])
total_counts = sum(word_counts[ele] for ele in word_counts)
total_counts -= word_counts[vocabulary_inv_list[0]]
background_array = np.zeros(vocab_sz)
for i in range(1, vocab_sz):
background_array[i] = word_counts[vocabulary_inv[i]] / total_counts
seed_docs, seed_label = pseudodocs(
word_sup_array, gamma, background_array, sequence_length, len_avg,
len_std, beta, alpha, vocabulary_inv, embedding_mat, centers,
kappa, 'cnn',
'./results/{}/{}/phase1/'.format(args.dataset, 'cnn'))
if args.sup_source == 'docs':
num_real_doc = len(sup_idx.flatten()) * int(1 + beta * 0.1)
real_seed_docs, real_seed_label = augment(x, sup_idx, num_real_doc)
seed_docs = np.concatenate((seed_docs, real_seed_docs), axis=0)
seed_label = np.concatenate((seed_label, real_seed_label), axis=0)
perm_seed = np.random.permutation(len(seed_label))
seed_docs = seed_docs[perm_seed]
seed_label = seed_label[perm_seed]
print('\n### Phase 2: pre-training with pseudo documents ###')
wstc.pretrain(x=seed_docs,
pretrain_labels=seed_label,
sup_idx=sup_idx,
optimizer=SGD(lr=0.1, momentum=0.9),
epochs=pretrain_epochs,
batch_size=args.batch_size,
def proceed_level(x, sequences, wstc, args, pretrain_epochs, self_lr, decay, update_interval,
delta, class_tree, level, expand_num, background_array, doc_length, sent_length, len_avg,
len_std, num_doc, interp_weight, vocabulary_inv, common_words):
print(f"\n### Proceeding level {level} ###")
dataset = args.dataset
sup_source = args.sup_source
maxiter = args.maxiter.split(',')
maxiter = int(maxiter[level])
batch_size = args.batch_size
parents = class_tree.find_at_level(level)
parents_names = [parent.name for parent in parents]
print(f'Nodes: {parents_names}')
for parent in parents:
# initialize classifiers in hierarchy
print("\n### Input preparation ###")
if class_tree.embedding is None:
train_class_embedding(x, vocabulary_inv, dataset_name=args.dataset, node=class_tree)
parent.embedding = class_tree.embedding
wstc.instantiate(class_tree=parent)
save_dir = f'./results/{dataset}/{sup_source}/level_{level}'
if parent.model is not None:
print("\n### Phase 1: vMF distribution fitting & pseudo document generation ###")
if args.pseudo == "bow":
print("Pseudo documents generation (Method: Bag-of-words)...")
seed_docs, seed_label = bow_pseudodocs(parent.children, expand_num, background_array, doc_length, len_avg,
len_std, num_doc, interp_weight, vocabulary_inv, parent.embedding, save_dir)
elif args.pseudo == "lstm":
print("Pseudo documents generation (Method: LSTM language model)...")
lm = train_lstm(sequences, common_words, sent_length, f'./{dataset}/lm', embedding_matrix=class_tree.embedding)
seed_docs, seed_label = lstm_pseudodocs(parent, expand_num, doc_length, len_avg, sent_length, len_std, num_doc,
interp_weight, vocabulary_inv, lm, common_words, save_dir)
print("Finished pseudo documents generation.")
num_real_doc = len(seed_docs) / 5
if sup_source == 'docs':
real_seed_docs, real_seed_label = augment(x, parent.children, num_real_doc)
print(f'Labeled docs {len(real_seed_docs)} + Pseudo docs {len(seed_docs)}')
seed_docs = np.concatenate((seed_docs, real_seed_docs), axis=0)
seed_label = np.concatenate((seed_label, real_seed_label), axis=0)
perm = np.random.permutation(len(seed_label))
seed_docs = seed_docs[perm]
seed_label = seed_label[perm]
print('\n### Phase 2: pre-training with pseudo documents ###')
print(f'Pretraining node {parent.name}')
wstc.pretrain(x=seed_docs, pretrain_labels=seed_label, model=parent.model,
optimizer=SGD(lr=0.1, momentum=0.9),
epochs=pretrain_epochs, batch_size=batch_size,
save_dir=save_dir, suffix=parent.name)
global_classifier = wstc.ensemble_classifier(level)
wstc.model.append(global_classifier)
t0 = time()
print("\n### Phase 3: self-training ###")
selftrain_optimizer = SGD(lr=self_lr, momentum=0.9, decay=decay)
wstc.compile(level, optimizer=selftrain_optimizer, loss='kld')
y_pred = wstc.fit(x, level=level, tol=delta, maxiter=maxiter, batch_size=batch_size,
update_interval=update_interval, save_dir=save_dir)
print(f'Self-training time: {time() - t0:.2f}s')
return y_pred
