="weight_clip", input_quantizer="ste_sign", activation="softmax", use_bias=False, ) ) return model def random_input(shape): for i, dim in enumerate(shape): if dim is None: shape[i] = np.random.randint(1, 4) data = 10 * np.random.random(shape) - 0.5 return data.astype("float32") # This is a fork of https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/keras/testing_utils.py#L72 # as recommended in https://github.com/tensorflow/tensorflow/issues/28601#issuecomment-492810252 def layer_test( layer_cls, kwargs=None, input_shape=None, input_dtype=None, input_data=None, expected_output=None, expected_output_dtype=None, should_run_eagerly=False, ): """Test routine for a layer with a single input and single output. Arguments: layer_cls: Layer class object. kwargs: Optional dictionary of keyword arguments for instantiating the layer. input_shape: Input shape tuple. input_dtype: Data type of the input data. input_data: Numpy array of input data. expected_output: Shape tuple for the expected shape of the output. expected_output_dtype: Data type expected for the output. Returns: The output data (Numpy array) returned by the layer, for additional checks to be done by the calling code. Raises: ValueError: if `input_shape is None`. """ if input_data is None: if input_shape is None: raise ValueError("input_shape is None") if not input_dtype: input_dtype = "float32" input_data_shape = list(input_shape) for i, e in enumerate(input_data_shape): if e is None: input_data_shape[i] = np.random.randint(1, 4) input_data = 10 * np.random.random(input_data_shape) if input_dtype[:5] == "float": input_data -= 0.5 input_data = input_data.astype(input_dtype) elif input_shape is None: input_shape = input_data.shape if input_dtype is None: input_dtype = input_data.dtype if expected_output_dtype is None: expected_output_dtype = input_dtype # instantiation kwargs = kwargs or {} layer = layer_cls(**kwargs) # test get_weights , set_weights at layer level weights = layer.get_weights() layer.set_weights(weights) # test in functional API x = tf.keras.layers.Input(shape=input_shape[1:], dtype=input_dtype) y = layer(x) if tf.keras.backend.dtype(y) != expected_output_dtype: raise AssertionError( "When testing layer %s, for input %s, found output " "dtype=%s but expected to find %s.\nFull kwargs: %s" % ( layer_cls.__name__, x, tf.keras.backend.dtype(y), expected_output_dtype, kwargs, ) ) # check shape inference model = tf.keras.models.Model(x, y) expected_output_shape = tuple( layer.compute_output_shape(tf.TensorShape(input_shape)).as_list() ) actual_output = model.predict(input_data) actual_output_shape = actual_output.shape for expected_dim, actual_dim in zip(expected_output_shape, actual_output_shape): if expected_dim is not None: if expected_dim != actual_dim: raise AssertionError( "When testing layer %s, for input %s, found output_shape=" "%s but expected to find %s.\nFull kwargs: %s" % ( layer_cls.__name__, x, actual_output_shape, expected_output_shape, kwargs, ) ) if expected_output is not None: np.testing.assert_allclose(actual_output, expected_output, rtol=1e-3) # test serialization, weight setting at model level model_config = model.get_config() recovered_model = tf.keras.models.Model.from_config(model_config) if model.weights: weights = model.get_weights() recovered_model.set_weights(weights) output = recovered_model.predict(input_data) np.testing.assert_allclose(output, actual_output, rtol=2e-3) # Recreate layer to prevent layer metrics from being configured multiple times. layer = layer_cls(**kwargs) # test training mode (e.g. useful for dropout tests) # Rebuild the model to avoid the graph being reused between predict() and # train(). This was causing some error for layer with Defun as it body. # See b/120160788 for more details. This should be mitigated after 2.0. model = tf.keras.models.Model(x, layer(x)) model.compile( "rmsprop", "mse", weighted_metrics=["acc"], run_eagerly=should_run_eagerly, ) model.train_on_batch(input_data, actual_output) # test as first layer in Sequential API layer_config = layer.get_config() layer_config["batch_input_shape"] = input_shape layer = layer.__class__.from_config(layer_config) model = tf.keras.models.Sequential() model.add(layer) actual_output = model.predict(input_data) actual_output_shape = actual_output.shape for expected_dim, actual_dim in zip(expected_output_shape, actual_output_shape): if expected_dim is not None: if expected_dim != actual_dim: raise AssertionError( "When testing layer %s **after deserialization**, " "for input %s, found output_shape=" "%s but expected to find inferred shape %s.\nFull kwargs: %s" % ( layer_cls.__name__, x, actual_output_shape, expected_output_shape, kwargs, ) ) if expected_output is not None: np.testing.assert_allclose(actual_output, expected_output, rtol=1e-3) # test serialization, weight setting at model level model_config = model.get_config() recovered_model = tf.keras.models.Sequential.from_config(model_config) if model.weights: weights = model.get_weights() recovered_model.set_weights(weights) output = recovered_model.predict(input_data) np.testing.assert_allclose(output, actual_output, rtol=2e-3) # for further checks in the caller function return actual_output
7,954
34.044053
117
py
larq
larq-main/larq/quantized_variable.py
"""Contains QuantizedVariable, a variable that can be quantized in the forward pass.""" from typing import Optional import tensorflow as tf from packaging import version from tensorflow.python.distribute.values import DistributedVariable from tensorflow.python.framework import ops from tensorflow.python.ops import resource_variable_ops from larq import context from larq.quantizers import QuantizerType # pytype: disable=import-error try: from tensorflow.python.distribute.ps_values import AggregatingVariable from tensorflow.python.types.core import Tensor as TensorType except ModuleNotFoundError: TensorType = object from tensorflow.python.distribute.values import AggregatingVariable # pytype: enable=import-error UNSPECIFIED = object() _SUPPORTS_TRACE_TYPE = version.parse(tf.__version__) >= version.parse("2.8") if _SUPPORTS_TRACE_TYPE: try: from tensorflow.types.experimental import TraceType except ImportError: from tensorflow.python.types.trace import TraceType class QuantizedVariableSpec(TraceType): """TraceType for QuantizedVariableSpec for tracing with tf.function. This class implements the Type for QuantizedVariable used in tracing. """ def __init__(self, value): self.latent_variable = value def is_subtype_of(self, other) -> bool: """If the other spec is the same as `self`, return True.""" return self == other def most_specific_common_supertype(self, others): """`self` is the common supertype if all input types match it.""" return self if all(self == other for other in others) else None def placeholder_value(self, placeholder_context=None): """Use the QuantizedVariable value itself as a placeholder.""" return self.latent_variable def _cast(self, value, _): return value def _to_tensors(self, value): return [] def __hash__(self) -> int: return hash(id(self.latent_variable)) def __eq__(self, other) -> bool: return self is other class QuantizedVariable(tf.Variable, TensorType): """A Variable that can be quantized in the forward pass in applicable contexts.""" def __init__( self, variable: tf.Variable, quantizer: Optional[QuantizerType] = None, precision: Optional[int] = None, op: Optional[tf.Operation] = UNSPECIFIED, ): """Creates an QuantizedVariable instance. # Arguments variable: A floating-point resource variable to wrap. quantizer: An optional quantizer to transform the floating-point variable to a fake quantized variable. precision: An optional integer defining the precision of the quantized variable. If `None`, `quantizer.precision` is used. op: An optional operation of this variable. """ if not resource_variable_ops.is_resource_variable(variable): raise ValueError( "`variable` must be of type `tf.ResourceVariable`, " f"but got `{type(variable)}`." ) if not (quantizer is None or callable(quantizer)): raise ValueError( "`quantizer` must be `callable` or `None`, " f"but got `{type(quantizer)}`." ) if not (precision is None or type(precision) == int): raise ValueError( "`precision` must be of type `int` or `None`, " f"but got `{type(precision)}`." ) self.latent_variable = variable self.quantizer = quantizer self.precision = precision or getattr(quantizer, "precision", None) self._op = op @classmethod def from_variable( cls, variable: tf.Variable, quantizer: Optional[QuantizerType] = None, precision: Optional[int] = None, op: Optional[tf.Operation] = UNSPECIFIED, ): """Creates a QuantizedVariable that wraps another variable. This typically just returns `QuantizedVariable(variable)`. But, if the variable is a DistributedVariable or one of its subclasses, we instead dynamically create a class that subclasses from both QuantizedVariable and variable.__class__. This is so the returned variable will still pass `isinstance(variable, variable.__class__)`, which is required for DistributedVariables and its subclasses to work properly. # Arguments variable: A floating-point resource variable to wrap. quantizer: An optional quantizer to transform the floating-point variable to a fake quantized variable. precision: An optional integer defining the precision of the quantized variable. If `None`, `quantizer.precision` is used. op: An optional operation of this variable. # Returns A QuantizedVariable that wraps the variable. """ if not isinstance(variable, (DistributedVariable, AggregatingVariable)): return cls(variable, quantizer, precision, op=op) class QuantizedDistributedVariable(cls, variable.__class__): """A QuantizedVariable that also subclasses from `variable.__class__`. `variable.__class__` is either a `DistributedVariable` or an `AggregatingVariable`. """ def get(self, *args, **kwargs): # For some reason this is needed to make unit `x + x` pass on TF 1.14 return self._quantize(self.latent_variable.get(*args, **kwargs)) return QuantizedDistributedVariable(variable, quantizer, precision, op=op) def _quantize(self, value): if self.quantizer and context.should_quantize(): return self.quantizer(value) return value def value(self): return self._quantize(self.latent_variable.value()) def read_value(self): return self._quantize(self.latent_variable.read_value()) def numpy(self): return self._quantize(self.latent_variable).numpy() def sparse_read(self, *args, **kwargs): return self._quantize(self.latent_variable.sparse_read(*args, **kwargs)) def gather_nd(self, *args, **kwargs): return self._quantize(self.latent_variable.gather_nd(*args, **kwargs)) def __getattr__(self, name): return getattr(self.latent_variable, name) def _dense_var_to_tensor(self, *args, **kwargs): return self._quantize( self.latent_variable._dense_var_to_tensor(*args, **kwargs) ) def eval(self, session=None): return self._quantize(self.latent_variable).eval(session=session) def initialized_value(self): return self._quantize(self.latent_variable.initialized_value()) @property def initial_value(self): return self._quantize(self.latent_variable.initial_value) def __tf_tensor__( self, dtype: Optional[tf.dtypes.DType] = None, name: Optional[str] = None ) -> tf.Tensor: return self._dense_var_to_tensor(dtype=dtype, name=name) def _should_act_as_resource_variable(self): """Pass resource_variable_ops.is_resource_variable check.""" pass @staticmethod def _get_name(obj) -> str: try: return obj.__name__ except AttributeError: return obj.__class__.__name__ def __repr__(self) -> str: repr_ = ( f"<{self.__class__.__name__} '{self.name}' " f"shape={self.shape} dtype={self.dtype.name}" ) if self.quantizer is not None: repr_ += f" quantizer={self._get_name(self.quantizer)}" if self.precision is not None: repr_ += f" precision={self.precision}" if tf.executing_eagerly() and not self._in_graph_mode: return f"{repr_} numpy={ops.numpy_text(self.read_value(), is_repr=True)}>" return f"{repr_}>" # Method delegations: We delegate the following methods to self.latent_variable. # Each of these methods simply calls the same method on self.latent_variable. The # base Variable raises NotImplementedError for most of these, so we must # override them. # # We do not define the following methods from Variable for the following # reasons: # * 'ref': Instead we inherit the definition from Variable. # If we defined and delegated to Variable, the ref of an QuantizedVariable # would be the same as the ref of the underlying variable, which would be # strange as they are different Python objects. def set_shape(self, *args, **kwargs): return self.latent_variable.set_shape(*args, **kwargs) @property def trainable(self): return self.latent_variable.trainable @property def synchronization(self): return self.latent_variable.synchronization @property def aggregation(self): return self.latent_variable.aggregation @property def constraint(self): return self.latent_variable.constraint def _apply_assign_update( self, update_fn, value, use_locking=None, name=None, read_value=True ): if ops.executing_eagerly_outside_functions(): assign_op = update_fn(value, use_locking, name, False) if read_value: return QuantizedVariable.from_variable( self.latent_variable, self.quantizer, self.precision, op=assign_op ) return assign_op # Fallback to wrapping the returned variable in graph mode if possible assign_var = update_fn(value, use_locking, name, read_value) if read_value and resource_variable_ops.is_resource_variable(assign_var): return QuantizedVariable.from_variable( assign_var, self.quantizer, self.precision ) return assign_var def _apply_update(self, update_fn, *args, **kwargs): update_var = update_fn(*args, **kwargs) if ops.executing_eagerly_outside_functions(): return self # Fallback to wrapping the returned variable in graph mode if possible if resource_variable_ops.is_resource_variable(update_var): return QuantizedVariable.from_variable( update_var, self.quantizer, self.precision ) return update_var def assign(self, value, use_locking=None, name=None, read_value=True): return self._apply_assign_update( self.latent_variable.assign, value, use_locking, name, read_value ) def assign_add(self, delta, use_locking=None, name=None, read_value=True): return self._apply_assign_update( self.latent_variable.assign_add, delta, use_locking, name, read_value ) def assign_sub(self, delta, use_locking=None, name=None, read_value=True): return self._apply_assign_update( self.latent_variable.assign_sub, delta, use_locking, name, read_value ) def scatter_sub(self, *args, **kwargs): return self._apply_update(self.latent_variable.scatter_sub, *args, **kwargs) def scatter_add(self, *args, **kwargs): return self._apply_update(self.latent_variable.scatter_add, *args, **kwargs) def scatter_max(self, *args, **kwargs): return self._apply_update(self.latent_variable.scatter_max, *args, **kwargs) def scatter_min(self, *args, **kwargs): return self._apply_update(self.latent_variable.scatter_min, *args, **kwargs) def scatter_mul(self, *args, **kwargs): return self._apply_update(self.latent_variable.scatter_mul, *args, **kwargs) def scatter_div(self, *args, **kwargs): return self._apply_update(self.latent_variable.scatter_div, *args, **kwargs) def scatter_update(self, *args, **kwargs): return self._apply_update(self.latent_variable.scatter_update, *args, **kwargs) def batch_scatter_update(self, *args, **kwargs): return self._apply_update( self.latent_variable.batch_scatter_update, *args, **kwargs ) def scatter_nd_sub(self, *args, **kwargs): return self._apply_update(self.latent_variable.scatter_nd_sub, *args, **kwargs) def scatter_nd_add(self, *args, **kwargs): return self._apply_update(self.latent_variable.scatter_nd_add, *args, **kwargs) def scatter_nd_update(self, *args, **kwargs): return self._apply_update( self.latent_variable.scatter_nd_update, *args, **kwargs ) def count_up_to(self, *args, **kwargs): return self.latent_variable.count_up_to(*args, **kwargs) def load(self, *args, **kwargs): return self.latent_variable.load(*args, **kwargs) @property def dtype(self): return self.latent_variable.dtype @property def name(self): return self.latent_variable.name @property def _shared_name(self): return self.latent_variable._shared_name @property def initializer(self): return self.latent_variable.initializer @property def device(self): return self.latent_variable.device @property def op(self): if self._op is not UNSPECIFIED: return self._op return self.latent_variable.op @property def graph(self): return self.latent_variable.graph @property def shape(self): return self.latent_variable.shape def get_shape(self): return self.latent_variable.get_shape() def __tf_tracing_type__(self, context): if _SUPPORTS_TRACE_TYPE: return QuantizedVariableSpec(self) return NotImplemented def _gather_saveables_for_checkpoint(self): # By delegating this method to the wrapped variable, checkpoints with # QuantizedVariables are identical to checkpoints with normal variables. # Therefore models checkpointed with QuantizedVariables can be restored on # models with normal variables, and vice versa. return self.latent_variable._gather_saveables_for_checkpoint() def _map_resources(self, *args): # By delegating this method to the wrapped variable, SavedModel with # QuantizedVariables are identical to SavedModel with normal variables. obj_map, resource_map = self.latent_variable._map_resources(*args) obj_map[self] = obj_map[self.latent_variable] return obj_map, resource_map def _export_to_saved_model_graph(self, object_map, tensor_map, options, **kwargs): # By delegating this method to the wrapped variable, SavedModel with # QuantizedVariables are identical to SavedModel with normal variables. resource_list = self.latent_variable._export_to_saved_model_graph( object_map, tensor_map, options, **kwargs ) object_map[self] = object_map[self.latent_variable] return resource_list # TODO: Maybe encode the fact the variable is an QuantizedVariable in to_proto(). def to_proto(self, *args, **kwargs): return self.latent_variable.to_proto(*args, **kwargs) def from_proto(self, *args, **kwargs): return self.latent_variable.from_proto(*args, **kwargs) # Delegate the private attributes _handle_name and _initializer_op to # self.latent_variable. SavedModel sets these attributes when loading a model. For # example, it sets _handle_name here: # https://github.com/tensorflow/tensorflow/blob/db26bd574fa95b5bdd53c08463dd19407cc0297e/tensorflow/python/keras/saving/saved_model/load.py#L211 # We need to expose these attributes on AutoCastVariable as well for # SavedModel to work properly. # TODO: Find a better way to support SavedModel. Exposing private attributes is # hacky and difficult to maintain. # For more info see https://github.com/tensorflow/tensorflow/commit/1fcda57f37c2ac854cabf1c3462eb14e39d36c60 @property def _handle_name(self): return self.latent_variable._handle_name @_handle_name.setter def _handle_name(self, handle_name): self.latent_variable._handle_name = handle_name @property def _initializer_op(self): return self.latent_variable._initializer_op @_initializer_op.setter def _initializer_op(self, initializer_op): self.latent_variable._initializer_op = initializer_op def _as_graph_element(self): if self.quantizer and context.should_quantize(): return self.quantizer(self.latent_variable) graph_element = self.latent_variable._as_graph_element() if graph_element is None: return self._op return graph_element QuantizedVariable._OverloadAllOperators() tf.register_tensor_conversion_function( QuantizedVariable, QuantizedVariable._dense_var_to_tensor ) try: ops.register_dense_tensor_like_type(QuantizedVariable) except AttributeError: pass
17,061
36.915556
148
py
larq
larq-main/larq/optimizers.py
"""Neural networks with extremely low-precision weights and activations, such as Binarized Neural Networks (BNNs), usually contain a mix of low-precision weights (e.g. 1-bit) and higher-precision weights (e.g. 8-bit, 16-bit, or 32-bit). Examples of this include the first and last layers of image classificiation models, which have higher-precision weights in most BNN architectures from the literature. Training a BNN, then, consists of optimizing both low-precision and higher-precision weights. In `larq`, we provide a mechanism to target different bit-precision variables with different optimizers using the `CaseOptimizer` class. Modeled after the [`tf.case`](https://www.tensorflow.org/api_docs/python/tf/case) signature, `CaseOptimizer` accepts pairs of predicates and optimizers. A predicate, given a variable, decides whether its optimizer should train that variable. A `CaseOptimizer` behaves much like any other [Keras optimizer](https://www.tensorflow.org/api_docs/python/tf/keras/optimizers), and once you instantiate it you can pass it to your `model.compile()` as usual. To instantiate a `CaseOptimzer`, pass one or a list of `(predicate, optimizer)` tuples, along with a `default` optimizer which trains any variables not claimed by another optimizer. A variable may not be claimed by more than one optimizer's predicate. !!! example ```python no_op_quantizer = lq.quantizers.NoOp(precision=1) layer = lq.layers.QuantDense(16, kernel_quantizer=no_op_quantizer) case_optimizer = lq.optimizers.CaseOptimizer( ( lq.optimizers.Bop.is_binary_variable, # predicate lq.optimizers.Bop(threshold=1e-6, gamma=1e-3), # optimizer ), default_optimizer=tf.keras.optimizers.Adam(0.01), ) ``` """ import warnings from copy import deepcopy from typing import Callable, Optional, Tuple import tensorflow as tf from packaging import version import larq as lq from larq import utils __all__ = ["Bop", "CaseOptimizer"] if version.parse(tf.__version__) >= version.parse("2.11"): from tensorflow.keras.optimizers.legacy import Optimizer # type: ignore else: from tensorflow.keras.optimizers import Optimizer # type: ignore # From https://github.com/keras-team/keras/blob/a8606fd45b760cce3e65727e9d62cae796c45930/keras/optimizer_v2/optimizer_v2.py#L1430-L1450 def _var_key(var): """Key for representing a primary variable, for looking up slots. In graph mode the name is derived from the var shared name. In eager mode the name is derived from the var unique id. If distribution strategy exists, get the primary variable first. Args: var: the variable. Returns: the unique name of the variable. """ # Get the distributed variable if it exists. if hasattr(var, "_distributed_container"): var = var._distributed_container() if var._in_graph_mode: return var._shared_name return var._unique_id @utils.register_keras_custom_object class CaseOptimizer(Optimizer): """An optmizer wrapper that applies different optimizers to a subset of variables. An optimizer is used to train a variable iff its accompanying predicate evaluates to `True`. For each variable, at most one optimizer's predicate may evaluate to `True`. If no optimizer's predicate evaluates to `True` for a variable, it is trained with the `default_optimizer`. If a variable is claimed by no optimizers and `default_optimizer == None`, the variable is not trained. # Arguments predicate_optimizer_pairs: One or more `(pred, tf.keras.optimizers.legacy.Optimizer)` pairs, where `pred` takes one `tf.Variable` as argument and returns `True` if the optimizer should be used for that variable, e.g. `pred(var) == True`. default_optimizer: A `tf.keras.optimizers.legacy.Optimizer` to be applied to any variable not claimed by any other optimizer. (Must be passed as keyword argument.) """ _HAS_AGGREGATE_GRAD = True def __init__( self, *predicate_optimizer_pairs: Tuple[Callable[[tf.Variable], bool], Optimizer], default_optimizer: Optional[Optimizer] = None, name: str = "optimizer_case", ): super().__init__(name=name) # Type checks for (predicate, optimizer) pairs for i, (predicate, optimizer) in enumerate(predicate_optimizer_pairs): if not callable(predicate): raise TypeError( f"Expected callable predicate at `predicate_optimizer_pairs[{i}][0]` but got `{type(predicate)}`." ) if not isinstance(optimizer, Optimizer): raise TypeError( f"Expected `tf.keras.optimizers.legacy.Optimizer` at `predicate_optimizer_pairs[{i}][1]` but got `{type(optimizer)}`." ) # Type check for default optimizers if default_optimizer is not None and not isinstance( default_optimizer, Optimizer ): raise TypeError( f"Expected `Optimizer` for `default_optimizer` but got `{type(default_optimizer)}`." ) self.pred_opt_pairs = predicate_optimizer_pairs self.default = default_optimizer self.var_opt_mapping = None # List of optimizers ending in `default_optimizer`, for easier internal access self.optimizers = [opt for (_, opt) in self.pred_opt_pairs] if self.default: self.optimizers.append(self.default) self.DEFAULT_OPT_INDEX = len(self.pred_opt_pairs) # Track optimizers to support reloading via tf.train.Checkpoint for i, optimizer in enumerate(self.optimizers): self._track_trackable(optimizer, name=f"optimizer_{i}") @property def weights(self): weights = [] for optimizer in self.optimizers: weights.extend(optimizer.weights) return weights @Optimizer.iterations.setter def iterations(self, variable): raise NotImplementedError("CaseOptimzer does not support setting iterations.") def apply_gradients(self, grads_and_vars, name: Optional[str] = None, **kwargs): """Apply gradients to variables for each optimizer. On the first call to `apply_gradients()`, compute the mapping from variables to optimizers and cache it in the `self.var_opt_mapping` dict for serialization and faster access. """ if self.var_opt_mapping is None: # Convert `grads_and_vars` to list so we can iterate multiple times over it grads_and_vars = list(grads_and_vars) self._compute_var_opt_mapping(grads_and_vars) # Split gradients and variables into a separate list for each optimizer grad_var_lists = [[] for _ in range(len(self.pred_opt_pairs) + 1)] for grad, var in grads_and_vars: var_key = _var_key(var) if var_key in self.var_opt_mapping: grad_var_lists[self.var_opt_mapping[var_key]].append((grad, var)) with tf.init_scope(): _ = self.iterations # This is only necessary in TF 2.0 and older, but doesn't hurt on newer versions for optimizer, opt_grads_and_vars in zip(self.optimizers, grad_var_lists): optimizer._create_slots([v for (_, v) in opt_grads_and_vars]) return tf.distribute.get_replica_context().merge_call( self._apply_gradients, args=(grad_var_lists, name), kwargs=kwargs ) def _apply_gradients(self, distribution, grad_var_lists, name, **kwargs): # Apply gradients to each optimizer with tf.name_scope(self._name): train_ops = [ distribution.extended.call_for_each_replica( optimizer.apply_gradients, args=(opt_grads_and_vars,), kwargs=kwargs ) for optimizer, opt_grads_and_vars in zip( self.optimizers, grad_var_lists ) ] return tf.group(*train_ops, name=name or "train_with_group") def get_config(self): optimizer_configs = [opt.get_config() for (_, opt) in self.pred_opt_pairs] default_config = self.default.get_config() config = { "optimizer_configs": [ {"class_name": optimizer_config["name"], "config": optimizer_config} for optimizer_config in optimizer_configs ], "default_config": { "class_name": default_config["name"], "config": default_config, }, "var_opt_mapping": self.var_opt_mapping, # serialized instead of `pred`s } return {**super().get_config(), **config} @classmethod def from_config(cls, original_config, custom_objects=None): config = deepcopy(original_config) case_optimizer = cls( *[ # `(pred, opt)` tuples ( lambda _: False, # placeholder callable (`pred` is not serialized) tf.keras.optimizers.deserialize( # optimizer `opt` opt_config, custom_objects=custom_objects ), ) for opt_config in config["optimizer_configs"] ], default_optimizer=tf.keras.optimizers.deserialize( config["default_config"], custom_objects=custom_objects ), ) # Since we no longer have the `pred`s, we set the mapping explicitly case_optimizer.var_opt_mapping = config["var_opt_mapping"] return case_optimizer def _compute_var_opt_mapping(self, grads_and_vars): """Compute a unique mapping from variables to optimizer indices.""" self.var_opt_mapping = {} for _, var in grads_and_vars: num_optimizers = 0 var_key = _var_key(var) # Find the optimizer(s) that want to claim this variable for optimizer_index, (predicate, _) in enumerate(self.pred_opt_pairs): if predicate(var): self.var_opt_mapping[var_key] = optimizer_index num_optimizers += 1 if num_optimizers > 1: raise ValueError(f"Variable `{var}` claimed by multiple optimizers.") if num_optimizers == 0: if self.default is not None: self.var_opt_mapping[var_key] = self.DEFAULT_OPT_INDEX else: warnings.warn( f"No `default_optimizer` provided to train variable `{var}`." ) # Make sure that each optimizer touches at least one variable for optimizer_index, (_, optimizer) in enumerate(self.pred_opt_pairs): if optimizer_index not in self.var_opt_mapping.values(): raise ValueError( f"Optimizer `{optimizer}` did not claim any variables." ) @utils.register_keras_custom_object class Bop(Optimizer): """Binary optimizer (Bop). Bop is a latent-free optimizer for Binarized Neural Networks (BNNs) and Binary Weight Networks (BWN). Bop maintains an exponential moving average of the gradients controlled by `gamma`. If this average exceeds the `threshold`, a weight is flipped. The hyperparameter `gamma` is somewhat analogues to the learning rate in SGD methods: a high `gamma` results in rapid convergence but also makes training more noisy. Note that the default `threshold` is not optimal for all situations. Setting the threshold too high results in little learning, while setting it too low results in overly noisy behaviour. !!! warning The `is_binary_variable` check of this optimizer will only target variables that have been explicitly marked as being binary using `NoOp(precision=1)`. !!! example ```python no_op_quantizer = lq.quantizers.NoOp(precision=1) layer = lq.layers.QuantDense(16, kernel_quantizer=no_op_quantizer) optimizer = lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, lq.optimizers.Bop()), default_optimizer=tf.keras.optimizers.Adam(0.01), # for FP weights ) ``` # Arguments threshold: magnitude of average gradient signal required to flip a weight. gamma: the adaptivity rate. name: name of the optimizer. # References - [Latent Weights Do Not Exist: Rethinking Binarized Neural Network Optimization](https://papers.nips.cc/paper/8971-latent-weights-do-not-exist-rethinking-binarized-neural-network-optimization) """ _HAS_AGGREGATE_GRAD = True def __init__( self, threshold: float = 1e-8, gamma: float = 1e-4, name: str = "Bop", **kwargs ): super().__init__(name=name, **kwargs) self._set_hyper("threshold", threshold) self._set_hyper("gamma", gamma) def _create_slots(self, var_list): for var in var_list: self.add_slot(var, "m") def _get_decayed_hyper(self, name: str, var_dtype): hyper = self._get_hyper(name, var_dtype) if isinstance(hyper, tf.keras.optimizers.schedules.LearningRateSchedule): local_step = tf.cast(self.iterations, var_dtype) hyper = tf.cast(hyper(local_step), var_dtype) return hyper def _resource_apply_dense(self, grad, var): var_dtype = var.dtype.base_dtype gamma = self._get_decayed_hyper("gamma", var_dtype) threshold = self._get_decayed_hyper("threshold", var_dtype) m = self.get_slot(var, "m") m_t = m.assign_add(gamma * (grad - m)) var_t = lq.math.sign(-tf.sign(var * m_t - threshold) * var) return var.assign(var_t).op def get_config(self): config = { "threshold": self._serialize_hyperparameter("threshold"), "gamma": self._serialize_hyperparameter("gamma"), } return {**super().get_config(), **config} @classmethod def from_config(cls, config, custom_objects=None): for hyper in ("gamma", "threshold"): if hyper in config and isinstance(config[hyper], dict): config[hyper] = tf.keras.optimizers.schedules.deserialize( config[hyper], custom_objects=custom_objects ) return cls(**config) @staticmethod def is_binary_variable(var: tf.Variable) -> bool: """Returns `True` for variables with `var.precision == 1`. This is an example of a predictate that can be used by the `CaseOptimizer`. # Arguments var: a `tf.Variable`. """ return getattr(var, "precision", 32) == 1
14,848
39.350543
201
py
larq
larq-main/larq/math_test.py
import numpy as np import pytest import tensorflow as tf import larq as lq from larq.testing_utils import generate_real_values_with_zeros @pytest.mark.parametrize("fn", [lq.math.sign]) def test_sign(fn): x = tf.keras.backend.placeholder(ndim=2) f = tf.keras.backend.function([x], [fn(x)]) binarized_values = np.random.choice([-1, 1], size=(2, 5)).astype(np.float32) result = f(binarized_values)[0] np.testing.assert_allclose(result, binarized_values) real_values = generate_real_values_with_zeros() result = f(real_values)[0] assert not np.any(result == 0) assert np.all(result[real_values < 0] == -1) assert np.all(result[real_values >= 0] == 1) zero_values = np.zeros((2, 5)) result = f(zero_values)[0] assert np.all(result == 1) @pytest.mark.parametrize("fn", [lq.math.heaviside]) def test_heaviside(fn): x = tf.keras.backend.placeholder(ndim=2) f = tf.keras.backend.function([x], [fn(x)]) binarized_values = np.random.choice([0, 1], size=(2, 5)) result = f([binarized_values])[0] np.testing.assert_allclose(result, binarized_values) real_values = generate_real_values_with_zeros() result = f([real_values])[0] assert np.all(result[real_values <= 0] == 0) assert np.all(result[real_values > 0] == 1)
1,299
31.5
80
py
larq
larq-main/larq/layers_base.py
import logging from typing import Optional import tensorflow as tf from larq import context, quantizers, utils from larq.quantized_variable import QuantizedVariable from larq.quantizers import NoOp, QuantizerType log = logging.getLogger(__name__) def _is_binary(quantizer): return getattr(quantizer, "precision", None) == 1 and not isinstance( quantizer, NoOp ) def _compute_padded_size(stride, dilation_rate, input_size, filter_size): if input_size is None: return None effective_filter_size = (filter_size - 1) * dilation_rate + 1 output_size = (input_size + stride - 1) // stride padded_size = (output_size - 1) * stride + effective_filter_size if tf.is_tensor(input_size): return tf.math.maximum(padded_size, input_size) return max(padded_size, input_size) def _compute_padding(stride, dilation_rate, input_size, filter_size): padded_size = _compute_padded_size(stride, dilation_rate, input_size, filter_size) total_padding = padded_size - input_size padding = total_padding // 2 return padding, padding + (total_padding % 2) class BaseLayer(tf.keras.layers.Layer): """Base class for defining quantized layers. `input_quantizer` is the element-wise quantization functions to use. If `input_quantizer=None` this layer is equivalent to `tf.keras.layers.Layer`. """ def __init__(self, *args, input_quantizer=None, **kwargs): self.input_quantizer = quantizers.get(input_quantizer) super().__init__(*args, **kwargs) def call(self, inputs): if self.input_quantizer: inputs = self.input_quantizer(inputs) with context.quantized_scope(True): return super().call(inputs) def get_config(self): return { **super().get_config(), "input_quantizer": quantizers.serialize(self.input_quantizer), } def _get_quantizer(self, name) -> Optional[QuantizerType]: """Get quantizer for given kernel name""" return None def _add_variable_with_custom_getter(self, name: str, **kwargs): quantizer = self._get_quantizer(name) if quantizer is None: return super()._add_variable_with_custom_getter(name, **kwargs) old_getter = kwargs.pop("getter") # Wrap `getter` with a version that returns a `QuantizedVariable`. def getter(*args, **kwargs): variable = old_getter(*args, **kwargs) return QuantizedVariable.from_variable(variable, quantizer) return super()._add_variable_with_custom_getter(name, getter=getter, **kwargs) class QuantizerBase(BaseLayer): """Base class for defining quantized layers with a single kernel. `kernel_quantizer` is the element-wise quantization functions to use. If `kernel_quantizer=None` this layer is equivalent to `BaseLayer`. """ def __init__(self, *args, kernel_quantizer=None, **kwargs): self.kernel_quantizer = quantizers.get_kernel_quantizer(kernel_quantizer) super().__init__(*args, **kwargs) if _is_binary(self.kernel_quantizer) and not self.kernel_constraint: log.warning( "Using a binary weight quantizer without setting `kernel_constraint` " "may result in starved weights (where the gradient is always zero)." ) def _get_quantizer(self, name: str) -> Optional[QuantizerType]: return self.kernel_quantizer if name == "kernel" else None def get_config(self): return { **super().get_config(), "kernel_quantizer": quantizers.serialize(self.kernel_quantizer), } class QuantizerBaseConv(tf.keras.layers.Layer): """Base class for defining quantized conv layers""" def __init__(self, *args, pad_values=0.0, **kwargs): self.pad_values = pad_values super().__init__(*args, **kwargs) is_zero_padding = not tf.is_tensor(self.pad_values) and self.pad_values == 0.0 self._is_native_padding = self.padding != "same" or is_zero_padding if self.padding == "causal" and not is_zero_padding: raise ValueError("Causal padding with `pad_values != 0` is not supported.") def _get_spatial_padding_same(self, shape): return [ _compute_padding(stride, dilation_rate, shape[i], filter_size) for i, (stride, dilation_rate, filter_size) in enumerate( zip(self.strides, self.dilation_rate, self.kernel_size) ) ] def _get_spatial_shape(self, input_shape): return ( input_shape[1:-1] if self.data_format == "channels_last" else input_shape[2:] ) def _get_padding_same(self, inputs): input_shape = inputs.shape if not input_shape[1:].is_fully_defined(): input_shape = tf.shape(inputs) padding = self._get_spatial_padding_same(self._get_spatial_shape(input_shape)) return ( [[0, 0], *padding, [0, 0]] if self.data_format == "channels_last" else [[0, 0], [0, 0], *padding] ) def _get_padding_same_shape(self, input_shape): spatial_input_shape = self._get_spatial_shape(input_shape) spatial_shape = [ _compute_padded_size(stride, dilation, size, filter_size) for size, stride, dilation, filter_size in zip( spatial_input_shape, self.strides, self.dilation_rate, self.kernel_size, ) ] if self.data_format == "channels_last": return tf.TensorShape([input_shape[0], *spatial_shape, input_shape[-1]]) return tf.TensorShape([*input_shape[:2], *spatial_shape]) def build(self, input_shape): if self._is_native_padding: super().build(input_shape) else: with utils.patch_object(self, "padding", "valid"): super().build(self._get_padding_same_shape(input_shape)) def call(self, inputs): if self._is_native_padding: return super().call(inputs) inputs = tf.pad( inputs, self._get_padding_same(inputs), constant_values=self.pad_values ) with utils.patch_object(self, "padding", "valid"): return super().call(inputs) def get_config(self): return { **super().get_config(), "pad_values": tf.keras.backend.get_value(self.pad_values), } class QuantizerDepthwiseBase(BaseLayer): """Base class for defining depthwise quantized layers `depthwise_quantizer` is the element-wise quantization functions to use. If `depthwise_quantizer=None` this layer is equivalent to `BaseLayer`. """ def __init__( self, *args, depthwise_quantizer: Optional[QuantizerType] = None, **kwargs, ): self.depthwise_quantizer = quantizers.get_kernel_quantizer(depthwise_quantizer) super().__init__(*args, **kwargs) if _is_binary(self.depthwise_quantizer) and not self.depthwise_constraint: log.warning( "Using a binary weight quantizer without setting `depthwise_constraint` " "may result in starved weights (where the gradient is always zero)." ) def _get_quantizer(self, name: str) -> Optional[QuantizerType]: return self.depthwise_quantizer if name == "depthwise_kernel" else None def get_config(self): return { **super().get_config(), "depthwise_quantizer": quantizers.serialize(self.depthwise_quantizer), } class QuantizerSeparableBase(BaseLayer): """Base class for defining separable quantized layers. `depthwise_quantizer` and `pointwise_quantizer` are the element-wise quantization functions to use. If all quantization functions are `None` this layer is equivalent to `BaseLayer`. """ def __init__( self, *args, depthwise_quantizer: Optional[QuantizerType] = None, pointwise_quantizer: Optional[QuantizerType] = None, **kwargs, ): self.depthwise_quantizer = quantizers.get_kernel_quantizer(depthwise_quantizer) self.pointwise_quantizer = quantizers.get_kernel_quantizer(pointwise_quantizer) super().__init__(*args, **kwargs) if _is_binary(self.depthwise_quantizer) and not self.depthwise_constraint: log.warning( "Using a binary `depthwise_quantizer` without setting `depthwise_constraint` " "may result in starved weights (where the gradient is always zero)." ) if _is_binary(self.pointwise_quantizer) and not self.pointwise_constraint: log.warning( "Using a binary `pointwise_quantizer` without setting `pointwise_constraint` " "may result in starved weights (where the gradient is always zero)." ) def _get_quantizer(self, name: str) -> Optional[QuantizerType]: if name == "depthwise_kernel": return self.depthwise_quantizer if name == "pointwise_kernel": return self.pointwise_quantizer return None def get_config(self): return { **super().get_config(), "depthwise_quantizer": quantizers.serialize(self.depthwise_quantizer), "pointwise_quantizer": quantizers.serialize(self.pointwise_quantizer), }
9,491
35.933852
94
py
larq
larq-main/larq/utils.py
from contextlib import contextmanager import tensorflow as tf def memory_as_readable_str(num_bits: int) -> str: """Generate a human-readable string for the memory size. 1 KiB = 1024 B; we use the binary prefix (KiB) [1,2] instead of the decimal prefix (KB) to avoid any confusion with multiplying by 1000 instead of 1024. [1] https://en.wikipedia.org/wiki/Binary_prefix [2] https://physics.nist.gov/cuu/Units/binary.html """ suffixes = ["B", "KiB", "MiB", "GiB"] num_bytes = num_bits / 8 for i, suffix in enumerate(suffixes): rounded = num_bytes / (1024**i) if rounded < 1024: break return f"{rounded:,.2f} {suffix}" def register_keras_custom_object(cls): """See https://github.com/tensorflow/addons/blob/master/tensorflow_addons/utils/keras_utils.py#L25""" tf.keras.utils.get_custom_objects()[cls.__name__] = cls return cls def register_alias(name: str): """A decorator to register a custom keras object under a given alias. !!! example ```python @utils.register_alias("degeneration") class Degeneration(tf.keras.metrics.Metric): pass ``` """ def register_func(cls): tf.keras.utils.get_custom_objects()[name] = cls return cls return register_func def set_precision(precision: int = 32): """A decorator to set the precision of a quantizer function # Arguments precision: An integer defining the precision of the output. """ def decorator(function): setattr(function, "precision", precision) return function return decorator @contextmanager def patch_object(object, name, value): """Temporarily overwrite attribute on object""" old_value = getattr(object, name) setattr(object, name, value) yield setattr(object, name, old_value)
1,874
25.408451
105
py
larq
larq-main/larq/models_test.py
import numpy as np import pytest import tensorflow as tf from packaging import version import larq as lq from larq.models import ModelProfile class ToyModel(tf.keras.Model): def __init__(self, **kwargs): super().__init__(**kwargs) self.conv = lq.layers.QuantConv2D( filters=32, kernel_size=(3, 3), kernel_quantizer="ste_sign", input_shape=(64, 64, 1), padding="same", ) self.pool = tf.keras.layers.GlobalAvgPool2D() self.dense = tf.keras.layers.Dense(10, activation="softmax") def call(self, inputs): return self.dense(self.pool(self.conv(inputs))) def get_functional_model(): input = tf.keras.Input((32, 32, 3)) x = lq.layers.QuantConv2D( filters=32, kernel_size=(3, 3), kernel_quantizer="ste_sign", padding="same", )(input) y, z = tf.split(x, 2, axis=-1) x = tf.concat([y, z], axis=-1) return tf.keras.Model(input, x, name="toy_model") def get_profile_model(): return tf.keras.models.Sequential( [ lq.layers.QuantConv2D( filters=32, kernel_size=(3, 3), kernel_quantizer="ste_sign", input_shape=(64, 64, 1), padding="same", ), tf.keras.layers.MaxPooling2D((2, 2)), lq.layers.QuantDepthwiseConv2D( kernel_size=3, strides=(3, 3), input_quantizer=lq.quantizers.SteTern(), depthwise_quantizer=lq.quantizers.SteTern(), padding="same", pad_values=1.0, use_bias=False, ), tf.keras.layers.BatchNormalization(scale=False), lq.layers.QuantSeparableConv2D( 32, (3, 3), input_quantizer="ste_sign", depthwise_quantizer="ste_sign", pointwise_quantizer="ste_sign", padding="same", ), tf.keras.layers.Flatten(), tf.keras.layers.Dense(10, trainable=False), ], ) def get_submodel_profile_model(start_index=2, end_index=5): # Same as above, but with a submodel as a layer model = get_profile_model() # Create submodel from e.g. the middle three layers submodel = tf.keras.models.Sequential( [layer for layer in model.layers[start_index:end_index]], ) return tf.keras.models.Sequential( [*model.layers[:start_index], submodel, *model.layers[end_index:]] ) def test_model_profile(): profile = ModelProfile(get_profile_model()) assert len(profile.layer_profiles) == 7 def test_layer_profile(): profile = ModelProfile(get_profile_model()) kernel_count = [ 32 * 3 * 3 * 1, 0, 32 * 3 * 3, 0, 32 * 3 * 3 * 1 + 32 * 1 * 1 * 32, 0, 32 * 11 * 11 * 10, ] bias_count = [32, 0, 0, 64, 32, 0, 10] param_count = [k + b for k, b in zip(kernel_count, bias_count)] memory = [ # bits * (c * w * h * b) + bits * bias 1 * (32 * 3 * 3 * 1) + 32 * 32, 0, 2 * (32 * 3 * 3), 32 * (2 * 32), 1 * (32 * 3 * 3 * 1 + 32 * 1 * 1 * 32) + 32 * 32, 0, 32 * (32 * 11 * 11 * 10 + 10), ] int8_fp_weights_mem = [ 1 * (32 * 3 * 3 * 1) + 8 * 32, 0, 2 * (32 * 3 * 3), 8 * (32 * 2), 1 * (32 * 3 * 3 * 1 + 32 * 1 * 1 * 32) + 8 * 32, 0, 8 * (32 * 11 * 11 * 10 + 10), ] fp_equiv_mem = [32 * n for n in param_count] input_precision = [None, None, 2, None, 1, None, None] output_shape = [ (-1, 64, 64, 32), (-1, 32, 32, 32), (-1, 11, 11, 32), (-1, 11, 11, 32), (-1, 11, 11, 32), (-1, 11 * 11 * 32), (-1, 10), ] output_pixels = [int(np.prod(os[1:-1])) for os in output_shape] unique_param_bidtwidths = [[1, 32], [], [2], [32], [1, 32], [], [32]] unique_op_precisions = [[32], [], [2], [], [1], [], [32]] mac_count = [params * pixels for params, pixels in zip(kernel_count, output_pixels)] bin_mac_count = [ mc if (1 in pb and ip == 1) else 0 for mc, pb, ip in zip(mac_count, unique_param_bidtwidths, input_precision) ] profiles = profile.layer_profiles for i in range(len(profiles)): print(f"Testing layer {i}...") assert profiles[i].input_precision == input_precision[i] assert profiles[i].output_shape == output_shape[i] assert profiles[i].output_pixels == output_pixels[i] assert profiles[i].weight_count() == param_count[i] assert profiles[i].unique_param_bidtwidths == unique_param_bidtwidths[i] assert profiles[i].unique_op_precisions == unique_op_precisions[i] assert profiles[i].memory == memory[i] assert profiles[i].fp_equivalent_memory == fp_equiv_mem[i] assert profiles[i].int8_fp_weights_memory == int8_fp_weights_mem[i] assert profiles[i].op_count("mac") == mac_count[i] assert profiles[i].op_count("mac", 1) == bin_mac_count[i] def test_layer_profile_1d(): model = tf.keras.models.Sequential( [ lq.layers.QuantConv1D( filters=32, kernel_size=3, input_shape=(64, 6), kernel_quantizer="ste_sign", padding="same", ), tf.keras.layers.MaxPooling1D(2), lq.layers.QuantSeparableConv1D( filters=16, kernel_size=3, input_quantizer="ste_sign", depthwise_quantizer="ste_sign", pointwise_quantizer="ste_sign", padding="same", ), tf.keras.layers.Flatten(), tf.keras.layers.Dense(10, trainable=False), ] ) profile = ModelProfile(model) kernel_count = [(32 * 3 * 6), 0, (32 * 3 + 16 * 32), 0, (16 * 32 * 10)] bias_count = [32, 0, 16, 0, 10] param_count = [k + b for k, b in zip(kernel_count, bias_count)] memory = [ # bits * (c * w * d) + bits * bias 1 * (32 * 3 * 6) + 32 * 32, 0, 1 * (32 * 3 + 16 * 32) + 32 * 16, 0, 32 * (32 * 16 * 10 + 10), ] int8_fp_weights_mem = [ 1 * (32 * 3 * 6) + 8 * 32, 0, 1 * (32 * 3 + 16 * 32) + 8 * 16, 0, 8 * (32 * 16 * 10 + 10), ] fp_equiv_mem = [32 * n for n in param_count] input_precision = [None, None, 1, None, None] output_shape = [ (-1, 64, 32), (-1, 32, 32), (-1, 32, 16), (-1, 32 * 16), (-1, 10), ] output_pixels = [int(np.prod(os[1:-1])) for os in output_shape] unique_param_bidtwidths = [[1, 32], [], [1, 32], [], [32]] unique_op_precisions = [[32], [], [1], [], [32]] mac_count = [params * pixels for params, pixels in zip(kernel_count, output_pixels)] bin_mac_count = [ mc if (1 in pb and ip == 1) else 0 for mc, pb, ip in zip(mac_count, unique_param_bidtwidths, input_precision) ] profiles = profile.layer_profiles for i in range(len(profiles)): print(f"Testing layer {i}...") assert profiles[i].input_precision == input_precision[i] assert profiles[i].output_shape == output_shape[i] assert profiles[i].output_pixels == output_pixels[i] assert profiles[i].weight_count() == param_count[i] assert profiles[i].unique_param_bidtwidths == unique_param_bidtwidths[i] assert profiles[i].unique_op_precisions == unique_op_precisions[i] assert profiles[i].memory == memory[i] assert profiles[i].fp_equivalent_memory == fp_equiv_mem[i] assert profiles[i].int8_fp_weights_memory == int8_fp_weights_mem[i] assert profiles[i].op_count("mac") == mac_count[i] assert profiles[i].op_count("mac", 1) == bin_mac_count[i] def test_summary(snapshot, capsys): model = get_profile_model() lq.models.summary(model) captured = capsys.readouterr() snapshot.assert_match(captured.out) # A model with no weights model = tf.keras.models.Sequential( [tf.keras.layers.Lambda(lambda x: tf.zeros(2), input_shape=(32, 32))] ) lq.models.summary(model) captured = capsys.readouterr() snapshot.assert_match(captured.out) def test_submodel_summary(capsys, snapshot): default_profile = ModelProfile(get_profile_model()) submodel = get_submodel_profile_model(start_index=2, end_index=5) submodel_profile = ModelProfile(submodel) submodel_layer_profile = submodel_profile.layer_profiles[2] # Assert that layer profile of the submodel "layer" matches the original layers profiles = default_profile.layer_profiles[2:5] assert submodel_layer_profile.input_precision == profiles[0].input_precision assert submodel_layer_profile.output_shape == profiles[-1].output_shape assert submodel_layer_profile.output_pixels == profiles[-1].output_pixels assert submodel_layer_profile.weight_count() == sum( (p.weight_count() for p in profiles) ) bitwidths = [] op_precisions = [] for p in profiles: bitwidths.extend(p.unique_param_bidtwidths) op_precisions.extend(p.unique_op_precisions) assert set(submodel_layer_profile.unique_param_bidtwidths) == set(bitwidths) assert set(submodel_layer_profile.unique_op_precisions) == set(op_precisions) assert submodel_layer_profile.memory == sum((p.memory for p in profiles)) assert submodel_layer_profile.fp_equivalent_memory == sum( (p.fp_equivalent_memory for p in profiles) ) assert submodel_layer_profile.int8_fp_weights_memory == sum( (p.int8_fp_weights_memory for p in profiles) ) assert submodel_layer_profile.op_count("mac") == sum( (p.op_count("mac") for p in profiles) ) assert submodel_layer_profile.op_count("mac", 1) == sum( (p.op_count("mac", 1) for p in profiles) ) # Assert that the total profile summary matches assert submodel_profile.generate_summary() == default_profile.generate_summary() # Snapshot the submodel profile itself to make sure it remains correct lq.models.summary(get_submodel_profile_model()) snapshot.assert_match(capsys.readouterr().out) def test_subclass_model_summary(snapshot, capsys): model = ToyModel() model.build((None, 32, 32, 3)) lq.models.summary(model) captured = capsys.readouterr() snapshot.assert_match(captured.out) def test_functional_model_summary(snapshot, capsys): lq.models.summary(get_functional_model()) captured = capsys.readouterr() key = "2.4+" if version.parse(tf.__version__) >= version.parse("2.3.9") else "<2.4" snapshot.assert_match(captured.out.lower(), key) def test_summary_invalid_model(): with pytest.raises(ValueError): lq.models.summary(tf.keras.Model()) def test_bitsize_invalid_key(): with pytest.raises(NotImplementedError): lq.models._bitsize_as_str(-1) def test_number_as_readable_str_large(): assert lq.models._number_as_readable_str(1e16) == "1.00E+16" @pytest.fixture(autouse=True) def run_around_tests(): tf.keras.backend.clear_session() yield
11,312
33.281818
88
py
larq
larq-main/larq/layers.py
"""Each Quantized Layer requires a `input_quantizer` and `kernel_quantizer` that describes the way of quantizing the activation of the previous layer and the weights respectively. If both `input_quantizer` and `kernel_quantizer` are `None` the layer is equivalent to a full precision layer. """ import tensorflow as tf from packaging import version from larq import utils from larq.layers_base import ( QuantizerBase, QuantizerBaseConv, QuantizerDepthwiseBase, QuantizerSeparableBase, ) @utils.register_keras_custom_object class QuantDense(QuantizerBase, tf.keras.layers.Dense): """Just your regular densely-connected quantized NN layer. `QuantDense` implements the operation: `output = activation(dot(input_quantizer(input), kernel_quantizer(kernel)) + bias)`, where `activation` is the element-wise activation function passed as the `activation` argument, `kernel` is a weights matrix created by the layer, and `bias` is a bias vector created by the layer (only applicable if `use_bias` is `True`). `input_quantizer` and `kernel_quantizer` are the element-wise quantization functions to use. If both quantization functions are `None` this layer is equivalent to `Dense`. !!! note "" If the input to the layer has a rank greater than 2, then it is flattened prior to the initial dot product with `kernel`. !!! example ```python # as first layer in a sequential model: model = Sequential() model.add( QuantDense( 32, input_quantizer="ste_sign", kernel_quantizer="ste_sign", kernel_constraint="weight_clip", input_shape=(16,), ) ) # now the model will take as input arrays of shape (*, 16) # and output arrays of shape (*, 32) # after the first layer, you don't need to specify # the size of the input anymore: model.add( QuantDense( 32, input_quantizer="ste_sign", kernel_quantizer="ste_sign", kernel_constraint="weight_clip", ) ) ``` # Arguments units: Positive integer, dimensionality of the output space. activation: Activation function to use. If you don't specify anything, no activation is applied (`a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. kernel_quantizer: Quantization function applied to the `kernel` weights matrix. kernel_initializer: Initializer for the `kernel` weights matrix. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer function applied to the `kernel` weights matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). kernel_constraint: Constraint function applied to the `kernel` weights matrix. bias_constraint: Constraint function applied to the bias vector. # Input shape N-D tensor with shape: `(batch_size, ..., input_dim)`. The most common situation would be a 2D input with shape `(batch_size, input_dim)`. # Output shape N-D tensor with shape: `(batch_size, ..., units)`. For instance, for a 2D input with shape `(batch_size, input_dim)`, the output would have shape `(batch_size, units)`. """ def __init__( self, units, activation=None, use_bias=True, input_quantizer=None, kernel_quantizer=None, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs, ): super().__init__( units, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, kernel_quantizer=kernel_quantizer, kernel_initializer=kernel_initializer, bias_initializer=bias_initializer, kernel_regularizer=kernel_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, kernel_constraint=kernel_constraint, bias_constraint=bias_constraint, **kwargs, ) @utils.register_keras_custom_object class QuantConv1D(QuantizerBase, QuantizerBaseConv, tf.keras.layers.Conv1D): """1D quantized convolution layer (e.g. temporal convolution). This layer creates a convolution kernel that is convolved with the layer input over a single spatial (or temporal) dimension to produce a tensor of outputs. `input_quantizer` and `kernel_quantizer` are the element-wise quantization functions to use. If both quantization functions are `None` this layer is equivalent to `Conv1D`. If `use_bias` is True, a bias vector is created and added to the outputs. Finally, if `activation` is not `None`, it is applied to the outputs as well. When using this layer as the first layer in a model, provide an `input_shape` argument (tuple of integers or `None`, e.g. `(10, 128)` for sequences of 10 vectors of 128-dimensional vectors, or `(None, 128)` for variable-length sequences of 128-dimensional vectors. # Arguments filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of a single integer, specifying the length of the 1D convolution window. strides: An integer or tuple/list of a single integer, specifying the stride length of the convolution. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"`, `"causal"` or `"same"` (case-insensitive). `"causal"` results in causal (dilated) convolutions, e.g. output[t] does not depend on input[t+1:]. Useful when modeling temporal data where the model should not violate the temporal order. See [WaveNet: A Generative Model for Raw Audio, section 2.1](https://arxiv.org/abs/1609.03499). pad_values: The pad value to use when `padding="same"`. data_format: A string, one of `channels_last` (default) or `channels_first`. dilation_rate: an integer or tuple/list of a single integer, specifying the dilation rate to use for dilated convolution. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any `strides` value != 1. groups: A positive integer specifying the number of groups in which the input is split along the channel axis. Each group is convolved separately with `filters / groups` filters. The output is the concatenation of all the `groups` results along the channel axis. Input channels and `filters` must both be divisible by `groups`. activation: Activation function to use. If you don't specify anything, no activation is applied (`a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. kernel_quantizer: Quantization function applied to the `kernel` weights matrix. kernel_initializer: Initializer for the `kernel` weights matrix. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer function applied to the `kernel` weights matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). kernel_constraint: Constraint function applied to the kernel matrix. bias_constraint: Constraint function applied to the bias vector. # Input shape 3D tensor with shape: `(batch_size, steps, input_dim)` # Output shape 3D tensor with shape: `(batch_size, new_steps, filters)`. `steps` value might have changed due to padding or strides. """ def __init__( self, filters, kernel_size, strides=1, padding="valid", pad_values=0.0, data_format="channels_last", dilation_rate=1, groups=1, activation=None, use_bias=True, input_quantizer=None, kernel_quantizer=None, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs, ): if groups != 1: if version.parse(tf.__version__) >= version.parse("2.3"): kwargs = {**kwargs, "groups": groups} else: raise ValueError( "`groups` != 1 requires TensorFlow version 2.3 or newer." ) super().__init__( filters, kernel_size, strides=strides, padding=padding, pad_values=pad_values, data_format=data_format, dilation_rate=dilation_rate, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, kernel_quantizer=kernel_quantizer, kernel_initializer=kernel_initializer, bias_initializer=bias_initializer, kernel_regularizer=kernel_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, kernel_constraint=kernel_constraint, bias_constraint=bias_constraint, **kwargs, ) @utils.register_keras_custom_object class QuantConv2D(QuantizerBase, QuantizerBaseConv, tf.keras.layers.Conv2D): """2D quantized convolution layer (e.g. spatial convolution over images). This layer creates a convolution kernel that is convolved with the layer input to produce a tensor of outputs. `input_quantizer` and `kernel_quantizer` are the element-wise quantization functions to use. If both quantization functions are `None` this layer is equivalent to `Conv2D`. If `use_bias` is True, a bias vector is created and added to the outputs. Finally, if `activation` is not `None`, it is applied to the outputs as well. When using this layer as the first layer in a model, provide the keyword argument `input_shape` (tuple of integers, does not include the sample axis), e.g. `input_shape=(128, 128, 3)` for 128x128 RGB pictures in `data_format="channels_last"`. # Arguments filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of 2 integers, specifying the height and width of the 2D convolution window. Can be a single integer to specify the same value for all spatial dimensions. strides: An integer or tuple/list of 2 integers, specifying the strides of the convolution along the height and width. Can be a single integer to specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). pad_values: The pad value to use when `padding="same"`. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch, height, width, channels)` while `channels_first` corresponds to inputs with shape `(batch, channels, height, width)`. It defaults to the `image_data_format` value found in your Keras config file at `~/.keras/keras.json`. If you never set it, then it will be "channels_last". dilation_rate: an integer or tuple/list of 2 integers, specifying the dilation rate to use for dilated convolution. Can be a single integer to specify the same value for all spatial dimensions. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any stride value != 1. groups: A positive integer specifying the number of groups in which the input is split along the channel axis. Each group is convolved separately with `filters / groups` filters. The output is the concatenation of all the `groups` results along the channel axis. Input channels and `filters` must both be divisible by `groups`. activation: Activation function to use. If you don't specify anything, no activation is applied (`a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. kernel_quantizer: Quantization function applied to the `kernel` weights matrix. kernel_initializer: Initializer for the `kernel` weights matrix. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer function applied to the `kernel` weights matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). kernel_constraint: Constraint function applied to the kernel matrix. bias_constraint: Constraint function applied to the bias vector. # Input shape 4D tensor with shape: `(samples, channels, rows, cols)` if data_format='channels_first' or 4D tensor with shape: `(samples, rows, cols, channels)` if data_format='channels_last'. # Output shape 4D tensor with shape: `(samples, filters, new_rows, new_cols)` if data_format='channels_first' or 4D tensor with shape: `(samples, new_rows, new_cols, filters)` if data_format='channels_last'. `rows` and `cols` values might have changed due to padding. """ def __init__( self, filters, kernel_size, strides=(1, 1), padding="valid", pad_values=0.0, data_format=None, dilation_rate=(1, 1), groups=1, activation=None, use_bias=True, input_quantizer=None, kernel_quantizer=None, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs, ): if groups != 1: if version.parse(tf.__version__) >= version.parse("2.3"): kwargs = {**kwargs, "groups": groups} else: raise ValueError( "`groups` != 1 requires TensorFlow version 2.3 or newer." ) super().__init__( filters, kernel_size, strides=strides, padding=padding, pad_values=pad_values, data_format=data_format, dilation_rate=dilation_rate, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, kernel_quantizer=kernel_quantizer, kernel_initializer=kernel_initializer, bias_initializer=bias_initializer, kernel_regularizer=kernel_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, kernel_constraint=kernel_constraint, bias_constraint=bias_constraint, **kwargs, ) @utils.register_keras_custom_object class QuantConv3D(QuantizerBase, QuantizerBaseConv, tf.keras.layers.Conv3D): """3D convolution layer (e.g. spatial convolution over volumes). This layer creates a convolution kernel that is convolved with the layer input to produce a tensor of outputs. `input_quantizer` and `kernel_quantizer` are the element-wise quantization functions to use. If both quantization functions are `None` this layer is equivalent to `Conv3D`. If `use_bias` is True, a bias vector is created and added to the outputs. Finally, if `activation` is not `None`, it is applied to the outputs as well. When using this layer as the first layer in a model, provide the keyword argument `input_shape` (tuple of integers, does not include the sample axis), e.g. `input_shape=(128, 128, 128, 1)` for 128x128x128 volumes with a single channel, in `data_format="channels_last"`. # Arguments filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of 3 integers, specifying the depth, height and width of the 3D convolution window. Can be a single integer to specify the same value for all spatial dimensions. strides: An integer or tuple/list of 3 integers, specifying the strides of the convolution along each spatial dimension. Can be a single integer to specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). pad_values: The pad value to use when `padding="same"`. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch, spatial_dim1, spatial_dim2, spatial_dim3, channels)` while `channels_first` corresponds to inputs with shape `(batch, channels, spatial_dim1, spatial_dim2, spatial_dim3)`. It defaults to the `image_data_format` value found in your Keras config file at `~/.keras/keras.json`. If you never set it, then it will be "channels_last". dilation_rate: an integer or tuple/list of 2 integers, specifying the dilation rate to use for dilated convolution. Can be a single integer to specify the same value for all spatial dimensions. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any stride value != 1. groups: A positive integer specifying the number of groups in which the input is split along the channel axis. Each group is convolved separately with `filters / groups` filters. The output is the concatenation of all the `groups` results along the channel axis. Input channels and `filters` must both be divisible by `groups`. activation: Activation function to use. If you don't specify anything, no activation is applied (`a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. kernel_quantizer: Quantization function applied to the `kernel` weights matrix. kernel_initializer: Initializer for the `kernel` weights matrix. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer function applied to the `kernel` weights matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). kernel_constraint: Constraint function applied to the kernel matrix. bias_constraint: Constraint function applied to the bias vector. # Input shape 5D tensor with shape: `(samples, channels, conv_dim1, conv_dim2, conv_dim3)` if data_format='channels_first' or 5D tensor with shape: `(samples, conv_dim1, conv_dim2, conv_dim3, channels)` if data_format='channels_last'. # Output shape 5D tensor with shape: `(samples, filters, new_conv_dim1, new_conv_dim2, new_conv_dim3)` if data_format='channels_first' or 5D tensor with shape: `(samples, new_conv_dim1, new_conv_dim2, new_conv_dim3, filters)` if data_format='channels_last'. `new_conv_dim1`, `new_conv_dim2` and `new_conv_dim3` values might have changed due to padding. """ def __init__( self, filters, kernel_size, strides=(1, 1, 1), padding="valid", pad_values=0.0, data_format=None, dilation_rate=(1, 1, 1), groups=1, activation=None, use_bias=True, input_quantizer=None, kernel_quantizer=None, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs, ): if groups != 1: if version.parse(tf.__version__) >= version.parse("2.3"): kwargs = {**kwargs, "groups": groups} else: raise ValueError( "`groups` != 1 requires TensorFlow version 2.3 or newer." ) super().__init__( filters, kernel_size, strides=strides, padding=padding, pad_values=pad_values, data_format=data_format, dilation_rate=dilation_rate, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, kernel_quantizer=kernel_quantizer, kernel_initializer=kernel_initializer, bias_initializer=bias_initializer, kernel_regularizer=kernel_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, kernel_constraint=kernel_constraint, bias_constraint=bias_constraint, **kwargs, ) @utils.register_keras_custom_object class QuantDepthwiseConv2D( QuantizerDepthwiseBase, QuantizerBaseConv, tf.keras.layers.DepthwiseConv2D ): """Quantized depthwise separable 2D convolution. Depthwise Separable convolutions consists in performing just the first step in a depthwise spatial convolution (which acts on each input channel separately). The `depth_multiplier` argument controls how many output channels are generated per input channel in the depthwise step. # Arguments kernel_size: An integer or tuple/list of 2 integers, specifying the height and width of the 2D convolution window. Can be a single integer to specify the same value for all spatial dimensions. strides: An integer or tuple/list of 2 integers, specifying the strides of the convolution along the height and width. Can be a single integer to specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `'valid'` or `'same'` (case-insensitive). pad_values: The pad value to use when `padding="same"`. depth_multiplier: The number of depthwise convolution output channels for each input channel. The total number of depthwise convolution output channels will be equal to `filters_in * depth_multiplier`. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch, height, width, channels)` while `channels_first` corresponds to inputs with shape `(batch, channels, height, width)`. It defaults to the `image_data_format` value found in your Keras config file at `~/.keras/keras.json`. If you never set it, then it will be 'channels_last'. dilation_rate: an integer or tuple/list of 2 integers, specifying the dilation rate to use for dilated convolution. Can be a single integer to specify the same value for all spatial dimensions. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any stride value != 1. activation: Activation function to use. If you don't specify anything, no activation is applied (ie. `a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. depthwise_quantizer: Quantization function applied to the `depthwise_kernel` weights matrix. depthwise_initializer: Initializer for the depthwise kernel matrix. bias_initializer: Initializer for the bias vector. depthwise_regularizer: Regularizer function applied to the depthwise kernel matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its 'activation'). depthwise_constraint: Constraint function applied to the depthwise kernel matrix. bias_constraint: Constraint function applied to the bias vector. # Input shape 4D tensor with shape: `[batch, channels, rows, cols]` if data_format='channels_first' or 4D tensor with shape: `[batch, rows, cols, channels]` if data_format='channels_last'. # Output shape 4D tensor with shape: `[batch, filters, new_rows, new_cols]` if data_format='channels_first' or 4D tensor with shape: `[batch, new_rows, new_cols, filters]` if data_format='channels_last'. `rows` and `cols` values might have changed due to padding. """ def __init__( self, kernel_size, strides=(1, 1), padding="valid", pad_values=0.0, depth_multiplier=1, data_format=None, dilation_rate=(1, 1), activation=None, use_bias=True, input_quantizer=None, depthwise_quantizer=None, depthwise_initializer="glorot_uniform", bias_initializer="zeros", depthwise_regularizer=None, bias_regularizer=None, activity_regularizer=None, depthwise_constraint=None, bias_constraint=None, **kwargs, ): super().__init__( kernel_size=kernel_size, strides=strides, padding=padding, pad_values=pad_values, depth_multiplier=depth_multiplier, data_format=data_format, dilation_rate=dilation_rate, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, depthwise_quantizer=depthwise_quantizer, depthwise_initializer=depthwise_initializer, bias_initializer=bias_initializer, depthwise_regularizer=depthwise_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, depthwise_constraint=depthwise_constraint, bias_constraint=bias_constraint, **kwargs, ) @utils.register_keras_custom_object class QuantSeparableConv1D( QuantizerSeparableBase, QuantizerBaseConv, tf.keras.layers.SeparableConv1D ): """Depthwise separable 1D quantized convolution. This layer performs a depthwise convolution that acts separately on channels, followed by a pointwise convolution that mixes channels. `input_quantizer`, `depthwise_quantizer` and `pointwise_quantizer` are the element-wise quantization functions to use. If all quantization functions are `None` this layer is equivalent to `SeparableConv1D`. If `use_bias` is True and a bias initializer is provided, it adds a bias vector to the output. It then optionally applies an activation function to produce the final output. # Arguments filters: Integer, the dimensionality of the output space (i.e. the number of filters in the convolution). kernel_size: A single integer specifying the spatial dimensions of the filters. strides: A single integer specifying the strides of the convolution. Specifying any `stride` value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"`, `"same"`, or `"causal"` (case-insensitive). pad_values: The pad value to use when `padding="same"`. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch, length, channels)` while `channels_first` corresponds to inputs with shape `(batch, channels, length)`. dilation_rate: A single integer, specifying the dilation rate to use for dilated convolution. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any stride value != 1. depth_multiplier: The number of depthwise convolution output channels for each input channel. The total number of depthwise convolution output channels will be equal to `num_filters_in * depth_multiplier`. activation: Activation function. Set it to None to maintain a linear activation. use_bias: Boolean, whether the layer uses a bias. input_quantizer: Quantization function applied to the input of the layer. depthwise_quantizer: Quantization function applied to the depthwise kernel. pointwise_quantizer: Quantization function applied to the pointwise kernel. depthwise_initializer: An initializer for the depthwise convolution kernel. pointwise_initializer: An initializer for the pointwise convolution kernel. bias_initializer: An initializer for the bias vector. If None, the default initializer will be used. depthwise_regularizer: Optional regularizer for the depthwise convolution kernel. pointwise_regularizer: Optional regularizer for the pointwise convolution kernel. bias_regularizer: Optional regularizer for the bias vector. activity_regularizer: Optional regularizer function for the output. depthwise_constraint: Optional projection function to be applied to the depthwise kernel after being updated by an `Optimizer` (e.g. used for norm constraints or value constraints for layer weights). The function must take as input the unprojected variable and must return the projected variable (which must have the same shape). Constraints are not safe to use when doing asynchronous distributed training. pointwise_constraint: Optional projection function to be applied to the pointwise kernel after being updated by an `Optimizer`. bias_constraint: Optional projection function to be applied to the bias after being updated by an `Optimizer`. trainable: Boolean, if `True` the weights of this layer will be marked as trainable (and listed in `layer.trainable_weights`). name: A string, the name of the layer. """ def __init__( self, filters, kernel_size, strides=1, padding="valid", pad_values=0.0, data_format=None, dilation_rate=1, depth_multiplier=1, activation=None, use_bias=True, input_quantizer=None, depthwise_quantizer=None, pointwise_quantizer=None, depthwise_initializer="glorot_uniform", pointwise_initializer="glorot_uniform", bias_initializer="zeros", depthwise_regularizer=None, pointwise_regularizer=None, bias_regularizer=None, activity_regularizer=None, depthwise_constraint=None, pointwise_constraint=None, bias_constraint=None, **kwargs, ): super().__init__( filters, kernel_size, strides=strides, padding=padding, pad_values=pad_values, data_format=data_format, dilation_rate=dilation_rate, depth_multiplier=depth_multiplier, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, depthwise_quantizer=depthwise_quantizer, pointwise_quantizer=pointwise_quantizer, depthwise_initializer=depthwise_initializer, pointwise_initializer=pointwise_initializer, bias_initializer=bias_initializer, depthwise_regularizer=depthwise_regularizer, pointwise_regularizer=pointwise_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, depthwise_constraint=depthwise_constraint, pointwise_constraint=pointwise_constraint, bias_constraint=bias_constraint, **kwargs, ) @utils.register_keras_custom_object class QuantSeparableConv2D( QuantizerSeparableBase, QuantizerBaseConv, tf.keras.layers.SeparableConv2D ): """Depthwise separable 2D convolution. Separable convolutions consist in first performing a depthwise spatial convolution (which acts on each input channel separately) followed by a pointwise convolution which mixes together the resulting output channels. The `depth_multiplier` argument controls how many output channels are generated per input channel in the depthwise step. `input_quantizer`, `depthwise_quantizer` and `pointwise_quantizer` are the element-wise quantization functions to use. If all quantization functions are `None` this layer is equivalent to `SeparableConv1D`. If `use_bias` is True and a bias initializer is provided, it adds a bias vector to the output. It then optionally applies an activation function to produce the final output. Intuitively, separable convolutions can be understood as a way to factorize a convolution kernel into two smaller kernels, or as an extreme version of an Inception block. # Arguments filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of 2 integers, specifying the height and width of the 2D convolution window. Can be a single integer to specify the same value for all spatial dimensions. strides: An integer or tuple/list of 2 integers, specifying the strides of the convolution along the height and width. Can be a single integer to specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). pad_values: The pad value to use when `padding="same"`. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch, height, width, channels)` while `channels_first` corresponds to inputs with shape `(batch, channels, height, width)`. It defaults to the `image_data_format` value found in your Keras config file at `~/.keras/keras.json`. If you never set it, then it will be "channels_last". dilation_rate: an integer or tuple/list of 2 integers, specifying the dilation rate to use for dilated convolution. Can be a single integer to specify the same value for all spatial dimensions. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any stride value != 1. depth_multiplier: The number of depthwise convolution output channels for each input channel. The total number of depthwise convolution output channels will be equal to `filters_in * depth_multiplier`. activation: Activation function to use. If you don't specify anything, no activation is applied (`a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. depthwise_quantizer: Quantization function applied to the depthwise kernel matrix. pointwise_quantizer: Quantization function applied to the pointwise kernel matrix. depthwise_initializer: Initializer for the depthwise kernel matrix. pointwise_initializer: Initializer for the pointwise kernel matrix. bias_initializer: Initializer for the bias vector. depthwise_regularizer: Regularizer function applied to the depthwise kernel matrix. pointwise_regularizer: Regularizer function applied to the pointwise kernel matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). depthwise_constraint: Constraint function applied to the depthwise kernel matrix. pointwise_constraint: Constraint function applied to the pointwise kernel matrix. bias_constraint: Constraint function applied to the bias vector.` # Input shape 4D tensor with shape: `(batch, channels, rows, cols)` if data_format='channels_first' or 4D tensor with shape: `(batch, rows, cols, channels)` if data_format='channels_last'. # Output shape 4D tensor with shape: `(batch, filters, new_rows, new_cols)` if data_format='channels_first' or 4D tensor with shape: `(batch, new_rows, new_cols, filters)` if data_format='channels_last'. `rows` and `cols` values might have changed due to padding. """ def __init__( self, filters, kernel_size, strides=(1, 1), padding="valid", pad_values=0.0, data_format=None, dilation_rate=(1, 1), depth_multiplier=1, activation=None, use_bias=True, input_quantizer=None, depthwise_quantizer=None, pointwise_quantizer=None, depthwise_initializer="glorot_uniform", pointwise_initializer="glorot_uniform", bias_initializer="zeros", depthwise_regularizer=None, pointwise_regularizer=None, bias_regularizer=None, activity_regularizer=None, depthwise_constraint=None, pointwise_constraint=None, bias_constraint=None, **kwargs, ): super().__init__( filters, kernel_size, strides=strides, padding=padding, pad_values=pad_values, data_format=data_format, dilation_rate=dilation_rate, depth_multiplier=depth_multiplier, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, depthwise_quantizer=depthwise_quantizer, pointwise_quantizer=pointwise_quantizer, depthwise_initializer=depthwise_initializer, pointwise_initializer=pointwise_initializer, bias_initializer=bias_initializer, depthwise_regularizer=depthwise_regularizer, pointwise_regularizer=pointwise_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, depthwise_constraint=depthwise_constraint, pointwise_constraint=pointwise_constraint, bias_constraint=bias_constraint, **kwargs, ) @utils.register_keras_custom_object class QuantConv2DTranspose(QuantizerBase, tf.keras.layers.Conv2DTranspose): """Transposed quantized convolution layer (sometimes called Deconvolution). The need for transposed convolutions generally arises from the desire to use a transformation going in the opposite direction of a normal convolution, i.e., from something that has the shape of the output of some convolution to something that has the shape of its input while maintaining a connectivity pattern that is compatible with said convolution. `input_quantizer` and `kernel_quantizer` are the element-wise quantization functions to use. If both quantization functions are `None` this layer is equivalent to `Conv2DTranspose`. When using this layer as the first layer in a model, provide the keyword argument `input_shape` (tuple of integers, does not include the sample axis), e.g. `input_shape=(128, 128, 3)` for 128x128 RGB pictures in `data_format="channels_last"`. # Arguments filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of 2 integers, specifying the height and width of the 2D convolution window. Can be a single integer to specify the same value for all spatial dimensions. strides: An integer or tuple/list of 2 integers, specifying the strides of the convolution along the height and width. Can be a single integer to specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). output_padding: An integer or tuple/list of 2 integers, specifying the amount of padding along the height and width of the output tensor. Can be a single integer to specify the same value for all spatial dimensions. The amount of output padding along a given dimension must be lower than the stride along that same dimension. If set to `None` (default), the output shape is inferred. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch, height, width, channels)` while `channels_first` corresponds to inputs with shape `(batch, channels, height, width)`. It defaults to the `image_data_format` value found in your Keras config file at `~/.keras/keras.json`. If you never set it, then it will be "channels_last". dilation_rate: an integer or tuple/list of 2 integers, specifying the dilation rate to use for dilated convolution. Can be a single integer to specify the same value for all spatial dimensions. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any stride value != 1. activation: Activation function to use. If you don't specify anything, no activation is applied (`a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. kernel_quantizer: Quantization function applied to the `kernel` weights matrix. kernel_initializer: Initializer for the `kernel` weights matrix. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer function applied to the `kernel` weights matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). kernel_constraint: Constraint function applied to the kernel matrix. bias_constraint: Constraint function applied to the bias vector. # Input shape 4D tensor with shape: `(batch, channels, rows, cols)` if data_format='channels_first' or 4D tensor with shape: `(batch, rows, cols, channels)` if data_format='channels_last'. # Output shape 4D tensor with shape: `(batch, filters, new_rows, new_cols)` if data_format='channels_first' or 4D tensor with shape: `(batch, new_rows, new_cols, filters)` if data_format='channels_last'. `rows` and `cols` values might have changed due to padding. # References - [A guide to convolution arithmetic for deep learning](https://arxiv.org/abs/1603.07285v1) - [Deconvolutional Networks](https://www.matthewzeiler.com/mattzeiler/deconvolutionalnetworks.pdf) """ def __init__( self, filters, kernel_size, strides=(1, 1), padding="valid", output_padding=None, data_format=None, dilation_rate=(1, 1), activation=None, use_bias=True, input_quantizer=None, kernel_quantizer=None, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs, ): super().__init__( filters, kernel_size, strides=strides, padding=padding, output_padding=output_padding, data_format=data_format, dilation_rate=dilation_rate, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, kernel_quantizer=kernel_quantizer, kernel_initializer=kernel_initializer, bias_initializer=bias_initializer, kernel_regularizer=kernel_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, kernel_constraint=kernel_constraint, bias_constraint=bias_constraint, **kwargs, ) @utils.register_keras_custom_object class QuantConv3DTranspose(QuantizerBase, tf.keras.layers.Conv3DTranspose): """Transposed quantized convolution layer (sometimes called Deconvolution). The need for transposed convolutions generally arises from the desire to use a transformation going in the opposite direction of a normal convolution, i.e., from something that has the shape of the output of some convolution to something that has the shape of its input while maintaining a connectivity pattern that is compatible with said convolution. `input_quantizer` and `kernel_quantizer` are the element-wise quantization functions to use. If both quantization functions are `None` this layer is equivalent to `Conv3DTranspose`. When using this layer as the first layer in a model, provide the keyword argument `input_shape` (tuple of integers, does not include the sample axis), e.g. `input_shape=(128, 128, 128, 3)` for a 128x128x128 volume with 3 channels if `data_format="channels_last"`. # Arguments filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of 3 integers, specifying the depth, height and width of the 3D convolution window. Can be a single integer to specify the same value for all spatial dimensions. strides: An integer or tuple/list of 3 integers, specifying the strides of the convolution along the depth, height and width. Can be a single integer to specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). output_padding: An integer or tuple/list of 3 integers, specifying the amount of padding along the depth, height, and width. Can be a single integer to specify the same value for all spatial dimensions. The amount of output padding along a given dimension must be lower than the stride along that same dimension. If set to `None` (default), the output shape is inferred. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch, depth, height, width, channels)` while `channels_first` corresponds to inputs with shape `(batch, channels, depth, height, width)`. It defaults to the `image_data_format` value found in your Keras config file at `~/.keras/keras.json`. If you never set it, then it will be "channels_last". dilation_rate: an integer or tuple/list of 2 integers, specifying the dilation rate to use for dilated convolution. Can be a single integer to specify the same value for all spatial dimensions. Currently, specifying any `dilation_rate` value != 1 is incompatible with specifying any stride value != 1. activation: Activation function to use. If you don't specify anything, no activation is applied (`a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. kernel_quantizer: Quantization function applied to the `kernel` weights matrix. kernel_initializer: Initializer for the `kernel` weights matrix. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer function applied to the `kernel` weights matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). kernel_constraint: Constraint function applied to the kernel matrix. bias_constraint: Constraint function applied to the bias vector. # Input shape 5D tensor with shape: `(batch, channels, depth, rows, cols)` if data_format='channels_first' or 5D tensor with shape: `(batch, depth, rows, cols, channels)` if data_format='channels_last'. # Output shape 5D tensor with shape: `(batch, filters, new_depth, new_rows, new_cols)` if data_format='channels_first' or 5D tensor with shape: `(batch, new_depth, new_rows, new_cols, filters)` if data_format='channels_last'. `depth` and `rows` and `cols` values might have changed due to padding. # References - [A guide to convolution arithmetic for deep learning](https://arxiv.org/abs/1603.07285v1) - [Deconvolutional Networks](https://www.matthewzeiler.com/mattzeiler/deconvolutionalnetworks.pdf) """ def __init__( self, filters, kernel_size, strides=(1, 1, 1), padding="valid", output_padding=None, data_format=None, dilation_rate=(1, 1, 1), activation=None, use_bias=True, input_quantizer=None, kernel_quantizer=None, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs, ): super().__init__( filters, kernel_size, strides=strides, padding=padding, output_padding=output_padding, data_format=data_format, dilation_rate=dilation_rate, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, kernel_quantizer=kernel_quantizer, kernel_initializer=kernel_initializer, bias_initializer=bias_initializer, kernel_regularizer=kernel_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, kernel_constraint=kernel_constraint, bias_constraint=bias_constraint, **kwargs, ) @utils.register_keras_custom_object class QuantLocallyConnected1D(QuantizerBase, tf.keras.layers.LocallyConnected1D): """Locally-connected quantized layer for 1D inputs. The `QuantLocallyConnected1D` layer works similarly to the `QuantConv1D` layer, except that weights are unshared, that is, a different set of filters is applied at each different patch of the input. `input_quantizer` and `kernel_quantizer` are the element-wise quantization functions to use. If both quantization functions are `None` this layer is equivalent to `LocallyConnected1D`. !!! example ```python # apply a unshared weight convolution 1d of length 3 to a sequence with # 10 timesteps, with 64 output filters model = Sequential() model.add(QuantLocallyConnected1D(64, 3, input_shape=(10, 32))) # now model.output_shape == (None, 8, 64) # add a new conv1d on top model.add(QuantLocallyConnected1D(32, 3)) # now model.output_shape == (None, 6, 32) ``` # Arguments filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of a single integer, specifying the length of the 1D convolution window. strides: An integer or tuple/list of a single integer, specifying the stride length of the convolution. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: Currently only supports `"valid"` (case-insensitive). `"same"` may be supported in the future. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch, length, channels)` while `channels_first` corresponds to inputs with shape `(batch, channels, length)`. It defaults to the `image_data_format` value found in your Keras config file at `~/.keras/keras.json`. If you never set it, then it will be "channels_last". activation: Activation function to use. If you don't specify anything, no activation is applied (`a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. kernel_quantizer: Quantization function applied to the `kernel` weights matrix. kernel_initializer: Initializer for the `kernel` weights matrix. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer function applied to the `kernel` weights matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). kernel_constraint: Constraint function applied to the kernel matrix. bias_constraint: Constraint function applied to the bias vector. implementation: implementation mode, either `1` or `2`. `1` loops over input spatial locations to perform the forward pass. It is memory-efficient but performs a lot of (small) ops. `2` stores layer weights in a dense but sparsely-populated 2D matrix and implements the forward pass as a single matrix-multiply. It uses a lot of RAM but performs few (large) ops. Depending on the inputs, layer parameters, hardware, and `tf.executing_eagerly()` one implementation can be dramatically faster (e.g. 50X) than another. It is recommended to benchmark both in the setting of interest to pick the most efficient one (in terms of speed and memory usage). Following scenarios could benefit from setting `implementation=2`: - eager execution; - inference; - running on CPU; - large amount of RAM available; - small models (few filters, small kernel); - using `padding=same` (only possible with `implementation=2`). # Input shape 3D tensor with shape: `(batch_size, steps, input_dim)` # Output shape 3D tensor with shape: `(batch_size, new_steps, filters)` `steps` value might have changed due to padding or strides. """ def __init__( self, filters, kernel_size, strides=1, padding="valid", data_format=None, activation=None, use_bias=True, input_quantizer=None, kernel_quantizer=None, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, implementation=1, **kwargs, ): super().__init__( filters, kernel_size, strides=strides, padding=padding, data_format=data_format, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, kernel_quantizer=kernel_quantizer, kernel_initializer=kernel_initializer, bias_initializer=bias_initializer, kernel_regularizer=kernel_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, kernel_constraint=kernel_constraint, bias_constraint=bias_constraint, implementation=implementation, **kwargs, ) @utils.register_keras_custom_object class QuantLocallyConnected2D(QuantizerBase, tf.keras.layers.LocallyConnected2D): """Locally-connected quantized layer for 2D inputs. The `QuantLocallyConnected2D` layer works similarly to the `QuantConv2D` layer, except that weights are unshared, that is, a different set of filters is applied at each different patch of the input. `input_quantizer` and `kernel_quantizer` are the element-wise quantization functions to use. If both quantization functions are `None` this layer is equivalent to `LocallyConnected2D`. !!! example ```python # apply a 3x3 unshared weights convolution with 64 output filters on a 32x32 image # with `data_format="channels_last"`: model = Sequential() model.add(QuantLocallyConnected2D(64, (3, 3), input_shape=(32, 32, 3))) # now model.output_shape == (None, 30, 30, 64) # notice that this layer will consume (30*30)*(3*3*3*64) + (30*30)*64 parameters # add a 3x3 unshared weights convolution on top, with 32 output filters: model.add(QuantLocallyConnected2D(32, (3, 3))) # now model.output_shape == (None, 28, 28, 32) ``` # Arguments filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of 2 integers, specifying the width and height of the 2D convolution window. Can be a single integer to specify the same value for all spatial dimensions. strides: An integer or tuple/list of 2 integers, specifying the strides of the convolution along the width and height. Can be a single integer to specify the same value for all spatial dimensions. padding: Currently only support `"valid"` (case-insensitive). `"same"` will be supported in future. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch, height, width, channels)` while `channels_first` corresponds to inputs with shape `(batch, channels, height, width)`. It defaults to the `image_data_format` value found in your Keras config file at `~/.keras/keras.json`. If you never set it, then it will be "channels_last". activation: Activation function to use. If you don't specify anything, no activation is applied (`a(x) = x`). use_bias: Boolean, whether the layer uses a bias vector. input_quantizer: Quantization function applied to the input of the layer. kernel_quantizer: Quantization function applied to the `kernel` weights matrix. kernel_initializer: Initializer for the `kernel` weights matrix. bias_initializer: Initializer for the bias vector. kernel_regularizer: Regularizer function applied to the `kernel` weights matrix. bias_regularizer: Regularizer function applied to the bias vector. activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). kernel_constraint: Constraint function applied to the kernel matrix. bias_constraint: Constraint function applied to the bias vector. implementation: implementation mode, either `1` or `2`. `1` loops over input spatial locations to perform the forward pass. It is memory-efficient but performs a lot of (small) ops. `2` stores layer weights in a dense but sparsely-populated 2D matrix and implements the forward pass as a single matrix-multiply. It uses a lot of RAM but performs few (large) ops. Depending on the inputs, layer parameters, hardware, and `tf.executing_eagerly()` one implementation can be dramatically faster (e.g. 50X) than another. It is recommended to benchmark both in the setting of interest to pick the most efficient one (in terms of speed and memory usage). Following scenarios could benefit from setting `implementation=2`: - eager execution; - inference; - running on CPU; - large amount of RAM available; - small models (few filters, small kernel); - using `padding=same` (only possible with `implementation=2`). # Input shape 4D tensor with shape: `(samples, channels, rows, cols)` if data_format='channels_first' or 4D tensor with shape: `(samples, rows, cols, channels)` if data_format='channels_last'. # Output shape 4D tensor with shape: `(samples, filters, new_rows, new_cols)` if data_format='channels_first' or 4D tensor with shape: `(samples, new_rows, new_cols, filters)` if data_format='channels_last'. `rows` and `cols` values might have changed due to padding. """ def __init__( self, filters, kernel_size, strides=(1, 1), padding="valid", data_format=None, activation=None, use_bias=True, input_quantizer=None, kernel_quantizer=None, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, implementation=1, **kwargs, ): super().__init__( filters, kernel_size, strides=strides, padding=padding, data_format=data_format, activation=activation, use_bias=use_bias, input_quantizer=input_quantizer, kernel_quantizer=kernel_quantizer, kernel_initializer=kernel_initializer, bias_initializer=bias_initializer, kernel_regularizer=kernel_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, kernel_constraint=kernel_constraint, bias_constraint=bias_constraint, implementation=implementation, **kwargs, )
65,598
46.535507
91
py
larq
larq-main/larq/callbacks_test.py
import math import numpy as np import pytest import tensorflow as tf from packaging import version from tensorflow.python.keras import testing_utils import larq as lq from larq import testing_utils as lq_testing_utils from larq.callbacks import HyperparameterScheduler if version.parse(tf.__version__) >= version.parse("2.11"): from tensorflow.keras.optimizers import legacy as optimizers # type: ignore else: from tensorflow.keras import optimizers # type: ignore class TestHyperparameterScheduler: def _create_data_and_model(self, train_samples=1000): np.random.seed(1337) (x_train, y_train), (x_test, y_test) = testing_utils.get_test_data( train_samples=train_samples, test_samples=0, input_shape=(10,), num_classes=2, ) y_train = tf.keras.utils.to_categorical(y_train) model = lq_testing_utils.get_small_bnn_model( x_train.shape[1], 20, y_train.shape[1] ) return x_train, y_train, model def test_normal_optimizer(self): x_train, y_train, model = self._create_data_and_model() model.compile( loss="categorical_crossentropy", optimizer=tf.keras.optimizers.Adam(0.01), metrics=["accuracy"], ) def scheduler(x): return 1.0 / (1.0 + x) # We shouldn' t need to specify the optimizer test_scheduler = HyperparameterScheduler( schedule=scheduler, hyperparameter="lr", verbose=1, ) num_epochs = 2 model.fit( x_train, y_train, epochs=num_epochs, batch_size=16, callbacks=[test_scheduler], verbose=0, ) np.testing.assert_almost_equal( tf.keras.backend.get_value(model.optimizer.lr), scheduler(num_epochs - 1), decimal=8, ) def test_per_step(self): train_samples = 20 x_train, y_train, model = self._create_data_and_model(train_samples) model.compile( loss="categorical_crossentropy", optimizer=tf.keras.optimizers.Adam(0.01), metrics=["accuracy"], ) def scheduler(x): return 1.0 / (1.0 + x) # Test that we don't accept incorrect `update_freq` with pytest.raises(ValueError): HyperparameterScheduler( schedule=scheduler, hyperparameter="lr", update_freq="wrong", ) # The actual scheduler we'll use test_scheduler = HyperparameterScheduler( schedule=scheduler, hyperparameter="lr", update_freq="step", verbose=1, ) num_epochs = 1 batch_size = 10 model.fit( x_train, y_train, epochs=num_epochs, batch_size=16, callbacks=[test_scheduler], verbose=0, ) np.testing.assert_almost_equal( tf.keras.backend.get_value(model.optimizer.lr), scheduler(math.ceil(train_samples / batch_size) - 1), decimal=8, ) def test_case_optimizer(self): x_train, y_train, model = self._create_data_and_model() bop = lq.optimizers.Bop(threshold=1e-6, gamma=1e-3) adam = optimizers.Adam(0.01) case_optimizer = lq.optimizers.CaseOptimizer( (lq.optimizers.Bop.is_binary_variable, bop), default_optimizer=adam, ) model.compile( loss="categorical_crossentropy", optimizer=case_optimizer, metrics=["accuracy"], ) def scheduler(x): return 1.0 / (1.0 + x) cbk_gamma_scheduler = HyperparameterScheduler( schedule=scheduler, optimizer=model.optimizer.optimizers[0], hyperparameter="gamma", verbose=1, ) cbk_threshold_scheduler = HyperparameterScheduler( schedule=scheduler, optimizer=model.optimizer.optimizers[0], hyperparameter="threshold", verbose=1, ) cbk_lr_scheduler = HyperparameterScheduler( schedule=scheduler, optimizer=model.optimizer.optimizers[1], hyperparameter="lr", verbose=1, ) num_epochs = 3 model.fit( x_train, y_train, epochs=num_epochs, batch_size=16, callbacks=[cbk_gamma_scheduler, cbk_lr_scheduler, cbk_threshold_scheduler], verbose=0, ) np.testing.assert_almost_equal( tf.keras.backend.get_value(model.optimizer.optimizers[0].gamma), scheduler(num_epochs - 1), decimal=8, ) np.testing.assert_almost_equal( tf.keras.backend.get_value(model.optimizer.optimizers[0].threshold), scheduler(num_epochs - 1), decimal=8, ) np.testing.assert_almost_equal( tf.keras.backend.get_value(model.optimizer.optimizers[1].lr), scheduler(num_epochs - 1), decimal=8, ) def test_wrong_param(self): x_train, y_train, model = self._create_data_and_model() model.compile( loss="categorical_crossentropy", optimizer=tf.keras.optimizers.Adam(0.01), metrics=["accuracy"], ) def scheduler(x): return 1.0 / (1.0 + x) wrong_scheduler = HyperparameterScheduler( schedule=scheduler, hyperparameter="invalid_param", verbose=1, ) with pytest.raises(ValueError): model.fit( x_train, y_train, epochs=1, batch_size=16, callbacks=[wrong_scheduler], verbose=0, )
6,015
27.647619
87
py
larq
larq-main/larq/constraints_test.py
import numpy as np import pytest import tensorflow as tf import larq as lq from larq.testing_utils import generate_real_values_with_zeros @pytest.mark.parametrize("name", ["weight_clip"]) def test_serialization(name): fn = tf.keras.constraints.get(name) ref_fn = getattr(lq.constraints, name)() assert fn.__class__ == ref_fn.__class__ config = tf.keras.constraints.serialize(fn) fn = tf.keras.constraints.deserialize(config) assert fn.__class__ == ref_fn.__class__ def test_clip(): real_values = generate_real_values_with_zeros() clip_instance = lq.constraints.weight_clip(clip_value=0.75) result = clip_instance(tf.keras.backend.variable(real_values)) result = tf.keras.backend.eval(result) np.testing.assert_allclose(result, np.clip(real_values, -0.75, 0.75))
811
31.48
73
py
larq
larq-main/larq/layers_test.py
import inspect import numpy as np import pytest import tensorflow as tf from packaging import version import larq as lq from larq import testing_utils PARAMS_ALL_LAYERS = [ (lq.layers.QuantDense, tf.keras.layers.Dense, (3, 2), dict(units=3)), ( lq.layers.QuantConv1D, tf.keras.layers.Conv1D, (2, 3, 7), dict(filters=2, kernel_size=3), ), ( lq.layers.QuantConv2D, tf.keras.layers.Conv2D, (2, 3, 7, 6), dict(filters=2, kernel_size=3), ), ( lq.layers.QuantConv3D, tf.keras.layers.Conv3D, (2, 3, 7, 6, 5), dict(filters=2, kernel_size=3), ), ( lq.layers.QuantConv2DTranspose, tf.keras.layers.Conv2DTranspose, (2, 3, 7, 6), dict(filters=2, kernel_size=3), ), ( lq.layers.QuantConv3DTranspose, tf.keras.layers.Conv3DTranspose, (2, 3, 7, 6, 5), dict(filters=2, kernel_size=3), ), ( lq.layers.QuantLocallyConnected1D, tf.keras.layers.LocallyConnected1D, (2, 8, 5), dict(filters=4, kernel_size=3), ), ( lq.layers.QuantLocallyConnected2D, tf.keras.layers.LocallyConnected2D, (8, 6, 10, 4), dict(filters=3, kernel_size=3), ), ] PARAMS_SEP_LAYERS = [ (lq.layers.QuantSeparableConv1D, tf.keras.layers.SeparableConv1D, (2, 3, 7)), (lq.layers.QuantSeparableConv2D, tf.keras.layers.SeparableConv2D, (2, 3, 7, 6)), ] class TestLayers: @pytest.mark.parametrize( "quantized_layer, layer, input_shape, kwargs", PARAMS_ALL_LAYERS ) def test_binarization( self, quantized_layer, layer, input_shape, kwargs, keras_should_run_eagerly ): input_data = testing_utils.random_input(input_shape) random_weight = np.random.random() - 0.5 with lq.context.metrics_scope(["flip_ratio"]): quant_output = testing_utils.layer_test( quantized_layer, kwargs=dict( **kwargs, kernel_quantizer="ste_sign", input_quantizer="ste_sign", kernel_initializer=tf.keras.initializers.constant(random_weight), ), input_data=input_data, should_run_eagerly=keras_should_run_eagerly, ) fp_model = tf.keras.models.Sequential( [ layer( **kwargs, kernel_initializer=tf.keras.initializers.constant( np.sign(random_weight) ), input_shape=input_shape[1:], ) ] ) np.testing.assert_allclose(quant_output, fp_model.predict(np.sign(input_data))) @pytest.mark.parametrize("quantized_layer, layer, input_shape", PARAMS_SEP_LAYERS) def test_separable_layers( self, quantized_layer, layer, input_shape, keras_should_run_eagerly ): input_data = testing_utils.random_input(input_shape) random_d_kernel = np.random.random() - 0.5 random_p_kernel = np.random.random() - 0.5 with lq.context.metrics_scope(["flip_ratio"]): quant_output = testing_utils.layer_test( quantized_layer, kwargs=dict( filters=3, kernel_size=3, depthwise_quantizer="ste_sign", pointwise_quantizer="ste_sign", input_quantizer="ste_sign", depthwise_initializer=tf.keras.initializers.constant( random_d_kernel ), pointwise_initializer=tf.keras.initializers.constant( random_p_kernel ), ), input_data=input_data, should_run_eagerly=keras_should_run_eagerly, ) fp_model = tf.keras.models.Sequential( [ layer( filters=3, kernel_size=3, depthwise_initializer=tf.keras.initializers.constant( np.sign(random_d_kernel) ), pointwise_initializer=tf.keras.initializers.constant( np.sign(random_p_kernel) ), input_shape=input_shape[1:], ) ] ) np.testing.assert_allclose(quant_output, fp_model.predict(np.sign(input_data))) def test_depthwise_layers(self, keras_should_run_eagerly): input_data = testing_utils.random_input((2, 3, 7, 6)) random_weight = np.random.random() - 0.5 with lq.context.metrics_scope(["flip_ratio"]): quant_output = testing_utils.layer_test( lq.layers.QuantDepthwiseConv2D, kwargs=dict( kernel_size=3, depthwise_quantizer="ste_sign", input_quantizer="ste_sign", depthwise_initializer=tf.keras.initializers.constant(random_weight), ), input_data=input_data, should_run_eagerly=keras_should_run_eagerly, ) fp_model = tf.keras.models.Sequential( [ tf.keras.layers.DepthwiseConv2D( kernel_size=3, depthwise_initializer=tf.keras.initializers.constant( np.sign(random_weight) ), input_shape=input_data.shape[1:], ) ] ) np.testing.assert_allclose(quant_output, fp_model.predict(np.sign(input_data))) @pytest.mark.parametrize( "layer_cls, input_dim", [ (lq.layers.QuantConv1D, 3), (lq.layers.QuantConv2D, 4), (lq.layers.QuantConv3D, 5), (lq.layers.QuantSeparableConv1D, 3), (lq.layers.QuantSeparableConv2D, 4), (lq.layers.QuantDepthwiseConv2D, 4), ], ) @pytest.mark.parametrize("dilation", [True, False]) def test_non_zero_padding_layers( self, mocker, layer_cls, input_dim, data_format, dilation ): inputs = np.zeros(np.random.randint(5, 20, size=input_dim), np.float32) kernel = tuple(np.random.randint(3, 7, size=input_dim - 2)) rand_tuple = tuple(np.random.randint(1, 4, size=input_dim - 2)) if not dilation and layer_cls in ( lq.layers.QuantSeparableConv2D, lq.layers.QuantDepthwiseConv2D, ): rand_tuple = int(rand_tuple[0]) kwargs = {"dilation_rate": rand_tuple} if dilation else {"strides": rand_tuple} args = (kernel,) if layer_cls == lq.layers.QuantDepthwiseConv2D else (2, kernel) ref_layer = layer_cls(*args, padding="same", **kwargs) spy = mocker.spy(tf, "pad") layer = layer_cls(*args, padding="same", pad_values=1.0, **kwargs) layer.build(inputs.shape) conv_op = getattr(layer, "_convolution_op", None) assert layer(inputs).shape == ref_layer(inputs).shape spy.assert_called_once_with(mocker.ANY, mocker.ANY, constant_values=1.0) assert conv_op == getattr(layer, "_convolution_op", None) @pytest.mark.parametrize( "layer_cls", [ lq.layers.QuantConv1D, lq.layers.QuantConv2D, lq.layers.QuantConv3D, lq.layers.QuantSeparableConv1D, lq.layers.QuantSeparableConv2D, lq.layers.QuantDepthwiseConv2D, ], ) @pytest.mark.parametrize("static", [True, False]) def test_non_zero_padding_shapes(self, layer_cls, data_format, static): layer = layer_cls( 16, 3, padding="same", pad_values=1.0, data_format=data_format ) input_shape = [32 if static else None] * layer.rank + [3] if data_format == "channels_first": input_shape = reversed(input_shape) input = tf.keras.layers.Input(shape=input_shape) layer(input) if static: for dim in layer.output_shape[1:]: assert dim is not None class TestLayerWarns: def test_layer_warns(self, caplog): lq.layers.QuantDense(5, kernel_quantizer="ste_sign") assert len(caplog.records) >= 1 assert "kernel_constraint" in caplog.text def test_layer_does_not_warn(self, caplog): lq.layers.QuantDense( 5, kernel_quantizer="ste_sign", kernel_constraint="weight_clip" ) assert "kernel_constraint" not in caplog.text def test_depthwise_layer_warns(self, caplog): lq.layers.QuantDepthwiseConv2D(5, depthwise_quantizer="ste_sign") assert len(caplog.records) >= 1 assert "depthwise_constraint" in caplog.text def test_depthwise_layer_does_not_warn(self, caplog): lq.layers.QuantDepthwiseConv2D( 5, depthwise_quantizer="ste_sign", depthwise_constraint="weight_clip" ) assert "depthwise_constraint" not in caplog.text def test_separable_layer_warns(self, caplog): lq.layers.QuantSeparableConv2D( 3, 3, depthwise_quantizer="ste_sign", pointwise_quantizer="ste_sign" ) assert "depthwise_constraint" in caplog.text assert "pointwise_constraint" in caplog.text def test_separable_layer_does_not_warn(self, caplog): lq.layers.QuantSeparableConv2D( 3, 3, depthwise_quantizer="ste_sign", pointwise_quantizer="ste_sign", depthwise_constraint="weight_clip", pointwise_constraint="weight_clip", ) assert caplog.records == [] def test_conv1d_non_zero_padding_raises(self): with pytest.raises(ValueError, match=r".*pad_values.*"): lq.layers.QuantConv1D(24, 3, padding="causal", pad_values=1.0) @pytest.mark.parametrize( "layer", [lq.layers.QuantConv1D, lq.layers.QuantConv2D, lq.layers.QuantConv3D] ) def test_groups(self, layer): if version.parse(tf.__version__) < version.parse("2.3"): with pytest.raises(ValueError, match=r".*groups.*"): layer(24, 3, groups=2) else: assert layer(24, 3, groups=2).groups == 2 @pytest.mark.parametrize( "quant_layer,layer", [ (lq.layers.QuantDense, tf.keras.layers.Dense), (lq.layers.QuantConv1D, tf.keras.layers.Conv1D), (lq.layers.QuantConv2D, tf.keras.layers.Conv2D), (lq.layers.QuantConv3D, tf.keras.layers.Conv3D), (lq.layers.QuantConv2DTranspose, tf.keras.layers.Conv2DTranspose), (lq.layers.QuantConv3DTranspose, tf.keras.layers.Conv3DTranspose), (lq.layers.QuantLocallyConnected1D, tf.keras.layers.LocallyConnected1D), (lq.layers.QuantLocallyConnected2D, tf.keras.layers.LocallyConnected2D), (lq.layers.QuantDepthwiseConv2D, tf.keras.layers.DepthwiseConv2D), ], ) def test_layer_kwargs(quant_layer, layer): quant_params = inspect.signature(quant_layer).parameters params = inspect.signature(layer).parameters quant_params_list = list(quant_params.keys()) params_list = list(params.keys()) ignored_params = [ "input_quantizer", "kernel_quantizer", "depthwise_quantizer", "pointwise_quantizer", "pad_values", ] if version.parse(tf.__version__) < version.parse("2.3"): ignored_params.append("groups") if layer in (tf.keras.layers.DepthwiseConv2D, tf.keras.layers.Conv3DTranspose): ignored_params.append("dilation_rate") for p in ignored_params: try: quant_params_list.remove(p) except ValueError: pass assert quant_params_list == params_list for param in params_list: assert quant_params.get(param).default == params.get(param).default # type: ignore
12,025
34.68546
91
py
larq
larq-main/larq/constraints.py
"""Functions from the `constraints` module allow setting constraints (eg. weight clipping) on network parameters during optimization. The penalties are applied on a per-layer basis. The exact API will depend on the layer, but the layers `QuantDense`, `QuantConv1D`, `QuantConv2D` and `QuantConv3D` have a unified API. These layers expose 2 keyword arguments: - `kernel_constraint` for the main weights matrix - `bias_constraint` for the bias. ```python import larq as lq lq.layers.QuantDense(64, kernel_constraint="weight_clip") lq.layers.QuantDense(64, kernel_constraint=lq.constraints.WeightClip(2.)) ``` """ from typing import Any, Mapping import tensorflow as tf from larq import utils @utils.register_keras_custom_object class WeightClip(tf.keras.constraints.Constraint): """Weight Clip constraint Constrains the weights incident to each hidden unit to be between `[-clip_value, clip_value]`. # Arguments clip_value: The value to clip incoming weights. """ def __init__(self, clip_value: float = 1): self.clip_value = clip_value def __call__(self, x: tf.Tensor) -> tf.Tensor: return tf.clip_by_value(x, -self.clip_value, self.clip_value) def get_config(self) -> Mapping[str, Any]: return {"clip_value": self.clip_value} # Aliases @utils.register_keras_custom_object class weight_clip(WeightClip): pass
1,392
25.283019
87
py
larq
larq-main/larq/models.py
import itertools from dataclasses import dataclass from typing import Any, Callable, Iterator, Mapping, Optional, Sequence, TypeVar, Union import numpy as np import tensorflow as tf from terminaltables import AsciiTable from larq import layers as lq_layers from larq.utils import memory_as_readable_str __all__ = ["summary"] mac_containing_layers = ( lq_layers.QuantConv2D, lq_layers.QuantSeparableConv2D, lq_layers.QuantDepthwiseConv2D, lq_layers.QuantDense, tf.keras.layers.Conv2D, tf.keras.layers.SeparableConv2D, tf.keras.layers.DepthwiseConv2D, tf.keras.layers.Dense, lq_layers.QuantConv1D, lq_layers.QuantSeparableConv1D, tf.keras.layers.Conv1D, tf.keras.layers.SeparableConv1D, ) op_count_supported_layer_types = ( tf.keras.layers.Flatten, tf.keras.layers.BatchNormalization, tf.keras.layers.MaxPool2D, tf.keras.layers.AveragePooling2D, tf.keras.layers.MaxPool1D, tf.keras.layers.AveragePooling1D, *mac_containing_layers, ) T = TypeVar("T") def _flatten(lst: Iterator[Iterator[T]]) -> Sequence[T]: return list(itertools.chain.from_iterable(lst)) def _bitsize_as_str(bitsize: int) -> str: bitsize_names = {8: "byte", 8 * 1024: "kB"} try: return bitsize_names[bitsize] except KeyError: raise NotImplementedError() def _number_as_readable_str(num: float) -> str: # The initial rounding here is necessary so that e.g. `999000` gets # formatted as `1.000 M` rather than `1000 k` num = float(f"{num:.3g}") # For numbers less than 1000, output them directly, stripping any trailing # zeros and decimal places. if num < 1000: return str(num).rstrip("0").rstrip(".") # For numbers that are at least 1000 trillion (1 quadrillion) format with # scientific notation (3 s.f. = 2 d.p. in scientific notation). if num >= 1e15: return f"{num:.2E}" # Count the magnitude. magnitude = 0 while abs(num) >= 1000 and magnitude < 4: magnitude += 1 num /= 1000.0 # ':.3g' formats the number with 3 significant figures, without stripping trailing # zeros. num = f"{num:.3g}".rstrip(".") unit = ["", " k", " M", " B", " T"][magnitude] return num + unit def _format_table_entry(x: float, units: int = 1) -> Union[float, str]: try: assert not np.isnan(x) if type(x) == str or x == 0 or units == 1: return x return x / units except Exception: return "?" def _normalize_shape(shape): return tuple(dim if dim else -1 for dim in shape) class WeightProfile: def __init__(self, weight, trainable: bool = True): self._weight = weight self.bitwidth = getattr(weight, "precision", 32) self.trainable = trainable @property def count(self) -> int: return int(np.prod(self._weight.shape.as_list())) @property def memory(self) -> int: return self.bitwidth * self.count @property def fp_equivalent_memory(self) -> int: return 32 * self.count @property def int8_fp_weights_memory(self) -> int: """Count any 32- or 16-bit weights as 8 bits instead.""" if self.bitwidth > 8: return self.count * 8 return self.bitwidth * self.count def is_bias(self) -> bool: return "bias" in self._weight.name @dataclass class OperationProfile: n: int precision: int op_type: str class LayerProfile: def __init__(self, layer: tf.keras.layers.Layer): self._layer = layer self.name = layer.name weights = layer.weights if isinstance(layer, tf.keras.layers.BatchNormalization): fused_pairs = [("beta", "moving_mean"), ("gamma", "moving_variance")] for pair in fused_pairs: names = [w.name.split("/")[-1].replace(":0", "") for w in weights] if pair[0] in names and pair[1] in names: weights.pop(names.index(pair[0])) self.weight_profiles = [ WeightProfile( weight, trainable=any(weight is w for w in layer.trainable_weights), ) for weight in weights ] self.op_profiles = [] if isinstance(layer, mac_containing_layers) and self.output_pixels: for p in self.weight_profiles: if not p.is_bias(): self.op_profiles.append( OperationProfile( n=p.count * self.output_pixels, precision=max(self.input_precision or 32, p.bitwidth), op_type="mac", ) ) @property def memory(self) -> int: return sum(p.memory for p in self.weight_profiles) @property def int8_fp_weights_memory(self) -> int: return sum(p.int8_fp_weights_memory for p in self.weight_profiles) @property def fp_equivalent_memory(self) -> int: return sum(p.fp_equivalent_memory for p in self.weight_profiles) def weight_count( self, bitwidth: Optional[int] = None, trainable: Optional[bool] = None ) -> int: count = 0 for p in self.weight_profiles: if (bitwidth is None or p.bitwidth == bitwidth) and ( trainable is None or p.trainable == trainable ): count += p.count return count def op_count( self, op_type: Optional[str] = None, precision: Optional[int] = None ) -> Optional[int]: if op_type != "mac": raise ValueError("Currently only counting of MAC-operations is supported.") if ( isinstance(self._layer, op_count_supported_layer_types) and self.output_pixels ): count = 0 for op in self.op_profiles: if (precision is None or op.precision == precision) and ( op_type is None or op.op_type == op_type ): count += op.n return count return None @property def input_precision(self) -> Optional[int]: try: return self._layer.input_quantizer.precision except AttributeError: return None @property def output_shape(self) -> Optional[Sequence[int]]: try: output_shape = self._layer.output_shape if isinstance(output_shape, list): if len(output_shape) == 1: return _normalize_shape(output_shape[0]) return [_normalize_shape(shape) for shape in output_shape] return _normalize_shape(output_shape) except AttributeError: return None @property def output_shape_str(self) -> str: try: return str(self.output_shape or "multiple") except RuntimeError: return "?" @property def output_pixels(self) -> Optional[int]: """Number of pixels for a single feature map (1 for fully connected layers).""" if not self.output_shape: return None if len(self.output_shape) == 4: return int(np.prod(self.output_shape[1:3])) if len(self.output_shape) == 3: return self.output_shape[1] if len(self.output_shape) == 2: return 1 raise NotImplementedError() @property def unique_param_bidtwidths(self) -> Sequence[int]: return sorted(set([p.bitwidth for p in self.weight_profiles])) @property def unique_op_precisions(self) -> Sequence[int]: return sorted(set([op.precision for op in self.op_profiles])) def generate_table_row( self, table_config: Mapping[str, Any] ) -> Sequence[Union[str, float]]: row = [self.name, self.input_precision or "-", self.output_shape_str] for i in table_config["param_bidtwidths"]: n = self.weight_count(i) n = _format_table_entry(n, table_config["param_units"]) row.append(n) row.append(_format_table_entry(self.memory, table_config["memory_units"])) for i in table_config["mac_precisions"]: n = self.op_count("mac", i) n = _format_table_entry(n, table_config["mac_units"]) row.append(n) return row class ModelProfile(LayerProfile): def __init__(self, model: tf.keras.models.Model): self.name = model.name def get_profile(layer): return ( LayerProfile(layer) if not isinstance(layer, tf.keras.models.Model) else ModelProfile(layer) ) self.layer_profiles = [get_profile(layer) for layer in model.layers] @property def memory(self) -> int: return sum(lp.memory for lp in self.layer_profiles) @property def int8_fp_weights_memory(self) -> int: return sum(lp.int8_fp_weights_memory for lp in self.layer_profiles) @property def fp_equivalent_memory(self) -> int: return sum(lp.fp_equivalent_memory for lp in self.layer_profiles) def weight_count( self, bitwidth: Optional[int] = None, trainable: Optional[bool] = None ) -> int: return sum(lp.weight_count(bitwidth, trainable) for lp in self.layer_profiles) def op_count( self, op_type: Optional[str] = None, bitwidth: Optional[int] = None ) -> int: return sum(lp.op_count(op_type, bitwidth) or 0 for lp in self.layer_profiles) @property def unique_param_bidtwidths(self) -> Sequence[int]: return sorted( set(_flatten(lp.unique_param_bidtwidths for lp in self.layer_profiles)) ) @property def unique_op_precisions(self) -> Sequence[int]: return sorted( set(_flatten(lp.unique_op_precisions for lp in self.layer_profiles)) ) @property def input_precision(self) -> Optional[int]: return self.layer_profiles[0].input_precision @property def output_shape(self) -> Optional[Sequence[int]]: return self.layer_profiles[-1].output_shape def _generate_table_header(self, table_config: Mapping[str, Any]) -> Sequence[str]: return [ "Layer", "Input prec.\n(bit)", "Outputs", *( f"# {i}-bit\nx {table_config['param_units']}" for i in table_config["param_bidtwidths"] ), f"Memory\n({_bitsize_as_str(table_config['memory_units'])})", *(f"{i}-bit MACs" for i in table_config["mac_precisions"]), ] def _generate_table_total( self, table_config: Mapping[str, Any] ) -> Sequence[Union[float, str]]: row = ["Total", "", ""] for i in table_config["param_bidtwidths"]: row.append( _format_table_entry(self.weight_count(i), table_config["param_units"]) ) row.append(_format_table_entry(self.memory, table_config["memory_units"])) for i in table_config["mac_precisions"]: row.append( _format_table_entry(self.op_count("mac", i), table_config["mac_units"]) ) return row def generate_table( self, include_macs: bool = True ) -> Sequence[Sequence[Union[float, str]]]: table_config = { "param_bidtwidths": self.unique_param_bidtwidths, "mac_precisions": self.unique_op_precisions if include_macs else [], "param_units": 1, "memory_units": 8 * 1024, "mac_units": 1, } table = [] table.append(self._generate_table_header(table_config)) for lp in self.layer_profiles: table.append(lp.generate_table_row(table_config)) table.append(self._generate_table_total(table_config)) return table def generate_summary( self, include_macs: bool = True ) -> Sequence[Sequence[Union[str, float]]]: summary = [ ["Total params", _number_as_readable_str(self.weight_count())], [ "Trainable params", _number_as_readable_str(self.weight_count(trainable=True)), ], [ "Non-trainable params", _number_as_readable_str(self.weight_count(trainable=False)), ], ["Model size", memory_as_readable_str(self.memory)], [ "Model size (8-bit FP weights)", memory_as_readable_str(self.int8_fp_weights_memory), ], ["Float-32 Equivalent", memory_as_readable_str(self.fp_equivalent_memory)], [ "Compression Ratio of Memory", self.memory / max(1e-8, self.fp_equivalent_memory), ], ] if include_macs: binarization_ratio = self.op_count("mac", 1) / max( 1, self.op_count(op_type="mac") ) ternarization_ratio = self.op_count("mac", 2) / max( 1, self.op_count(op_type="mac") ) summary.append( [ "Number of MACs", _number_as_readable_str(self.op_count(op_type="mac")), ] ) if binarization_ratio > 0: summary.append( ["Ratio of MACs that are binarized", f"{binarization_ratio:.4f}"] ) if ternarization_ratio > 0: summary.append( ["Ratio of MACs that are ternarized", f"{ternarization_ratio:.4f}"] ) return summary def sanitize_table(table_data: Sequence[Sequence[Any]]) -> Sequence[Sequence[str]]: return [ [f"{v:.2f}" if type(v) == float else str(v) for v in row] for row in table_data ] class LayersTable(AsciiTable): def __init__(self, table_data, title=None): super().__init__(sanitize_table(table_data), title=title) self.inner_column_border = False self.justify_columns = { i: "left" if i == 0 else "right" for i in range(len(table_data[0])) } self.inner_footing_row_border = True self.inner_heading_row_border = True class SummaryTable(AsciiTable): def __init__(self, table_data, title=None): super().__init__(sanitize_table(table_data), title=title) self.inner_column_border = False self.inner_heading_row_border = False def summary( model: tf.keras.models.Model, print_fn: Optional[Callable[[str], Any]] = None, include_macs: bool = True, ) -> None: """Prints a string summary of the network. The summary includes the following information per layer: - input precision, - output dimension, - weight count (broken down by bidtwidth), - memory footprint in kilobytes (`8*1024` 1-bit weights = 1 kB), - number of multiply-accumulate (MAC) operations broken down by precision (*optional & expermental*). A single MAC operation contains both a multiplication and an addition. The precision of a MAC operation is defined as the maximum bitwidth of its inputs. Additionally, the following overall statistics for the model are supplied: - total number of weights, - total number of trainable weights, - total number of non-trainable weights, - model size, - model size (8-bit FP weights): memory footprint if FP weights were 8 bit, - float-32 equivalent size: memory footprint if all weights were 32 bit, - compression ratio achieved by quantizing weights, - total number of MAC operations, - ratio of MAC operations that is binarized and can be accelated with XNOR-gates. # Arguments model: model instance. print_fn: Print function to use. Defaults to `print`. You can set it to a custom function in order to capture the string summary. include_macs: whether or not to include the number of MAC-operations in the summary. # Raises ValueError: if called before the model is built. """ if not model.built: raise ValueError( "This model has not yet been built. Build the model first by calling " "`model.build()` or calling `model.fit()` with some data, or specify an " "`input_shape` argument in the first layer(s) for automatic build." ) if not print_fn: print_fn = print model_profile = ModelProfile(model) print_fn( LayersTable(model_profile.generate_table(), title=f"{model.name} stats").table ) print_fn( SummaryTable( model_profile.generate_summary(include_macs), title=f"{model.name} summary" ).table )
16,824
31.861328
105
py
larq
larq-main/larq/metrics.py
"""We add metrics specific to extremely quantized networks using a `larq.context.metrics_scope` rather than through the `metrics` parameter of `model.compile()`, where most common metrics reside. This is because, to calculate metrics like the `flip_ratio`, we need a layer's kernel or activation and not just the `y_true` and `y_pred` that Keras passes to metrics defined in the usual way. """ import numpy as np import tensorflow as tf from larq import utils @utils.register_alias("flip_ratio") @utils.register_keras_custom_object class FlipRatio(tf.keras.metrics.Metric): """Computes the mean ratio of changed values in a given tensor. !!! example ```python m = metrics.FlipRatio() m.update_state((1, 1)) # result: 0 m.update_state((2, 2)) # result: 1 m.update_state((1, 2)) # result: 0.75 print('Final result: ', m.result().numpy()) # Final result: 0.75 ``` # Arguments name: Name of the metric. values_dtype: Data type of the tensor for which to track changes. dtype: Data type of the moving mean. """ def __init__(self, values_dtype="int8", name="flip_ratio", dtype=None): super().__init__(name=name, dtype=dtype) self.built = False self.values_dtype = tf.as_dtype(values_dtype) def build(self, input_shape): self._previous_values = self.add_weight( "previous_values", shape=input_shape, dtype=self.values_dtype, initializer=tf.keras.initializers.zeros, aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA, ) self.total = self.add_weight( "total", initializer=tf.keras.initializers.zeros, aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA, ) self.count = self.add_weight( "count", initializer=tf.keras.initializers.zeros, aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA, ) self._size = tf.cast(np.prod(input_shape), self.dtype) self.built = True def update_state(self, values, sample_weight=None): values = tf.cast(values, self.values_dtype) if not self.built: with tf.name_scope(self.name), tf.init_scope(): self.build(values.shape) unchanged_values = tf.math.count_nonzero( tf.equal(self._previous_values, values) ) flip_ratio = 1 - ( tf.cast(unchanged_values, self.dtype) / tf.cast(self._size, self.dtype) ) update_total_op = self.total.assign_add(flip_ratio * tf.sign(self.count)) with tf.control_dependencies([update_total_op]): update_count_op = self.count.assign_add(1) with tf.control_dependencies([update_count_op]): return self._previous_values.assign(values) def result(self): return tf.compat.v1.div_no_nan(self.total, self.count - 1) def reset_state(self): tf.keras.backend.batch_set_value( [(v, 0) for v in self.variables if v is not self._previous_values] ) def reset_states(self): self.reset_state() # For backwards compatibility with < 2.5 def get_config(self): return {**super().get_config(), "values_dtype": self.values_dtype.name}
3,341
34.935484
86
py
larq
larq-main/larq/activations.py
"""Activations can either be used through an `Activation` layer, or through the `activation` argument supported by all forward layers: ```python import tensorflow as tf import larq as lq model.add(lq.layers.QuantDense(64)) model.add(tf.keras.layers.Activation('hard_tanh')) ``` This is equivalent to: ```python model.add(lq.layers.QuantDense(64, activation='hard_tanh')) ``` You can also pass an element-wise TensorFlow function as an activation: ```python model.add(lq.layers.QuantDense(64, activation=lq.activations.hard_tanh)) ``` """ import tensorflow as tf from larq import utils @utils.register_keras_custom_object def hard_tanh(x: tf.Tensor) -> tf.Tensor: """Hard tanh activation function. ```plot-activation activations.hard_tanh ``` # Arguments x: Input tensor. # Returns Hard tanh activation. """ return tf.clip_by_value(x, -1, 1) @utils.register_keras_custom_object def leaky_tanh(x: tf.Tensor, alpha: float = 0.2) -> tf.Tensor: r"""Leaky tanh activation function. Similar to hard tanh, but with non-zero slopes as in leaky ReLU. ```plot-activation activations.leaky_tanh ``` # Arguments x: Input tensor. alpha: Slope of the activation function outside of [-1, 1]. # Returns Leaky tanh activation. """ return ( tf.clip_by_value(x, -1, 1) + (tf.math.maximum(x, 1) - 1) * alpha + (tf.math.minimum(x, -1) + 1) * alpha )
1,481
21.119403
79
py
larq
larq-main/larq/quantizers_test.py
import functools import numpy as np import pytest import tensorflow as tf from packaging import version import larq as lq from larq import testing_utils class DummyTrainableQuantizer(tf.keras.layers.Layer): """Used to test whether we can set layers as quantizers without any throws.""" _custom_metrics = None def build(self, input_shape): self.dummy_weight = self.add_weight("dummy_weight", trainable=True) super().build(input_shape) def call(self, inputs): return self.dummy_weight * inputs class TestCommonFunctionality: """Test functionality common to all quantizers, like serialization and usage.""" @pytest.mark.parametrize("module", [lq.quantizers, tf.keras.activations]) @pytest.mark.parametrize( "name,ref_cls", [ ("ste_sign", lq.quantizers.SteSign), ("approx_sign", lq.quantizers.ApproxSign), ("ste_heaviside", lq.quantizers.SteHeaviside), ("magnitude_aware_sign", lq.quantizers.MagnitudeAwareSign), ("swish_sign", lq.quantizers.SwishSign), ("ste_tern", lq.quantizers.SteTern), ], ) def test_serialization(self, module, name, ref_cls): if module == tf.keras.activations and ( version.parse(tf.__version__) < version.parse("2.13") ): # New serialisation in Keras doesn't support using quantizers strings as activations fn = module.get(name) assert fn.__class__ == ref_cls fn = module.get(ref_cls()) assert fn.__class__ == ref_cls assert type(fn.precision) == int if module == tf.keras.activations and ( version.parse(tf.__version__) < version.parse("1.15") ): pytest.skip( "TensorFlow < 1.15 does not support Quantizer classes as activations" ) config = module.serialize(fn) fn = module.deserialize(config) assert fn.__class__ == ref_cls assert type(fn.precision) == int def test_noop_serialization(self): fn = lq.quantizers.get(lq.quantizers.NoOp(precision=1)) assert fn.__class__ == lq.quantizers.NoOp assert fn.precision == 1 config = lq.quantizers.serialize(fn) fn = lq.quantizers.deserialize(config) assert fn.__class__ == lq.quantizers.NoOp assert fn.precision == 1 def test_invalid_usage(self): with pytest.raises(ValueError): lq.quantizers.get(42) with pytest.raises(ValueError): lq.quantizers.get("unknown") with pytest.raises(ValueError): lq.quantizers.DoReFa(k_bit=2, mode="unknown") f = lq.quantizers.DoReFa(k_bit=2, mode="activations") f.mode = "unknown" with pytest.raises(ValueError): f.call([0.0]) @pytest.mark.parametrize("quantizer", ["input_quantizer", "kernel_quantizer"]) def test_layer_as_quantizer(self, quantizer, keras_should_run_eagerly): """Test whether a keras.layers.Layer can be used as quantizer.""" input_data = testing_utils.random_input((1, 10)) model = tf.keras.Sequential( [lq.layers.QuantDense(1, **{quantizer: DummyTrainableQuantizer()})] ) model.compile(optimizer="sgd", loss="mse", run_eagerly=keras_should_run_eagerly) model.fit(input_data, np.ones((1,)), epochs=1) assert any(["dummy_weight" in var.name for var in model.trainable_variables]) class TestQuantization: """Test binarization and ternarization.""" @pytest.mark.parametrize( "fn", [ "ste_sign", lq.quantizers.SteSign(), "approx_sign", lq.quantizers.ApproxSign(), "swish_sign", lq.quantizers.SwishSign(), ], ) def test_xnor_binarization(self, fn): x = tf.keras.backend.placeholder(ndim=2) f = tf.keras.backend.function([x], [lq.quantizers.get(fn)(x)]) binarized_values = np.random.choice([-1, 1], size=(2, 5)) result = f([binarized_values])[0] np.testing.assert_allclose(result, binarized_values) real_values = testing_utils.generate_real_values_with_zeros() result = f([real_values])[0] assert not np.any(result == 0) assert np.all(result[real_values < 0] == -1) assert np.all(result[real_values >= 0] == 1) zero_values = np.zeros((2, 5)) result = f([zero_values])[0] assert np.all(result == 1) @pytest.mark.parametrize("fn", ["ste_heaviside", lq.quantizers.SteHeaviside()]) def test_and_binarization(self, fn): x = tf.keras.backend.placeholder(ndim=2) f = tf.keras.backend.function([x], [lq.quantizers.get(fn)(x)]) binarized_values = np.random.choice([0, 1], size=(2, 5)) result = f([binarized_values])[0] np.testing.assert_allclose(result, binarized_values) real_values = testing_utils.generate_real_values_with_zeros() result = f([real_values])[0] assert np.all(result[real_values <= 0] == 0) assert np.all(result[real_values > 0] == 1) @pytest.mark.usefixtures("eager_mode") def test_magnitude_aware_sign_binarization(self): a = np.random.uniform(-2, 2, (3, 2, 2, 3)) x = tf.Variable(a) y = lq.quantizers.MagnitudeAwareSign()(x) assert y.shape == x.shape # check sign np.testing.assert_allclose(tf.sign(y).numpy(), np.sign(a)) # check magnitude np.testing.assert_allclose( tf.reduce_mean(tf.abs(y), axis=[0, 1, 2]).numpy(), [np.mean(np.reshape(np.abs(a[:, :, :, i]), [-1])) for i in range(3)], ) @pytest.mark.parametrize( "fn", [ "ste_tern", lq.quantizers.SteTern(), lq.quantizers.SteTern(ternary_weight_networks=True), lq.quantizers.SteTern(threshold_value=np.random.uniform(0.01, 0.8)), ], ) def test_ternarization_basic(self, fn): x = tf.keras.backend.placeholder(ndim=2) f = tf.keras.backend.function([x], [lq.quantizers.get(fn)(x)]) ternarized_values = np.random.choice([-1, 0, 1], size=(4, 10)) result = f([ternarized_values])[0] np.testing.assert_allclose(result, ternarized_values) assert not np.any(result > 1) assert not np.any(result < -1) assert np.any(result == -1) assert np.any(result == 1) assert np.any(result == 0) real_values = testing_utils.generate_real_values_with_zeros() result = f([real_values])[0] assert not np.any(result > 1) assert not np.any(result < -1) assert np.any(result == -1) assert np.any(result == 1) assert np.any(result == 0) @pytest.mark.parametrize("fn", ["ste_tern", lq.quantizers.SteTern()]) def test_ternarization_with_default_threshold(self, fn): x = tf.keras.backend.placeholder(ndim=2) test_threshold = 0.05 # This is the default f = tf.keras.backend.function([x], [lq.quantizers.get(fn)(x)]) real_values = testing_utils.generate_real_values_with_zeros() result = f([real_values])[0] assert np.all(result[real_values > test_threshold] == 1) assert np.all(result[real_values < -test_threshold] == -1) assert np.all(result[np.abs(real_values) < test_threshold] == 0) assert not np.any(result > 1) assert not np.any(result < -1) def test_ternarization_with_custom_threshold(self): x = tf.keras.backend.placeholder(ndim=2) test_threshold = np.random.uniform(0.01, 0.8) fn = lq.quantizers.SteTern(threshold_value=test_threshold) f = tf.keras.backend.function([x], [fn(x)]) real_values = testing_utils.generate_real_values_with_zeros() result = f([real_values])[0] assert np.all(result[real_values > test_threshold] == 1) assert np.all(result[real_values < -test_threshold] == -1) assert np.all(result[np.abs(real_values) < test_threshold] == 0) assert not np.any(result > 1) assert not np.any(result < -1) def test_ternarization_with_ternary_weight_networks(self): x = tf.keras.backend.placeholder(ndim=2) real_values = testing_utils.generate_real_values_with_zeros() test_threshold = 0.7 * np.sum(np.abs(real_values)) / np.size(real_values) fn = lq.quantizers.SteTern(ternary_weight_networks=True) f = tf.keras.backend.function([x], [fn(x)]) result = f([real_values])[0] assert np.all(result[real_values > test_threshold] == 1) assert np.all(result[real_values < -test_threshold] == -1) assert np.all(result[np.abs(real_values) < test_threshold] == 0) assert not np.any(result > 1) assert not np.any(result < -1) @pytest.mark.parametrize("k_bit", [1, 2, 4, 6, 8]) @pytest.mark.parametrize("mode", ["activations", "weights"]) def test_dorefa_quantize(self, k_bit, mode): x = tf.keras.backend.placeholder(ndim=2) f = tf.keras.backend.function([x], [lq.quantizers.DoReFa(k_bit, mode)(x)]) real_values = testing_utils.generate_real_values_with_zeros() result = f([real_values])[0] n = 2**k_bit - 1 if mode == "weights": # Create the preprocessed and scaled stimulus, which is then ready to # go through the same test like for the activation quantizer divider = np.amax(np.abs(np.tanh(real_values))) real_values = np.tanh(real_values) / divider real_values = (real_values / 2.0) + 0.5 # The results, which are currently on [-1, 1] range get the same # scaling, so they behave like they were created on the activation # range and can be tested like that result = result / 2.0 + 0.5 assert not np.any(result > 1) assert not np.any(result < 0) for i in range(n + 1): np.testing.assert_allclose( result[ (real_values > (2 * i - 1) / (2 * n)) & (real_values < (2 * i + 1) / (2 * n)) ], i / n, atol=1e-6, ) @pytest.mark.usefixtures("eager_mode") class TestGradients: """Test gradients for different quantizers.""" @pytest.mark.parametrize( "fn", [ lq.quantizers.SteSign(clip_value=None), lq.quantizers.SteTern(clip_value=None), lq.quantizers.SteHeaviside(clip_value=None), ], ) def test_identity_ste_grad(self, fn): x = testing_utils.generate_real_values_with_zeros(shape=(8, 3, 3, 16)) tf_x = tf.Variable(x) with tf.GradientTape() as tape: activation = fn(tf_x) grad = tape.gradient(activation, tf_x) np.testing.assert_allclose(grad.numpy(), np.ones_like(x)) @pytest.mark.parametrize( "fn", [ lq.quantizers.SteSign(), lq.quantizers.SteTern(), lq.quantizers.SteHeaviside(), ], ) def test_ste_grad(self, fn): @np.vectorize def ste_grad(x): if np.abs(x) <= 1: return 1.0 return 0.0 x = testing_utils.generate_real_values_with_zeros(shape=(8, 3, 3, 16)) tf_x = tf.Variable(x) with tf.GradientTape() as tape: activation = fn(tf_x) grad = tape.gradient(activation, tf_x) np.testing.assert_allclose(grad.numpy(), ste_grad(x)) # Test with and without default threshold def test_swish_grad(self): def swish_grad(x, beta): return ( beta * (2 - beta * x * np.tanh(beta * x / 2)) / (1 + np.cosh(beta * x)) ) x = testing_utils.generate_real_values_with_zeros(shape=(8, 3, 3, 16)) tf_x = tf.Variable(x) with tf.GradientTape() as tape: activation = lq.quantizers.SwishSign()(tf_x) grad = tape.gradient(activation, tf_x) np.testing.assert_allclose(grad.numpy(), swish_grad(x, beta=5.0)) with tf.GradientTape() as tape: activation = lq.quantizers.SwishSign(beta=10.0)(tf_x) grad = tape.gradient(activation, tf_x) np.testing.assert_allclose(grad.numpy(), swish_grad(x, beta=10.0)) def test_approx_sign_grad(self): @np.vectorize def approx_sign_grad(x): if np.abs(x) <= 1: return 2 - 2 * np.abs(x) return 0.0 x = testing_utils.generate_real_values_with_zeros(shape=(8, 3, 3, 16)) tf_x = tf.Variable(x) with tf.GradientTape() as tape: activation = lq.quantizers.ApproxSign()(tf_x) grad = tape.gradient(activation, tf_x) np.testing.assert_allclose(grad.numpy(), approx_sign_grad(x)) def test_magnitude_aware_sign_grad(self): a = np.random.uniform(-2, 2, (3, 2, 2, 3)) x = tf.Variable(a) with tf.GradientTape() as tape: y = lq.quantizers.MagnitudeAwareSign()(x) grad = tape.gradient(y, x) scale_vector = [ np.mean(np.reshape(np.abs(a[:, :, :, i]), [-1])) for i in range(3) ] np.testing.assert_allclose( grad.numpy(), np.where(abs(a) < 1, np.ones(a.shape) * scale_vector, 0) ) @pytest.mark.parametrize("mode", ["activations", "weights"]) def test_dorefa_ste_grad(self, mode): @np.vectorize def ste_grad(x): if x <= 1 and x >= 0: return 1.0 return 0.0 def tanh_grad(x): # 1/(cosh**2) is the derivative of tanh. The gradients of the # scaling operations cancel each other and the gradient of the # quantizek function is supposed to be 1 everywhere, because it # is used on its linear region only. tanh does all the limiting. dividend = np.amax(np.abs(np.tanh(x))) return 1 / (np.cosh(x) ** 2.0) / dividend expected_gradient = ste_grad if mode == "activations" else tanh_grad x = testing_utils.generate_real_values_with_zeros(shape=(8, 3, 3, 16)) tf_x = tf.Variable(x) with tf.GradientTape() as tape: activation = lq.quantizers.DoReFa(2, mode)(tf_x) grad = tape.gradient(activation, tf_x) np.testing.assert_allclose(grad.numpy(), expected_gradient(x)) @pytest.mark.parametrize( "quantizer", [ ("ste_sign", lq.quantizers.SteSign), ("approx_sign", lq.quantizers.ApproxSign), ("ste_heaviside", lq.quantizers.SteHeaviside), ("swish_sign", lq.quantizers.SwishSign), ("magnitude_aware_sign", lq.quantizers.MagnitudeAwareSign), ("ste_tern", lq.quantizers.SteTern), ("dorefa_quantizer", lq.quantizers.DoReFa), ("dorefa_quantizer", functools.partial(lq.quantizers.DoReFa, mode="weights")), ], ) def test_metrics(quantizer): quantizer_str, quantizer_cls = quantizer # No metric model = tf.keras.models.Sequential( [lq.layers.QuantDense(3, kernel_quantizer=quantizer_str, input_shape=(32,))] ) model.compile(loss="mse", optimizer="sgd") assert len(model.layers[0]._metrics) == 0 # Metric added using scope with lq.context.metrics_scope(["flip_ratio"]): model = tf.keras.models.Sequential( [lq.layers.QuantDense(3, kernel_quantizer=quantizer_str, input_shape=(32,))] ) model.compile(loss="mse", optimizer="sgd") if version.parse(tf.__version__) > version.parse("1.14"): assert len(model.layers[0].kernel_quantizer._metrics) == 1 else: # In TF1.14, call() gets called twice, resulting in having an extra initial # metrics copy. assert len(model.layers[0].kernel_quantizer._metrics) == 2 # Metric added explicitly to quantizer model = tf.keras.models.Sequential( [ lq.layers.QuantDense( 3, kernel_quantizer=quantizer_cls(metrics=["flip_ratio"]), input_shape=(32,), ) ] ) model.compile(loss="mse", optimizer="sgd") if version.parse(tf.__version__) > version.parse("1.14"): assert len(model.layers[0].kernel_quantizer._metrics) == 1 else: # In TF1.14, call() gets called twice, resulting in having an extra initial # metrics copy. assert len(model.layers[0].kernel_quantizer._metrics) == 2 def test_get_kernel_quantizer_assigns_metrics(): with lq.context.metrics_scope(["flip_ratio"]): ste_sign = lq.quantizers.get_kernel_quantizer("ste_sign") assert "flip_ratio" in lq.context.get_training_metrics() assert isinstance(ste_sign, lq.quantizers.SteSign) assert "flip_ratio" in ste_sign._custom_metrics def test_get_kernel_quantizer_accepts_function(): custom_quantizer = lq.quantizers.get_kernel_quantizer(lambda x: x) assert callable(custom_quantizer) assert not hasattr(custom_quantizer, "_custom_metrics") def test_backwards_compat_aliases(): assert lq.quantizers.DoReFaQuantizer == lq.quantizers.DoReFa assert lq.quantizers.NoOpQuantizer == lq.quantizers.NoOp
17,283
37.238938
96
py
larq
larq-main/larq/activations_test.py
import numpy as np import pytest import tensorflow as tf import larq as lq from larq.testing_utils import generate_real_values_with_zeros @pytest.mark.parametrize("name", ["hard_tanh", "leaky_tanh"]) def test_serialization(name): fn = tf.keras.activations.get(name) ref_fn = getattr(lq.activations, name) assert fn == ref_fn config = tf.keras.activations.serialize(fn) fn = tf.keras.activations.deserialize(config) assert fn == ref_fn def test_hard_tanh(): real_values = generate_real_values_with_zeros() x = tf.keras.backend.placeholder(ndim=2) f = tf.keras.backend.function([x], [lq.activations.hard_tanh(x)]) result = f([real_values])[0] np.testing.assert_allclose(result, np.clip(real_values, -1, 1)) def test_leaky_tanh(): @np.vectorize def leaky_tanh(x, alpha): if x <= -1: return -1 + alpha * (x + 1) elif x <= 1: return x else: return 1 + alpha * (x - 1) real_values = generate_real_values_with_zeros() x = tf.keras.backend.placeholder(ndim=2) f = tf.keras.backend.function([x], [lq.activations.leaky_tanh(x)]) result = f([real_values])[0] np.testing.assert_allclose(result, leaky_tanh(real_values, alpha=0.2))
1,259
29
74
py
larq
larq-main/larq/quantized_variable_test.py
import numpy as np import pytest import tensorflow as tf from numpy.testing import assert_almost_equal, assert_array_equal from packaging import version from tensorflow.python.distribute.values import DistributedVariable from larq import context, testing_utils from larq.quantized_variable import QuantizedVariable from larq.testing_utils import evaluate def get_var(val, dtype=None, name=None): return tf.compat.v1.Variable(val, use_resource=True, dtype=dtype, name=name) def test_inheritance(distribute_scope): variable = get_var(3.0) quantized_variable = QuantizedVariable.from_variable(variable) assert isinstance(quantized_variable, QuantizedVariable) assert isinstance(quantized_variable, tf.Variable) assert isinstance(quantized_variable, DistributedVariable) is distribute_scope @pytest.mark.usefixtures("eager_and_graph_mode", "distribute_scope") def test_read(): x = QuantizedVariable.from_variable(get_var(3.5), quantizer=lambda x: 2 * x) evaluate(x.initializer) assert evaluate(x) == 3.5 assert evaluate(x.value()) == 3.5 assert evaluate(x.read_value()) == 3.5 assert evaluate(tf.identity(x)) == 3.5 with context.quantized_scope(True): assert evaluate(x) == 7 assert evaluate(x.value()) == 7 assert evaluate(x.read_value()) == 7 assert evaluate(tf.identity(x)) == 7 @pytest.mark.usefixtures("eager_and_graph_mode") def test_sparse_reads(): x = QuantizedVariable.from_variable(get_var([1.0, 2.0]), quantizer=lambda x: 2 * x) evaluate(x.initializer) assert evaluate(x.sparse_read([0])) == 1 assert evaluate(x.gather_nd([0])) == 1 with context.quantized_scope(True): assert evaluate(x.sparse_read([0])) == 2 assert evaluate(x.gather_nd([0])) == 2 @pytest.mark.usefixtures("eager_and_graph_mode", "distribute_scope") def test_read_nested_scopes(): x = QuantizedVariable.from_variable(get_var(3.5), quantizer=lambda x: 2 * x) evaluate(x.initializer) with context.quantized_scope(True): assert evaluate(x.read_value()) == 7 with context.quantized_scope(False): assert evaluate(x.read_value()) == 3.5 assert evaluate(x.read_value()) == 7 @pytest.mark.usefixtures("eager_and_graph_mode") def test_method_delegations(distribute_scope): x = QuantizedVariable.from_variable(get_var(3.5), quantizer=lambda x: 2 * x) with context.quantized_scope(True): evaluate(x.initializer) assert evaluate(x.value()) == 7 assert evaluate(x.read_value()) == 7 assert x.trainable if version.parse(tf.__version__) > version.parse("1.14"): assert x.synchronization == x.latent_variable.synchronization assert x.aggregation == x.latent_variable.aggregation if version.parse(tf.__version__) < version.parse("2.13"): assert evaluate(x.initialized_value()) == 7 if not tf.executing_eagerly(): if not distribute_scope: # These functions are not supported for DistributedVariables x.load(4.5) assert x.eval() == 9 assert evaluate(x.initial_value) == 7 assert x.op == x.latent_variable.op assert x.graph == x.latent_variable.graph if not distribute_scope: # These attributes are not supported for DistributedVariables assert x.constraint is None assert x.initializer == x.latent_variable.initializer def apply_and_read(x, fn, args): evaluate(fn(*args)) return evaluate(x) assert apply_and_read(x, x.assign, [4]) == 8 assert apply_and_read(x, x.assign_add, [1]) == 10 assert apply_and_read(x, x.assign_sub, [1.5]) == 7 assert x.name == x.latent_variable.name assert x.device == x.latent_variable.device assert x.shape == () assert x.get_shape() == () try: x.set_shape(()) assert x.shape == () except NotImplementedError: pass @pytest.mark.usefixtures("eager_and_graph_mode") def test_scatter_method_delegations(): x = QuantizedVariable.from_variable(get_var([3.5, 4]), quantizer=lambda x: 2 * x) evaluate(x.initializer) with context.quantized_scope(True): assert_array_equal(evaluate(x.value()), [7, 8]) def slices(val, index): return tf.IndexedSlices(values=val, indices=index) assert_array_equal(evaluate(x.scatter_sub(slices(0.5, 0))), [6, 8]) assert_array_equal(evaluate(x.scatter_add(slices(0.5, 0))), [7, 8]) if version.parse(tf.__version__) > version.parse("1.14"): assert_array_equal(evaluate(x.scatter_max(slices(4.5, 1))), [7, 9]) assert_array_equal(evaluate(x.scatter_min(slices(4.0, 1))), [7, 8]) assert_array_equal(evaluate(x.scatter_mul(slices(2.0, 1))), [7, 16]) assert_array_equal(evaluate(x.scatter_div(slices(2.0, 1))), [7, 8]) assert_array_equal(evaluate(x.scatter_update(slices(2.0, 1))), [7, 4]) assert_array_equal(evaluate(x.scatter_nd_sub([[0], [1]], [0.5, 1.0])), [6, 2]) assert_array_equal(evaluate(x.scatter_nd_add([[0], [1]], [0.5, 1.0])), [7, 4]) assert_array_equal( evaluate(x.scatter_nd_update([[0], [1]], [0.5, 1.0])), [1, 2] ) assert_array_equal(evaluate(x.batch_scatter_update(slices([2.0], [1]))), [1, 4]) @pytest.mark.usefixtures("eager_and_graph_mode", "distribute_scope") def test_overloads(quantized): if quantized: x = QuantizedVariable.from_variable(get_var(3.5), quantizer=lambda x: 2 * x) else: x = QuantizedVariable.from_variable(get_var(7.0)) evaluate(x.initializer) assert_almost_equal(8, evaluate(x + 1)) assert_almost_equal(10, evaluate(3 + x)) assert_almost_equal(14, evaluate(x + x)) assert_almost_equal(5, evaluate(x - 2)) assert_almost_equal(6, evaluate(13 - x)) assert_almost_equal(0, evaluate(x - x)) assert_almost_equal(14, evaluate(x * 2)) assert_almost_equal(21, evaluate(3 * x)) assert_almost_equal(49, evaluate(x * x)) assert_almost_equal(3.5, evaluate(x / 2)) assert_almost_equal(1.5, evaluate(10.5 / x)) assert_almost_equal(3, evaluate(x // 2)) assert_almost_equal(2, evaluate(15 // x)) assert_almost_equal(1, evaluate(x % 2)) assert_almost_equal(2, evaluate(16 % x)) assert evaluate(x < 12) assert evaluate(x <= 12) assert not evaluate(x > 12) assert not evaluate(x >= 12) assert not evaluate(12 < x) assert not evaluate(12 <= x) assert evaluate(12 > x) assert evaluate(12 >= x) assert_almost_equal(343, evaluate(pow(x, 3))) assert_almost_equal(128, evaluate(pow(2, x))) assert_almost_equal(-7, evaluate(-x)) assert_almost_equal(7, evaluate(abs(x))) @pytest.mark.usefixtures("eager_mode") def test_tensor_equality(quantized): if quantized: x = QuantizedVariable.from_variable( get_var([3.5, 4.0, 4.5]), quantizer=lambda x: 2 * x ) else: x = QuantizedVariable.from_variable(get_var([7.0, 8.0, 9.0])) evaluate(x.initializer) assert_array_equal(evaluate(x), [7.0, 8.0, 9.0]) if version.parse(tf.__version__) >= version.parse("2"): assert_array_equal(x == [7.0, 8.0, 10.0], [True, True, False]) assert_array_equal(x != [7.0, 8.0, 10.0], [False, False, True]) @pytest.mark.usefixtures("eager_and_graph_mode") def test_assign(quantized, distribute_scope): x = QuantizedVariable.from_variable( get_var(0.0, tf.float64), quantizer=lambda x: 2 * x ) evaluate(x.initializer) latent_value = 3.14 value = latent_value * 2 if quantized else latent_value # Assign doesn't correctly return a quantized variable in graph mode if a strategy is used if tf.executing_eagerly() or not distribute_scope or not quantized: # Assign float32 values lv = tf.constant(latent_value, dtype=tf.float64) assert_almost_equal(evaluate(x.assign(lv)), value) assert_almost_equal(evaluate(x.assign_add(lv)), value * 2) assert_almost_equal(evaluate(x.assign_sub(lv)), value) # Assign Python floats assert_almost_equal(evaluate(x.assign(0.0)), 0.0) assert_almost_equal(evaluate(x.assign(latent_value)), value) assert_almost_equal(evaluate(x.assign_add(latent_value)), value * 2) assert_almost_equal(evaluate(x.assign_sub(latent_value)), value) # Use the tf.assign functions instead of the var.assign methods. assert_almost_equal(evaluate(tf.compat.v1.assign(x, 0.0)), 0.0) assert_almost_equal(evaluate(tf.compat.v1.assign(x, latent_value)), value) assert_almost_equal( evaluate(tf.compat.v1.assign_add(x, latent_value)), value * 2 ) assert_almost_equal(evaluate(tf.compat.v1.assign_sub(x, latent_value)), value) # Assign multiple times if version.parse(tf.__version__) >= version.parse("2.2") and ( tf.executing_eagerly() or not distribute_scope ): assign = x.assign(0.0) assert_almost_equal(evaluate(assign), 0.0) assert_almost_equal(evaluate(assign.assign(latent_value)), value) assert_almost_equal( evaluate(x.assign_add(latent_value).assign_add(latent_value)), value * 3 ) assert_almost_equal(evaluate(x), value * 3) assert_almost_equal( evaluate(x.assign_sub(latent_value).assign_sub(latent_value)), value ) assert_almost_equal(evaluate(x), value) # Assign with read_value=False assert_almost_equal(evaluate(x.assign(0.0)), 0.0) assert evaluate(x.assign(latent_value, read_value=False)) is None assert_almost_equal(evaluate(x), value) assert evaluate(x.assign_add(latent_value, read_value=False)) is None assert_almost_equal(evaluate(x), 2 * value) assert evaluate(x.assign_sub(latent_value, read_value=False)) is None assert_almost_equal(evaluate(x), value) @pytest.mark.usefixtures("eager_mode", "distribute_scope") def test_assign_tf_function(quantized): x = QuantizedVariable.from_variable(get_var(0.0), quantizer=lambda x: 2 * x) @tf.function def run_assign(): return x.assign(1.0).assign_add(3.0).assign_add(3.0).assign_sub(2.0) assert_almost_equal(evaluate(run_assign()), 10.0 if quantized else 5.0) @pytest.mark.usefixtures("eager_and_graph_mode", "distribute_scope") def test_assign_op(): x = QuantizedVariable.from_variable(get_var(0.0), quantizer=lambda x: 2 * x) @tf.function def func(): assert x.assign(1.0).op is not None assert x.assign_add(1.0).op is not None assert x.assign_sub(1.0).op is not None func() @pytest.mark.usefixtures("eager_mode", "distribute_scope") def test_tf_function_control_dependencies(quantized): x = QuantizedVariable.from_variable(get_var(0.0), quantizer=lambda x: 2 * x) @tf.function def func(): update = x.assign_add(1.0) with tf.control_dependencies([update]): x.assign_add(1.0) func() assert_almost_equal(evaluate(x), 4.0 if quantized else 2.0) def test_tf_function_with_variable_and_quantized_variable(): variable = get_var(tf.ones(2, 2)) quantized_variable = QuantizedVariable.from_variable(variable) @tf.function def f(x): return x + 1 f(variable) f(quantized_variable) @pytest.mark.usefixtures("eager_and_graph_mode") def test_checkpoint(tmp_path): x = QuantizedVariable.from_variable(get_var(0.0), quantizer=lambda x: 2 * x) evaluate(x.initializer) evaluate(x.assign(123.0)) checkpoint = tf.train.Checkpoint(x=x) save_path = checkpoint.save(tmp_path) evaluate(x.assign(234.0)) checkpoint.restore(save_path).assert_consumed().run_restore_ops() assert isinstance(x, QuantizedVariable) assert evaluate(x) == 123.0 with context.quantized_scope(True): assert evaluate(x) == 123.0 * 2 @pytest.mark.usefixtures("distribute_scope") def test_invalid_wrapped_usage(): with pytest.raises(ValueError, match="`variable` must be of type"): QuantizedVariable.from_variable(tf.constant([1.0])) with pytest.raises(ValueError, match="`quantizer` must be `callable` or `None`"): QuantizedVariable.from_variable(get_var([1.0]), 1) # type: ignore with pytest.raises(ValueError, match="`precision` must be of type `int` or `None`"): QuantizedVariable.from_variable(get_var([1.0]), precision=1.0) # type: ignore @pytest.mark.usefixtures("eager_and_graph_mode") def test_repr(snapshot): x = get_var(0.0, name="x") class Quantizer: def __call__(self, x): return x snapshot.assert_match( repr(QuantizedVariable.from_variable(x, quantizer=lambda x: 2 * x)) ) snapshot.assert_match( repr(QuantizedVariable.from_variable(x, quantizer=Quantizer())) ) snapshot.assert_match(repr(QuantizedVariable.from_variable(x, precision=1))) @pytest.mark.usefixtures("eager_mode") @pytest.mark.parametrize("should_quantize", [True, False]) def test_optimizer(should_quantize): x = QuantizedVariable.from_variable(get_var(1.0), quantizer=lambda x: -x) opt = tf.keras.optimizers.SGD(1.0) def loss(): with context.quantized_scope(should_quantize): return x + 1.0 @tf.function def f(): opt.minimize(loss, var_list=[x]) f() if should_quantize: assert evaluate(x) == 2.0 with context.quantized_scope(should_quantize): assert evaluate(x) == -2.0 else: assert evaluate(x) == 0.0 @pytest.mark.skipif( version.parse(tf.__version__) < version.parse("2"), reason="Requires TensorFlow 2", ) def test_saved_model(tmp_path): model_path = str(tmp_path / "model") x = np.random.normal(size=(4, 32)) model = testing_utils.get_small_bnn_model(x.shape[1], 16, 10) weights = model.get_weights() model.save(model_path, save_format="tf") reloaded_model = tf.keras.models.load_model(model_path) reloaded_weights = reloaded_model.get_weights() assert_almost_equal(reloaded_model.predict(x), model.predict(x)) assert len(reloaded_weights) == len(weights) for reloaded_weight, weight in zip(reloaded_weights, weights): assert_almost_equal(reloaded_weight, weight)
14,405
37.31383
94
py
DAC2018
DAC2018-master/setup.py
from distutils.core import setup, Extension module = Extension('mypack',extra_compile_args=['-std=c++11'], include_dirs=['/usr/local/cuda/include'], sources = ['Detector.cpp'],extra_objects = ['./plugin.o', './kernel.o'], extra_link_args=['-lnvinfer', '-lnvcaffe_parser', '-lcudnn']) setup(name = 'mypack', version='1.0', description="Detector package", ext_modules = [module])
388
54.571429
142
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/aspect_category_model/capsnet.py
import torch from torch import nn import torch.nn.functional as F from torch.nn import init from src.module.utils.constants import PAD_INDEX, INF from src.module.utils.sentence_clip import sentence_clip from src.module.attention.dot_attention import DotAttention from src.module.attention.scaled_dot_attention import ScaledDotAttention from src.module.attention.bilinear_attention import BilinearAttention from src.module.attention.tanh_bilinear_attention import TanhBilinearAttention from src.module.attention.concat_attention import ConcatAttention from src.module.attention.tanh_concat_attention import TanhConcatAttention from src.module.attention.mlp_attention import MlpAttention import numpy as np from src.module.utils.squash import squash class CapsuleNetwork(nn.Module): def __init__(self, embedding, aspect_embedding, hidden_size, capsule_size, dropout, num_categories): super(CapsuleNetwork, self).__init__() self.embedding = embedding self.aspect_embedding = aspect_embedding embed_size = embedding.embedding_dim self.capsule_size = capsule_size self.aspect_transform = nn.Sequential( nn.Linear(embed_size, capsule_size), nn.Dropout(dropout) ) self.sentence_transform = nn.Sequential( nn.Linear(hidden_size, capsule_size), nn.Dropout(dropout) ) self.norm_attention = BilinearAttention(capsule_size, capsule_size) self.guide_capsule = nn.Parameter( torch.Tensor(num_categories, capsule_size) ) self.guide_weight = nn.Parameter( torch.Tensor(capsule_size, capsule_size) ) self.scale = nn.Parameter(torch.tensor(5.0)) self.capsule_projection = nn.Linear(capsule_size, capsule_size * num_categories) self.dropout = dropout self.num_categories = num_categories self._reset_parameters() def _reset_parameters(self): init.xavier_uniform_(self.guide_capsule) init.xavier_uniform_(self.guide_weight) def load_sentiment(self, path): sentiment = np.load(path) e1 = np.mean(sentiment) d1 = np.std(sentiment) e2 = 0 d2 = np.sqrt(2.0 / (sentiment.shape[0] + sentiment.shape[1])) sentiment = (sentiment - e1) / d1 * d2 + e2 self.guide_capsule.data.copy_(torch.tensor(sentiment)) def forward(self, sentence, aspect): # get lengths and masks sentence = sentence_clip(sentence) sentence_mask = (sentence != PAD_INDEX) # embedding sentence = self.embedding(sentence) sentence = F.dropout(sentence, p=self.dropout, training=self.training) aspect = self.embedding(aspect) aspect = F.dropout(aspect, p=self.dropout, training=self.training) # sentence encode layer sentence = self._sentence_encode(sentence, aspect) # primary capsule layer sentence = self.sentence_transform(sentence) primary_capsule = squash(sentence, dim=-1) # aspect capsule layer aspect = self.aspect_transform(aspect) aspect_capsule = squash(aspect, dim=-1) # aspect aware normalization norm_weight = self.norm_attention.get_attention_weights(aspect_capsule, primary_capsule, sentence_mask) # capsule guided routing category_capsule = self._capsule_guided_routing(primary_capsule, norm_weight) category_capsule_norm = torch.sqrt(torch.sum(category_capsule * category_capsule, dim=-1, keepdim=False)) return category_capsule_norm def _sentence_encode(self, sentence, aspect, mask=None): raise NotImplementedError('_sentence_encode method is not implemented.') def _capsule_guided_routing(self, primary_capsule, norm_weight): guide_capsule = squash(self.guide_capsule) guide_matrix = primary_capsule.matmul(self.guide_weight).matmul(guide_capsule.transpose(0, 1)) guide_matrix = F.softmax(guide_matrix, dim=-1) guide_matrix = guide_matrix * norm_weight.unsqueeze(-1) * self.scale # (batch_size, time_step, num_categories) category_capsule = guide_matrix.transpose(1, 2).matmul(primary_capsule) category_capsule = F.dropout(category_capsule, p=self.dropout, training=self.training) category_capsule = squash(category_capsule) return category_capsule
4,384
45.648936
119
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/aspect_category_model/recurrent_capsnet.py
import torch from torch import nn import torch.nn.functional as F from src.aspect_category_model.capsnet import CapsuleNetwork class RecurrentCapsuleNetwork(CapsuleNetwork): def __init__(self, embedding, aspect_embedding, num_layers, bidirectional, capsule_size, dropout, num_categories): super(RecurrentCapsuleNetwork, self).__init__( embedding=embedding, aspect_embedding=aspect_embedding, hidden_size=embedding.embedding_dim * (2 if bidirectional else 1), capsule_size=capsule_size, dropout=dropout, num_categories=num_categories ) embed_size = embedding.embedding_dim self.rnn = nn.GRU( input_size=embed_size * 2, hidden_size=embed_size, num_layers=num_layers, bidirectional=bidirectional, batch_first=True ) self.bidirectional = bidirectional def _sentence_encode(self, sentence, aspect, mask=None): batch_size, time_step, embed_size = sentence.size() aspect_aware_sentence = torch.cat(( sentence, aspect.unsqueeze(1).expand(batch_size, time_step, embed_size) ), dim=-1) output, _ = self.rnn(aspect_aware_sentence) if self.bidirectional: sentence = sentence.unsqueeze(-1).expand(batch_size, time_step, embed_size, 2) sentence = sentence.contiguous().view(batch_size, time_step, embed_size * 2) output = output + sentence output = F.dropout(output, p=self.dropout, training=self.training) return output
1,597
41.052632
118
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/aspect_category_model/bert_capsnet.py
import torch from torch import nn import torch.nn.functional as F from torch.nn import init from src.module.utils.constants import PAD_INDEX, INF from src.module.utils.sentence_clip import sentence_clip from src.module.attention.dot_attention import DotAttention from src.module.attention.scaled_dot_attention import ScaledDotAttention from src.module.attention.bilinear_attention import BilinearAttention from src.module.attention.tanh_bilinear_attention import TanhBilinearAttention from src.module.attention.concat_attention import ConcatAttention from src.module.attention.tanh_concat_attention import TanhConcatAttention from src.module.attention.mlp_attention import MlpAttention from src.aspect_term_model.capsnet import squash import numpy as np class BertCapsuleNetwork(nn.Module): def __init__(self, bert, bert_size, capsule_size, dropout, num_categories): super(BertCapsuleNetwork, self).__init__() self.bert = bert self.bert_size = bert_size self.capsule_size = capsule_size self.aspect_transform = nn.Sequential( nn.Linear(bert_size, capsule_size), nn.Dropout(dropout) ) self.sentence_transform = nn.Sequential( nn.Linear(bert_size, capsule_size), nn.Dropout(dropout) ) self.norm_attention = BilinearAttention(capsule_size, capsule_size) self.guide_capsule = nn.Parameter( torch.Tensor(num_categories, capsule_size) ) self.guide_weight = nn.Parameter( torch.Tensor(capsule_size, capsule_size) ) self.scale = nn.Parameter(torch.tensor(5.0)) self.capsule_projection = nn.Linear(bert_size, bert_size * num_categories) self.dropout = dropout self.num_categories = num_categories self._reset_parameters() def _reset_parameters(self): init.xavier_uniform_(self.guide_capsule) init.xavier_uniform_(self.guide_weight) def load_sentiment(self, path): sentiment = np.load(path) e1 = np.mean(sentiment) d1 = np.std(sentiment) e2 = 0 d2 = np.sqrt(2.0 / (sentiment.shape[0] + sentiment.shape[1])) sentiment = (sentiment - e1) / d1 * d2 + e2 self.guide_capsule.data.copy_(torch.tensor(sentiment)) def forward(self, bert_token, bert_segment): # BERT encoding encoder_layer, _ = self.bert(bert_token, bert_segment, output_all_encoded_layers=False) batch_size, segment_len = bert_segment.size() max_segment_len = bert_segment.argmax(dim=-1, keepdim=True) batch_arrange = torch.arange(segment_len).unsqueeze(0).expand(batch_size, segment_len).to(bert_segment.device) segment_mask = batch_arrange <= max_segment_len sentence_mask = segment_mask & (1 - bert_segment).byte() aspect_mask = bert_segment sentence_lens = sentence_mask.long().sum(dim=1, keepdim=True) aspect_lens = aspect_mask.long().sum(dim=1, keepdim=True) aspect = encoder_layer.masked_fill(aspect_mask.unsqueeze(-1) == 0, 0) aspect = aspect.sum(dim=1, keepdim=False) / aspect_lens.float() # sentence encode layer max_len = sentence_lens.max().item() sentence = encoder_layer[:, 0: max_len].contiguous() sentence_mask = sentence_mask[:, 0: max_len].contiguous() sentence = sentence.masked_fill(sentence_mask.unsqueeze(-1) == 0, 0) # primary capsule layer sentence = self.sentence_transform(sentence) primary_capsule = squash(sentence, dim=-1) # aspect capsule layer aspect = self.aspect_transform(aspect) aspect_capsule = squash(aspect, dim=-1) # aspect aware normalization norm_weight = self.norm_attention.get_attention_weights(aspect_capsule, primary_capsule, sentence_mask) # capsule guided routing category_capsule = self._capsule_guided_routing(primary_capsule, norm_weight) category_capsule_norm = torch.sqrt(torch.sum(category_capsule * category_capsule, dim=-1, keepdim=False)) return category_capsule_norm def _capsule_guided_routing(self, primary_capsule, norm_weight): guide_capsule = squash(self.guide_capsule) guide_matrix = primary_capsule.matmul(self.guide_weight).matmul(guide_capsule.transpose(0, 1)) guide_matrix = F.softmax(guide_matrix, dim=-1) guide_matrix = guide_matrix * norm_weight.unsqueeze(-1) * self.scale # (batch_size, time_step, num_categories) category_capsule = guide_matrix.transpose(1, 2).matmul(primary_capsule) category_capsule = F.dropout(category_capsule, p=self.dropout, training=self.training) category_capsule = squash(category_capsule) return category_capsule
4,772
48.206186
119
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/aspect_term_model/capsnet.py
import torch from torch import nn import torch.nn.functional as F from torch.nn import init from src.module.utils.constants import PAD_INDEX, INF from src.module.utils.sentence_clip import sentence_clip from src.module.attention.dot_attention import DotAttention from src.module.attention.scaled_dot_attention import ScaledDotAttention from src.module.attention.bilinear_attention import BilinearAttention from src.module.attention.tanh_bilinear_attention import TanhBilinearAttention from src.module.attention.concat_attention import ConcatAttention from src.module.attention.tanh_concat_attention import TanhConcatAttention from src.module.attention.mlp_attention import MlpAttention import numpy as np from src.module.utils.squash import squash class CapsuleNetwork(nn.Module): def __init__(self, embedding, hidden_size, capsule_size, dropout, num_categories): super(CapsuleNetwork, self).__init__() self.embedding = embedding embed_size = embedding.embedding_dim self.capsule_size = capsule_size self.aspect_transform = nn.Sequential( nn.Linear(embed_size, capsule_size), nn.Dropout(dropout) ) self.sentence_transform = nn.Sequential( nn.Linear(hidden_size, capsule_size), nn.Dropout(dropout) ) self.norm_attention = BilinearAttention(capsule_size, capsule_size) self.guide_capsule = nn.Parameter( torch.Tensor(num_categories, capsule_size) ) self.guide_weight = nn.Parameter( torch.Tensor(capsule_size, capsule_size) ) self.scale = nn.Parameter(torch.tensor(4.0)) self.capsule_projection = nn.Linear(capsule_size, capsule_size * num_categories) self.dropout = dropout self.num_categories = num_categories self._reset_parameters() def _reset_parameters(self): init.xavier_uniform_(self.guide_capsule) init.xavier_uniform_(self.guide_weight) def load_sentiment(self, path): sentiment = np.load(path) e1 = np.mean(sentiment) d1 = np.std(sentiment) e2 = 0 d2 = np.sqrt(2.0 / (sentiment.shape[0] + sentiment.shape[1])) sentiment = (sentiment - e1) / d1 * d2 + e2 self.guide_capsule.data.copy_(torch.tensor(sentiment)) def forward(self, sentence, aspect): # get lengths and masks sentence = sentence_clip(sentence) aspect = sentence_clip(aspect) sentence_mask = (sentence != PAD_INDEX) aspect_mask = (aspect != PAD_INDEX) # sentence_lens = sentence_mask.long().sum(dim=1, keepdim=True) aspect_lens = aspect_mask.long().sum(dim=1, keepdim=True) # embedding sentence = self.embedding(sentence) sentence = F.dropout(sentence, p=self.dropout, training=self.training) aspect = self.embedding(aspect) aspect = F.dropout(aspect, p=self.dropout, training=self.training) # aspect average pooling aspect = aspect.masked_fill(aspect_mask.unsqueeze(-1) == 0, 0) aspect = aspect.sum(dim=1, keepdim=False) / aspect_lens.float() # sentence encode layer sentence = self._sentence_encode(sentence, aspect) # primary capsule layer sentence = self.sentence_transform(sentence) primary_capsule = squash(sentence, dim=-1) # aspect capsule layer aspect = self.aspect_transform(aspect) aspect_capsule = squash(aspect, dim=-1) # aspect aware normalization norm_weight = self.norm_attention.get_attention_weights(aspect_capsule, primary_capsule, sentence_mask) # capsule guided routing category_capsule = self._capsule_guided_routing(primary_capsule, norm_weight) category_capsule_norm = torch.sqrt(torch.sum(category_capsule * category_capsule, dim=-1, keepdim=False)) return category_capsule_norm def _sentence_encode(self, sentence, aspect, mask=None): raise NotImplementedError('_sentence_encode method is not implemented.') def _capsule_guided_routing(self, primary_capsule, norm_weight): guide_capsule = squash(self.guide_capsule) guide_matrix = primary_capsule.matmul(self.guide_weight).matmul(guide_capsule.transpose(0, 1)) guide_matrix = F.softmax(guide_matrix, dim=-1) guide_matrix = guide_matrix * norm_weight.unsqueeze(-1) * self.scale # (batch_size, time_step, num_categories) category_capsule = guide_matrix.transpose(1, 2).matmul(primary_capsule) category_capsule = F.dropout(category_capsule, p=self.dropout, training=self.training) category_capsule = squash(category_capsule) return category_capsule
4,714
46.15
119
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/aspect_term_model/recurrent_capsnet.py
import torch from torch import nn import torch.nn.functional as F from src.aspect_term_model.capsnet import CapsuleNetwork class RecurrentCapsuleNetwork(CapsuleNetwork): def __init__(self, embedding, num_layers, bidirectional, capsule_size, dropout, num_categories): super(RecurrentCapsuleNetwork, self).__init__( embedding=embedding, hidden_size=embedding.embedding_dim * (2 if bidirectional else 1), capsule_size=capsule_size, dropout=dropout, num_categories=num_categories ) embed_size = embedding.embedding_dim self.rnn = nn.GRU( input_size=embed_size * 2, hidden_size=embed_size, num_layers=num_layers, bidirectional=bidirectional, batch_first=True ) self.bidirectional = bidirectional def _sentence_encode(self, sentence, aspect, mask=None): batch_size, time_step, embed_size = sentence.size() aspect_aware_sentence = torch.cat(( sentence, aspect.unsqueeze(1).expand(batch_size, time_step, embed_size) ), dim=-1) output, _ = self.rnn(aspect_aware_sentence) if self.bidirectional: sentence = sentence.unsqueeze(-1).expand(batch_size, time_step, embed_size, 2) sentence = sentence.contiguous().view(batch_size, time_step, embed_size * 2) output = output + sentence output = F.dropout(output, p=self.dropout, training=self.training) return output
1,528
40.324324
100
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/aspect_term_model/bert_capsnet.py
import torch from torch import nn import torch.nn.functional as F from torch.nn import init from src.module.utils.constants import PAD_INDEX, INF from src.module.utils.sentence_clip import sentence_clip from src.module.attention.dot_attention import DotAttention from src.module.attention.scaled_dot_attention import ScaledDotAttention from src.module.attention.bilinear_attention import BilinearAttention from src.module.attention.tanh_bilinear_attention import TanhBilinearAttention from src.module.attention.concat_attention import ConcatAttention from src.module.attention.tanh_concat_attention import TanhConcatAttention from src.module.attention.mlp_attention import MlpAttention from src.module.utils.squash import squash import numpy as np import time class BertCapsuleNetwork(nn.Module): def __init__(self, bert, bert_size, capsule_size, dropout, num_categories): super(BertCapsuleNetwork, self).__init__() self.bert = bert self.bert_size = bert_size self.capsule_size = capsule_size self.aspect_transform = nn.Sequential( nn.Linear(bert_size, capsule_size), nn.Dropout(dropout) ) self.sentence_transform = nn.Sequential( nn.Linear(bert_size, capsule_size), nn.Dropout(dropout) ) self.norm_attention = BilinearAttention(capsule_size, capsule_size) self.guide_capsule = nn.Parameter( torch.Tensor(num_categories, capsule_size) ) self.guide_weight = nn.Parameter( torch.Tensor(capsule_size, capsule_size) ) self.scale = nn.Parameter(torch.tensor(5.0)) self.capsule_projection = nn.Linear(bert_size, bert_size * num_categories) self.dropout = dropout self.num_categories = num_categories self._reset_parameters() def _reset_parameters(self): init.xavier_uniform_(self.guide_capsule) init.xavier_uniform_(self.guide_weight) def load_sentiment(self, path): sentiment = np.load(path) e1 = np.mean(sentiment) d1 = np.std(sentiment) e2 = 0 d2 = np.sqrt(2.0 / (sentiment.shape[0] + sentiment.shape[1])) sentiment = (sentiment - e1) / d1 * d2 + e2 self.guide_capsule.data.copy_(torch.tensor(sentiment)) def forward(self, bert_token, bert_segment): # BERT encoding encoder_layer, _ = self.bert(bert_token, bert_segment, output_all_encoded_layers=False) batch_size, segment_len = bert_segment.size() max_segment_len = bert_segment.argmax(dim=-1, keepdim=True) batch_arrange = torch.arange(segment_len).unsqueeze(0).expand(batch_size, segment_len).to(bert_segment.device) segment_mask = batch_arrange <= max_segment_len sentence_mask = segment_mask & (1 - bert_segment).byte() aspect_mask = bert_segment sentence_lens = sentence_mask.long().sum(dim=1, keepdim=True) # aspect average pooling aspect_lens = aspect_mask.long().sum(dim=1, keepdim=True) aspect = encoder_layer.masked_fill(aspect_mask.unsqueeze(-1) == 0, 0) aspect = aspect.sum(dim=1, keepdim=False) / aspect_lens.float() # sentence encode layer max_len = sentence_lens.max().item() sentence = encoder_layer[:, 0: max_len].contiguous() sentence_mask = sentence_mask[:, 0: max_len].contiguous() sentence = sentence.masked_fill(sentence_mask.unsqueeze(-1) == 0, 0) # primary capsule layer sentence = self.sentence_transform(sentence) primary_capsule = squash(sentence, dim=-1) aspect = self.aspect_transform(aspect) aspect_capsule = squash(aspect, dim=-1) # aspect aware normalization norm_weight = self.norm_attention.get_attention_weights(aspect_capsule, primary_capsule, sentence_mask) # capsule guided routing category_capsule = self._capsule_guided_routing(primary_capsule, norm_weight) category_capsule_norm = torch.sqrt(torch.sum(category_capsule * category_capsule, dim=-1, keepdim=False)) return category_capsule_norm def _capsule_guided_routing(self, primary_capsule, norm_weight): guide_capsule = squash(self.guide_capsule) guide_matrix = primary_capsule.matmul(self.guide_weight).matmul(guide_capsule.transpose(0, 1)) guide_matrix = F.softmax(guide_matrix, dim=-1) guide_matrix = guide_matrix * norm_weight.unsqueeze(-1) * self.scale # (batch_size, time_step, num_categories) category_capsule = guide_matrix.transpose(1, 2).matmul(primary_capsule) category_capsule = F.dropout(category_capsule, p=self.dropout, training=self.training) category_capsule = squash(category_capsule) return category_capsule
4,780
47.785714
119
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/module/attention/concat_attention.py
import torch from torch import nn from torch.nn import init from src.module.attention.attention import Attention class ConcatAttention(Attention): def __init__(self, query_size, key_size, dropout=0): super(ConcatAttention, self).__init__(dropout) self.query_weights = nn.Parameter(torch.Tensor(query_size, 1)) self.key_weights = nn.Parameter(torch.Tensor(key_size, 1)) init.xavier_uniform_(self.query_weights) init.xavier_uniform_(self.key_weights) def _score(self, query, key): """ query: FloatTensor (batch_size, num_queries, query_size) key: FloatTensor (batch_size, time_step, key_size) """ batch_size, num_queries, time_step = query.size(0), query.size(1), key.size(1) query = query.matmul(self.query_weights).expand(batch_size, num_queries, time_step) key = key.matmul(self.key_weights).transpose(1, 2).expand(batch_size, num_queries, time_step) score = query + key return score
1,007
41
101
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/module/attention/bilinear_attention.py
import torch from torch import nn from torch.nn import init from src.module.attention.attention import Attention class BilinearAttention(Attention): def __init__(self, query_size, key_size, dropout=0): super(BilinearAttention, self).__init__(dropout) self.weights = nn.Parameter(torch.FloatTensor(query_size, key_size)) init.xavier_uniform_(self.weights) def _score(self, query, key): """ query: FloatTensor (batch_size, num_queries, query_size) key: FloatTensor (batch_size, time_step, key_size) """ score = query.matmul(self.weights).matmul(key.transpose(1, 2)) return score
659
33.736842
76
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/module/attention/tanh_bilinear_attention.py
import torch from torch import nn from torch.nn import init from src.module.attention.attention import Attention class TanhBilinearAttention(Attention): def __init__(self, query_size, key_size, dropout=0): super(TanhBilinearAttention, self).__init__(dropout) self.weights = nn.Parameter(torch.FloatTensor(query_size, key_size)) init.xavier_uniform_(self.weights) self.bias = nn.Parameter(torch.zeros(1)) def _score(self, query, key): """ query: FloatTensor (batch_size, num_queries, query_size) key: FloatTensor (batch_size, time_step, key_size) """ score = torch.tanh(query.matmul(self.weights).matmul(key.transpose(1, 2)) + self.bias) return score
740
36.05
94
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/module/attention/tanh_concat_attention.py
import torch from torch import nn from torch.nn import init from src.module.attention.attention import Attention class TanhConcatAttention(Attention): def __init__(self, query_size, key_size, dropout=0): super(TanhConcatAttention, self).__init__(dropout) self.query_weights = nn.Parameter(torch.Tensor(query_size, 1)) self.key_weights = nn.Parameter(torch.Tensor(key_size, 1)) init.xavier_uniform_(self.query_weights) init.xavier_uniform_(self.key_weights) def _score(self, query, key): """ query: FloatTensor (batch_size, num_queries, query_size) key: FloatTensor (batch_size, time_step, key_size) """ batch_size, num_queries, time_step = query.size(0), query.size(1), key.size(1) query = query.matmul(self.query_weights).expand(batch_size, num_queries, time_step) key = key.matmul(self.key_weights).transpose(1, 2).expand(batch_size, num_queries, time_step) score = query + key score = torch.tanh(score) return score
1,049
41
101
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/module/attention/multi_head_attention.py
from torch import nn from torch.nn import init import math class MultiHeadAttention(nn.Module): def __init__(self, attention, num_heads, hidden_size, key_size='default', value_size='default', out_size='default'): key_size = hidden_size // num_heads if key_size == 'default' else key_size value_size = hidden_size // num_heads if value_size == 'default' else value_size out_size = hidden_size if out_size == 'default' else out_size super(MultiHeadAttention, self).__init__() self.num_heads = num_heads self.key_size = key_size self.value_size = value_size self.query_projection = nn.Linear(hidden_size, num_heads * key_size) self.key_projection = nn.Linear(hidden_size, num_heads * key_size) self.value_projection = nn.Linear(hidden_size, num_heads * value_size) init.normal_(self.query_projection.weight, mean=0, std=math.sqrt(2.0 / (hidden_size + key_size))) init.normal_(self.key_projection.weight, mean=0, std=math.sqrt(2.0 / (hidden_size + key_size))) init.normal_(self.value_projection.weight, mean=0, std=math.sqrt(2.0 / (hidden_size + value_size))) self.output_projection = nn.Linear(num_heads * value_size, out_size) init.xavier_normal_(self.output_projection.weight) self.attention = attention def forward(self, query, key, value, mask=None): """ query: FloatTensor (batch_size, hidden_size) or (batch_size, num_queries, hidden_size) key: FloatTensor (batch_size, time_step, hidden_size) value: FloatTensor (batch_size, time_step, hidden_size) mask: ByteTensor (batch_size, time_step) or ByteTensor (batch_size, num_queries, time_step) subsequent_mask: ByteTensor (num_queries, time_step) """ single_query = False if len(query.size()) == 2: query = query.unsqueeze(1) single_query = True if mask is not None: if len(mask.size()) == 2: mask = mask.unsqueeze(1) else: assert mask.size(1) == query.size(1) num_heads, key_size, value_size = self.num_heads, self.key_size, self.value_size batch_size, num_queries, time_step = query.size(0), query.size(1), key.size(1) query = self.query_projection(query).view(batch_size, num_queries, num_heads, key_size) key = self.key_projection(key).view(batch_size, time_step, num_heads, key_size) value = self.value_projection(value).view(batch_size, time_step, num_heads, value_size) if mask is not None: if len(mask.size()) == 2: mask = mask.unsqueeze(0).repeat(num_heads, 1, 1).view(-1, time_step) else: mask = mask.unsqueeze(0).repeat(num_heads, 1, 1, 1).view(-1, num_queries, time_step) query = query.permute(2, 0, 1, 3).contiguous().view(-1, num_queries, key_size) key = key.permute(2, 0, 1, 3).contiguous().view(-1, time_step, key_size) value = value.permute(2, 0, 1, 3).contiguous().view(-1, time_step, value_size) output = self.attention(query, key, value, mask) output = output.view(num_heads, batch_size, num_queries, value_size) output = output.permute(1, 2, 0, 3).contiguous().view(batch_size, num_queries, -1) output = self.output_projection(output) if single_query: output = output.squeeze(1) return output
3,451
55.590164
120
py
MAMS-for-ABSA
MAMS-for-ABSA-master/src/module/attention/attention.py
from torch import nn import torch.nn.functional as F from src.module.utils import constants class Attention(nn.Module): """ The base class of attention. """ def __init__(self, dropout): super(Attention, self).__init__() self.dropout = dropout def forward(self, query, key, value, mask=None): """ query: FloatTensor (batch_size, query_size) or FloatTensor (batch_size, num_queries, query_size) key: FloatTensor (batch_size, time_step, key_size) value: FloatTensor (batch_size, time_step, hidden_size) mask: ByteTensor (batch_size, time_step) or ByteTensor (batch_size, num_queries, time_step) """ single_query = False if len(query.size()) == 2: query = query.unsqueeze(1) single_query = True if mask is not None: if len(mask.size()) == 2: mask = mask.unsqueeze(1) else: assert mask.size(1) == query.size(1) score = self._score(query, key) # FloatTensor (batch_size, num_queries, time_step) weights = self._weights_normalize(score, mask) weights = F.dropout(weights, p=self.dropout, training=self.training) output = weights.matmul(value) if single_query: output = output.squeeze(1) return output def _score(self, query, key): raise NotImplementedError('Attention score method is not implemented.') def _weights_normalize(self, score, mask): if not mask is None: score = score.masked_fill(mask == 0, -constants.INF) weights = F.softmax(score, dim=-1) return weights def get_attention_weights(self, query, key, mask=None): single_query = False if len(query.size()) == 2: query = query.unsqueeze(1) single_query = True if mask is not None: if len(mask.size()) == 2: mask = mask.unsqueeze(1) else: assert mask.size(1) == query.size(1) score = self._score(query, key) # FloatTensor (batch_size, num_queries, time_step) weights = self._weights_normalize(score, mask) weights = F.dropout(weights, p=self.dropout, training=self.training) if single_query: weights = weights.squeeze(1) return weights
2,355
37
104
py