vectorizer = dataset.vectorizer
# Create embeddings
embeddings = None
if args.use_med_embeddings:
words = vectorizer.paper_vocab.token_to_idx.keys()
embeddings = make_embeddings_matrix(words=words)
print ("<Embeddings(words={0}, dim={1})>".format(np.shape(embeddings)[0], np.shape(embeddings)[1]))
del(med_w2v)
# Initialize model
model = PapersModel(embedding_dim=args.embedding_dim,
num_embeddings=len(vectorizer.paper_vocab),
num_input_channels=args.embedding_dim, filter_sizes=args.filter_sizes,
num_channels=args.num_filters, hidden_dim=args.hidden_dim,
num_classes=len(vectorizer.label_vocab),
dropout_p=args.dropout_p, pretrained_embeddings=embeddings, # pretrained_embeddings=None,
padding_idx=vectorizer.paper_vocab.mask_index)
print(model.named_modules)
#%% Train
trainer = Trainer(dataset=dataset, model=model,
model_state_file=args.model_state_file,
save_dir=args.save_dir,
device=args.device,
shuffle=args.shuffle, num_epochs=args.num_epochs,
batch_size=args.batch_size, learning_rate=args.learning_rate,
early_stopping_criteria=args.early_stopping_criteria)
print("Using CUDA: {}".format(args.cuda))
#%% Using embeddings
#%% Initialization
dataset = PapersDataset.load_dataset_and_make_vectorizer(df=split_df, cutoff=args.cutoff)
dataset.save_vectorizer(args.vectorizer_file)
vectorizer = dataset.vectorizer
model = PapersModel(embedding_dim=args.embedding_dim,
num_embeddings=len(vectorizer.title_vocab),
num_input_channels=args.embedding_dim,
num_channels=args.num_filters, hidden_dim=args.hidden_dim,
num_classes=len(vectorizer.category_vocab),
dropout_p=args.dropout_p, pretrained_embeddings=None,
padding_idx=vectorizer.title_vocab.mask_index)
print (model.named_modules)
#%% Train
trainer = Trainer(dataset=dataset, model=model,
model_state_file=args.model_state_file,
save_dir=args.save_dir, device=args.device,
shuffle=args.shuffle, num_epochs=args.num_epochs,
batch_size=args.batch_size, learning_rate=args.learning_rate,
early_stopping_criteria=args.early_stopping_criteria)
trainer.run_train_loop()
embeddings = None
if args.use_med_embeddings:
words = vectorizer.paper_vocab.token_to_idx.keys()
embeddings = make_embeddings_matrix(words=words)
print("<Embeddings(words={0}, dim={1})>".format(
np.shape(embeddings)[0],
np.shape(embeddings)[1]))
del (med_w2v)
# Initialize model
model = PapersModel(embedding_dim=args.embedding_dim,
num_embeddings=len(vectorizer.paper_vocab),
rnn_hidden_dim=args.rnn_hidden_dim,
hidden_dim=args.hidden_dim,
output_dim=len(vectorizer.label_vocab),
num_layers=args.num_layers,
bidirectional=args.bidirectional,
dropout_p=args.dropout_p,
pretrained_embeddings=embeddings,
padding_idx=vectorizer.paper_vocab.mask_index)
print(model.named_modules)
###### Train ######
trainer = Trainer(dataset=dataset,
model=model,
model_state_file=args.model_state_file,
save_dir=args.save_dir,
device=args.device,
shuffle=args.shuffle,
num_epochs=args.num_epochs,
batch_size=args.batch_size,