with tf.Graph().as_default() as g:
tf.train.create_global_step(g)
features, labels = input_fn()
labels["alignment"] = None // Add alignment key to force the model to return attention.
outputs, _ = self._model(
features,
labels,
After Change
if checkpoint_path is None:
raise ValueError("could not find a trained model in %s" % self._config["model_dir"])
model = copy.deepcopy(self._model)
with tf.Graph().as_default():
dataset = model.examples_inputter.make_evaluation_dataset(
features_file,
predictions_file,
self._config["score"]["batch_size"],
num_threads=self._config["score"].get("num_threads"),
prefetch_buffer_size=self._config["score"].get("prefetch_buffer_size"))
iterator = dataset.make_initializable_iterator()
features, labels = iterator.get_next()
labels["alignment"] = None // Add alignment key to force the model to return attention.
outputs, _ = model(
features,
labels,
self._config["params"],
tf.estimator.ModeKeys.EVAL)
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=outputs["logits"], labels=labels["ids_out"])
weights = tf.sequence_mask(labels["length"], dtype=cross_entropy.dtype)
masked_cross_entropy = cross_entropy * weights
scores = tf.reduce_sum(masked_cross_entropy, axis=1)
results = {
"attention": outputs["attention"],
"cross_entropy": cross_entropy,
"score": scores,
"tokens": labels["tokens"],
"length": labels["length"] - 1 // -1 for the special token.
}
if output_file:
stream = io.open(output_file, encoding="utf-8", mode="w")
else:
stream = sys.stdout
with tf.train.MonitoredSession(
session_creator=tf.train.ChiefSessionCreator(
checkpoint_filename_with_path=checkpoint_path,
config=self._session_config)) as sess:
sess.run(iterator.initializer)
while not sess.should_stop():
for batch in misc.extract_batches(sess.run(results)):
tokens = batch["tokens"][:batch["length"]]
sentence = self._model.labels_inputter.tokenizer.detokenize(tokens)
