def compute_dev_loss(self):
loss_builder = LossBuilder()
trg_words_cnt = 0
for src, trg in zip(self.dev_src, self.dev_trg):
dy.renew_cg()
standard_loss = self.model.calc_loss(src, trg)
loss_builder.add_loss("loss", standard_loss)
trg_words_cnt += self.logger.count_trg_words(trg)
loss_builder.compute()
return trg_words_cnt, LossScore(loss_builder.sum() / trg_words_cnt)
def training_step(self, src, trg):
"""
Performs forward pass, backward pass, parameter update for the given minibatch
"""
loss_builder = LossBuilder()
standard_loss = self.model.calc_loss(src, trg, self.loss_calculator)
if standard_loss.__class__ == LossBuilder:
loss = None
for loss_name, loss_expr in standard_loss.loss_nodes:
loss_builder.add_loss(loss_name, loss_expr)
loss = loss_expr if not loss else loss + loss_expr
standard_loss = loss
else:
loss_builder.add_loss("loss", standard_loss)
additional_loss = self.model.calc_additional_loss(
dy.nobackprop(-standard_loss))
if additional_loss != None:
loss_builder.add_loss("additional_loss", additional_loss)
loss_value = loss_builder.compute()
self.logger.update_epoch_loss(src, trg, loss_builder)
self.logger.report_train_process()
return loss_value
def one_epoch(self, update_weights=True):
"""
:param update_weights: Whether to perform backward pass & update weights (useful for debugging)
"""
self.logger.new_epoch()
if self.args["reload_command"] is not None:
self._augment_data_next_epoch()
self.model.set_train(update_weights)
order = list(range(0, len(self.train_src)))
np.random.shuffle(order)
for batch_num in order:
src = self.train_src[batch_num]
trg = self.train_trg[batch_num]
# Loss calculation
dy.renew_cg()
loss_builder = LossBuilder()
standard_loss = self.model.calc_loss(src, trg)
if standard_loss.__class__ == LossBuilder:
loss = None
for loss_name, loss_expr in standard_loss.loss_nodes:
loss_builder.add_loss(loss_name, loss_expr)
loss = loss_expr if not loss else loss + loss_expr
standard_loss = loss
else:
loss_builder.add_loss("loss", standard_loss)
additional_loss = self.model.calc_additional_loss(
dy.nobackprop(-standard_loss))
if additional_loss != None:
loss_builder.add_loss("additional_loss", additional_loss)
# Log the loss sum
loss_value = loss_builder.compute()
self.logger.update_epoch_loss(src, trg, loss_builder)
if update_weights:
loss_value.backward()
self.trainer.update()
# Devel reporting
self.logger.report_train_process()
if self.logger.should_report_dev():
self.dev_evaluation()
self.model.new_epoch()
def training_step(self, src, trg):
"""
Performs forward pass, backward pass, parameter update for the given minibatch
"""
loss_builder = LossBuilder()
standard_loss = self.model.calc_loss(src, trg, self.loss_calculator)
additional_loss = self.model.calc_additional_loss(standard_loss)
loss_builder.add_loss("standard_loss", standard_loss)
loss_builder.add_loss("additional_loss", additional_loss)
loss_value = loss_builder.compute()
self.logger.update_epoch_loss(src, trg, loss_builder.get_loss_stats())
self.logger.report_train_process()
return loss_value
