def test_predict(self):
task_config = tagging.TaggingConfig(
model=tagging.ModelConfig(encoder=self._encoder_config),
train_data=self._train_data_config,
class_names=["O", "B-PER", "I-PER"])
task = tagging.TaggingTask(task_config)
model = task.build_model()
test_data_path = os.path.join(self.get_temp_dir(), "test.tf_record")
seq_length = 16
num_examples = 100
_create_fake_dataset(test_data_path,
seq_length=seq_length,
num_labels=len(task_config.class_names),
num_examples=num_examples)
test_data_config = tagging_dataloader.TaggingDataConfig(
input_path=test_data_path,
seq_length=seq_length,
is_training=False,
global_batch_size=16,
drop_remainder=False,
include_sentence_id=True)
results = tagging.predict(task, test_data_config, model)
self.assertLen(results, num_examples)
self.assertLen(results[0], 3)
def bert_tagging() -> cfg.ExperimentConfig:
"""BERT tagging task."""
config = cfg.ExperimentConfig(
task=tagging.TaggingConfig(
train_data=tagging_dataloader.TaggingDataConfig(),
validation_data=tagging_dataloader.TaggingDataConfig(
is_training=False, drop_remainder=False)),
trainer=cfg.TrainerConfig(
optimizer_config=optimization.OptimizationConfig({
'optimizer': {
'type': 'adamw',
'adamw': {
'weight_decay_rate':
0.01,
'exclude_from_weight_decay':
['LayerNorm', 'layer_norm', 'bias'],
}
},
'learning_rate': {
'type': 'polynomial',
'polynomial': {
'initial_learning_rate': 8e-5,
'end_learning_rate': 0.0,
}
},
'warmup': {
'type': 'polynomial'
}
})),
restrictions=[
'task.train_data.is_training != None',
'task.validation_data.is_training != None',
])
return config
def test_task_with_hub(self):
hub_module_url = self._export_bert_tfhub()
config = tagging.TaggingConfig(hub_module_url=hub_module_url,
model=self._encoder_config,
class_names=["O", "B-PER", "I-PER"],
train_data=self._train_data_config)
self._run_task(config)
def test_task_with_hub(self):
hub_module_url = self._export_bert_tfhub()
config = tagging.TaggingConfig(hub_module_url=hub_module_url,
model=self._encoder_config,
num_classes=4,
train_data=self._train_data_config)
self._run_task(config)
def test_tagging(self, use_v2_feature_names, output_encoder_outputs):
if use_v2_feature_names:
input_word_ids_field = "input_word_ids"
input_type_ids_field = "input_type_ids"
else:
input_word_ids_field = "input_ids"
input_type_ids_field = "segment_ids"
hidden_size = 768
num_classes = 3
config = tagging.TaggingConfig(
model=tagging.ModelConfig(encoder=encoders.EncoderConfig(
bert=encoders.BertEncoderConfig(hidden_size=hidden_size,
num_layers=1))),
class_names=["class_0", "class_1", "class_2"])
task = tagging.TaggingTask(config)
model = task.build_model()
params = serving_modules.Tagging.Params(
parse_sequence_length=10,
use_v2_feature_names=use_v2_feature_names,
output_encoder_outputs=output_encoder_outputs)
export_module = serving_modules.Tagging(params=params, model=model)
functions = export_module.get_inference_signatures({
"serve":
"serving_default",
"serve_examples":
"serving_examples"
})
dummy_ids = tf.ones((10, 10), dtype=tf.int32)
outputs = functions["serving_default"](input_word_ids=dummy_ids,
input_mask=dummy_ids,
input_type_ids=dummy_ids)
self.assertEqual(outputs["logits"].shape, (10, 10, num_classes))
if output_encoder_outputs:
self.assertEqual(outputs["encoder_outputs"].shape,
(10, 10, hidden_size))
dummy_ids = tf.ones((10, ), dtype=tf.int32)
examples = _create_fake_serialized_examples({
input_word_ids_field:
dummy_ids,
"input_mask":
dummy_ids,
input_type_ids_field:
dummy_ids
})
outputs = functions["serving_examples"](examples)
self.assertEqual(outputs["logits"].shape, (10, 10, num_classes))
if output_encoder_outputs:
self.assertEqual(outputs["encoder_outputs"].shape,
(10, 10, hidden_size))
with self.assertRaises(ValueError):
_ = export_module.get_inference_signatures({"foo": None})
def test_task_with_fit(self):
config = tagging.TaggingConfig(
network=self._encoder_config,
train_data=self._train_data_config,
num_classes=3)
task = tagging.TaggingTask(config)
model = task.build_model()
model = task.compile_model(
model,
optimizer=tf.keras.optimizers.SGD(lr=0.1),
train_step=task.train_step,
metrics=[tf.keras.metrics.SparseCategoricalAccuracy(name="accuracy")])
dataset = task.build_inputs(config.train_data)
logs = model.fit(dataset, epochs=1, steps_per_epoch=2)
self.assertIn("loss", logs.history)
self.assertIn("accuracy", logs.history)
def test_tagging(self, output_encoder_outputs):
hidden_size = 768
num_classes = 3
config = tagging.TaggingConfig(
model=tagging.ModelConfig(
encoder=encoders.EncoderConfig(
bert=encoders.BertEncoderConfig(
hidden_size=hidden_size, num_layers=1))),
class_names=["class_0", "class_1", "class_2"])
task = tagging.TaggingTask(config)
model = task.build_model()
ckpt = tf.train.Checkpoint(model=model)
ckpt_path = ckpt.save(self.get_temp_dir())
export_module_cls = export_savedmodel.lookup_export_module(task)
serving_params = {
"parse_sequence_length": 10,
}
params = export_module_cls.Params(
**serving_params, output_encoder_outputs=output_encoder_outputs)
export_module = export_module_cls(params=params, model=model)
export_dir = export_savedmodel_util.export(
export_module,
function_keys={
"serve": "serving_default",
"serve_examples": "serving_examples"
},
checkpoint_path=ckpt_path,
export_savedmodel_dir=self.get_temp_dir())
imported = tf.saved_model.load(export_dir)
self.assertCountEqual(imported.signatures.keys(),
["serving_default", "serving_examples"])
serving_fn = imported.signatures["serving_default"]
dummy_ids = tf.ones((1, 5), dtype=tf.int32)
inputs = dict(
input_word_ids=dummy_ids,
input_mask=dummy_ids,
input_type_ids=dummy_ids)
outputs = serving_fn(**inputs)
self.assertEqual(outputs["logits"].shape, (1, 5, num_classes))
if output_encoder_outputs:
self.assertEqual(outputs["encoder_outputs"].shape, (1, 5, hidden_size))
def test_task(self):
# Saves a checkpoint.
encoder = encoders.instantiate_encoder_from_cfg(self._encoder_config)
ckpt = tf.train.Checkpoint(encoder=encoder)
saved_path = ckpt.save(self.get_temp_dir())
config = tagging.TaggingConfig(init_checkpoint=saved_path,
model=self._encoder_config,
train_data=self._train_data_config,
class_names=["O", "B-PER", "I-PER"])
task = tagging.TaggingTask(config)
model = task.build_model()
metrics = task.build_metrics()
dataset = task.build_inputs(config.train_data)
iterator = iter(dataset)
optimizer = tf.keras.optimizers.SGD(lr=0.1)
task.train_step(next(iterator), model, optimizer, metrics=metrics)
task.validation_step(next(iterator), model, metrics=metrics)
task.initialize(model)
def test_seqeval_metrics(self):
config = tagging.TaggingConfig(model=self._encoder_config,
train_data=self._train_data_config,
class_names=["O", "B-PER", "I-PER"])
task = tagging.TaggingTask(config)
model = task.build_model()
dataset = task.build_inputs(config.train_data)
iterator = iter(dataset)
strategy = tf.distribute.get_strategy()
distributed_outputs = strategy.run(functools.partial(
task.validation_step, model=model),
args=(next(iterator), ))
outputs = tf.nest.map_structure(strategy.experimental_local_results,
distributed_outputs)
aggregated = task.aggregate_logs(step_outputs=outputs)
aggregated = task.aggregate_logs(state=aggregated,
step_outputs=outputs)
self.assertCountEqual({"f1", "precision", "recall", "accuracy"},
task.reduce_aggregated_logs(aggregated).keys())
