def main(argv):
args = parse_args(argv)
set_package_verbosity(args.debug)
data_collection, meta_data = data_factory.preprocess(args, return_meta=True)
tokenizer = meta_data.tokenizer
metric_calcuators = []
if args.fed:
metric_calcuators.append(FEDMetrics(data_collection, tokenizer, sample_size=args.fed))
if args.bleu:
metric_calcuators.append(
BLEUMetrics(
data_collection,
max_gram=args.bleu,
eos_idx=tokenizer.eos_idx,
cache_dir=meta_data.cache_dir,
),
)
with open(args.eval_path, 'r') as f_in:
texts = [line.rstrip() for line in f_in.readlines()]
tokens = tokenizer.texts_to_array(texts)
metrics = {}
for calc in metric_calcuators:
metrics.update(calc.calculate(tokens=tokens, texts=texts))
print(
f"{os.path.basename(args.eval_path)},",
*[f"{key}: {val:.5f}" for key, val in metrics.items()],
sep="\n",
)
def main(argv):
args = parse_args(argv)
set_package_verbosity(args.debug)
data_collection = data_factory.preprocess(args)
with tf.Session(config=get_tf_config_proto()).as_default():
signature = load_serving_signature(
os.path.join(args.model_path, args.version_number))
generator = PerplexityCalculator.from_signature(
signature['perplexity'])
for tag, dataset in data_collection.items():
print(f"Evaluate {tag} perplexity:")
perplexity = generator.perplexity(dataset.ids)
print(f"Perplexity = {perplexity}")
def main(argv):
args = parse_args(argv)
set_package_verbosity(args.debug)
with tf.Session(config=get_tf_config_proto()).as_default():
signature = load_serving_signature(os.path.join(args.model_path, args.version_number))
tokenizer = Tokenizer.load(os.path.join(os.path.dirname(args.model_path), 'tokenizer.json'))
generator = TextGenerator.from_signature(signature['generate'], tokenizer=tokenizer)
print(f"Generate sentences to '{args.export_path}'")
with open(args.export_path, 'w') as f_out:
f_out.writelines([
sentence + "\n"
for sentence in generator.generate_texts(args.samples)
])
def main(argv, base_tag=None, checkpoint=None, override_namespace=None):
args = parse_args(argv)
if override_namespace:
args.__dict__.update(override_namespace.__dict__)
set_package_verbosity(args.debug)
with logging_block("Set global random seed"):
set_global_random_seed(args.random_seed)
with logging_block("Preprocess data"):
data_collection, meta_data = data_factory.preprocess(args,
return_meta=True)
data_collection.summary()
meta_data.summary()
data_generator = DataGenerator(data_collection.train,
batch_size=args.batch_size)
with logging_block("Prepare Generator"):
generator = generator_factory.create(args, meta_data)
with logging_block("Prepare Generator Trainer"):
trainer = trainer_factory.create(args, meta_data, generator)
trainer.summary()
with logging_block("Prepare Callback"):
if base_tag is None:
base_tag = f"{args.dataset}@{time.strftime('%Y%m%d-%H%M%S')}"
data_generator.callback = CallbackFactory(
trainer=trainer,
generator=generator,
data_collection=data_collection,
meta_data=meta_data,
tags=args.tags + [base_tag],
).create_by_args(args)
with tf.Session(config=get_tf_config_proto(args.jit)) as sess:
if checkpoint:
print(f"Restore from checkpoint: {checkpoint}")
tf.train.Saver().restore(sess, save_path=checkpoint)
data_generator.skip_epochs(int(checkpoint.split('-')[-1]))
else:
tf.global_variables_initializer().run()
for batch_data in data_generator.iter_batch_until(
n_epochs=args.epochs,
logs={'arg_string':
" ".join(argv)}, # for callback.on_train_begin()
):
trainer.fit_batch(batch_data)