len(dataset.out_vocab[0]),
args.n_layers,
args.dropout,
)
# Initialize optimizers and criterion
# encoder_optimizer = optim.Adam(encoder.parameters(), lr=args.learning_rate)
# decoder_optimizer = optim.Adam(decoder.parameters(), lr=args.learning_rate * decoder_learning_ratio)
encoder_optimizer = optim.Adadelta(encoder.parameters())
decoder_optimizer = optim.Adadelta(decoder.parameters())
criterion = nn.CrossEntropyLoss()
# Move models to GPU
if args.USE_CUDA:
encoder.cuda()
decoder.cuda()
# train(dataset,
# args.batch_size,
# args.n_epochs,
# encoder,
# decoder,
# encoder_optimizer,
# decoder_optimizer,
# criterion,
# 'checkpoints/pov',
# lang)
# evaluate
# find the last encoder state
def train(x,
y,
optimizer=optim.Adam,
criterion=nn.MSELoss(),
n_steps=100,
attn_model="general",
hidden_size=128,
n_layers=1,
dropout=0,
batch_size=50,
elr=0.001,
dlr=0.005,
clip=50.0,
print_every=10,
teacher_forcing_ratio=lambda x: 1 if x < 10 else 0):
# Configure training/optimization
encoder_learning_rate = elr
decoder_learning_ratio = dlr
# Initialize models
encoder = EncoderRNN(1, hidden_size, n_layers, dropout=dropout)
decoder = LuongAttnDecoderRNN(attn_model,
1,
hidden_size,
n_layers,
dropout=dropout)
# Initialize optimizers and criterion
encoder_optimizer = optimizer(encoder.parameters(),
lr=encoder_learning_rate)
decoder_optimizer = optimizer(decoder.parameters(),
lr=decoder_learning_ratio)
# Move models to GPU
if USE_CUDA:
encoder.cuda()
decoder.cuda()
# Begin!
print_loss_total = 0
step = 0
while step < n_steps:
step += 1
# Get training data for this cycle
batch_idx = np.random.randint(0, x.shape[1], batch_size)
input_batches, target_batches = x[:, batch_idx], y[:, batch_idx]
# Run the train function
loss, _ = _train(input_batches,
target_batches,
encoder,
decoder,
encoder_optimizer,
decoder_optimizer,
criterion,
teacher_forcing_ratio=teacher_forcing_ratio(step),
clip=clip)
# print(np.mean(np.square((output.data.cpu().numpy() - series[-20:, batch_idx]))))
# Keep track of loss
print_loss_total += loss
if step % print_every == 0:
print_loss_avg = print_loss_total / print_every
print_loss_total = 0
print_summary = '(%d %d%%) %.4f' % (step, step / n_steps * 100,
print_loss_avg)
print(print_summary)
return encoder, decoder