embedding_size = embedding.embedding_matrix.shape[1]
print('Building model.')
model = Manhattan_LSTM(args.data_name,
args.hidden_size,
embedding.embedding_matrix,
use_embedding=True,
train_embedding=True)
if use_cuda: model = model.cuda()
model.init_weights()
print("Training Network.")
train_network = Train_Network(model, data.index2word)
run_iterations = Run_Iterations(args.data_name,
train_network,
data.x_train,
data.y_train,
data.index2word,
args.batch_size,
args.num_iters,
args.learning_rate,
tracking_pair=args.tracking_pair,
x_val=data.x_val,
y_val=data.y_val)
run_iterations.train_iters()
run_iterations.evaluate_randomly()
torch.save(model.state_dict(), './manhattan_lstm.pt')
if use_cuda:
encoder = encoder.cuda()
decoder = decoder.cuda()
print("Training Network.")
train_network = Train_Network(encoder,
decoder,
index2word,
num_layers=num_layers)
run_iterations = Run_Iterations(train_network,
train_in_seq,
train_out_seq,
train_lengths,
train_speakers,
train_addressees,
index2word,
batch_size,
num_iters,
learning_rate,
tracking_pair=tracking_pair,
dev_in_seq=dev_in_seq,
dev_out_seq=dev_out_seq,
dev_input_lengths=dev_lengths,
dev_speakers=dev_speakers,
dev_addressees=dev_addressees)
run_iterations.train_iters()
run_iterations.evaluate_randomly()
decoder = decoder.cuda()
print("Training Network.")
train_network = Train_Network(encoder,
decoder,
data_p.index2word,
num_layers=args.num_layers)
run_iterations = Run_Iterations(train_network,
data_p.x_train,
data_p.y_train,
data_p.train_lengths,
data_p.index2word,
args.batch_size,
args.num_iters,
args.learning_rate,
fold_size=args.fold_size,
track_minor=args.track_minor,
tracking_pair=args.tracking_pair,
val_in_seq=data_p.x_val,
val_out_seq=data_p.y_val,
val_lengths=data_p.val_lengths)
run_iterations.train_iters()
run_iterations.evaluate_randomly()
if not args.encoder_parameters.endswith('.pt'):
args.encoder_parameters += '.pt'
if not args.decoder_parameters.endswith('.pt'):
args.decoder_parameters += '.pt'
torch.save(encoder.state_dict(), args.encoder_parameters)
print('--------------------------------------\n')
print('Reading Data.')
data = Data(args.data_name, args.data_file, args.training_ratio, args.max_len)
print('\n')
print('Number of training samples :', len(data.x_train))
print('Number of validation samples :', len(data.x_val))
print('Maximum sequence length :', args.max_len)
print('\n')
print('Building Embedding Matrix')
embedding = Get_Embedding(args.embd_file, data.word2index,args)
embedding_size = embedding.embedding_matrix.shape[1]
print('Building model.')
model = Manhattan_LSTM(embedding.embedding_matrix, args)
if use_cuda: model = model.cuda()
model.init_weights() # set weights to be same in lstm1 and lstm2
print("Training Network.")
train_network = Train_Network(model, data.index2word)
run_iterations = Run_Iterations(train_network, data.x_train, data.x_train_char, data.y_train, data.index2word,
args.batch_size, args.num_iters, args.learning_rate,
tracking_pair=args.tracking_pair, x_val=data.x_val, x_val_char = data.x_val_char, y_val=data.y_val)
run_iterations.train_iters()
#run_iterations.evaluate_randomly()
torch.save(model.state_dict(), './manhattan_lstm.pt')