def test_encoder_decoder_from_pretrained(self):
load_weight_prefix = TFEncoderDecoderModel.load_weight_prefix
config = self.get_encoder_decoder_config()
encoder_tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
decoder_tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
input_ids = encoder_tokenizer("who sings does he love me with reba", return_tensors="tf").input_ids
decoder_input_ids = decoder_tokenizer("Linda Davis", return_tensors="tf").input_ids
with tempfile.TemporaryDirectory() as tmp_dirname:
# Since most of HF's models don't have pretrained cross-attention layers, they are randomly
# initialized even if we create models using `from_pretrained` method.
# For the tests, the decoder need to be a model with pretrained cross-attention layers.
# So we create pretrained models (without `load_weight_prefix`), save them, and later,
# we load them using `from_pretrained`.
# (we don't need to do this for encoder, but let's make the code more similar between encoder/decoder)
encoder = TFAutoModel.from_pretrained("bert-base-uncased", name="encoder")
# It's necessary to specify `add_cross_attention=True` here.
decoder = TFAutoModelForCausalLM.from_pretrained(
"bert-base-uncased", is_decoder=True, add_cross_attention=True, name="decoder"
)
pretrained_encoder_dir = os.path.join(tmp_dirname, "pretrained_encoder")
pretrained_decoder_dir = os.path.join(tmp_dirname, "pretrained_decoder")
encoder.save_pretrained(pretrained_encoder_dir)
decoder.save_pretrained(pretrained_decoder_dir)
del encoder
del decoder
enc_dec_model = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
pretrained_encoder_dir,
pretrained_decoder_dir,
)
# check that the from pretrained methods work
enc_dec_model.save_pretrained(tmp_dirname)
enc_dec_model = TFEncoderDecoderModel.from_pretrained(tmp_dirname)
output = enc_dec_model(input_ids, decoder_input_ids=decoder_input_ids, labels=decoder_input_ids)
loss_pretrained = output.loss
del enc_dec_model
# Create the model using `__init__` with loaded ``pretrained`` encoder / decoder
encoder = TFAutoModel.from_pretrained(
pretrained_encoder_dir, load_weight_prefix=load_weight_prefix, name="encoder"
)
decoder = TFAutoModelForCausalLM.from_pretrained(
pretrained_decoder_dir, load_weight_prefix=load_weight_prefix, name="decoder"
)
enc_dec_model = TFEncoderDecoderModel(config=config, encoder=encoder, decoder=decoder)
output = enc_dec_model(input_ids, decoder_input_ids=decoder_input_ids, labels=decoder_input_ids)
loss_init = output.loss
max_diff = np.max(np.abs(loss_pretrained - loss_init))
expected_diff = 0.0
self.assertAlmostEqual(max_diff, expected_diff, places=4)
def test_model_for_causal_lm(self):
for model_name in TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
config = AutoConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, GPT2Config)
model = TFAutoModelForCausalLM.from_pretrained(model_name)
model, loading_info = TFAutoModelForCausalLM.from_pretrained(model_name, output_loading_info=True)
self.assertIsNotNone(model)
self.assertIsInstance(model, TFGPT2LMHeadModel)
def test_small_model_tf(self):
tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-small")
model = TFAutoModelForCausalLM.from_pretrained("microsoft/DialoGPT-small")
conversation_agent = ConversationalPipeline(model=model, tokenizer=tokenizer)
conversation = Conversation("hello")
output = conversation_agent(conversation)
self.assertEqual(output, Conversation(past_user_inputs=["hello"], generated_responses=["Hi"]))
def __init__(
self,
tokenizer_name: str,
model_name: str,
utterance_window: int = 4,
):
self.tokenizer = GPT2TokenizerFast.from_pretrained(tokenizer_name)
self.model = TFAutoModelForCausalLM.from_pretrained(model_name,
from_pt=True)
def main():
# region Argument Parsing
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TFTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Sanity checks
if data_args.dataset_name is None and data_args.train_file is None and data_args.validation_file is None:
raise ValueError(
"Need either a dataset name or a training/validation file.")
else:
if data_args.train_file is not None:
extension = data_args.train_file.split(".")[-1]
assert extension in [
"csv", "json", "txt"
], "`train_file` should be a csv, json or txt file."
if data_args.validation_file is not None:
extension = data_args.validation_file.split(".")[-1]
assert extension in [
"csv", "json", "txt"
], "`validation_file` should be a csv, json or txt file."
if training_args.output_dir is not None:
training_args.output_dir = Path(training_args.output_dir)
os.makedirs(training_args.output_dir, exist_ok=True)
if isinstance(
training_args.strategy,
tf.distribute.TPUStrategy) and not data_args.pad_to_max_length:
logger.warning("We are training on TPU - forcing pad_to_max_length")
data_args.pad_to_max_length = True
# endregion
# region Checkpoints
# Detecting last checkpoint.
checkpoint = None
if len(os.listdir(training_args.output_dir)
) > 0 and not training_args.overwrite_output_dir:
config_path = training_args.output_dir / CONFIG_NAME
weights_path = training_args.output_dir / TF2_WEIGHTS_NAME
if config_path.is_file() and weights_path.is_file():
checkpoint = training_args.output_dir
logger.info(
f"Checkpoint detected, resuming training from checkpoint in {training_args.output_dir}. To avoid this"
" behavior, change the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
else:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to continue regardless.")
# endregion
# region Setup logging
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(logging.INFO)
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
# endregion
# If passed along, set the training seed now.
if training_args.seed is not None:
set_seed(training_args.seed)
# region Load datasets
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
)
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
)
else:
data_files = {}
dataset_args = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
dataset_args["keep_linebreaks"] = data_args.keep_linebreaks
raw_datasets = load_dataset(extension,
data_files=data_files,
**dataset_args)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# endregion
# region Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
# endregion
# region Dataset preprocessing
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
if data_args.max_seq_length is None:
max_seq_length = tokenizer.model_max_length
if max_seq_length > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
)
max_seq_length = 1024
else:
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length,
tokenizer.model_max_length)
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
def tokenize_function(examples):
return tokenizer(examples[text_column_name])
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on dataset",
)
block_size = tokenizer.model_max_length
if block_size > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can reduce that value by passing --block_size xxx."
)
block_size = 1024
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {
k: sum(examples[k], [])
for k in examples.keys()
}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
if total_length >= block_size:
total_length = (total_length // block_size) * block_size
# Split by chunks of max_len.
result = {
k:
[t[i:i + block_size] for i in range(0, total_length, block_size)]
for k, t in concatenated_examples.items()
}
result["labels"] = result["input_ids"].copy()
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
# for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
# to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
desc=f"Grouping texts in chunks of {block_size}",
)
train_dataset = lm_datasets["train"]
if data_args.validation_file is not None:
eval_dataset = lm_datasets["validation"]
else:
logger.info(
f"Validation file not found: using {data_args.validation_split_percentage}% of the dataset as validation as provided in data_args"
)
train_indices, val_indices = train_test_split(
list(range(len(train_dataset))),
test_size=data_args.validation_split_percentage / 100)
eval_dataset = train_dataset.select(val_indices)
train_dataset = train_dataset.select(train_indices)
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(range(
data_args.max_train_samples))
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# endregion
with training_args.strategy.scope():
# region Prepare model
if checkpoint is not None:
model = TFAutoModelForCausalLM.from_pretrained(checkpoint,
config=config)
elif model_args.model_name_or_path:
model = TFAutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path, config=config)
else:
logger.info("Training new model from scratch")
model = TFAutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# endregion
# region TF Dataset preparation
num_replicas = training_args.strategy.num_replicas_in_sync
train_generator = partial(sample_generator, train_dataset, tokenizer)
train_signature = {
feature: tf.TensorSpec(shape=(None, ), dtype=tf.int64)
for feature in train_dataset.features
if feature != "special_tokens_mask"
}
train_sig = (train_signature, train_signature["labels"])
options = tf.data.Options()
options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
tf_train_dataset = (tf.data.Dataset.from_generator(
train_generator,
output_signature=train_sig).with_options(options).batch(
batch_size=num_replicas *
training_args.per_device_train_batch_size,
drop_remainder=True).repeat(int(
training_args.num_train_epochs)))
eval_generator = partial(sample_generator, eval_dataset, tokenizer)
eval_signature = {
feature: tf.TensorSpec(shape=(None, ), dtype=tf.int64)
for feature in eval_dataset.features
if feature != "special_tokens_mask"
}
eval_sig = (eval_signature, eval_signature["labels"])
tf_eval_dataset = (tf.data.Dataset.from_generator(
eval_generator,
output_signature=eval_sig).with_options(options).batch(
batch_size=num_replicas *
training_args.per_device_eval_batch_size,
drop_remainder=True).repeat(int(
training_args.num_train_epochs)))
# endregion
# region Optimizer and loss
batches_per_epoch = len(train_dataset) // (
num_replicas * training_args.per_device_train_batch_size)
# Bias and layernorm weights are automatically excluded from the decay
optimizer, lr_schedule = create_optimizer(
init_lr=training_args.learning_rate,
num_train_steps=int(training_args.num_train_epochs *
batches_per_epoch),
num_warmup_steps=training_args.warmup_steps,
adam_beta1=training_args.adam_beta1,
adam_beta2=training_args.adam_beta2,
adam_epsilon=training_args.adam_epsilon,
weight_decay_rate=training_args.weight_decay,
)
def dummy_loss(y_true, y_pred):
return tf.reduce_mean(y_pred)
model.compile(optimizer=optimizer, loss={"loss": dummy_loss})
# endregion
# region Training and validation
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {training_args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size = {training_args.per_device_train_batch_size * num_replicas}"
)
history = model.fit(
tf_train_dataset,
validation_data=tf_eval_dataset,
epochs=int(training_args.num_train_epochs),
steps_per_epoch=len(train_dataset) //
(training_args.per_device_train_batch_size * num_replicas),
callbacks=[
SavePretrainedCallback(output_dir=training_args.output_dir)
],
)
try:
train_perplexity = math.exp(history.history["loss"][-1])
except OverflowError:
train_perplexity = math.inf
try:
validation_perplexity = math.exp(history.history["val_loss"][-1])
except OverflowError:
validation_perplexity = math.inf
logger.info(f" Final train loss: {history.history['loss'][-1]:.3f}")
logger.info(f" Final train perplexity: {train_perplexity:.3f}")
logger.info(
f" Final validation loss: {history.history['val_loss'][-1]:.3f}")
logger.info(
f" Final validation perplexity: {validation_perplexity:.3f}")
# endregion
if training_args.output_dir is not None:
model.save_pretrained(training_args.output_dir)
if training_args.push_to_hub:
# You'll probably want to include some of your own metadata here!
model.push_to_hub()
def main():
# region Argument Parsing
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_clm", model_args, data_args, framework="tensorflow")
# Sanity checks
if data_args.dataset_name is None and data_args.train_file is None and data_args.validation_file is None:
raise ValueError("Need either a dataset name or a training/validation file.")
else:
if data_args.train_file is not None:
extension = data_args.train_file.split(".")[-1]
assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, json or txt file."
if data_args.validation_file is not None:
extension = data_args.validation_file.split(".")[-1]
assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, json or txt file."
if training_args.output_dir is not None:
training_args.output_dir = Path(training_args.output_dir)
os.makedirs(training_args.output_dir, exist_ok=True)
# endregion
# region Checkpoints
# Detecting last checkpoint.
checkpoint = None
if len(os.listdir(training_args.output_dir)) > 0 and not training_args.overwrite_output_dir:
config_path = training_args.output_dir / CONFIG_NAME
weights_path = training_args.output_dir / TF2_WEIGHTS_NAME
if config_path.is_file() and weights_path.is_file():
checkpoint = training_args.output_dir
logger.info(
f"Checkpoint detected, resuming training from checkpoint in {training_args.output_dir}. To avoid this"
" behavior, change the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
else:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to continue regardless."
)
# endregion
# region Setup logging
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(logging.INFO)
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
# endregion
# If passed along, set the training seed now.
if training_args.seed is not None:
set_seed(training_args.seed)
# region Load datasets
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
data_files = {}
dataset_args = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = (
data_args.train_file.split(".")[-1]
if data_args.train_file is not None
else data_args.validation_file.split(".")[-1]
)
if extension == "txt":
extension = "text"
dataset_args["keep_linebreaks"] = data_args.keep_linebreaks
raw_datasets = load_dataset(
extension,
data_files=data_files,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
**dataset_args,
)
# If no validation data is there, validation_split_percentage will be used to divide the dataset.
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
extension,
data_files=data_files,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
**dataset_args,
)
raw_datasets["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
**dataset_args,
)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# endregion
# region Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
# endregion
# region Dataset preprocessing
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
def tokenize_function(examples):
return tokenizer(examples[text_column_name])
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on dataset",
)
if data_args.block_size is None:
block_size = tokenizer.model_max_length
if block_size > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --block_size xxx."
)
block_size = 1024
else:
if data_args.block_size > tokenizer.model_max_length:
logger.warning(
f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
)
block_size = min(data_args.block_size, tokenizer.model_max_length)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
if total_length >= block_size:
total_length = (total_length // block_size) * block_size
# Split by chunks of max_len.
result = {
k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
for k, t in concatenated_examples.items()
}
result["labels"] = result["input_ids"].copy()
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
# for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
# to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
desc=f"Grouping texts in chunks of {block_size}",
)
train_dataset = lm_datasets["train"]
if data_args.validation_file is not None:
eval_dataset = lm_datasets["validation"]
else:
logger.info(
f"Validation file not found: using {data_args.validation_split_percentage}% of the dataset as validation"
" as provided in data_args"
)
train_indices, val_indices = train_test_split(
list(range(len(train_dataset))), test_size=data_args.validation_split_percentage / 100
)
eval_dataset = train_dataset.select(val_indices)
train_dataset = train_dataset.select(train_indices)
if data_args.max_train_samples is not None:
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
if data_args.max_eval_samples is not None:
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(max_eval_samples))
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), min(3, len(train_dataset))):
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
# endregion
with training_args.strategy.scope():
# region Prepare model
if checkpoint is not None:
model = TFAutoModelForCausalLM.from_pretrained(checkpoint, config=config)
elif model_args.model_name_or_path:
model = TFAutoModelForCausalLM.from_pretrained(model_args.model_name_or_path, config=config)
else:
logger.info("Training new model from scratch")
model = TFAutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# endregion
# region TF Dataset preparation
num_replicas = training_args.strategy.num_replicas_in_sync
options = tf.data.Options()
options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
# model.prepare_tf_dataset() wraps a Hugging Face dataset in a tf.data.Dataset which is ready to use in
# training. This is the recommended way to use a Hugging Face dataset when training with Keras. You can also
# use the lower-level dataset.to_tf_dataset() method, but you will have to specify things like column names
# yourself if you use this method, whereas they are automatically inferred from the model input names when
# using model.prepare_tf_dataset()
# For more info see the docs:
# https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.TFPreTrainedModel.prepare_tf_dataset
# https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset.to_tf_dataset
tf_train_dataset = model.prepare_tf_dataset(
train_dataset,
shuffle=True,
batch_size=num_replicas * training_args.per_device_train_batch_size,
).with_options(options)
tf_eval_dataset = model.prepare_tf_dataset(
eval_dataset,
shuffle=False,
batch_size=num_replicas * training_args.per_device_eval_batch_size,
drop_remainder=True,
).with_options(options)
# endregion
# region Optimizer and loss
num_train_steps = len(tf_train_dataset) * int(training_args.num_train_epochs)
if training_args.warmup_steps > 0:
num_warmup_steps = training_args.warmup_steps
elif training_args.warmup_ratio > 0:
num_warmup_steps = int(num_train_steps * training_args.warmup_ratio)
else:
num_warmup_steps = 0
# Bias and layernorm weights are automatically excluded from the decay
optimizer, lr_schedule = create_optimizer(
init_lr=training_args.learning_rate,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
adam_beta1=training_args.adam_beta1,
adam_beta2=training_args.adam_beta2,
adam_epsilon=training_args.adam_epsilon,
weight_decay_rate=training_args.weight_decay,
adam_global_clipnorm=training_args.max_grad_norm,
)
# no user-specified loss = will use the model internal loss
model.compile(optimizer=optimizer, jit_compile=training_args.xla)
# endregion
# region Preparing push_to_hub and model card
push_to_hub_model_id = training_args.push_to_hub_model_id
model_name = model_args.model_name_or_path.split("/")[-1]
if not push_to_hub_model_id:
if data_args.dataset_name is not None:
push_to_hub_model_id = f"{model_name}-finetuned-{data_args.dataset_name}"
else:
push_to_hub_model_id = f"{model_name}-finetuned-clm"
model_card_kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "text-generation"}
if data_args.dataset_name is not None:
model_card_kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
model_card_kwargs["dataset_args"] = data_args.dataset_config_name
model_card_kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
model_card_kwargs["dataset"] = data_args.dataset_name
if training_args.push_to_hub:
callbacks = [
PushToHubCallback(
output_dir=training_args.output_dir,
model_id=push_to_hub_model_id,
organization=training_args.push_to_hub_organization,
token=training_args.push_to_hub_token,
tokenizer=tokenizer,
**model_card_kwargs,
)
]
else:
callbacks = []
# endregion
# region Training and validation
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {training_args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
logger.info(f" Total train batch size = {training_args.per_device_train_batch_size * num_replicas}")
# For long training runs, you may wish to use the PushToHub() callback here to save intermediate checkpoints
# to the Hugging Face Hub rather than just pushing the finished model.
# See https://huggingface.co/docs/transformers/main_classes/keras_callbacks#transformers.PushToHubCallback
history = model.fit(
tf_train_dataset,
validation_data=tf_eval_dataset,
epochs=int(training_args.num_train_epochs),
callbacks=callbacks,
)
train_loss = history.history["loss"][-1]
try:
train_perplexity = math.exp(train_loss)
except OverflowError:
train_perplexity = math.inf
logger.info(f" Final train loss: {train_loss:.3f}")
logger.info(f" Final train perplexity: {train_perplexity:.3f}")
validation_loss = history.history["val_loss"][-1]
try:
validation_perplexity = math.exp(validation_loss)
except OverflowError:
validation_perplexity = math.inf
logger.info(f" Final validation loss: {validation_loss:.3f}")
logger.info(f" Final validation perplexity: {validation_perplexity:.3f}")
if training_args.output_dir is not None:
output_eval_file = os.path.join(training_args.output_dir, "all_results.json")
results_dict = dict()
results_dict["train_loss"] = train_loss
results_dict["train_perplexity"] = train_perplexity
results_dict["eval_loss"] = validation_loss
results_dict["eval_perplexity"] = validation_perplexity
with open(output_eval_file, "w") as writer:
writer.write(json.dumps(results_dict))
# endregion
if training_args.output_dir is not None and not training_args.push_to_hub:
# If we're not pushing to hub, at least save a local copy when we're done
model.save_pretrained(training_args.output_dir)
import pickle, random
import tensorflow as tf
from itertools import chain
from transformers import AutoTokenizer, TFAutoModelForCausalLM
datafile = 'personachat_self_personalities.pickle'
model_path = 'personachat-distilgpt2'
N = 20 # Maximum number of words to output
k = 10 # Top-k items to select from
p = 0.8 # Top-p cumulative probability to select from
# Load the trained model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = TFAutoModelForCausalLM.from_pretrained(model_path, use_cache=True, return_dict=False)
bos = tokenizer.bos_token_id
bot, person = tokenizer.encode('<bot>', '<person>')
# Load personalities dataset
with open(datafile, 'rb') as fp:
personalities = pickle.load(fp)
def feed(input_ids, speaker, past=None):
return model({
'input_ids': tf.constant(input_ids),
'token_type_ids': tf.fill(len(input_ids), int(speaker == person))
}, past)
def new_chat():
from transformers import T5Tokenizer
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--model_dir",
type=str,
default="../data/model/huggingface_model")
args = parser.parse_args()
random.seed(42)
tokenizer = T5Tokenizer.from_pretrained(args.model_dir, extra_ids=0)
print(len(tokenizer))
pt_model = AutoModelForCausalLM.from_pretrained(args.model_dir)
tf_model = TFAutoModelForCausalLM.from_pretrained(args.model_dir)
prompt = "誰も到達していないArtificial Intelligenceの高みへ、ともに"
with amp.autocast():
pt_tensor = tokenizer(prompt, return_tensors="pt")["input_ids"]
output_sequences = pt_model.generate(
input_ids=pt_tensor,
max_length=50 + pt_tensor.size(1),
top_p=0.95,
top_k=50,
do_sample=True,
early_stopping=True,
bos_token_id=tokenizer.bos_token_id,
eos_token_id=tokenizer.eos_token_id,
pad_token_id=tokenizer.pad_token_id,
num_return_sequences=1)