def test_network_invocation(self):
config = bert.BertPretrainerConfig(
encoder=encoders.TransformerEncoderConfig(vocab_size=10,
num_layers=1))
_ = bert.instantiate_bertpretrainer_from_cfg(config)
# Invokes with classification heads.
config = bert.BertPretrainerConfig(
encoder=encoders.TransformerEncoderConfig(vocab_size=10,
num_layers=1),
cls_heads=[
bert.ClsHeadConfig(inner_dim=10,
num_classes=2,
name="next_sentence")
])
_ = bert.instantiate_bertpretrainer_from_cfg(config)
with self.assertRaises(ValueError):
config = bert.BertPretrainerConfig(
encoder=encoders.TransformerEncoderConfig(vocab_size=10,
num_layers=1),
cls_heads=[
bert.ClsHeadConfig(inner_dim=10,
num_classes=2,
name="next_sentence"),
bert.ClsHeadConfig(inner_dim=10,
num_classes=2,
name="next_sentence")
])
_ = bert.instantiate_bertpretrainer_from_cfg(config)
class ELECTRAPretrainerConfig(base_config.Config):
"""ELECTRA pretrainer configuration."""
num_masked_tokens: int = 76
sequence_length: int = 512
num_classes: int = 2
discriminator_loss_weight: float = 50.0
generator_encoder: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
discriminator_encoder: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
cls_heads: List[bert.ClsHeadConfig] = dataclasses.field(default_factory=list)
def setUp(self):
super(QuestionAnsweringTaskTest, self).setUp()
self._encoder_config = encoders.TransformerEncoderConfig(
vocab_size=30522, num_layers=1)
self._train_data_config = bert.QADataConfig(input_path="dummy",
seq_length=128,
global_batch_size=1)
def test_task(self):
config = masked_lm.MaskedLMConfig(
init_checkpoint=self.get_temp_dir(),
model=bert.BertPretrainerConfig(
encoders.TransformerEncoderConfig(vocab_size=30522,
num_layers=1),
num_masked_tokens=20,
cls_heads=[
bert.ClsHeadConfig(inner_dim=10,
num_classes=2,
name="next_sentence")
]),
train_data=pretrain_dataloader.BertPretrainDataConfig(
input_path="dummy",
max_predictions_per_seq=20,
seq_length=128,
global_batch_size=1))
task = masked_lm.MaskedLMTask(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.
ckpt = tf.train.Checkpoint(model=model, **model.checkpoint_items)
ckpt.save(config.init_checkpoint)
task.initialize(model)
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)
class QuestionAnsweringConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
network: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
def get_model_config(self, num_classes):
return bert.BertPretrainerConfig(
encoder=encoders.TransformerEncoderConfig(vocab_size=30522,
num_layers=1),
num_masked_tokens=0,
cls_heads=[
bert.ClsHeadConfig(inner_dim=10,
num_classes=num_classes,
name="sentence_prediction")
])
def test_network_invocation(self):
config = electra.ELECTRAPretrainerConfig(
generator_encoder=encoders.TransformerEncoderConfig(
vocab_size=10, num_layers=1),
discriminator_encoder=encoders.TransformerEncoderConfig(
vocab_size=10, num_layers=2),
)
_ = electra.instantiate_pretrainer_from_cfg(config)
# Invokes with classification heads.
config = electra.ELECTRAPretrainerConfig(
generator_encoder=encoders.TransformerEncoderConfig(
vocab_size=10, num_layers=1),
discriminator_encoder=encoders.TransformerEncoderConfig(
vocab_size=10, num_layers=2),
cls_heads=[
bert.ClsHeadConfig(
inner_dim=10, num_classes=2, name="next_sentence")
])
_ = electra.instantiate_pretrainer_from_cfg(config)
class QuestionAnsweringConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
n_best_size: int = 20
max_answer_length: int = 30
null_score_diff_threshold: float = 0.0
model: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
def test_checkpoint_items(self):
config = bert.BertPretrainerConfig(
encoder=encoders.TransformerEncoderConfig(vocab_size=10,
num_layers=1),
cls_heads=[
bert.ClsHeadConfig(inner_dim=10,
num_classes=2,
name="next_sentence")
])
encoder = bert.instantiate_bertpretrainer_from_cfg(config)
self.assertSameElements(encoder.checkpoint_items.keys(),
["encoder", "next_sentence.pooler_dense"])
def setUp(self):
super(SentencePredictionTaskTest, self).setUp()
self._network_config = bert.BertPretrainerConfig(
encoder=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")
])
self._train_data_config = bert.SentencePredictionDataConfig(
input_path="dummy", seq_length=128, global_batch_size=1)
class TaggingConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
model: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
num_classes: int = 0
# The ignored label id will not contribute to loss.
# A word may be tokenized into multiple word_pieces tokens, and we usually
# assign the real label id for the first token of the word, and
# `ignore_label_id` for the remaining tokens.
ignore_label_id: int = 0
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
class TaggingConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
model: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
# The number of real labels. Note that a word may be tokenized into
# multiple word_pieces tokens, and we asssume the real label id (non-negative)
# is assigned to the first token of the word, and a negative label id is
# assigned to the remaining tokens. The negative label id will not contribute
# to loss and metrics.
num_classes: int = 0
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
class TaggingConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
model: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
# The real class names, the order of which should match real label id.
# Note that a word may be tokenized into multiple word_pieces tokens, and
# we asssume the real label id (non-negative) is assigned to the first token
# of the word, and a negative label id is assigned to the remaining tokens.
# The negative label id will not contribute to loss and metrics.
class_names: Optional[List[str]] = None
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
def test_task_with_hub(self):
hub_module_url = self._export_bert_tfhub()
config = sentence_prediction.SentencePredictionConfig(
hub_module_url=hub_module_url,
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=10))
self._run_task(config)
def setUp(self):
super(QuestionAnsweringTaskTest, self).setUp()
self._encoder_config = encoders.TransformerEncoderConfig(
vocab_size=30522, num_layers=1)
self._train_data_config = bert.QADataConfig(input_path="dummy",
seq_length=128,
global_batch_size=1)
val_data = {
"version":
"1.1",
"data": [{
"paragraphs": [{
"context":
"Sky is blue.",
"qas": [{
"question":
"What is blue?",
"id":
"1234",
"answers": [{
"text": "Sky",
"answer_start": 0
}, {
"text": "Sky",
"answer_start": 0
}, {
"text": "Sky",
"answer_start": 0
}]
}]
}]
}]
}
self._val_input_path = os.path.join(self.get_temp_dir(),
"val_data.json")
with tf.io.gfile.GFile(self._val_input_path, "w") as writer:
writer.write(json.dumps(val_data, indent=4) + "\n")
self._test_vocab = os.path.join(self.get_temp_dir(), "vocab.txt")
with tf.io.gfile.GFile(self._test_vocab, "w") as writer:
writer.write("[PAD]\n[UNK]\n[CLS]\n[SEP]\n[MASK]\nsky\nis\nblue\n")
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)
class ModelConfig(base_config.Config):
"""A base span labeler configuration."""
encoder: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
class ModelConfig(base_config.Config):
"""A base span labeler configuration."""
encoder: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
head_dropout: float = 0.1
head_initializer_range: float = 0.02
class ModelConfig(base_config.Config):
"""A classifier/regressor configuration."""
num_classes: int = 0
use_encoder_pooler: bool = False
encoder: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
class BertPretrainerConfig(base_config.Config):
"""BERT encoder configuration."""
num_masked_tokens: int = 76
encoder: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
cls_heads: List[ClsHeadConfig] = dataclasses.field(default_factory=list)
def setUp(self):
super(TaggingTest, self).setUp()
self._encoder_config = encoders.TransformerEncoderConfig(
vocab_size=30522, num_layers=1)
self._train_data_config = tagging_data_loader.TaggingDataConfig(
input_path="dummy", seq_length=128, global_batch_size=1)
def get_model_config(self, num_classes):
return sentence_prediction.ModelConfig(
encoder=encoders.TransformerEncoderConfig(
vocab_size=30522, num_layers=1),
num_classes=num_classes)
