def val_step(inp, tar, epoch, inp_shape, tar_shape, batch, create_summ):
tar_inp = tar[:, :-1]
tar_real = tar[:, 1:]
enc_padding_mask, combined_mask, dec_padding_mask = create_masks(
inp, tar_inp)
predictions, attention_weights, dec_output = transformer(
inp, tar_inp, False, enc_padding_mask, combined_mask, dec_padding_mask)
if config.copy_gen:
predictions = pointer_generator(dec_output,
predictions,
attention_weights,
inp,
inp_shape,
tar_shape,
batch,
training=False)
loss = loss_function(tar_real, predictions)
validation_loss(loss)
validation_accuracy(tar_real, predictions)
if create_summ:
rouge, bert = tf_write_summary(tar_real, predictions, inp[:, 1:],
epoch)
else:
rouge, bert = (1.0, 1.0)
return (rouge, bert)
def val_step(input_ids,
input_mask,
input_segment_ids,
target_ids_,
target_mask,
target_segment_ids,
target_ids,
draft_mask,
refine_mask,
step,
create_summ):
(draft_predictions, draft_attention_weights,
refine_predictions, refine_attention_weights) = model(
input_ids, input_mask, input_segment_ids,
target_ids_, target_mask, target_segment_ids,
False
)
draft_summary_loss = loss_function(target_ids[:, 1:, :], draft_predictions, draft_mask)
refine_summary_loss = loss_function(target_ids[:, :-1, :], refine_predictions, refine_mask)
loss = draft_summary_loss + refine_summary_loss
loss = tf.reduce_mean(loss)
validation_loss(loss)
validation_accuracy(target_ids_[:, :-1], refine_predictions)
if create_summ:
rouge, bert = tf_write_summary(target_ids_[:, :-1], refine_predictions, step)
else:
rouge, bert = (1.0, 1.0)
return (rouge, bert)
def val_step(inp, tar, epoch, create_summ):
draft_predictions, draft_attention_weights, draft_dec_output = draft_summary_model(
inp, tar, training=False)
refine_predictions, refine_attention_weights, refine_dec_output = refine_summary_model(
inp, tar, training=False)
draft_summary_loss = loss_function(tar[0][:, 1:, :], draft_predictions)
refine_summary_loss = loss_function(tar[0][:, :-1, :], refine_predictions)
loss = draft_summary_loss + refine_summary_loss
validation_loss(loss)
validation_accuracy(tar_real, predictions)
if create_summ:
rouge, bert = tf_write_summary(tar_real, predictions, inp[:, 1:],
epoch)
else:
rouge, bert = (1.0, 1.0)
return (rouge, bert)
def val_step(input_ids, target_ids_, step, create_summ):
validation_accuracy.reset_states()
from decode_text import predict_using_sampling
(preds_draft_summary, _, refine_predictions,
_) = predict_using_sampling(input_ids,
refine_decoder_sampling_type='greedy',
temperature=0.9,
p=0.8,
k=7)
validation_accuracy(
target_ids_[:, 1:],
tf.one_hot(refine_predictions[:, 1:], depth=config.input_vocab_size))
if create_summ:
rouge, bert = tf_write_summary(target_ids_[:, 1:],
refine_predictions[:, 1:], step)
else:
rouge, bert = (1.0, 1.0)
return (rouge, bert)
def val_step(inp, tar, epoch, create_summ):
tar_inp = tar[:, :-1]
tar_real = tar[:, 1:]
enc_padding_mask, combined_mask, dec_padding_mask = create_masks(inp, tar_inp)
predictions, attention_weights, dec_output = model(
inp,
tar_inp,
enc_padding_mask,
combined_mask,
dec_padding_mask,
training=False
)
loss = loss_function(tar_real, predictions)
validation_loss(loss)
validation_accuracy(tar_real, predictions)
if create_summ:
rouge, bert = tf_write_summary(tar_real, predictions, inp[:, 1:], epoch)
else:
rouge, bert = (1.0, 1.0)
return (rouge, bert)
