732 lines
32 KiB
Python
732 lines
32 KiB
Python
#!/usr/bin/env python
|
|
# coding=utf-8
|
|
# Copyright 2021 The HuggingFace Team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
"""
|
|
Fine-tuning the library models for translation.
|
|
"""
|
|
# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
|
|
|
|
import json
|
|
import logging
|
|
import os
|
|
import sys
|
|
import warnings
|
|
from dataclasses import dataclass, field
|
|
from typing import Optional
|
|
|
|
import datasets
|
|
import evaluate
|
|
import numpy as np
|
|
import tensorflow as tf
|
|
from datasets import load_dataset
|
|
|
|
import transformers
|
|
from transformers import (
|
|
AutoConfig,
|
|
AutoTokenizer,
|
|
DataCollatorForSeq2Seq,
|
|
HfArgumentParser,
|
|
KerasMetricCallback,
|
|
M2M100Tokenizer,
|
|
MBart50Tokenizer,
|
|
MBart50TokenizerFast,
|
|
MBartTokenizer,
|
|
MBartTokenizerFast,
|
|
PushToHubCallback,
|
|
TFAutoModelForSeq2SeqLM,
|
|
TFTrainingArguments,
|
|
create_optimizer,
|
|
set_seed,
|
|
)
|
|
from transformers.trainer_utils import get_last_checkpoint
|
|
from transformers.utils import check_min_version, send_example_telemetry
|
|
from transformers.utils.versions import require_version
|
|
|
|
|
|
# region Dependencies and constants
|
|
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
|
|
check_min_version("4.41.0.dev0")
|
|
|
|
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/summarization/requirements.txt")
|
|
|
|
logger = logging.getLogger(__name__)
|
|
MULTILINGUAL_TOKENIZERS = [MBartTokenizer, MBartTokenizerFast, MBart50Tokenizer, MBart50TokenizerFast, M2M100Tokenizer]
|
|
# endregion
|
|
|
|
|
|
# region Arguments
|
|
@dataclass
|
|
class ModelArguments:
|
|
"""
|
|
Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
|
|
"""
|
|
|
|
model_name_or_path: str = field(
|
|
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
|
|
)
|
|
config_name: Optional[str] = field(
|
|
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
|
|
)
|
|
tokenizer_name: Optional[str] = field(
|
|
default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
|
|
)
|
|
cache_dir: Optional[str] = field(
|
|
default=None,
|
|
metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
|
|
)
|
|
use_fast_tokenizer: bool = field(
|
|
default=True,
|
|
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
|
|
)
|
|
model_revision: str = field(
|
|
default="main",
|
|
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
|
|
)
|
|
token: str = field(
|
|
default=None,
|
|
metadata={
|
|
"help": (
|
|
"The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
|
|
"generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
|
|
)
|
|
},
|
|
)
|
|
use_auth_token: bool = field(
|
|
default=None,
|
|
metadata={
|
|
"help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead."
|
|
},
|
|
)
|
|
trust_remote_code: bool = field(
|
|
default=False,
|
|
metadata={
|
|
"help": (
|
|
"Whether or not to allow for custom models defined on the Hub in their own modeling files. This option "
|
|
"should only be set to `True` for repositories you trust and in which you have read the code, as it will "
|
|
"execute code present on the Hub on your local machine."
|
|
)
|
|
},
|
|
)
|
|
|
|
|
|
@dataclass
|
|
class DataTrainingArguments:
|
|
"""
|
|
Arguments pertaining to what data we are going to input our model for training and eval.
|
|
"""
|
|
|
|
source_lang: str = field(default=None, metadata={"help": "Source language id for translation."})
|
|
target_lang: str = field(default=None, metadata={"help": "Target language id for translation."})
|
|
dataset_name: Optional[str] = field(
|
|
default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
|
|
)
|
|
dataset_config_name: Optional[str] = field(
|
|
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
|
|
)
|
|
train_file: Optional[str] = field(
|
|
default=None, metadata={"help": "The input training data file (a jsonlines or csv file)."}
|
|
)
|
|
validation_file: Optional[str] = field(
|
|
default=None,
|
|
metadata={
|
|
"help": (
|
|
"An optional input evaluation data file to evaluate the metrics (rouge) on (a jsonlines or csv file)."
|
|
)
|
|
},
|
|
)
|
|
test_file: Optional[str] = field(
|
|
default=None,
|
|
metadata={
|
|
"help": "An optional input test data file to evaluate the metrics (rouge) on (a jsonlines or csv file)."
|
|
},
|
|
)
|
|
overwrite_cache: bool = field(
|
|
default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
|
|
)
|
|
preprocessing_num_workers: Optional[int] = field(
|
|
default=None,
|
|
metadata={"help": "The number of processes to use for the preprocessing."},
|
|
)
|
|
max_source_length: Optional[int] = field(
|
|
default=1024,
|
|
metadata={
|
|
"help": (
|
|
"The maximum total input sequence length after tokenization. Sequences longer "
|
|
"than this will be truncated, sequences shorter will be padded."
|
|
)
|
|
},
|
|
)
|
|
max_target_length: Optional[int] = field(
|
|
default=128,
|
|
metadata={
|
|
"help": (
|
|
"The maximum total sequence length for target text after tokenization. Sequences longer "
|
|
"than this will be truncated, sequences shorter will be padded."
|
|
)
|
|
},
|
|
)
|
|
val_max_target_length: Optional[int] = field(
|
|
default=None,
|
|
metadata={
|
|
"help": (
|
|
"The maximum total sequence length for validation target text after tokenization. Sequences longer "
|
|
"than this will be truncated, sequences shorter will be padded. Will default to `max_target_length`. "
|
|
"This argument is also used to override the ``max_length`` param of ``model.generate``, which is used "
|
|
"during ``evaluate`` and ``predict``."
|
|
)
|
|
},
|
|
)
|
|
pad_to_max_length: bool = field(
|
|
default=False,
|
|
metadata={
|
|
"help": (
|
|
"Whether to pad all samples to model maximum sentence length. "
|
|
"If False, will pad the samples dynamically when batching to the maximum length in the batch. More "
|
|
"efficient on GPU but very bad for TPU."
|
|
)
|
|
},
|
|
)
|
|
max_train_samples: Optional[int] = field(
|
|
default=None,
|
|
metadata={
|
|
"help": (
|
|
"For debugging purposes or quicker training, truncate the number of training examples to this "
|
|
"value if set."
|
|
)
|
|
},
|
|
)
|
|
max_eval_samples: Optional[int] = field(
|
|
default=None,
|
|
metadata={
|
|
"help": (
|
|
"For debugging purposes or quicker training, truncate the number of evaluation examples to this "
|
|
"value if set."
|
|
)
|
|
},
|
|
)
|
|
max_predict_samples: Optional[int] = field(
|
|
default=None,
|
|
metadata={
|
|
"help": (
|
|
"For debugging purposes or quicker training, truncate the number of prediction examples to this "
|
|
"value if set."
|
|
)
|
|
},
|
|
)
|
|
num_beams: Optional[int] = field(
|
|
default=1,
|
|
metadata={
|
|
"help": (
|
|
"Number of beams to use for evaluation. This argument will be passed to ``model.generate``, "
|
|
"which is used during ``evaluate`` and ``predict``."
|
|
)
|
|
},
|
|
)
|
|
ignore_pad_token_for_loss: bool = field(
|
|
default=True,
|
|
metadata={
|
|
"help": "Whether to ignore the tokens corresponding to padded labels in the loss computation or not."
|
|
},
|
|
)
|
|
source_prefix: Optional[str] = field(
|
|
default=None, metadata={"help": "A prefix to add before every source text (useful for T5 models)."}
|
|
)
|
|
forced_bos_token: Optional[str] = field(
|
|
default=None,
|
|
metadata={
|
|
"help": (
|
|
"The token to force as the first generated token after the :obj:`decoder_start_token_id`.Useful for"
|
|
" multilingual models like :doc:`mBART <../model_doc/mbart>` where the first generated token needs to"
|
|
" be the target language token.(Usually it is the target language token)"
|
|
)
|
|
},
|
|
)
|
|
|
|
def __post_init__(self):
|
|
if self.dataset_name is None and self.train_file is None and self.validation_file is None:
|
|
raise ValueError("Need either a dataset name or a training/validation file.")
|
|
else:
|
|
if self.train_file is not None:
|
|
extension = self.train_file.split(".")[-1]
|
|
assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
|
|
if self.validation_file is not None:
|
|
extension = self.validation_file.split(".")[-1]
|
|
assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
|
|
if self.val_max_target_length is None:
|
|
self.val_max_target_length = self.max_target_length
|
|
|
|
|
|
# endregion
|
|
|
|
|
|
def main():
|
|
# region Argument parsing
|
|
# 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, TFTrainingArguments))
|
|
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
|
|
# If we pass only one argument to the script and it's the path to a json file,
|
|
# let's parse it to get our arguments.
|
|
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
|
|
else:
|
|
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
|
|
|
if model_args.use_auth_token is not None:
|
|
warnings.warn(
|
|
"The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead.",
|
|
FutureWarning,
|
|
)
|
|
if model_args.token is not None:
|
|
raise ValueError("`token` and `use_auth_token` are both specified. Please set only the argument `token`.")
|
|
model_args.token = model_args.use_auth_token
|
|
|
|
# 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_translation", model_args, data_args, framework="tensorflow")
|
|
# endregion
|
|
|
|
# region 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)
|
|
datasets.utils.logging.set_verbosity(logging.INFO)
|
|
transformers.utils.logging.set_verbosity(logging.INFO)
|
|
|
|
# Log on each process the small summary:
|
|
logger.info(f"Training/evaluation parameters {training_args}")
|
|
# endregion
|
|
|
|
# region 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."
|
|
)
|
|
# endregion
|
|
|
|
# Set seed before initializing model.
|
|
set_seed(training_args.seed)
|
|
|
|
# region Load datasets
|
|
# Get the datasets: you can either provide your own CSV/JSON 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 first column for the full texts and the second column for the
|
|
# summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
|
|
#
|
|
# 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,
|
|
token=model_args.token,
|
|
)
|
|
else:
|
|
data_files = {}
|
|
if data_args.train_file is not None:
|
|
data_files["train"] = data_args.train_file
|
|
extension = data_args.train_file.split(".")[-1]
|
|
if data_args.validation_file is not None:
|
|
data_files["validation"] = data_args.validation_file
|
|
extension = data_args.validation_file.split(".")[-1]
|
|
raw_datasets = load_dataset(
|
|
extension,
|
|
data_files=data_files,
|
|
cache_dir=model_args.cache_dir,
|
|
token=model_args.token,
|
|
)
|
|
# 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
|
|
# endregion
|
|
|
|
# region Load model config and tokenizer
|
|
#
|
|
# Distributed training:
|
|
# The .from_pretrained methods guarantee that only one local process can concurrently
|
|
# download model & vocab.
|
|
|
|
config = AutoConfig.from_pretrained(
|
|
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
|
|
cache_dir=model_args.cache_dir,
|
|
revision=model_args.model_revision,
|
|
token=model_args.token,
|
|
trust_remote_code=model_args.trust_remote_code,
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained(
|
|
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
|
|
cache_dir=model_args.cache_dir,
|
|
use_fast=model_args.use_fast_tokenizer,
|
|
revision=model_args.model_revision,
|
|
token=model_args.token,
|
|
trust_remote_code=model_args.trust_remote_code,
|
|
)
|
|
|
|
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
|
|
# endregion
|
|
|
|
# region Dataset preprocessing
|
|
# We need to tokenize inputs and targets.
|
|
if training_args.do_train:
|
|
column_names = raw_datasets["train"].column_names
|
|
elif training_args.do_eval:
|
|
column_names = raw_datasets["validation"].column_names
|
|
else:
|
|
logger.info("There is nothing to do. Please pass `do_train`, and/or `do_eval`.")
|
|
return
|
|
|
|
column_names = raw_datasets["train"].column_names
|
|
|
|
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
|
|
# ignore those attributes).
|
|
if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
|
|
assert data_args.target_lang is not None and data_args.source_lang is not None, (
|
|
f"{tokenizer.__class__.__name__} is a multilingual tokenizer which requires --source_lang and "
|
|
"--target_lang arguments."
|
|
)
|
|
tokenizer.src_lang = data_args.source_lang
|
|
tokenizer.tgt_lang = data_args.target_lang
|
|
forced_bos_token_id = (
|
|
tokenizer.lang_code_to_id[data_args.forced_bos_token] if data_args.forced_bos_token is not None else None
|
|
)
|
|
|
|
# Get the language codes for input/target.
|
|
source_lang = data_args.source_lang.split("_")[0]
|
|
target_lang = data_args.target_lang.split("_")[0]
|
|
|
|
padding = "max_length" if data_args.pad_to_max_length else False
|
|
|
|
# Temporarily set max_target_length for training.
|
|
max_target_length = data_args.max_target_length
|
|
padding = "max_length" if data_args.pad_to_max_length else False
|
|
|
|
def preprocess_function(examples):
|
|
inputs = [ex[source_lang] for ex in examples["translation"]]
|
|
targets = [ex[target_lang] for ex in examples["translation"]]
|
|
inputs = [prefix + inp for inp in inputs]
|
|
model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, padding=padding, truncation=True)
|
|
|
|
# Tokenize targets with the `text_target` keyword argument
|
|
labels = tokenizer(text_target=targets, max_length=max_target_length, padding=padding, truncation=True)
|
|
|
|
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
|
|
# padding in the loss.
|
|
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
|
|
labels["input_ids"] = [
|
|
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
|
|
]
|
|
|
|
model_inputs["labels"] = labels["input_ids"]
|
|
return model_inputs
|
|
|
|
if training_args.do_train:
|
|
if "train" not in raw_datasets:
|
|
raise ValueError("--do_train requires a train dataset")
|
|
train_dataset = raw_datasets["train"]
|
|
if data_args.max_train_samples is not None:
|
|
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
|
|
train_dataset = train_dataset.select(range(max_train_samples))
|
|
train_dataset = train_dataset.map(
|
|
preprocess_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 train dataset",
|
|
)
|
|
else:
|
|
train_dataset = None
|
|
|
|
if training_args.do_eval:
|
|
max_target_length = data_args.val_max_target_length
|
|
if "validation" not in raw_datasets:
|
|
raise ValueError("--do_eval requires a validation dataset")
|
|
eval_dataset = raw_datasets["validation"]
|
|
if data_args.max_eval_samples is not None:
|
|
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
|
|
eval_dataset = eval_dataset.select(range(max_eval_samples))
|
|
eval_dataset = eval_dataset.map(
|
|
preprocess_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 validation dataset",
|
|
)
|
|
else:
|
|
eval_dataset = None
|
|
# endregion
|
|
|
|
with training_args.strategy.scope():
|
|
# region Prepare model
|
|
model = TFAutoModelForSeq2SeqLM.from_pretrained(
|
|
model_args.model_name_or_path,
|
|
config=config,
|
|
cache_dir=model_args.cache_dir,
|
|
revision=model_args.model_revision,
|
|
token=model_args.token,
|
|
trust_remote_code=model_args.trust_remote_code,
|
|
)
|
|
|
|
# We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
|
|
# on a small vocab and want a smaller embedding size, remove this test.
|
|
embeddings = model.get_input_embeddings()
|
|
|
|
# Matt: This is a temporary workaround as we transition our models to exclusively using Keras embeddings.
|
|
# As soon as the transition is complete, all embeddings should be keras.Embeddings layers, and
|
|
# the weights will always be in embeddings.embeddings.
|
|
if hasattr(embeddings, "embeddings"):
|
|
embedding_size = embeddings.embeddings.shape[0]
|
|
else:
|
|
embedding_size = embeddings.weight.shape[0]
|
|
if len(tokenizer) > embedding_size:
|
|
model.resize_token_embeddings(len(tokenizer))
|
|
|
|
if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
|
|
model.config.forced_bos_token_id = forced_bos_token_id
|
|
# endregion
|
|
|
|
# region Set decoder_start_token_id
|
|
if model.config.decoder_start_token_id is None and isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
|
|
assert (
|
|
data_args.target_lang is not None and data_args.source_lang is not None
|
|
), "mBart requires --target_lang and --source_lang"
|
|
if isinstance(tokenizer, MBartTokenizer):
|
|
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[data_args.target_lang]
|
|
else:
|
|
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(data_args.target_lang)
|
|
|
|
if model.config.decoder_start_token_id is None:
|
|
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
|
|
# endregion
|
|
|
|
# region Prepare TF Dataset objects
|
|
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
|
|
data_collator = DataCollatorForSeq2Seq(
|
|
tokenizer,
|
|
model=model,
|
|
label_pad_token_id=label_pad_token_id,
|
|
pad_to_multiple_of=64, # Reduce the number of unique shapes for XLA, especially for generation
|
|
return_tensors="np",
|
|
)
|
|
num_replicas = training_args.strategy.num_replicas_in_sync
|
|
total_train_batch_size = training_args.per_device_train_batch_size * num_replicas
|
|
total_eval_batch_size = training_args.per_device_eval_batch_size * num_replicas
|
|
|
|
dataset_options = tf.data.Options()
|
|
dataset_options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
|
|
|
|
# model.prepare_tf_dataset() wraps a Hugging Face dataset in a tf.data.Dataset which is ready to use in
|
|
# training. This is the recommended way to use a Hugging Face dataset when training with Keras. You can also
|
|
# use the lower-level dataset.to_tf_dataset() method, but you will have to specify things like column names
|
|
# yourself if you use this method, whereas they are automatically inferred from the model input names when
|
|
# using model.prepare_tf_dataset()
|
|
# For more info see the docs:
|
|
# https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.TFPreTrainedModel.prepare_tf_dataset
|
|
# https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset.to_tf_dataset
|
|
|
|
tf_train_dataset = model.prepare_tf_dataset(
|
|
train_dataset,
|
|
collate_fn=data_collator,
|
|
batch_size=total_train_batch_size,
|
|
shuffle=True,
|
|
).with_options(dataset_options)
|
|
tf_eval_dataset = model.prepare_tf_dataset(
|
|
eval_dataset, collate_fn=data_collator, batch_size=total_eval_batch_size, shuffle=False
|
|
).with_options(dataset_options)
|
|
# endregion
|
|
|
|
# region Optimizer and LR scheduling
|
|
num_train_steps = int(len(tf_train_dataset) * training_args.num_train_epochs)
|
|
if training_args.warmup_steps > 0:
|
|
num_warmup_steps = training_args.warmup_steps
|
|
elif training_args.warmup_ratio > 0:
|
|
num_warmup_steps = int(num_train_steps * training_args.warmup_ratio)
|
|
else:
|
|
num_warmup_steps = 0
|
|
if training_args.do_train:
|
|
optimizer, lr_schedule = create_optimizer(
|
|
init_lr=training_args.learning_rate,
|
|
num_train_steps=num_train_steps,
|
|
num_warmup_steps=num_warmup_steps,
|
|
adam_beta1=training_args.adam_beta1,
|
|
adam_beta2=training_args.adam_beta2,
|
|
adam_epsilon=training_args.adam_epsilon,
|
|
weight_decay_rate=training_args.weight_decay,
|
|
adam_global_clipnorm=training_args.max_grad_norm,
|
|
)
|
|
else:
|
|
optimizer = "sgd" # Just write anything because we won't be using it
|
|
# endregion
|
|
|
|
# region Metric and postprocessing
|
|
if training_args.do_eval:
|
|
metric = evaluate.load("sacrebleu", cache_dir=model_args.cache_dir)
|
|
|
|
if data_args.val_max_target_length is None:
|
|
data_args.val_max_target_length = data_args.max_target_length
|
|
|
|
gen_kwargs = {
|
|
"max_length": data_args.val_max_target_length,
|
|
"num_beams": data_args.num_beams,
|
|
"no_repeat_ngram_size": 0, # Not supported under XLA right now, and some models set it by default
|
|
}
|
|
|
|
def postprocess_text(preds, labels):
|
|
preds = [pred.strip() for pred in preds]
|
|
labels = [[label.strip()] for label in labels]
|
|
|
|
return preds, labels
|
|
|
|
def compute_metrics(preds):
|
|
predictions, labels = preds
|
|
if isinstance(predictions, tuple):
|
|
predictions = predictions[0]
|
|
decoded_preds = tokenizer.batch_decode(predictions, skip_special_tokens=True)
|
|
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
|
|
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
|
|
decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
|
|
metrics = metric.compute(predictions=decoded_preds, references=decoded_labels)
|
|
return {"bleu": metrics["score"]}
|
|
|
|
# The KerasMetricCallback allows metrics that are too complex to write as standard Keras metrics
|
|
# to be computed each epoch. Any Python code can be included in the metric_fn. This is especially
|
|
# useful for metrics like BLEU and ROUGE that perform string comparisons on decoded model outputs.
|
|
# For more information, see the docs at
|
|
# https://huggingface.co/docs/transformers/main_classes/keras_callbacks#transformers.KerasMetricCallback
|
|
|
|
metric_callback = KerasMetricCallback(
|
|
metric_fn=compute_metrics,
|
|
eval_dataset=tf_eval_dataset,
|
|
predict_with_generate=True,
|
|
use_xla_generation=True,
|
|
generate_kwargs=gen_kwargs,
|
|
)
|
|
callbacks = [metric_callback]
|
|
else:
|
|
callbacks = []
|
|
|
|
# endregion
|
|
|
|
# region Preparing push_to_hub and model card
|
|
push_to_hub_model_id = training_args.push_to_hub_model_id
|
|
model_name = model_args.model_name_or_path.split("/")[-1]
|
|
if not push_to_hub_model_id:
|
|
push_to_hub_model_id = f"{model_name}-finetuned-{data_args.source_lang}-{data_args.target_lang}"
|
|
|
|
model_card_kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "translation"}
|
|
if data_args.dataset_name is not None:
|
|
model_card_kwargs["dataset_tags"] = data_args.dataset_name
|
|
if data_args.dataset_config_name is not None:
|
|
model_card_kwargs["dataset_args"] = data_args.dataset_config_name
|
|
model_card_kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
|
|
else:
|
|
model_card_kwargs["dataset"] = data_args.dataset_name
|
|
|
|
languages = [l for l in [data_args.source_lang, data_args.target_lang] if l is not None]
|
|
if len(languages) > 0:
|
|
model_card_kwargs["language"] = languages
|
|
|
|
if training_args.push_to_hub:
|
|
# Because this training can be quite long, we save once per epoch.
|
|
callbacks.append(
|
|
PushToHubCallback(
|
|
output_dir=training_args.output_dir,
|
|
hub_model_id=push_to_hub_model_id,
|
|
hub_token=training_args.push_to_hub_token,
|
|
tokenizer=tokenizer,
|
|
**model_card_kwargs,
|
|
)
|
|
)
|
|
# endregion
|
|
|
|
# region Training
|
|
eval_metrics = None
|
|
# Transformers models compute the right loss for their task by default when labels are passed, and will
|
|
# use this for training unless you specify your own loss function in compile().
|
|
model.compile(optimizer=optimizer, jit_compile=training_args.xla)
|
|
|
|
if training_args.do_train:
|
|
logger.info("***** Running training *****")
|
|
logger.info(f" Num examples = {len(train_dataset)}")
|
|
logger.info(f" Num Epochs = {training_args.num_train_epochs}")
|
|
logger.info(f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
|
|
logger.info(f" Total train batch size = {total_train_batch_size}")
|
|
logger.info(f" Total optimization steps = {num_train_steps}")
|
|
|
|
if training_args.xla and not data_args.pad_to_max_length:
|
|
logger.warning(
|
|
"XLA training may be slow at first when --pad_to_max_length is not set "
|
|
"until all possible shapes have been compiled."
|
|
)
|
|
|
|
history = model.fit(tf_train_dataset, epochs=int(training_args.num_train_epochs), callbacks=callbacks)
|
|
eval_metrics = {key: val[-1] for key, val in history.history.items()}
|
|
# endregion
|
|
|
|
# region Validation
|
|
if training_args.do_eval and not training_args.do_train:
|
|
# Compiling generation with XLA yields enormous speedups, see https://huggingface.co/blog/tf-xla-generate
|
|
@tf.function(jit_compile=True)
|
|
def generate(**kwargs):
|
|
return model.generate(**kwargs)
|
|
|
|
if training_args.do_eval:
|
|
logger.info("Evaluation...")
|
|
for batch, labels in tf_eval_dataset:
|
|
batch.update(gen_kwargs)
|
|
generated_tokens = generate(**batch)
|
|
if isinstance(generated_tokens, tuple):
|
|
generated_tokens = generated_tokens[0]
|
|
decoded_preds = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
|
|
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
|
|
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
|
|
decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
|
|
|
|
metric.add_batch(predictions=decoded_preds, references=decoded_labels)
|
|
|
|
eval_metrics = metric.compute()
|
|
logger.info({"bleu": eval_metrics["score"]})
|
|
# endregion
|
|
|
|
if training_args.output_dir is not None and eval_metrics is not None:
|
|
output_eval_file = os.path.join(training_args.output_dir, "all_results.json")
|
|
with open(output_eval_file, "w") as writer:
|
|
writer.write(json.dumps(eval_metrics))
|
|
|
|
if training_args.output_dir is not None and not training_args.push_to_hub:
|
|
# If we're not pushing to hub, at least save a local copy when we're done
|
|
model.save_pretrained(training_args.output_dir)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|