def test(self):
cumulative_metrics = Metrics.empty(mode="test")
for update_info in self.test_updates():
cumulative_metrics += update_info.metrics
print(cumulative_metrics)
def train(self, evaluate_every=100):
test_updates = self.test_updates()
cumulative_train_metrics = Metrics.empty(mode="train")
cumulative_evaluate_metrics = Metrics.empty(mode="eval")
for update_info in self.train_and_evaluate_updates(
evaluate_every=evaluate_every):
if update_info.from_train:
cumulative_train_metrics += update_info.metrics
print(
f"{update_info.batch.ix} \t| {update_info.metrics.loss} \t= {update_info.model_loss} \t+ {update_info.fooling_loss} \t| {update_info.discriminator_loss}"
)
if update_info.batch.ix % 200 == 0:
with torch.no_grad():
predicted = update_info.decoded_inferred_texts[
0].replace('\n', ' ').strip('❟ ❟ ❟')
headline = update_info.batch.orig_headline[0].replace(
'\n', ' ').lower().strip()
text = update_info.batch.orig_text[0].replace(
'\n', ' ').lower().strip()
print(
f"{update_info.batch.ix}\n\nTEXT:\n{text} \n\nHEADLINE:\n{headline} \n\nPREDICTED SUMMARY:\n{predicted}"
)
if update_info.batch.ix % 10 == 0:
self.writer.add_scalar('loss/train',
cumulative_train_metrics.loss,
update_info.batch.ix)
self.writer.add_scalar('model-loss/train',
cumulative_train_metrics.loss,
update_info.batch.ix)
self.writer.add_scalar('fooling-loss/train',
cumulative_train_metrics.loss,
update_info.batch.ix)
cumulative_train_metrics = Metrics.empty(mode="train")
if update_info.from_evaluate:
cumulative_evaluate_metrics += update_info.metrics
if len(cumulative_evaluate_metrics) == 10:
with torch.no_grad():
predicted = update_info.decoded_inferred_texts[
0].replace('\n', ' ').strip('❟ ❟ ❟')
headline = update_info.batch.orig_headline[0].replace(
'\n', ' ').lower().strip()
text = update_info.batch.orig_text[0].replace(
'\n', ' ').lower().strip()
print(
f"{update_info.batch.ix}\n\nEVAL TEXT:\n{text} \n\nEVAL HEADLINE:\n{headline} \n\nEVAL PREDICTED SUMMARY:\n{predicted}"
)
self.writer.add_text(
'text/eval', text,
int(update_info.batch.ix / evaluate_every))
self.writer.add_text(
'headline/eval', headline,
int(update_info.batch.ix / evaluate_every))
self.writer.add_text(
'predicted/eval', predicted,
int(update_info.batch.ix / evaluate_every))
self.writer.add_scalar(
'rouge-1/eval',
cumulative_evaluate_metrics.rouge_score,
int(update_info.batch.ix / evaluate_every))
cumulative_evaluate_metrics = Metrics.empty(mode="eval")
print(f"Eval: {update_info.metrics.loss}")
if update_info.batch.ix % 600 == 0 and update_info.batch.ix != 0:
print(f"Saving checkpoint")
self.save_checkpoint()