def main(
training_objective: TrainingObjective,
test: bool = False,
overfit: float = 0,
max_epochs: int = 1000,
checkpoint: str = None,
):
config: VisualBERTConfig = VisualBERTConfig()
# Base BERT model
config.tokenizer = AutoTokenizer.from_pretrained(
"google/bert_uncased_L-4_H-512_A-8"
# "bert-base-uncased"
)
config.training_objective = training_objective
bert_model_name = "google/bert_uncased_L-8_H-768_A-12"
config.hidden_size = 768
if training_objective == TrainingObjective.MaskedLanguageModelling:
config.bert_model = AutoModelForMaskedLM.from_pretrained(
bert_model_name)
elif training_objective == TrainingObjective.Captioning:
c = AutoConfig.from_pretrained(bert_model_name)
c.is_decoder = True
config.bert_model = AutoModelForCausalLM.from_config(c)
if checkpoint is None:
model = VisualBERT(config)
else:
model = VisualBERT.load_from_checkpoint(checkpoint, config=config)
data = ConceptualCaptions()
data.prepare_data()
data.setup()
logger = None
fast_dev_run = test & (overfit == 0)
if test is not True:
logger = WandbLogger(
project="final-year-project",
offline=False,
log_model=True,
save_dir=work_dir,
)
callbacks = [CheckpointEveryNSteps(50000)]
trainer = pl.Trainer(gpus=1,
fast_dev_run=fast_dev_run,
default_root_dir=work_dir,
log_every_n_steps=10,
logger=logger,
max_epochs=max_epochs,
overfit_batches=overfit,
callbacks=callbacks
# check_val_every_n_epoch=1000 if overfit > 0 else 1,
)
trainer.fit(model, data)
def __init__(
self,
config: Optional[PretrainedConfig] = None,
encoder: Optional[PreTrainedModel] = None,
decoder: Optional[PreTrainedModel] = None,
):
assert config is not None or (
encoder is not None and decoder is not None
), "Either a configuration or an Encoder and a decoder has to be provided"
if config is None:
config = EncoderDecoderConfig.from_encoder_decoder_configs(
encoder.config, decoder.config)
else:
assert isinstance(
config,
self.config_class), "config: {} has to be of type {}".format(
config, self.config_class)
# initialize with config
super().__init__(config)
if encoder is None:
from transformers import AutoModel
encoder = AutoModel.from_config(config.encoder)
if decoder is None:
from transformers import AutoModelForCausalLM
decoder = AutoModelForCausalLM.from_config(config.decoder)
self.encoder = encoder
self.decoder = decoder
assert (
self.encoder.get_output_embeddings() is None
), "The encoder {} should not have a LM Head. Please use a model without LM Head"
def train_function(train_dataset, eval_dataset=None, **config):
model_config = AutoConfig.from_pretrained(model_checkpoint)
model = AutoModelForCausalLM.from_config(model_config)
print("Initializing TrainingArguments...")
# The checkpoints will be moved to Ray Tune results
# directory automatically
training_dir = tempfile.mkdtemp()
training_args = TrainingArguments(
training_dir,
evaluation_strategy="epoch",
num_train_epochs=num_epochs,
learning_rate=2e-5,
weight_decay=0.01,
disable_tqdm=True,
save_strategy="epoch",
# Required to avoid an exception
no_cuda=not torch.cuda.is_available(),
)
print("Initializing Trainer...")
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
)
print("Trainer initialized! Starting training...")
return trainer
def __init__(self, config: Funnel_gpt2_VAE_Config):
super().__init__(config=config)
self.decoder = AutoModelForCausalLM.from_config(
config.transformer_decoder)
self.decoder_start_token_id = self.config.transformer_decoder.bos_token_id
assert (
self.decoder_start_token_id is not None
), "`self.config.transformer_decoder.bos_token_id` has to be defined."
def test():
with tempfile.TemporaryDirectory() as tmpdir:
model_config = AutoConfig.from_pretrained(model_checkpoint)
model = AutoModelForCausalLM.from_config(model_config)
tokenizer = AutoTokenizer.from_pretrained(tokenizer_checkpoint)
checkpoint = HuggingFaceCheckpoint.from_model(model,
tokenizer,
path=tmpdir)
predictor = HuggingFacePredictor.from_checkpoint(
checkpoint,
task="text-generation",
)
predictions = predictor.predict(prompts)
assert len(predictions) == 3
def __init__(self, tokenizer, pretrained, maxlen=512, share_embedding=False, **kwargs):
super().__init__()
self.tokenizer = tokenizer
self.pretrained = pretrained
decoder_config = copy.deepcopy(pretrained.config)
decoder_config.is_decoder = True
decoder_config.add_cross_attention = True
self.model = AutoModelForCausalLM.from_config(decoder_config) # decoder
if share_embedding:
decoder_base_model_prefix = self.model.base_model_prefix
self._tie_encoder_decoder_weights(
self.pretrained, self.model._modules[decoder_base_model_prefix], self.model.base_model_prefix
)
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.maxlen = maxlen
print('Using device:', self.device)
self.model.to(self.device)
self.encoder_hidden = None
def train_function(train_dataset, eval_dataset=None, **config):
model_config = AutoConfig.from_pretrained(model_checkpoint)
model = AutoModelForCausalLM.from_config(model_config)
training_args = TrainingArguments(
f"{model_checkpoint}-wikitext2",
evaluation_strategy="epoch",
num_train_epochs=config.get("epochs", 3),
learning_rate=2e-5,
weight_decay=0.01,
disable_tqdm=True,
no_cuda=True,
save_strategy=config.get("save_strategy", "no"),
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
)
return trainer
def test_predict(preprocessor, tmpdir):
if preprocessor:
preprocessor = DummyPreprocessor()
else:
preprocessor = None
model_config = AutoConfig.from_pretrained(model_checkpoint)
model = AutoModelForCausalLM.from_config(model_config)
predictor = HuggingFacePredictor(
pipeline=pipeline(
task="text-generation",
model=model,
tokenizer=AutoTokenizer.from_pretrained(tokenizer_checkpoint),
),
preprocessor=preprocessor,
)
predictions = predictor.predict(prompts)
assert len(predictions) == 3
if preprocessor:
assert hasattr(predictor.preprocessor, "_batch_transformed")
def __init__(
self,
config: Optional[PretrainedConfig] = None,
encoder: Optional[PreTrainedModel] = None,
decoder: Optional[PreTrainedModel] = None,
):
if config is None and (encoder is None or decoder is None):
raise ValueError(
"Either a configuration or an encoder and a decoder has to be provided."
)
if config is None:
config = EncoderDecoderConfig.from_encoder_decoder_configs(
encoder.config, decoder.config)
else:
if not isinstance(config, self.config_class):
raise ValueError(
f"Config: {config} has to be of type {self.config_class}")
if config.decoder.cross_attention_hidden_size is not None:
if config.decoder.cross_attention_hidden_size != config.encoder.hidden_size:
raise ValueError(
"If `cross_attention_hidden_size` is specified in the decoder's configuration, "
"it has to be equal to the encoder's `hidden_size`. "
f"Got {config.decoder.cross_attention_hidden_size} for `config.decoder.cross_attention_hidden_size` "
f"and {config.encoder.hidden_size} for `config.encoder.hidden_size`."
)
# initialize with config
super().__init__(config)
if encoder is None:
from transformers import AutoModel
encoder = AutoModel.from_config(config.encoder)
if decoder is None:
from transformers import AutoModelForCausalLM
decoder = AutoModelForCausalLM.from_config(config.decoder)
self.encoder = encoder
self.decoder = decoder
if self.encoder.config.to_dict() != self.config.encoder.to_dict():
logger.warning(
f"Config of the encoder: {self.encoder.__class__} is overwritten by shared encoder config: {self.config.encoder}"
)
if self.decoder.config.to_dict() != self.config.decoder.to_dict():
logger.warning(
f"Config of the decoder: {self.decoder.__class__} is overwritten by shared decoder config: {self.config.decoder}"
)
# make sure that the individual model's config refers to the shared config
# so that the updates to the config will be synced
self.encoder.config = self.config.encoder
self.decoder.config = self.config.decoder
# encoder outputs might need to be projected to different dimension for decoder
if (self.encoder.config.hidden_size != self.decoder.config.hidden_size
and self.decoder.config.cross_attention_hidden_size is None):
self.enc_to_dec_proj = nn.Linear(self.encoder.config.hidden_size,
self.decoder.config.hidden_size)
if self.encoder.get_output_embeddings() is not None:
raise ValueError(
f"The encoder {self.encoder} should not have a LM Head. Please use a model without LM Head"
)
# tie encoder, decoder weights if config set accordingly
self.tie_weights()
def __init__(
self,
model: str = None,
model_folder: str = None,
config: Union[str, GPT2Config] = None,
vocab_file: str = None,
merges_file: str = None,
tokenizer_file: str = None,
schema_tokens: List[str] = None,
schema_return: List[str] = None,
cache_dir: str = "aitextgen",
tf_gpt2: str = None,
to_gpu: bool = False,
to_fp16: bool = False,
verbose: bool = False,
gradient_checkpointing: bool = False,
bos_token: str = None,
eos_token: str = None,
unk_token: str = None,
**kwargs,
) -> None:
if not verbose:
for module in [
"transformers.file_utils",
"transformers.configuration_utils",
"transformers.tokenization_utils",
"filelock",
"transformers.modeling_gpt2",
]:
logging.getLogger(module).setLevel(logging.WARN)
logging.getLogger("transformers.modeling_utils").setLevel(
logging.ERROR)
if tf_gpt2:
self.openai_tf_gpt2 = tf_gpt2
# Download + convert the TF weights if a PyTorch model has not been created
if not os.path.isfile(
os.path.join(cache_dir, f"pytorch_model_{tf_gpt2}.bin")):
assert tf_gpt2 in [
"124M",
"355M",
"774M",
"1558M",
], "Invalid TensorFlow GPT-2 model size."
logger.info(
f"Downloading the {tf_gpt2} GPT-2 TensorFlow weights/config "
+ "from Google's servers")
download_gpt2(cache_dir, tf_gpt2)
logger.info(
f"Converting the {tf_gpt2} GPT-2 TensorFlow weights to PyTorch."
)
config_path = os.path.join(cache_dir, tf_gpt2, "hparams.json")
convert_gpt2_checkpoint_to_pytorch(
os.path.join(cache_dir, tf_gpt2),
config_path,
cache_dir,
)
os.rename(
os.path.join(cache_dir, "pytorch_model.bin"),
os.path.join(cache_dir, f"pytorch_model_{tf_gpt2}.bin"),
)
os.rename(
os.path.join(cache_dir, "config.json"),
os.path.join(cache_dir, f"config_{tf_gpt2}.json"),
)
logger.info(f"Loading {tf_gpt2} GPT-2 model from /{cache_dir}.")
model = os.path.join(cache_dir, f"pytorch_model_{tf_gpt2}.bin")
config = os.path.join(cache_dir, f"config_{tf_gpt2}.json")
self.model = GPT2LMHeadModel.from_pretrained(model, config=config)
elif model_folder:
# A folder is provided containing pytorch_model.bin and config.json
assert os.path.exists(
os.path.join(model_folder, "pytorch_model.bin")
), f"There is no pytorch_model.bin in /{model_folder}."
assert os.path.exists(os.path.join(
model_folder,
"config.json")), f"There is no config.json in /{model_folder}."
logger.info(
f"Loading model from provided weights and config in /{model_folder}."
)
self.model = AutoModelForCausalLM.from_pretrained(
model_folder, local_files_only=True)
elif config:
# Manually construct a model from scratch
logger.info("Constructing model from provided config.")
if isinstance(config, str):
config = AutoConfig.from_pretrained(config)
self.model = AutoModelForCausalLM.from_config(config=config)
else:
# Download and cache model from Huggingface
if os.path.isdir(cache_dir) and len(os.listdir(cache_dir)) > 0:
logger.info(
f"Loading {model or 'gpt2'} model from /{cache_dir}.")
else:
logger.info(
f"Downloading {model or 'gpt2'} model to /{cache_dir}.")
self.model = AutoModelForCausalLM.from_pretrained(
model or "gpt2", cache_dir=cache_dir)
if model and "gpt2" not in model:
logger.info(f"Using the tokenizer for {model}.")
self.tokenizer = GPT2TokenizerFast.from_pretrained(
model,
cache_dir=cache_dir,
)
logger.info(self)
if gradient_checkpointing or tf_gpt2 in ["355M", "774M", "1558M"]:
logger.info("Gradient checkpointing enabled for model training.")
setattr(self.model.config, "gradient_checkpointing", True)
setattr(self.model.config, "use_cache", False)
if schema_tokens:
setattr(self.model.config, "schema_tokens", schema_tokens)
if schema_tokens:
setattr(self.model.config, "schema_return", schema_return)
if self.tokenizer is None:
# Update tokenizer settings (if not set already)
args = locals()
custom_tokenizer = False
for attr in [
"vocab_file",
"merges_file",
"tokenizer_file",
"bos_token",
"eos_token",
"unk_token",
]:
if args[attr] is not None:
custom_tokenizer = True
setattr(self, attr, args[attr])
if custom_tokenizer:
logger.info("Using a custom tokenizer.")
else:
logger.info("Using the default GPT-2 Tokenizer.")
if tokenizer_file:
# load the custom GPT-2 tokenizer from a serialized tokenizer.
# GPT-Neo uses the GPT-2 tokenizer.
self.tokenizer = GPT2TokenizerFast(
vocab_file=None,
merges_file=None,
tokenizer_file=tokenizer_file,
bos_token=self.bos_token,
eos_token=self.eos_token,
unk_token=self.unk_token,
pad_token=self.pad_token,
)
else:
self.tokenizer = GPT2TokenizerFast(
vocab_file=self.vocab_file,
merges_file=self.merges_file,
bos_token=self.bos_token,
eos_token=self.eos_token,
unk_token=self.unk_token,
pad_token=self.pad_token,
verbose=False,
)
if not custom_tokenizer:
# https://github.com/huggingface/transformers/issues/10202
self.tokenizer.add_special_tokens(
{"additional_special_tokens": ["<|endoftext|>"]})
self.tokenizer.padding_side = "left"
if to_gpu:
if to_fp16:
logger.warn(
"Currently, FP16 text generation results in random output. "
+
"You may want to avoid using to_fp16 for the time being.")
self.to_fp16()
self.to_gpu()
from arguments import InitializationArguments
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, HfArgumentParser
# Configuration
parser = HfArgumentParser(InitializationArguments)
args = parser.parse_args()
# Load codeparrot tokenizer trained for Python code tokenization
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name)
# Config: "scale_attn_by_layer_idx" and "reorder_and_upcast_attn" are Mistral stability tweaks
config_kwargs = {
"vocab_size": len(tokenizer),
"scale_attn_by_inverse_layer_idx": True,
"reorder_and_upcast_attn": True,
}
# Load model config (GPT-2 large in this case)
config = AutoConfig.from_pretrained(args.config_name, **config_kwargs)
# Initialize new model with config
model = AutoModelForCausalLM.from_config(config)
# Save model to the hub
model.save_pretrained(args.model_name, push_to_hub=args.push_to_hub)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
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(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# 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.
# datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
# if "validation" not in datasets.keys():
# datasets["validation"] = load_dataset(
# data_args.dataset_name,
# data_args.dataset_config_name,
# split=f"train[:{data_args.validation_split_percentage}%]",
# )
# datasets["train"] = load_dataset(
# data_args.dataset_name,
# data_args.dataset_config_name,
# split=f"train[{data_args.validation_split_percentage}%:]",
# )
# else:
# data_files = {}
# 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"
# datasets = load_dataset(extension, data_files=data_files)
datasets = load_from_disk(dataset_path=data_args.dataset_name)
# 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.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
**config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
**config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
tokenizer_kwargs = {
"cache_dir": model_args.cache_dir,
"use_fast": model_args.use_fast_tokenizer,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name,
**tokenizer_kwargs)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, **tokenizer_kwargs)
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."
)
if tokenizer.pad_token_id is None and tokenizer.eos_token_id is not None:
logger.warning(
f"Setting `pad_token_id` to `eos_token_id`:{tokenizer.eos_token_id}."
)
tokenizer.pad_token = tokenizer.eos_token
if model_args.model_name_or_path:
model = AutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
def tokenize_function(examples):
def tok_f_ids(string):
return tokenizer(string, return_attention_mask=False)["input_ids"]
texts, texts_a, texts_b = [], [], []
unprompted_texts = examples["text"]
prompting_instances = examples["prompting_instances"]
for ump_text, ppt_instances in zip(unprompted_texts,
prompting_instances):
if ppt_instances:
for i, p, o in zip(ppt_instances["input"],
ppt_instances["prompt"],
ppt_instances["output"]):
texts.append([])
texts_a.append(
tok_f_ids(i) \
+ [tokenizer.eos_token_id] \
+ tok_f_ids(p) \
+ [tokenizer.eos_token_id]
)
texts_b.append(tok_f_ids(o))
else:
texts.append(tok_f_ids(ump_text))
texts_a.append([])
texts_b.append([])
return {
"text_full": texts,
"text_a": texts_a,
"text_b": texts_b,
}
datasets = datasets.shuffle()
logger.info("Mapping dataset to tokenized dataset.", )
tokenized_datasets = 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,
)
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):
texts = examples["text_full"]
texts_a = examples["text_a"]
texts_b = examples["text_b"]
result = {
"input_ids": [],
"labels": [],
"attention_mask": [],
"length": [],
}
n = int(block_size / 2)
for t, t_a, t_b in zip(texts, texts_a, texts_b):
if t == []:
cut_t_a = t_a[-n:]
cut_t_b = t_b[:n]
if len(cut_t_b) < 20:
continue
result["input_ids"].append(cut_t_a + cut_t_b)
result["labels"].append([-100] * len(cut_t_a) + cut_t_b)
else:
total_length = len(t)
total_length = (total_length // block_size) * block_size
for i in range(0, total_length, block_size):
sub_seq = t[i:i + block_size]
result["input_ids"].append(sub_seq)
result["labels"].append(sub_seq)
for i in result["labels"]:
result["attention_mask"].append([1] * len(i))
result["length"].append(len(i))
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
logger.info("Chunking tokenized dataset.")
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=tokenized_datasets["train"].column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_train:
if "train" not in tokenized_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = lm_datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
if training_args.do_eval:
if "validation" not in tokenized_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = lm_datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it.
data_collator=MyDataCollatorWithPadding(tokenizer=tokenizer,
padding=True),
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif model_args.model_name_or_path is not None and os.path.isdir(
model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
perplexity = math.exp(metrics["eval_loss"])
metrics["perplexity"] = perplexity
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if data_args.eval_data_file is None and training_args.do_eval:
raise ValueError(
"Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
"or remove the --do_eval argument.")
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
f"Use --overwrite_output_dir to overcome.")
if data_args.finetune_on_class and data_args.train_with_targets:
raise ValueError(
"Pretraining with flow labels and fine-tuning on the class simultaneously not supported."
)
if not model_args.model_name_or_path and not model_args.quantizer_path:
raise ValueError(
"Either model or quantizer checkpoint path must be specified")
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.model_name_or_path:
config = GPT2Config.from_pretrained(model_args.model_name_or_path,
cache_dir=model_args.cache_dir)
else:
config = GPT2Config.from_json_file(model_args.config_name)
logger.warning(
"You are instantiating a new config instance from scratch.")
if model_args.model_name_or_path:
tokenizer = PacketTokenizer.from_pretrained(
model_args.model_name_or_path)
else:
tokenizer = PacketTokenizer.from_pretrained(model_args.quantizer_path)
if model_args.model_name_or_path:
model = GPT2LMHeadModel.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
if data_args.block_size <= 0:
data_args.block_size = tokenizer.max_len
# Our input block size will be the max possible for the model
else:
data_args.block_size = min(data_args.block_size, tokenizer.max_len)
# Get datasets
train_dataset = get_dataset(
data_args, tokenizer=tokenizer) if training_args.do_train else None
eval_dataset = get_dataset(
data_args, tokenizer=tokenizer,
evaluate=True) if training_args.do_eval else None
model.resize_token_embeddings(len(tokenizer))
print(model)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
data_collator=PretrainCollator(tokenizer),
train_dataset=train_dataset,
eval_dataset=eval_dataset,
prediction_loss_only=True,
)
# Training
if training_args.do_train:
model_path = (model_args.model_name_or_path
if model_args.model_name_or_path is not None
and os.path.isdir(model_args.model_name_or_path) else
None)
trainer.train(model_path=model_path)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
result = {"perplexity": perplexity}
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_lm.txt")
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
results.update(result)
return results
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name,
token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# 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 args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
args.dataset_name,
args.dataset_config_name,
split=f"train[:{args.validation_split_percentage}%]",
)
raw_datasets["train"] = load_dataset(
args.dataset_name,
args.dataset_config_name,
split=f"train[{args.validation_split_percentage}%:]",
)
else:
data_files = {}
dataset_args = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
dataset_args["keep_linebreaks"] = not args.no_keep_linebreaks
raw_datasets = load_dataset(extension,
data_files=data_files,
**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[:{args.validation_split_percentage}%]",
**dataset_args,
)
raw_datasets["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{args.validation_split_percentage}%:]",
**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.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.config_name)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
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."
)
if args.model_name_or_path:
model = AutoModelForCausalLM.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# 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])
with accelerator.main_process_first():
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on dataset",
)
if 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 args.block_size > tokenizer.model_max_length:
logger.warning(
f"The block_size passed ({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(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
with accelerator.main_process_first():
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=args.preprocessing_num_workers,
load_from_cache_file=not args.overwrite_cache,
desc=f"Grouping texts in chunks of {block_size}",
)
train_dataset = lm_datasets["train"]
eval_dataset = lm_datasets["validation"]
# 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]}.")
# DataLoaders creation:
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=default_data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataloader = DataLoader(eval_dataset,
collate_fn=default_data_collator,
batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader)
# On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
if accelerator.distributed_type == DistributedType.TPU:
model.tie_weights()
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps /
num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps),
disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(
train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
losses = []
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch)
loss = outputs.loss
losses.append(
accelerator.gather(loss.repeat(
args.per_device_eval_batch_size)))
losses = torch.cat(losses)
losses = losses[:len(eval_dataset)]
try:
perplexity = math.exp(torch.mean(losses))
except OverflowError:
perplexity = float("inf")
logger.info(f"epoch {epoch}: perplexity: {perplexity}")
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}",
blocking=False,
auto_lfs_prune=True)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training",
auto_lfs_prune=True)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
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(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# 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.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
if "validation" not in datasets.keys():
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
)
else:
data_files = {}
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
# why there is no test?
# if data_args.test_file is not None:
# data_files["test"] = data_args.test_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"
script = '/scratch/jx880/capstone/transformers/examples/language-modeling/MINDuser_masked.py'
datasets = load_dataset(script, data_files=data_files)
# 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.
data_arrays = datasets['train']['input_ids']
data_arrays.extend(datasets['validation']['input_ids'])
pad_size = max(list(map(len, data_arrays)))
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
**config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
**config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
config.n_layer = 6
config.vocab_size = 50257
config.sep_token_id = 0
config.bos_token_id = 0
config.eos_token_id = 0
config.n_embd = 256
logger.info(config)
# tokenizer_kwargs = {
# "cache_dir": model_args.cache_dir,
# "use_fast": model_args.use_fast_tokenizer,
# "revision": model_args.model_revision,
# "use_auth_token": True if model_args.use_auth_token else None,
# }
# if model_args.tokenizer_name:
# tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
# elif model_args.model_name_or_path:
# tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
# 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."
# )
# if model_args.model_name_or_path:
# model = AutoModelForCausalLM.from_pretrained(
# model_args.model_name_or_path,
# from_tf=bool(".ckpt" in model_args.model_name_or_path),
# config=config,
# cache_dir=model_args.cache_dir,
# revision=model_args.model_revision,
# use_auth_token=True if model_args.use_auth_token else None,
# )
# else:
# logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
logger.info("***** Model After Reconfiguration *****")
logger.info(model.config)
# model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
# if training_args.do_train:
# column_names = datasets["train"].column_names
# else:
# column_names = datasets["validation"].column_names
# text_column_name = "text" if "text" in column_names else column_names[0]
def tokenize(examples):
processed_example = examples.copy()
# print(examples)
processed_example['input_ids'] += [0] * (pad_size -
len(examples['input_ids']))
mask = [1 for i in range(len(examples['input_ids']))
] + [0] * (pad_size - len(examples['input_ids']))
processed_example['attention_mask'] = mask
return processed_example
# def tokenize_function(examples):
# return tokenize(examples)
tokenized_datasets = datasets.map(
tokenize,
batched=False,
num_proc=data_args.preprocessing_num_workers,
# remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# if data_args.block_size is None:
# block_size = tokenizer.model_max_length
# if block_size > 1024:
# logger.warn(
# 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.warn(
# 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.
# print(examples)
concatenated_examples = {}
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.
block_size = 1
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=False, # why True doesn't work?
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=lm_datasets["train"] if training_args.do_train else None,
eval_dataset=lm_datasets["validation"]
if training_args.do_eval else None,
tokenizer=None,
# Data collator will default to DataCollatorWithPadding, so we change it.
data_collator=default_data_collator,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif model_args.model_name_or_path is not None and os.path.isdir(
model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
trainer.log_metrics("eval", results)
trainer.save_metrics("eval", results)
return results
def main():
if is_wandb_available():
import wandb
parser = HfArgumentParser(
(ModelArguments, DatasetArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
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(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome.")
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if is_main_process(training_args.local_rank) else logging.WARN,
)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
datasets = load_from_disk(data_args.dataset_path)
# Load pretrained model and tokenizer
#
# 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,
cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
cache_dir=model_args.cache_dir)
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,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer)
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."
)
if model_args.model_name_or_path:
model = AutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].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 = datasets.map(
tokenize_function,
batched=True,
num_proc=None,
remove_columns=column_names,
load_from_cache_file=not False,
)
block_size = tokenizer.model_max_length
if block_size > 1024:
logger.warn(
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
# 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.
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=None,
load_from_cache_file=not False,
)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=lm_datasets["train"] if training_args.do_train else None,
eval_dataset=lm_datasets["validation"]
if training_args.do_eval else None,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it.
data_collator=default_data_collator,
)
# Add Callbacks
savvi_callback = SavviCallback()
trainer.add_callback(savvi_callback)
# Training
if training_args.do_train:
model_path = (model_args.model_name_or_path if
(model_args.model_name_or_path is not None
and os.path.isdir(model_args.model_name_or_path)) else
None)
trainer.train(model_path=model_path)
trainer.save_model() # Saves the tokenizer too for easy upload
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_clm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
savvihub.log(step=0, row={'perplexity': results["perplexity"]})
return results
def main():
args = parse_args()
if args.shard_optimizer_state > 0 and not args.skip_full_optimizer:
raise ValueError(
"If shard_optimizer_state is enabled, skip_full_optimizer must also be enabled. Full optimizer saving is currently not supported under optimizer state sharding."
)
if args.partition_assignment != "" and args.manual_partition == 0:
print("[Warning] partition_assignment is set, enable manual_partition")
args.manual_partition = 1
# any value here is overriden by the config set in notebook when launching the sagemaker job
smp_config = {
"ddp": True,
"tensor_parallel_degree": args.tensor_parallel_degree,
"pipeline_parallel_degree": args.pipeline_parallel_degree,
"microbatches": args.microbatches,
# if activation_checkpointing true checkpoints transformer layers below
"checkpoint_attentions":
False if args.activation_checkpointing else True,
"shard_optimizer_state": args.shard_optimizer_state > 0,
"prescaled_batch": args.prescaled_batch > 0,
"offload_activations": args.offload_activations > 0,
"optimize": args.optimize,
"auto_partition": False if args.manual_partition else True,
"default_partition": 0,
"static_mode": args.static_mode > 0,
"fast_mode": args.fast_mode > 0,
}
if args.smp_version < 110:
smp_config["fp16_params"] = args.fp16 > 0
else:
smp_config["fp16"] = args.fp16 > 0
smp_config["delayed_parameter_initialization"] = args.delayed_param > 0
smp_config["placement_strategy"] = args.placement_strategy
smp_config[
"activation_loading_horizon"] = args.activation_loading_horizon
smp_config["skip_tracing"] = args.skip_tracing > 0
if args.active_microbatches is not None:
smp_config["active_microbatches"] = args.active_microbatches
smp.init(smp_config)
if smp.rank() == 0:
print("Arguments:", args.__dict__)
print(f"Transformers version: {transformers.__version__}")
print(
f"smdistributed.modelparallel version: {smdistributed.modelparallel.__version__}"
)
print(f"smdistributed config: {smp_config}")
if args.save_final_full_model and smp.rank() == 0:
print(
f"[Warning] Note that save_final_full_model only saves the final model at the end of all steps. It does not save optimizer state. Optimizer state is only saved with partial models which are saved at checkpointing_freq during training. If you want to restart training you need partial checkpoints."
)
if args.partition_assignment != "":
partition_assignment = args.partition_assignment.split(",")
assert (
len(partition_assignment) == smp.pp_size()
), f"partition_assignment must have the same size as pipeline parallel degree, but getting {len(partition_assignment)} vs {smp.pp_size()}"
if smp.rank() == 0 or (smp.local_rank() == 0 and args.use_fsx == 0):
for path in [args.model_dir, args.checkpoint_dir]:
if not os.path.exists(path):
os.makedirs(path, exist_ok=True)
model_config = GPT2Config(
vocab_size=args.vocab_size,
n_positions=args.max_context_width,
n_embd=args.hidden_width,
n_layer=args.num_layers,
n_head=args.num_heads,
n_inner=None,
activation_function="gelu_new",
resid_pdrop=args.resid_pdrop,
embd_pdrop=args.embd_pdrop,
attn_pdrop=args.attn_pdrop,
layer_norm_epsilon=1e-05,
initializer_range=0.02,
summary_type="cls_index",
summary_use_proj=True,
summary_activation=None,
summary_proj_to_labels=True,
summary_first_dropout=args.summary_first_pdrop,
# gradient_checkpointing=args.gradient_checkpointing > 0,
use_cache=False,
bos_token_id=50256,
eos_token_id=50256,
return_dict=True,
)
# the following improves start-up time by skipping proper initialization
# of weights in the original model. this is not a problem because DistributedModel
# will override those weights anyway when tensor_parallel_degree > 1.
if smp.tp_size() > 1:
from transformers.modeling_utils import PreTrainedModel
PreTrainedModel.init_weights = lambda x: None
set_seed(args.seed)
if args.enable_memory_profiling > 0:
memory_status_cpu(msg="before model creation")
if args.smp_version < 110:
if args.fp16:
torch.set_default_dtype(torch.float16)
with smp.tensor_parallelism(
enabled=smp.tp_size() > 1,
attention_in_fp32=args.attention_in_fp32 > 0):
with smp.delay_param_initialization(
enabled=(smp.tp_size() > 1 and args.delayed_param > 0)):
model = AutoModelForCausalLM.from_config(model_config)
else:
with smp.model_creation(
tensor_parallelism=smp.tp_size() > 1,
attention_in_fp32=args.attention_in_fp32 > 0,
query_key_layer_scaling=args.query_key_layer_scaling > 0,
fused_softmax=args.fused_softmax > 0,
fused_bias_gelu=args.fused_bias_gelu > 0,
dtype=torch.float16
if args.fp16 else torch.get_default_dtype(),
):
model = AutoModelForCausalLM.from_config(model_config)
if args.smp_version < 110 and args.fp16:
model = FP16_Module(model)
if args.enable_memory_profiling > 0:
memory_status_cpu(msg="after model creation")
num_params = sum([np.prod(p.size()) for p in model.parameters()])
if smp.rank() == 0:
print(f"# total parameters: {num_params}")
# smdistributed: Set the device to the GPU ID used by the current process.
# Input tensors should be transferred to this device.
torch.cuda.set_device(smp.local_rank())
device = torch.device("cuda")
if not args.same_seed:
# Set seed by tp_rank to prevent weights from being the same on different tp_ranks
set_seed(args.seed + smp.tp_rank())
# smdistributed: Use the DistributedModel container to provide the model
# to be partitioned across different ranks. For the rest of the script,
# the returned DistributedModel object should be used in place of
# the model provided for DistributedModel class instantiation.
if args.smp_version < 110 and args.fp16:
torch.set_default_dtype(torch.float16)
if args.enable_memory_profiling > 0:
memory_status_cpu(msg="before dist model creation")
model = smp.DistributedModel(model, trace_device="gpu")
if args.enable_memory_profiling > 0:
memory_status_cpu(msg="after dist model creation")
if args.smp_version < 110:
if smp.tp_size() > 1:
transformer_layers = model.module.module.module.transformer.seq_layers
else:
transformer_layers = model.module.module.module.transformer.h
else:
m = model.get_module()
if smp.tp_size() > 1:
transformer_layers = m.transformer.seq_layers
else:
transformer_layers = m.transformer.h
if args.manual_partition:
print(f"Manual partition enabled")
if args.partition_assignment != "":
get_num_layers = lambda x: int(partition_assignment[x])
total_layers = sum(
[get_num_layers(pp_rank) for pp_rank in range(smp.pp_size())])
assert (
total_layers == args.num_layers
), f"partition_assignment must have the same total transformer layers as model, but getting {total_layers} vs {args.num_layers}"
else:
# evenly distribute layers across all partitions
div, rem = divmod(args.num_layers, smp.pp_size())
get_num_layers = lambda x: (div + 1
if x >= smp.pp_size() - rem else div)
assignments = []
# (TODO) This is required for 175B otherwise a hang for partition "8,17,17,18,18,18"
# Need further investigation
# for pp_rank in reversed(range(smp.pp_size())):
for pp_rank in range(smp.pp_size()):
nl = get_num_layers(pp_rank)
print(f"{nl} layers assigned to partition {pp_rank}")
assignments += [pp_rank for _ in range(nl)]
for i, c in enumerate(transformer_layers.children()):
smp.set_partition(c, assignments[i])
if args.smp_version < 110:
iter_model = model
# Build parameter groups (weight decay and non-decay).
while isinstance(iter_model, (DistributedDataParallel, FP16_Module)):
iter_model = iter_model.module
else:
iter_model = m
param_groups = get_param_groups_by_weight_decay(iter_model)
if args.use_adamw > 0:
optimizer = optim.AdamW(param_groups,
betas=(args.beta1, args.beta2),
lr=args.lr,
weight_decay=args.weight_decay)
else:
optimizer = optim.Adam(param_groups,
betas=(args.beta1, args.beta2),
lr=args.lr,
weight_decay=args.weight_decay)
if args.activation_checkpointing:
kwargs = {}
if isinstance(transformer_layers, nn.Sequential):
kwargs["pack_args_as_tuple"] = True
kwargs["strategy"] = args.activation_strategy
smp.set_activation_checkpointing(transformer_layers, **kwargs)
if args.smp_version < 110:
optimizer = FP16_Optimizer(
model,
optimizer,
static_loss_scale=None,
dynamic_loss_scale=True,
use_smp=True,
dynamic_loss_args={
"scale_window": 1000,
"min_scale": 1,
"delayed_shift": 2
},
params_have_main_grad=False,
shard_optimizer_state=args.shard_optimizer_state > 0,
)
optimizer = smp.DistributedOptimizer(optimizer)
model.register_post_step_hook(
lambda model, optimizer: optimizer.init_master_params())
else:
optimizer = smp.DistributedOptimizer(
optimizer,
static_loss_scale=None,
dynamic_loss_scale=True,
dynamic_loss_args={
"scale_window": 1000,
"min_scale": 1,
"delayed_shift": 2
},
)
lr_scheduler = get_learning_rate_scheduler(optimizer, args)
if args.enable_memory_profiling > 0:
model.register_post_partition_hook(
lambda model, optimizer: memory_status(msg="After_partition"))
# load after wrapping model and optimizer with smp Distributed...
if args.load_full or args.load_partial:
if args.load_partial and args.load_full:
print(
"Since both --load_partial and --load_full set, will try to load from full checkpoint."
"If the intention is to load from partial checkpoint, please don't set --load_full"
)
partial = not args.load_full
path = args.checkpoint_dir if partial else args.model_dir
translate_from_hf = not partial
model, optimizer, total_steps, start_train_path_index, start_batch_index = load_model_and_optimizer(
path,
model,
optimizer,
lr_scheduler,
partial,
args,
translate_from_hf=translate_from_hf,
seq_length=args.max_context_width,
load_model=True,
load_optimizer=args.load_partial > 0,
num_params=num_params,
)
else:
total_steps = 0
start_train_path_index = 0
start_batch_index = 0
start = time.time()
total_steps, throughput, loss = train(
model,
optimizer,
lr_scheduler,
model_config,
start_train_path_index,
start_batch_index,
num_params,
total_steps,
args,
)
time_to_train = time.time() - start
if args.ci:
print(f"[SMP_METRIC]__GPT2__Time_to_train__{time_to_train}")
print(f"[SMP_METRIC]__GPT2__samples/second__{throughput}")
print(f"[SMP_METRIC]__GPT2__Loss__{loss}")
if not args.load_partial and not args.load_full:
assert time_to_train < args.time_to_train
assert throughput > args.throughput
if args.loss:
assert loss < args.loss
if args.save_final_full_model:
# saves full model at the end
base_path = f"trained_gpt_nparams-{num_params}_steps-{total_steps}.pt"
out_path = os.path.join(args.model_dir, base_path)
if smp.rdp_rank() == 0:
save(
out_path,
model,
optimizer,
lr_scheduler,
model_config,
num_params,
total_steps,
-1,
args,
partial=False,
translate_to_hf=smp.tp_size() > 1,
seq_length=args.max_context_width,
)
smp.barrier()
if smp.rank() == 0:
print("SMP training finished successfully")
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
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()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# 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.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
if "validation" not in datasets.keys():
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
)
else:
data_files = {}
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"
datasets = load_dataset(extension, data_files=data_files)
# 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.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
tokenizer_kwargs = {
"cache_dir": model_args.cache_dir,
"use_fast": model_args.use_fast_tokenizer,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
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."
)
if model_args.model_name_or_path:
model = AutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].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 = 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,
)
if data_args.block_size is None:
block_size = tokenizer.model_max_length
if block_size > 1024:
logger.warn(
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.warn(
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: 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.
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,
)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=lm_datasets["train"] if training_args.do_train else None,
eval_dataset=lm_datasets["validation"] if training_args.do_eval else None,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it.
data_collator=default_data_collator,
)
# Training
if training_args.do_train:
model_path = (
model_args.model_name_or_path
if (model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path))
else None
)
train_result = trainer.train(model_path=model_path)
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_result.metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir, "eval_results_clm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in sorted(results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
args = parse_args()
# 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_no_trainer", args)
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
# If we're using tracking, we also need to initialize it here and it will by default pick up all supported trackers
# in the environment
accelerator_log_kwargs = {}
if args.with_tracking:
accelerator_log_kwargs["log_with"] = args.report_to
accelerator_log_kwargs["logging_dir"] = args.output_dir
accelerator = Accelerator(
gradient_accumulation_steps=args.gradient_accumulation_steps,
**accelerator_log_kwargs)
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name,
token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
with open(os.path.join(args.output_dir, ".gitignore"),
"w+") as gitignore:
if "step_*" not in gitignore:
gitignore.write("step_*\n")
if "epoch_*" not in gitignore:
gitignore.write("epoch_*\n")
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# 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 args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
args.dataset_name,
args.dataset_config_name,
split=f"train[:{args.validation_split_percentage}%]",
)
raw_datasets["train"] = load_dataset(
args.dataset_name,
args.dataset_config_name,
split=f"train[{args.validation_split_percentage}%:]",
)
else:
data_files = {}
dataset_args = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
dataset_args["keep_linebreaks"] = not args.no_keep_linebreaks
raw_datasets = load_dataset(extension,
data_files=data_files,
**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[:{args.validation_split_percentage}%]",
**dataset_args,
)
raw_datasets["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{args.validation_split_percentage}%:]",
**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.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.config_name)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
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."
)
if args.model_name_or_path:
model = AutoModelForCausalLM.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# 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])
with accelerator.main_process_first():
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on dataset",
)
if 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 args.block_size > tokenizer.model_max_length:
logger.warning(
f"The block_size passed ({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(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
with accelerator.main_process_first():
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=args.preprocessing_num_workers,
load_from_cache_file=not args.overwrite_cache,
desc=f"Grouping texts in chunks of {block_size}",
)
train_dataset = lm_datasets["train"]
eval_dataset = lm_datasets["validation"]
# 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]}.")
# DataLoaders creation:
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=default_data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataloader = DataLoader(eval_dataset,
collate_fn=default_data_collator,
batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "layer_norm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
optimizer = torch.optim.AdamW(optimizer_grouped_parameters,
lr=args.learning_rate)
# On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
if accelerator.distributed_type == DistributedType.TPU:
model.tie_weights()
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps *
args.gradient_accumulation_steps,
num_training_steps=args.max_train_steps *
args.gradient_accumulation_steps,
)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader, lr_scheduler)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps /
num_update_steps_per_epoch)
# Figure out how many steps we should save the Accelerator states
if hasattr(args.checkpointing_steps, "isdigit"):
checkpointing_steps = args.checkpointing_steps
if args.checkpointing_steps.isdigit():
checkpointing_steps = int(args.checkpointing_steps)
else:
checkpointing_steps = None
# We need to initialize the trackers we use, and also store our configuration.
# We initialize the trackers only on main process because `accelerator.log`
# only logs on main process and we don't want empty logs/runs on other processes.
if args.with_tracking:
if accelerator.is_main_process:
experiment_config = vars(args)
# TensorBoard cannot log Enums, need the raw value
experiment_config["lr_scheduler_type"] = experiment_config[
"lr_scheduler_type"].value
accelerator.init_trackers("clm_no_trainer", experiment_config)
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps),
disable=not accelerator.is_local_main_process)
completed_steps = 0
starting_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint is not None or args.resume_from_checkpoint != "":
accelerator.print(
f"Resumed from checkpoint: {args.resume_from_checkpoint}")
accelerator.load_state(args.resume_from_checkpoint)
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the most recent checkpoint
dirs = [f.name for f in os.scandir(os.getcwd()) if f.is_dir()]
dirs.sort(key=os.path.getctime)
path = dirs[
-1] # Sorts folders by date modified, most recent checkpoint is the last
# Extract `epoch_{i}` or `step_{i}`
training_difference = os.path.splitext(path)[0]
if "epoch" in training_difference:
starting_epoch = int(training_difference.replace("epoch_", "")) + 1
resume_step = None
else:
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace(
"step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
# update the progress_bar if load from checkpoint
progress_bar.update(starting_epoch * num_update_steps_per_epoch)
completed_steps = starting_epoch * num_update_steps_per_epoch
for epoch in range(starting_epoch, args.num_train_epochs):
model.train()
if args.with_tracking:
total_loss = 0
for step, batch in enumerate(train_dataloader):
# We need to skip steps until we reach the resumed step
if args.resume_from_checkpoint and epoch == starting_epoch:
if resume_step is not None and step < resume_step:
if step % args.gradient_accumulation_steps == 0:
progress_bar.update(1)
completed_steps += 1
continue
with accelerator.accumulate(model):
outputs = model(**batch)
loss = outputs.loss
# We keep track of the loss at each epoch
if args.with_tracking:
total_loss += loss.detach().float()
accelerator.backward(loss)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
# Checks if the accelerator has performed an optimization step behind the scenes
if accelerator.sync_gradients:
progress_bar.update(1)
completed_steps += 1
if isinstance(checkpointing_steps, int):
if completed_steps % checkpointing_steps == 0:
output_dir = f"step_{completed_steps }"
if args.output_dir is not None:
output_dir = os.path.join(args.output_dir, output_dir)
accelerator.save_state(output_dir)
if completed_steps >= args.max_train_steps:
break
model.eval()
losses = []
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch)
loss = outputs.loss
losses.append(
accelerator.gather_for_metrics(
loss.repeat(args.per_device_eval_batch_size)))
losses = torch.cat(losses)
try:
eval_loss = torch.mean(losses)
perplexity = math.exp(eval_loss)
except OverflowError:
perplexity = float("inf")
logger.info(
f"epoch {epoch}: perplexity: {perplexity} eval_loss: {eval_loss}")
if args.with_tracking:
accelerator.log(
{
"perplexity": perplexity,
"eval_loss": eval_loss,
"train_loss": total_loss.item() / len(train_dataloader),
"epoch": epoch,
"step": completed_steps,
},
step=completed_steps,
)
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
args.output_dir,
is_main_process=accelerator.is_main_process,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}",
blocking=False,
auto_lfs_prune=True)
if args.checkpointing_steps == "epoch":
output_dir = f"epoch_{epoch}"
if args.output_dir is not None:
output_dir = os.path.join(args.output_dir, output_dir)
accelerator.save_state(output_dir)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
args.output_dir,
is_main_process=accelerator.is_main_process,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training",
auto_lfs_prune=True)
with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
json.dump({"perplexity": perplexity}, f)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
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(
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
logger.info(f"Training/evaluation parameters {training_args}")
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Set seed before initializing model.
set_seed(training_args.seed)
# 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)
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,
)
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,
)
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,
**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,
**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,
**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.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
**config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
**config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if model_args.config_overrides is not None:
logger.info(f"Overriding config: {model_args.config_overrides}")
config.update_from_string(model_args.config_overrides)
logger.info(f"New config: {config}")
tokenizer_kwargs = {
"cache_dir": model_args.cache_dir,
"use_fast": model_args.use_fast_tokenizer,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name,
**tokenizer_kwargs)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, **tokenizer_kwargs)
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."
)
if model_args.model_name_or_path:
model = AutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
model = AutoModelForCausalLM.from_config(config)
n_params = sum(
dict((p.data_ptr(), p.numel())
for p in model.parameters()).values())
logger.info(
f"Training new model from scratch - Total size={n_params/2**20:.2f}M params"
)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = raw_datasets["train"].column_names
else:
column_names = raw_datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
# since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
tok_logger = transformers.utils.logging.get_logger(
"transformers.tokenization_utils_base")
def tokenize_function(examples):
with CaptureLogger(tok_logger) as cl:
output = tokenizer(examples[text_column_name])
# clm input could be much much longer than block_size
if "Token indices sequence length is longer than the" in cl.out:
tok_logger.warning(
"^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
)
return output
with training_args.main_process_first(desc="dataset map tokenization"):
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
with training_args.main_process_first(desc="grouping texts together"):
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}",
)
if training_args.do_train:
if "train" not in tokenized_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = lm_datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
if training_args.do_eval:
if "validation" not in tokenized_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = lm_datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it.
data_collator=default_data_collator,
)
# Training
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
try:
perplexity = math.exp(metrics["eval_loss"])
except OverflowError:
perplexity = float("inf")
metrics["perplexity"] = perplexity
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "text-generation"
}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs[
"dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
def __init__(
self,
config: Optional[PretrainedConfig] = None,
image_encoder: Optional[PreTrainedModel] = None,
command_encoder: Optional[PreTrainedModel] = None,
decoder: Optional[PreTrainedModel] = None,
):
if config is None and (image_encoder is None or command_encoder is None
or decoder is None):
raise ValueError(
"Either a configuration or both encoders and a decoder have to be provided."
)
if config is None:
config = MultimodalEncoderDecoderConfig.from_encoder_decoder_configs(
image_encoder.config, command_encoder.config, decoder.config)
else:
if not isinstance(config, self.config_class):
raise ValueError(
f"Config: {config} has to be of type {self.config_class}")
if config.decoder.cross_attention_hidden_size is not None:
if config.decoder.cross_attention_hidden_size != config.image_encoder.hidden_size or \
config.decoder.cross_attention_hidden_size != config.command_encoder.hidden_size:
raise ValueError(
"If `cross_attention_hidden_size` is specified in the decoder's configuration, "
"it has to be equal to the sum of encoder's `hidden_size`s. "
f"Got {config.decoder.cross_attention_hidden_size} for `config.decoder.cross_attention_hidden_size` "
f"and {config.image_encoder.hidden_size} for `config.image_encoder.hidden_size`"
f"and {config.command_encoder.hidden_size} for `config.command_encoder.hidden_size`."
)
# initialize with config
super().__init__(config)
if image_encoder is None:
from transformers import AutoModel
image_encoder = AutoModel.from_config(config.image_encoder)
if command_encoder is None:
from transformers import AutoModel
command_encoder = AutoModel.from_config(config.command_encoder)
if decoder is None:
from transformers import AutoModelForCausalLM
decoder = AutoModelForCausalLM.from_config(config.decoder)
self.image_encoder = image_encoder
self.command_encoder = command_encoder
self.decoder = decoder
#if self.encoder.config.to_dict() != self.config.encoder.to_dict():
# logger.warning(
# f"Config of the encoder: {self.encoder.__class__} is overwritten by shared encoder config: {self.config.encoder}"
# )
#if self.decoder.config.to_dict() != self.config.decoder.to_dict():
# logger.warning(
# f"Config of the decoder: {self.decoder.__class__} is overwritten by shared decoder config: {self.config.decoder}"
# )
# make sure that the individual model's config refers to the shared config
# so that the updates to the config will be synced
self.image_encoder.config = self.config.image_encoder
self.command_encoder.config = self.config.command_encoder
self.decoder.config = self.config.decoder