def calc_loss(self, translator, src, trg):
search_outputs = translator.generate_search_output(src, self.search_strategy)
sign = -1 if self.inv_eval else 1
total_loss = FactoredLossExpr()
for search_output in search_outputs:
self.eval_score = []
for trg_i, sample_i in zip(trg, search_output.word_ids):
# Removing EOS
sample_i = self.remove_eos(sample_i.tolist())
ref_i = trg_i.words[:trg_i.len_unpadded()]
score = self.evaluation_metric.evaluate_one_sent(ref_i, sample_i)
self.eval_score.append(sign * score)
self.reward = dy.inputTensor(self.eval_score, batched=True)
# Composing losses
loss = FactoredLossExpr()
if self.baseline is not None:
baseline_loss = []
losses = []
for state, logsoft, mask in zip(search_output.state,
search_output.logsoftmaxes,
search_output.mask):
bs_score = self.baseline.transform(state)
baseline_loss.append(dy.squared_distance(self.reward, bs_score))
loss_i = dy.cmult(logsoft, self.reward - bs_score)
valid = list(np.nonzero(mask)[0])
losses.append(dy.cmult(loss_i, dy.inputTensor(mask, batched=True)))
loss.add_loss("reinforce", dy.sum_elems(dy.esum(losses)))
loss.add_loss("reinf_baseline", dy.sum_elems(dy.esum(baseline_loss)))
else:
loss.add_loss("reinforce", dy.sum_elems(dy.cmult(self.true_score, dy.esum(logsofts))))
total_loss.add_factored_loss_expr(loss)
return loss
def calc_loss(self,
model: 'model_base.ConditionedModel',
src: Union[sent.Sentence, 'batchers.Batch'],
trg: Union[sent.Sentence, 'batchers.Batch']):
total_loss = FactoredLossExpr()
for loss, weight in zip(self.losses, self.loss_weight):
total_loss.add_factored_loss_expr(loss.calc_loss(model, src, trg) * weight)
return total_loss
def calc_loss(self, src, trg, infer_prediction=False):
event_trigger.start_sent(src)
if not batchers.is_batched(src):
src = batchers.mark_as_batch([src])
if not batchers.is_batched(trg):
trg = batchers.mark_as_batch([trg])
src_words = np.array([[Vocab.SS] + x.words for x in src])
batch_size, src_len = src_words.shape
if isinstance(src.mask, type(None)):
src_mask = np.zeros((batch_size, src_len), dtype=np.int)
else:
src_mask = np.concatenate([
np.zeros((batch_size, 1), dtype=np.int),
src.mask.np_arr.astype(np.int)
],
axis=1)
src_embeddings = self.sentence_block_embed(
self.src_embedder.embeddings, src_words, src_mask)
src_embeddings = self.make_input_embedding(src_embeddings, src_len)
trg_words = np.array(
list(map(lambda x: [Vocab.SS] + x.words[:-1], trg)))
batch_size, trg_len = trg_words.shape
if isinstance(trg.mask, type(None)):
trg_mask = np.zeros((batch_size, trg_len), dtype=np.int)
else:
trg_mask = trg.mask.np_arr.astype(np.int)
trg_embeddings = self.sentence_block_embed(
self.trg_embedder.embeddings, trg_words, trg_mask)
trg_embeddings = self.make_input_embedding(trg_embeddings, trg_len)
xx_mask = self.make_attention_mask(src_mask, src_mask)
xy_mask = self.make_attention_mask(trg_mask, src_mask)
yy_mask = self.make_attention_mask(trg_mask, trg_mask)
yy_mask *= self.make_history_mask(trg_mask)
z_blocks = self.encoder.transduce(src_embeddings, xx_mask)
h_block = self.decoder(trg_embeddings, z_blocks, xy_mask, yy_mask)
if infer_prediction:
y_len = h_block.dim()[0][1]
last_col = dy.pick(h_block, dim=1, index=y_len - 1)
logits = self.decoder.output(last_col)
return logits
ref_list = list(
itertools.chain.from_iterable(map(lambda x: x.words, trg)))
concat_t_block = (1 -
trg_mask.ravel()).reshape(-1) * np.array(ref_list)
loss = self.decoder.output_and_loss(h_block, concat_t_block)
return FactoredLossExpr({"mle": loss})
def calc_loss(self,
model: 'model_base.ConditionedModel',
src: Union[sent.Sentence, 'batcher.Batch'],
trg: Union[sent.Sentence, 'batcher.Batch']):
loss_builder = FactoredLossExpr()
for _ in range(self.repeat):
standard_loss = self.child_loss.calc_loss(model, src, trg)
additional_loss = event_trigger.calc_additional_loss(trg, model, standard_loss)
loss_builder.add_factored_loss_expr(standard_loss)
loss_builder.add_factored_loss_expr(additional_loss)
return loss_builder
def calc_nll(self, src: Union[batchers.Batch, sent.Sentence], trg: Union[batchers.Batch, sent.Sentence]) -> dy.Expression:
sub_losses = collections.defaultdict(list)
for model in self.models:
for loss_name, loss in model.calc_nll(src, trg).expr_factors.items():
sub_losses[loss_name].append(loss)
model_loss = FactoredLossExpr()
for loss_name, losslist in sub_losses.items():
# TODO: dy.average(losslist) _or_ dy.esum(losslist) / len(self.models) ?
# -- might not be the same if not all models return all losses
model_loss.add_loss(loss_name, dy.average(losslist))
return model_loss
def calc_loss(self,
model: 'model_base.ConditionedModel',
src: Union[sent.Sentence, 'batchers.Batch'],
trg: Union[sent.Sentence, 'batchers.Batch']):
assert hasattr(model, "attender") and hasattr(model.attender, "attention_vecs"), \
"Must be called after MLELoss with models that have attender."
masked_attn = model.attender.attention_vecs
if trg.mask is not None:
trg_mask = 1-(trg.mask.np_arr.transpose())
masked_attn = [dy.cmult(attn, dy.inputTensor(mask, batched=True)) for attn, mask in zip(masked_attn, trg_mask)]
loss = self.global_fertility(masked_attn)
return FactoredLossExpr({"global_fertility": loss})
def calc_loss(self, translator, src, trg):
batch_size = trg.batch_size()
uniques = [set() for _ in range(batch_size)]
deltas = []
probs = []
sign = -1 if self.inv_eval else 1
search_outputs = translator.generate_search_output(
src, self.search_strategy)
for search_output in search_outputs:
logprob = search_output.logsoftmaxes
sample = search_output.word_ids
attentions = search_output.attentions
logprob = dy.esum(logprob) * self.alpha
# Calculate the evaluation score
eval_score = np.zeros(batch_size, dtype=float)
mask = np.zeros(batch_size, dtype=float)
for j in range(batch_size):
ref_j = self.remove_eos(trg[j].words)
hyp_j = self.remove_eos(sample[j].tolist())
if self.unique_sample:
hash_val = hash(tuple(hyp_j))
if len(hyp_j) == 0 or hash_val in uniques[j]:
mask[j] = -1e20 # represents negative infinity
continue
else:
uniques[j].add(hash_val)
# Calc evaluation score
eval_score[j] = self.evaluation_metric.evaluate_one_sent(
ref_j, hyp_j) * sign
# Appending the delta and logprob of this sample
prob = logprob + dy.inputTensor(mask, batched=True)
deltas.append(dy.inputTensor(eval_score, batched=True))
probs.append(prob)
sample_prob = dy.softmax(dy.concatenate(probs))
deltas = dy.concatenate(deltas)
risk = dy.sum_elems(dy.cmult(sample_prob, deltas))
### Debug
#print(sample_prob.npvalue().transpose()[0])
#print(deltas.npvalue().transpose()[0])
#print("----------------------")
### End debug
return FactoredLossExpr({"risk": risk})
def on_calc_additional_loss(self, trg, generator, generator_loss):
if self.policy_learning is None:
return None
reward = FactoredLossExpr()
reward.add_loss("generator",
-dy.inputTensor(generator_loss.value(), batched=True))
if self.length_prior is not None:
reward.add_loss('length_prior',
self.length_prior.log_ll(self.seg_size_unpadded))
reward_value = reward.value()
if trg.batch_size() == 1:
reward_value = [reward_value]
reward_tensor = dy.inputTensor(reward_value, batched=True)
### Calculate losses
try:
return self.policy_learning.calc_loss(reward_tensor)
finally:
self.reward = reward
if self.train and self.reporter is not None:
self.reporter.report_process(self)
def on_calc_additional_loss(self, trg, generator, generator_loss):
if self.policy_learning is None:
return None
trg_counts = dy.inputTensor([t.len_unpadded() for t in trg],
batched=True)
reward = FactoredLossExpr()
# Adding all reward from the translator
for loss_key, loss_value in generator_loss.get_nobackprop_loss().items(
):
if loss_key == 'mle':
reward.add_loss('mle', dy.cdiv(-loss_value, trg_counts))
else:
reward.add_loss(loss_key, -loss_value)
if self.length_prior is not None:
reward.add_loss('seg_lp',
self.length_prior.log_ll(self.seg_size_unpadded))
reward = dy.inputTensor(reward.value(), batched=True)
### Calculate losses
try:
return self.policy_learning.calc_loss(reward)
finally:
self.reward = reward
if self.reporter is not None:
self.reporter.report_process(self)
def calc_loss(self,
model: 'model_base.ConditionedModel',
src: Union[sent.Sentence, 'batchers.Batch'],
trg: Union[sent.Sentence, 'batchers.Batch']):
loss = model.calc_nll(src, trg)
return FactoredLossExpr({"mle": loss})