def eval_step(input_ids,
target_ids,
):
target_inp = target_ids[:, :-1]
_, combined_mask, dec_padding_mask = create_masks(input_ids, target_inp)
(draft_predictions, draft_attention_weights,
refine_predictions, refine_attention_weights) = Model(
input_ids,
dec_padding_mask=dec_padding_mask,
target_ids=target_inp,
look_ahead_mask=combined_mask,
training=False
)
loss, target = loss_function(target_ids,
draft_predictions,
refine_predictions,
Model
)
train_loss(loss)
log.info(Model.summary())
if config.save_initial_weights:
initial_weights = os.path.join(config.initial_weights,'initial_weights')
Model.save_weights(initial_weights)
return loss
def batch_run_check(batch, start):
if config.run_tensorboard:
with train_output_sequence_writer.as_default():
tf.summary.scalar('train_loss', train_loss.result(), step=batch)
tf.summary.scalar('train_accuracy',
train_accuracy.result(),
step=batch)
if config.display_model_summary:
log.info(Model.summary())
log.info(batch_zero.format(time.time() - start))
config.display_model_summary = False
log.info(
batch_run_details.format(train_loss.result(), train_accuracy.result()))
return train_loss.result()
def batch_run_check(batch, start_time, bert_f1_score):
if config.run_tensorboard:
with train_output_sequence_writer.as_default():
tf.summary.scalar('train_loss', train_loss.result(), step=batch)
if config.display_model_summary:
log.info(Model.summary())
log.info(batch_zero.format(time.time() - start_time))
config.display_model_summary = False
log.info(
batch_run_details.format(
tf.debugging.assert_all_finite(train_loss.result(),
message="NaN's or Inf's.",
name='NAN_assertion'),
bert_f1_score.numpy()))
def eval_step(input_ids, target_ids_, target_ids, draft_mask, refine_mask):
(draft_predictions, draft_attention_weights, refine_predictions,
refine_attention_weights) = Model(input_ids, target_ids_, False)
draft_output_sequence_loss = loss_function(target_ids[:, 1:, :],
draft_predictions, draft_mask)
if config.use_refine_decoder:
refine_output_sequence_loss = loss_function(target_ids[:, :-1, :],
refine_predictions,
refine_mask)
else:
refine_output_sequence_loss = 0
regularization_loss = tf.add_n(Model.losses)
loss = draft_output_sequence_loss + refine_output_sequence_loss + regularization_loss
log.info(Model.summary())
if config.save_initial_weights:
initial_weights = os.path.join(config.initial_weights,
'initial_weights')
Model.save_weights(initial_weights)
return loss