'''
loss_function = nn.NLLLoss(reduction = 'sum', ignore_index = de_vocab.item2index['_PAD_'])
en_optimizer = optim.Adam(encoder.parameters(), lr = 1e-3, weight_decay = 0)
de_optimizer = optim.Adam(decoder.parameters(), lr = 1e-3, weight_decay = 0)
if use_cuda:
encoder = encoder.cuda()
decoder = decoder.cuda()
ones_matrix = ones_matrix.cuda()
loss_function = loss_function.cuda()
for epoch in range(20):
pl.reset()
encoder.train()
decoder.train()
total_loss = torch.Tensor([0])
total_token = 0
for batch_idx, (en_seq, en_seq_len, de_seq, de_seq_len) in enumerate(pl.gen_pairs(batch_size)):
en_optimizer.zero_grad()
de_optimizer.zero_grad()
en_seq = torch.LongTensor(en_seq)
de_seq = torch.LongTensor(de_seq)
if use_cuda:
en_seq = en_seq.cuda()
de_seq = de_seq.cuda()
tst_tgt_t = torch.LongTensor(tst_tgt_p)
trn_src_t = torch.LongTensor(trn_src_p)
trn_tgt_t = torch.LongTensor(trn_tgt_p)
enc = Encoder(len(vocab), 100, 100, 2, 'cuda', vocab[pad])
dec = Decoder(len(vocab), 100, 100, 2, 'cuda', vocab[pad], vocab[sos],
vocab[eos], vocab[unk])
enc.to('cuda')
dec.to('cuda')
opt_enc = torch.optim.Adam(enc.parameters())
opt_dec = torch.optim.Adam(dec.parameters())
n_batch = len(trn_src_p) // batch_size
for e in range(epochs):
enc.train()
dec.train()
epoch_loss = 0
for i in range(n_batch):
opt_enc.zero_grad()
opt_dec.zero_grad()
lengths = torch.LongTensor(l_trn_src[batch_size * i:batch_size *
(i + 1)])
out, h_n = enc(trn_src_t[batch_size * i:batch_size * (i + 1)], lengths)
output = dec.teacher_force(
trn_tgt_t[batch_size * i:batch_size * (i + 1)].reshape(
[batch_size, tgt_max, 1]), h_n,
torch.LongTensor(l_trn_tgt[batch_size * i:batch_size * (i + 1)]))
loss = 0
for o, l, t in zip(output,
l_trn_tgt[batch_size * i:batch_size * (i + 1)],