def update(self, batch, eval=False):
inputs, orig_idx = unpack_batch(batch, self.use_cuda)
src, src_mask, tgt_in, tgt_out, pos, edits = inputs
if eval:
self.model.eval()
else:
self.model.train()
self.optimizer.zero_grad()
log_probs, edit_logits = self.model(src, src_mask, tgt_in, pos)
if self.args.get('edit', False):
assert edit_logits is not None
loss = self.crit(log_probs.view(-1, self.vocab['char'].size), tgt_out.view(-1), \
edit_logits, edits)
else:
loss = self.crit(log_probs.view(-1, self.vocab['char'].size),
tgt_out.view(-1))
loss_val = loss.data.item()
if eval:
return loss_val
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
self.args['max_grad_norm'])
self.optimizer.step()
return loss_val
def update(self, batch, eval=False):
inputs, orig_idx, word_orig_idx, sentlens, wordlens = unpack_batch(
batch, self.use_cuda)
word, word_mask, wordchars, wordchars_mask, upos, xpos, ufeats, pretrained, lemma, head, deprel = inputs
if eval:
self.model.eval()
else:
self.model.train()
self.optimizer.zero_grad()
print("In trainer.update, setting model.training to TRUE")
self.model.training = True
print("model.training: ", self.model.training)
loss, _ = self.model(word, word_mask, wordchars, wordchars_mask, upos,
xpos, ufeats, pretrained, lemma, head, deprel,
word_orig_idx, sentlens, wordlens)
loss_val = loss.data.item()
if eval:
return loss_val
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
self.args['max_grad_norm'])
self.optimizer.step()
return loss_val
def update(self, batch, eval=False):
inputs, orig_idx = unpack_batch(batch, self.use_cuda)
src, src_mask, tgt_in, tgt_out = inputs
if eval:
self.model.eval()
else:
self.model.train()
self.optimizer.zero_grad()
log_probs, _ = self.model(src, src_mask, tgt_in)
loss = self.crit(log_probs.view(-1, self.vocab.size), tgt_out.view(-1))
loss_val = loss.data.item()
if eval:
return loss_val
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
self.args['max_grad_norm'])
self.optimizer.step()
return loss_val
def update(self, batch, eval=False):
inputs, orig_idx, word_orig_idx, sentlens, wordlens = unpack_batch(batch, self.use_cuda)
word, word_mask, wordchars, wordchars_mask, upos, xpos, ufeats, pretrained, lemma, next_word, prev_word = inputs
if eval:
self.model.eval()
else:
self.model.train()
self.optimizer.zero_grad()
loss, preds = self.model(word, word_mask, wordchars, wordchars_mask, upos, xpos, ufeats,
pretrained, lemma, next_word, prev_word, word_orig_idx, sentlens, wordlens)
loss_val = loss.data.item()
if eval:
return loss_val
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.args['max_grad_norm'])
self.optimizer.step()
return loss_val
def update(self, batch, eval=False):
inputs, orig_idx, word_orig_idx, char_orig_idx, sentlens, wordlens, charlens, charoffsets = unpack_batch(
batch, self.use_cuda)
word, word_mask, wordchars, wordchars_mask, chars, tags = inputs
if eval:
self.model.eval()
else:
self.model.train()
self.optimizer.zero_grad()
loss, _, _ = self.model(word, word_mask, wordchars, wordchars_mask,
tags, word_orig_idx, sentlens, wordlens, chars,
charoffsets, charlens, char_orig_idx)
loss_val = loss.data.item()
if eval:
return loss_val
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
self.args['max_grad_norm'])
self.optimizer.step()
return loss_val
def update(self, dp_batch, lm_batch, eval=False):
inputs, orig_idx, word_orig_idx, sentlens, wordlens = depparse_unpack(
dp_batch, self.use_cuda)
word, word_mask, wordchars, wordchars_mask, upos, xpos, ufeats, pretrained, lemma, head, deprel = inputs
if eval:
self.model.eval()
else:
self.model.train()
self.optimizer.zero_grad()
depparse_loss, _ = self.model(word, word_mask, wordchars,
wordchars_mask, upos, xpos, ufeats,
pretrained, lemma, head, deprel,
word_orig_idx, sentlens, wordlens)
inputs, orig_idx, word_orig_idx, sentlens, wordlens = lm_unpack(
lm_batch, self.use_cuda)
word, word_mask, wordchars, wordchars_mask, upos, xpos, ufeats, pretrained, lemma, next_word, prev_word = inputs
lm_loss, _ = self.model.lm_forward(word, word_mask, wordchars,
wordchars_mask, upos, xpos, ufeats,
pretrained, lemma, next_word,
prev_word, word_orig_idx, sentlens,
wordlens)
loss = depparse_loss * self.args['lambd'] + lm_loss
# print(np.exp(lm_loss.item()))
if not eval:
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
self.args['max_grad_norm'])
self.optimizer.step()
depparse_loss = depparse_loss.item()
lm_loss = lm_loss.item()
loss = loss.item()
return depparse_loss, lm_loss, loss
