hidden_size=config.hidden_size,
n_layers=config.n_layers,
dropout_p=config.dropout,
max_length=config.max_length)
# Let criterion cannot count PAD as right prediction, because PAD is easy to predict.
loss_weight = torch.ones(len(loader.text.vocab))
loss_weight[data_loader.PAD] = 0
criterion = nn.NLLLoss(weight=loss_weight, size_average=False)
print(model)
print(criterion)
if config.gpu_id >= 0:
model.cuda(config.gpu_id)
criterion.cuda(config.gpu_id)
if config.n_epochs > 0:
trainer.train_epoch(model, criterion, loader.train_iter,
loader.valid_iter, config)
if config.gen > 0:
total_gen = 0
while total_gen < config.gen:
current_gen = min(config.batch_size, config.gen - total_gen)
_, indice = model.search(batch_size=current_gen)
total_gen += current_gen
lines = to_text(indice, loader.text.vocab)
print('\n'.join(lines))
