def test_prediction(self, num_classes):
task_config = sentence_prediction.SentencePredictionConfig(
model=self.get_model_config(num_classes=num_classes),
train_data=self._train_data_config)
task = sentence_prediction.SentencePredictionTask(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_classes=num_classes,
num_examples=num_examples)
test_data_config = (
sentence_prediction_dataloader.SentencePredictionDataConfig(
input_path=test_data_path,
seq_length=seq_length,
is_training=False,
label_type="int" if num_classes > 1 else "float",
global_batch_size=16,
drop_remainder=False,
include_example_id=True))
predictions = sentence_prediction.predict(task, test_data_config, model)
self.assertLen(predictions, num_examples)
for prediction in predictions:
self.assertEqual(prediction.dtype,
tf.int64 if num_classes > 1 else tf.float32)
def test_task(self, init_cls_pooler):
# Saves a checkpoint.
pretrain_cfg = bert.PretrainerConfig(
encoder=encoders.EncoderConfig(
bert=encoders.BertEncoderConfig(vocab_size=30522, num_layers=1)),
cls_heads=[
bert.ClsHeadConfig(
inner_dim=768, num_classes=2, name="next_sentence")
])
pretrain_model = masked_lm.MaskedLMTask(None).build_model(pretrain_cfg)
# The model variables will be created after the forward call.
_ = pretrain_model(pretrain_model.inputs)
ckpt = tf.train.Checkpoint(
model=pretrain_model, **pretrain_model.checkpoint_items)
init_path = ckpt.save(self.get_temp_dir())
# Creates the task.
config = sentence_prediction.SentencePredictionConfig(
init_checkpoint=init_path,
model=self.get_model_config(num_classes=2),
train_data=self._train_data_config,
init_cls_pooler=init_cls_pooler)
task = sentence_prediction.SentencePredictionTask(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.initialize(model)
task.train_step(next(iterator), model, optimizer, metrics=metrics)
task.validation_step(next(iterator), model, metrics=metrics)
def test_metrics_and_losses(self, num_classes):
config = sentence_prediction.SentencePredictionConfig(
init_checkpoint=self.get_temp_dir(),
model=self.get_model_config(num_classes),
train_data=self._train_data_config)
task = sentence_prediction.SentencePredictionTask(config)
model = task.build_model()
metrics = task.build_metrics()
if num_classes == 1:
self.assertIsInstance(metrics[0], tf.keras.metrics.MeanSquaredError)
else:
self.assertIsInstance(metrics[0],
tf.keras.metrics.SparseCategoricalAccuracy)
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)
logs = task.validation_step(next(iterator), model, metrics=metrics)
loss = logs["loss"].numpy()
if num_classes == 1:
self.assertGreater(loss, 1.0)
else:
self.assertLess(loss, 1.0)
def test_sentence_prediction(self):
config = sentence_prediction.SentencePredictionConfig(
model=sentence_prediction.ModelConfig(
encoder=encoders.EncoderConfig(
bert=encoders.BertEncoderConfig(vocab_size=30522,
num_layers=1)),
num_classes=2))
task = sentence_prediction.SentencePredictionTask(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 = {"inputs_only": False}
params = export_module_cls.Params(**serving_params)
export_module = export_module_cls(params=params, model=model)
export_dir = export_savedmodel_util.export(
export_module,
function_keys=["serve"],
checkpoint_path=ckpt_path,
export_savedmodel_dir=self.get_temp_dir())
imported = tf.saved_model.load(export_dir)
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)
ref_outputs = model(inputs)
outputs = serving_fn(**inputs)
self.assertAllClose(ref_outputs, outputs["outputs"])
self.assertEqual(outputs["outputs"].shape, (1, 2))
def test_task(self):
config = sentence_prediction.SentencePredictionConfig(
init_checkpoint=self.get_temp_dir(),
model=self.get_model_config(2),
train_data=self._train_data_config)
task = sentence_prediction.SentencePredictionTask(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)
# Saves a checkpoint.
pretrain_cfg = bert.BertPretrainerConfig(
encoder=encoders.TransformerEncoderConfig(
vocab_size=30522, num_layers=1),
cls_heads=[
bert.ClsHeadConfig(
inner_dim=10, num_classes=3, name="next_sentence")
])
pretrain_model = bert.instantiate_pretrainer_from_cfg(pretrain_cfg)
ckpt = tf.train.Checkpoint(
model=pretrain_model, **pretrain_model.checkpoint_items)
ckpt.save(config.init_checkpoint)
task.initialize(model)
def test_sentence_prediction_text(self, inputs_only):
vocab_file_path = os.path.join(self.get_temp_dir(), "vocab.txt")
_create_fake_vocab_file(vocab_file_path)
config = sentence_prediction.SentencePredictionConfig(
model=sentence_prediction.ModelConfig(
encoder=encoders.EncoderConfig(bert=encoders.BertEncoderConfig(
vocab_size=30522, num_layers=1)),
num_classes=2))
task = sentence_prediction.SentencePredictionTask(config)
model = task.build_model()
params = serving_modules.SentencePrediction.Params(
inputs_only=inputs_only,
parse_sequence_length=10,
text_fields=["foo", "bar"],
vocab_file=vocab_file_path)
export_module = serving_modules.SentencePrediction(params=params,
model=model)
examples = _create_fake_serialized_examples({
"foo": b"hello world",
"bar": b"hello world"
})
functions = export_module.get_inference_signatures({
"serve_text_examples":
"serving_default",
})
outputs = functions["serving_default"](examples)
self.assertEqual(outputs["outputs"].shape, (10, 2))
def test_task_with_fit(self):
config = sentence_prediction.SentencePredictionConfig(
model=self.get_model_config(2), train_data=self._train_data_config)
task = sentence_prediction.SentencePredictionTask(config)
model = task.build_model()
model = task.compile_model(model,
optimizer=tf.keras.optimizers.SGD(lr=0.1),
train_step=task.train_step,
metrics=task.build_metrics())
dataset = task.build_inputs(config.train_data)
logs = model.fit(dataset, epochs=1, steps_per_epoch=2)
self.assertIn("loss", logs.history)
def _run_task(self, config):
task = sentence_prediction.SentencePredictionTask(config)
model = task.build_model()
metrics = task.build_metrics()
strategy = tf.distribute.get_strategy()
dataset = strategy.experimental_distribute_datasets_from_function(
functools.partial(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)
return task.validation_step(next(iterator), model, metrics=metrics)
def _run_task(self, config):
task = sentence_prediction.SentencePredictionTask(config)
model = task.build_model()
metrics = task.build_metrics()
strategy = tf.distribute.get_strategy()
dataset = orbit.utils.make_distributed_dataset(strategy,
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)
def _run_task(self, config):
task = sentence_prediction.SentencePredictionTask(config)
model = task.build_model()
metrics = task.build_metrics()
strategy = tf.distribute.get_strategy()
dataset = strategy.distribute_datasets_from_function(
functools.partial(task.build_inputs, config.train_data))
iterator = iter(dataset)
optimizer = tf.keras.optimizers.SGD(learning_rate=0.1)
task.train_step(next(iterator), model, optimizer, metrics=metrics)
model.save(os.path.join(self.get_temp_dir(), "saved_model"))
return task.validation_step(next(iterator), model, metrics=metrics)
def test_np_metrics(self, metric_type, num_classes):
config = sentence_prediction.SentencePredictionConfig(
metric_type=metric_type,
init_checkpoint=self.get_temp_dir(),
model=self.get_model_config(num_classes),
train_data=self._train_data_config)
task = sentence_prediction.SentencePredictionTask(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.assertIn(metric_type, task.reduce_aggregated_logs(aggregated))
def test_task(self):
config = sentence_prediction.SentencePredictionConfig(
network=bert.BertPretrainerConfig(
encoders.TransformerEncoderConfig(vocab_size=30522,
num_layers=1),
num_masked_tokens=0,
cls_heads=[
bert.ClsHeadConfig(inner_dim=10,
num_classes=3,
name="sentence_prediction")
]),
train_data=bert.BertSentencePredictionDataConfig(
input_path="dummy", seq_length=128, global_batch_size=1))
task = sentence_prediction.SentencePredictionTask(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)
def test_sentence_prediction(self, use_v2_feature_names):
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"
config = sentence_prediction.SentencePredictionConfig(
model=sentence_prediction.ModelConfig(
encoder=encoders.EncoderConfig(bert=encoders.BertEncoderConfig(
vocab_size=30522, num_layers=1)),
num_classes=2))
task = sentence_prediction.SentencePredictionTask(config)
model = task.build_model()
params = serving_modules.SentencePrediction.Params(
inputs_only=True,
parse_sequence_length=10,
use_v2_feature_names=use_v2_feature_names)
export_module = serving_modules.SentencePrediction(params=params,
model=model)
functions = export_module.get_inference_signatures({
"serve":
"serving_default",
"serve_examples":
"serving_examples"
})
self.assertSameElements(functions.keys(),
["serving_default", "serving_examples"])
dummy_ids = tf.ones((10, 10), dtype=tf.int32)
outputs = functions["serving_default"](dummy_ids)
self.assertEqual(outputs["outputs"].shape, (10, 2))
params = serving_modules.SentencePrediction.Params(
inputs_only=False,
parse_sequence_length=10,
use_v2_feature_names=use_v2_feature_names)
export_module = serving_modules.SentencePrediction(params=params,
model=model)
functions = export_module.get_inference_signatures({
"serve":
"serving_default",
"serve_examples":
"serving_examples"
})
outputs = functions["serving_default"](input_word_ids=dummy_ids,
input_mask=dummy_ids,
input_type_ids=dummy_ids)
self.assertEqual(outputs["outputs"].shape, (10, 2))
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["outputs"].shape, (10, 2))
with self.assertRaises(ValueError):
_ = export_module.get_inference_signatures({"foo": None})
