def get_eval_dataloader(self,
eval_dataset: Optional[Dataset] = None
) -> DataLoader:
"""
Returns the evaluation :class:`~torch.utils.data.DataLoader`.
Args:
eval_dataset (:obj:`Dataset`, `optional`):
If provided, will override `self.eval_dataset`.
"""
if eval_dataset is None and self.eval_dataset is None:
raise ValueError("Trainer: evaluation requires an eval_dataset.")
eval_dataset = eval_dataset if eval_dataset is not None else self.eval_dataset
if is_torch_tpu_available():
sampler = SequentialDistributedSampler(
eval_dataset,
num_replicas=xm.xrt_world_size(),
rank=xm.get_ordinal())
elif self.args.local_rank != -1:
sampler = SequentialDistributedSampler(eval_dataset)
else:
sampler = SequentialSampler(eval_dataset)
data_loader = DataLoader(
eval_dataset,
sampler=sampler
if not isinstance(self.train_dataset, IterableDataset) else None,
batch_size=self.args.eval_batch_size,
collate_fn=self.data_collator,
drop_last=self.args.dataloader_drop_last,
)
return data_loader
def get_test_dataloader(self, test_dataset: Dataset) -> DataLoader:
"""
Returns the test :class:`~torch.utils.data.DataLoader`.
Args:
test_dataset (obj:`Dataset`): The test dataset to use.
"""
# We use the same batch_size as for eval.
if is_torch_tpu_available():
sampler = SequentialDistributedSampler(
test_dataset,
num_replicas=xm.xrt_world_size(),
rank=xm.get_ordinal())
elif self.args.local_rank != -1:
sampler = SequentialDistributedSampler(test_dataset)
else:
sampler = SequentialSampler(test_dataset)
data_loader = DataLoader(
test_dataset,
sampler=sampler
if not isinstance(self.train_dataset, IterableDataset) else None,
batch_size=self.args.eval_batch_size,
collate_fn=self.data_collator,
drop_last=self.args.dataloader_drop_last,
)
return data_loader
def _get_world_size(self, local_rank: int):
world_size = 1
if is_torch_tpu_available():
raise NotImplementedError()
# world_size = xm.xrt_world_size()
elif local_rank != -1:
# noinspection PyUnresolvedReferences
world_size = torch.distributed.get_world_size()
return max(1, world_size)
def _gather_and_numpify(self, tensors):
"""
Gather value of `tensors` (tensor or list/tuple of nested tensors) and convert them to numpy before
concatenating them to `gathered`
"""
if tensors is None:
return
if is_torch_tpu_available():
# tensors = nested_xla_mesh_reduce(tensors, name)
raise NotImplementedError()
elif self.local_rank != -1:
tensors = distributed_concat(tensors)
return nested_numpify(tensors)
def get_dataloader(self, dataset, batch_size):
if dataset is None:
raise ValueError("Trainer: requires a dataset.")
if transformers.is_torch_tpu_available():
sampler = get_tpu_sampler(dataset)
elif not isinstance(dataset, IterableDataset):
sampler = (RandomSampler(dataset) if self.args.local_rank == -1
else DistributedSampler(dataset))
else:
sampler = None
data_loader = DataLoader(
dataset,
batch_size=batch_size,
sampler=sampler,
collate_fn=self.data_collator,
)
return data_loader
def get_single_train_dataloader(self, task_name, train_dataset):
"""
Create a single-task data loader that also yields task names
"""
if self.train_dataset is None:
raise ValueError("Trainer: training requires a train_dataset.")
if False and is_torch_tpu_available():
train_sampler = get_tpu_sampler(train_dataset)
else:
train_sampler = (RandomSampler(train_dataset)
if self.args.local_rank == -1 else
DistributedSampler(train_dataset))
data_loader = DataLoaderWithTaskname(
task_name=task_name,
data_loader=DataLoader(
train_dataset,
batch_size=self.args.train_batch_size,
sampler=train_sampler,
collate_fn=self.data_collator.collate_batch,
),
)
return data_loader
def get_train_dataloader(self) -> DataLoader:
"""
Returns the training :class:`~torch.utils.data.DataLoader`.
"""
if self.train_dataset is None:
raise ValueError("Trainer: training requires a train_dataset.")
if is_torch_tpu_available():
train_sampler = get_tpu_sampler(self.train_dataset)
else:
train_sampler = (RandomSampler(self.train_dataset)
if self.args.local_rank == -1 else
DistributedSampler(self.train_dataset))
data_loader = DataLoader(
self.train_dataset,
batch_size=self.args.train_batch_size,
sampler=train_sampler
if not isinstance(self.train_dataset, IterableDataset) else None,
collate_fn=self.data_collator,
drop_last=self.args.dataloader_drop_last,
)
return data_loader
def train(self, model_path=None, dev_objective=None):
"""
Main training entry point.
The training logic is directly borrowed from transformers.Trainer (version 3.0.2).
Add early stopping.
"""
self.best_dir = None
self.objective = -float("inf")
self.dev_objective = dev_objective if dev_objective is not None else default_dev_objective
# Data loading.
train_dataloader = self.get_train_dataloader()
num_update_steps_per_epoch = len(
train_dataloader) // self.args.gradient_accumulation_steps
if num_update_steps_per_epoch == 0:
num_update_steps_per_epoch = 1
if self.args.max_steps > 0:
t_total = self.args.max_steps
num_train_epochs = self.args.max_steps // num_update_steps_per_epoch + int(
self.args.max_steps % num_update_steps_per_epoch > 0)
else:
t_total = int(
len(train_dataloader) //
self.args.gradient_accumulation_steps *
self.args.num_train_epochs)
num_train_epochs = self.args.num_train_epochs
self.create_optimizer_and_scheduler(num_training_steps=t_total)
optimizer = self.optimizer
scheduler = self.lr_scheduler
# Check if saved optimizer or scheduler states exist
if (model_path is not None
and os.path.isfile(os.path.join(model_path, "optimizer.pt"))
and os.path.isfile(os.path.join(model_path, "scheduler.pt"))):
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(model_path, "optimizer.pt"),
map_location=self.args.device))
scheduler.load_state_dict(
torch.load(os.path.join(model_path, "scheduler.pt")))
model = self.model
if self.args.fp16 and _use_apex:
if not transformers.is_apex_available():
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(
model, optimizer, opt_level=self.args.fp16_opt_level)
# Multi-gpu training (should be after apex fp16 initialization)
if self.args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if self.args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[self.args.local_rank],
output_device=self.args.local_rank,
find_unused_parameters=True,
)
# Train
if transformers.is_torch_tpu_available():
total_train_batch_size = self.args.train_batch_size * xm.xrt_world_size(
)
else:
total_train_batch_size = (self.args.train_batch_size *
self.args.gradient_accumulation_steps *
(torch.distributed.get_world_size()
if self.args.local_rank != -1 else 1))
logger.info("***** Running training *****")
logger.info(" Num examples = %d", self.num_examples(train_dataloader))
logger.info(" Num Epochs = %d", num_train_epochs)
logger.info(" Instantaneous batch size per device = %d",
self.args.per_device_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
total_train_batch_size)
logger.info(" Gradient Accumulation steps = %d",
self.args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
self.global_step = 0
self.epoch = 0
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if model_path is not None:
# set global_step to global_step of last saved checkpoint from model path
try:
self.global_step = int(model_path.split("-")[-1].split("/")[0])
epochs_trained = self.global_step // (
len(train_dataloader) //
self.args.gradient_accumulation_steps)
steps_trained_in_current_epoch = self.global_step % (
len(train_dataloader) //
self.args.gradient_accumulation_steps)
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from epoch %d",
epochs_trained)
logger.info(" Continuing training from global step %d",
self.global_step)
logger.info(
" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch)
except ValueError:
self.global_step = 0
logger.info(" Starting fine-tuning.")
tr_loss = torch.tensor(0.0).to(self.args.device)
logging_loss_scalar = 0.0
model.zero_grad()
train_iterator = trange(epochs_trained,
int(num_train_epochs),
desc="Epoch",
disable=not self.is_local_master())
for epoch in train_iterator:
if isinstance(train_dataloader, DataLoader) and isinstance(
train_dataloader.sampler, DistributedSampler):
train_dataloader.sampler.set_epoch(epoch)
if transformers.is_torch_tpu_available():
parallel_loader = pl.ParallelLoader(
train_dataloader,
[self.args.device]).per_device_loader(self.args.device)
epoch_iterator = tqdm(parallel_loader,
desc="Iteration",
disable=not self.is_local_master())
else:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=True)
# Reset the past mems state at the beginning of each epoch if necessary.
if self.args.past_index >= 0:
self._past = None
for step, inputs in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
tr_loss += self.training_step(model, inputs)
if (step + 1) % self.args.gradient_accumulation_steps == 0 or (
# last step in epoch but step is always smaller than gradient_accumulation_steps
len(epoch_iterator) <=
self.args.gradient_accumulation_steps and
(step + 1) == len(epoch_iterator)):
if self.args.fp16 and _use_native_amp:
self.scaler.unscale_(optimizer)
norm = torch.nn.utils.clip_grad_norm_(
model.parameters(), self.args.max_grad_norm)
elif self.args.fp16:
norm = torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer),
self.args.max_grad_norm)
else:
norm = torch.nn.utils.clip_grad_norm_(
model.parameters(), self.args.max_grad_norm)
if transformers.is_torch_tpu_available():
xm.optimizer_step(optimizer)
elif self.args.fp16 and _use_native_amp:
self.scaler.step(optimizer)
self.scaler.update()
else:
optimizer.step()
scheduler.step()
model.zero_grad()
self.global_step += 1
self.epoch = epoch + (step + 1) / len(epoch_iterator)
if (self.args.logging_steps > 0
and self.global_step % self.args.logging_steps
== 0) or (self.global_step == 1
and self.args.logging_first_step):
logs = {}
tr_loss_scalar = tr_loss.item()
logs["loss"] = (tr_loss_scalar - logging_loss_scalar
) / self.args.logging_steps
logs["norm"] = norm.item()
# backward compatibility for pytorch schedulers
logs["learning_rate"] = (
scheduler.get_last_lr()[0]
if version.parse(torch.__version__) >=
version.parse("1.4") else scheduler.get_lr()[0])
logging_loss_scalar = tr_loss_scalar
self.log(logs)
# ----------------------------------------------------------------------
# BEGIN CHANGES.
# ----------------------------------------------------------------------
metrics = None
if self.args.evaluate_during_training and self.global_step % self.args.eval_steps == 0:
output = self.evaluate()
metrics = output.metrics
objective = self.dev_objective(metrics)
if objective > self.objective:
logger.info(
"Best dev result: {}".format(objective))
self.objective = objective
self.save_model(self.args.output_dir)
# ----------------------------------------------------------------------
# END CHANGES.
# ----------------------------------------------------------------------
if self.args.max_steps > 0 and self.global_step > self.args.max_steps:
epoch_iterator.close()
break
if self.args.max_steps > 0 and self.global_step > self.args.max_steps:
train_iterator.close()
break
if self.args.tpu_metrics_debug or self.args.debug:
# tpu-comment: Logging debug metrics for PyTorch/XLA (compile, execute times, ops, etc.)
xm.master_print(met.metrics_report())
if self.args.past_index and hasattr(self, "_past"):
# Clean the state at the end of training
delattr(self, "_past")
logger.info(
"\n\nTraining completed. Do not forget to share your model on huggingface.co/models =)\n\n"
)
return TrainOutput(self.global_step,
tr_loss / self.global_step), self.objective
def train(self,
model_name,
batch_size,
sequence_length,
trace_memory=False):
try:
config = self.config_dict[model_name]
if self.args.torchscript:
config.torchscript = True
model = MODEL_WITH_LM_HEAD_MAPPING[config.__class__](config)
model.to(self.args.device)
model.train()
# encoder-decoder has vocab size saved differently
vocab_size = config.vocab_size if hasattr(
config, "vocab_size") else config.encoder.vocab_size
input_ids = torch.randint(vocab_size,
(batch_size, sequence_length),
dtype=torch.long,
device=self.args.device)
if self.args.torchscript:
raise NotImplementedError(
"Training for torchscript is currently not implemented")
else:
train_model = model
def compute_loss_and_backprob_encoder():
loss = train_model(input_ids, labels=input_ids)[0]
loss.backward()
train_model.zero_grad()
def compute_loss_and_backprob_encoder_decoder():
loss = train_model(input_ids,
decoder_input_ids=input_ids,
labels=input_ids)[0]
loss.backward()
train_model.zero_grad()
_train = (compute_loss_and_backprob_encoder_decoder
if config.is_encoder_decoder else
compute_loss_and_backprob_encoder)
if trace_memory is True:
if self.args.trace_memory_line_by_line:
trace = start_memory_tracing("transformers")
if self.args.n_gpu > 0:
# gpu
# clear gpu cache
torch.cuda.empty_cache()
if hasattr(torch.cuda, "max_memory_reserved"):
torch.cuda.reset_peak_memory_stats()
else:
logger.info(
"Please consider updating PyTorch to version 1.4 to get more accuracy on GPU memory usage"
)
torch.cuda.reset_max_memory_cached()
# calculate loss and do backpropagation
_train()
elif not self.args.no_tpu and is_torch_tpu_available():
# tpu
raise NotImplementedError(
"Memory Benchmarking is currently not implemented for TPU. Please disable memory benchmarking with `args.no_memory=True`"
)
else:
# cpu
memory_bytes = measure_peak_memory_cpu(_train)
memory = Memory(memory_bytes) if isinstance(
memory_bytes, int) else memory_bytes
if self.args.trace_memory_line_by_line:
summary = stop_memory_tracing(trace)
else:
summary = None
if self.args.n_gpu > 0:
# gpu
if hasattr(torch.cuda, "max_memory_reserved"):
memory = Memory(torch.cuda.max_memory_reserved())
else:
logger.info(
"Please consider updating PyTorch to version 1.4 to get more accuracy on GPU memory usage"
)
memory = Memory(torch.cuda.max_memory_cached())
memory = Memory(torch.cuda.max_memory_reserved())
return memory, summary
else:
if (not self.args.no_tpu
and is_torch_tpu_available()) or self.args.torchscript:
# run additional 10 times to stabilize compilation for tpu and torchscript
logger.info(
"Do inference on TPU or torchscript. Running model 5 times to stabilize compilation"
)
timeit.repeat(
_train,
repeat=1,
number=5,
)
# as written in https://docs.python.org/2/library/timeit.html#timeit.Timer.repeat, min should be taken rather than the average
runtimes = timeit.repeat(
_train,
repeat=self.args.repeat,
number=10,
)
if not self.args.no_tpu and is_torch_tpu_available(
) and self.args.tpu_print_metrics:
import torch_xla.debug.metrics as met
self.print_fn(met.metrics_report())
return min(runtimes) / 10.0
except RuntimeError as e:
self.print_fn("Doesn't fit on GPU. {}".format(e))
if trace_memory:
return "N/A", None
else:
return "N/A"
#
# 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.
"""
A subclass of `Trainer` specific to Question-Answering tasks
"""
from transformers import Trainer, is_torch_tpu_available
from transformers.trainer_utils import PredictionOutput
if is_torch_tpu_available(check_device=False):
import torch_xla.core.xla_model as xm
import torch_xla.debug.metrics as met
class QuestionAnsweringTrainer(Trainer):
def __init__(self,
*args,
eval_examples=None,
post_process_function=None,
**kwargs):
super().__init__(*args, **kwargs)
self.eval_examples = eval_examples
self.post_process_function = post_process_function
def evaluate(self,
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(
)
# 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_mlm", model_args, data_args)
# 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}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
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. You can easily tweak this
# behavior (see below)
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
data_files = {}
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]
if extension == "txt":
extension = "text"
raw_datasets = load_dataset(
extension,
data_files=data_files,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
# If no validation data is there, validation_split_percentage will be used to divide the dataset.
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
extension,
data_files=data_files,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
raw_datasets["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
# 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 = AutoModelForMaskedLM.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 = AutoModelForMaskedLM.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 = 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]
if data_args.max_seq_length is None:
max_seq_length = tokenizer.model_max_length
if max_seq_length > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
)
max_seq_length = 1024
else:
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length,
tokenizer.model_max_length)
if data_args.line_by_line:
# When using line_by_line, we just tokenize each nonempty line.
padding = "max_length" if data_args.pad_to_max_length else False
def tokenize_function(examples):
# Remove empty lines
examples[text_column_name] = [
line for line in examples[text_column_name]
if len(line) > 0 and not line.isspace()
]
return tokenizer(
examples[text_column_name],
padding=padding,
truncation=True,
max_length=max_seq_length,
# We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
# receives the `special_tokens_mask`.
return_special_tokens_mask=True,
)
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=[text_column_name],
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on dataset line_by_line",
)
else:
# Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
# We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
# efficient when it receives the `special_tokens_mask`.
def tokenize_function(examples):
return tokenizer(examples[text_column_name],
return_special_tokens_mask=True)
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 every text in dataset",
)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of
# max_seq_length.
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 >= max_seq_length:
total_length = (total_length //
max_seq_length) * max_seq_length
# Split by chunks of max_len.
result = {
k: [
t[i:i + max_seq_length]
for i in range(0, total_length, max_seq_length)
]
for k, t in concatenated_examples.items()
}
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"):
tokenized_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 {max_seq_length}",
)
if training_args.do_train:
if "train" not in tokenized_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = tokenized_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))
if training_args.do_eval:
if "validation" not in tokenized_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = tokenized_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))
def preprocess_logits_for_metrics(logits, labels):
if isinstance(logits, tuple):
# Depending on the model and config, logits may contain extra tensors,
# like past_key_values, but logits always come first
logits = logits[0]
return logits.argmax(dim=-1)
metric = load_metric("accuracy")
def compute_metrics(eval_preds):
preds, labels = eval_preds
# preds have the same shape as the labels, after the argmax(-1) has been calculated
# by preprocess_logits_for_metrics
labels = labels.reshape(-1)
preds = preds.reshape(-1)
mask = labels != -100
labels = labels[mask]
preds = preds[mask]
return metric.compute(predictions=preds, references=labels)
# Data collator
# This one will take care of randomly masking the tokens.
pad_to_multiple_of_8 = data_args.line_by_line and training_args.fp16 and not data_args.pad_to_max_length
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer,
mlm_probability=data_args.mlm_probability,
pad_to_multiple_of=8 if pad_to_multiple_of_8 else None,
)
# 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=data_collator,
compute_metrics=compute_metrics
if training_args.do_eval and not is_torch_tpu_available() else None,
preprocess_logits_for_metrics=preprocess_logits_for_metrics
if training_args.do_eval and not is_torch_tpu_available() else None,
)
# 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": "fill-mask"
}
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 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. 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 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("Model parameters %s", model_args)
logger.info("Data parameters %s", data_args)
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.config_name:
config = BertConfig.from_pretrained(model_args.config_name,
cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = BertConfig.from_pretrained(model_args.model_name_or_path,
cache_dir=model_args.cache_dir)
else:
config = BertConfig()
logger.warning(
"You are instantiating a new config instance from scratch.")
config.loss_fct = model_args.loss_fct
if model_args.tokenizer_name:
tokenizer = BertTokenizer.from_pretrained(
model_args.tokenizer_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
tokenizer = BertTokenizer.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported, but 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 = BertForTagGeneration.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 = BertForTagGeneration.from_config(config)
# add vocab for special tokens and hashtags
special_tokens = ['<img>', '<loc>', '<time>']
num_added_special_toks = tokenizer.add_tokens(special_tokens)
print('We have added', num_added_special_toks, 'special tokens')
tokenizer.img_token = '<img>'
tokenizer.loc_token = '<loc>'
tokenizer.time_token = '<time>'
print(tokenizer.convert_tokens_to_ids(special_tokens))
assert tokenizer.img_token == '<img>'
assert tokenizer.loc_token == '<loc>'
assert tokenizer.time_token == '<time>'
with open(data_args.tag_list) as f:
tag_list = f.readlines()
tag_list = ' '.join(tag_list).replace('\n', '').split()
num_added_toks = tokenizer.add_tokens(tag_list)
print('tag_list:', data_args.tag_list)
print('We have added', num_added_toks, 'tokens for hashtags')
print('total vocab_size:', len(tokenizer))
model.resize_token_embeddings(len(tokenizer))
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)
neptune_project_name = 'junmokang/bertinsta'
neptune_experiment_name = 'bertinsta-generation'
if not training_args.do_eval:
if is_torch_tpu_available():
if xm.get_ordinal() == 0:
neptune.init(neptune_project_name)
neptune.create_experiment(name=neptune_project_name,
params=training_args.__dict__)
else:
neptune.init(neptune_project_name)
neptune.create_experiment(name=neptune_project_name,
params=training_args.__dict__)
# Get datasets
train_dataset = get_dataset(
data_args, tokenizer=tokenizer,
loss_fct=model_args.loss_fct) if training_args.do_train else None
eval_dataset = get_dataset(
data_args, tokenizer=tokenizer,
evaluate=True) if training_args.do_eval else None
data_collator = DataCollatorForTagGeneration(config.vocab_size,
loss_fct=model_args.loss_fct)
training_args.per_device_eval_batch_size = 1 # force eval_batch as 1
# Initialize our Trainer
trainer = Trainer(model=model,
args=training_args,
neptune=neptune,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
data_collator=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)
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
if training_args.do_eval:
logger.info("*** Evaluate ***")
dataloader = trainer.get_eval_dataloader(eval_dataset)
# multi-gpu eval
if training_args.n_gpu > 1:
model = torch.nn.DataParallel(model)
description = "Evaluation"
batch_size = dataloader.batch_size
logger.info("***** Running %s *****", description)
logger.info(" Num examples = %d", len(dataloader.dataset))
logger.info(" Batch size = %d", batch_size)
model.eval()
if is_torch_tpu_available():
dataloader = pl.ParallelLoader(
dataloader,
[training_args.device]).per_device_loader(training_args.device)
results = {}
grouping_results = {}
# interaction_matrix = np.zeros((6, 6)) # feature interaction
beam_width = 1
top_k = 10
# tag to contexts mapping
context_list = [
'emotion', 'mood', 'location', 'time', 'object', 'activity',
'event', 'others'
]
context2ids = {c: [] for c in context_list}
if data_args.tag2contexts:
with open(data_args.tag2contexts) as f:
tag2contexts = json.load(f)
for tag, contexts in tag2contexts.items():
for c in contexts:
context2ids[c].append(tag)
for c in context_list:
context2ids[c] = tokenizer.convert_tokens_to_ids(
context2ids[c])
for eid, example in enumerate(tqdm(dataloader, desc=description)):
generated_tags = beam_decode(beam_width, top_k, model, example,
tokenizer, data_args.block_size,
training_args.device)
# generated_tags = beam_decode(beam_width, top_k, model, example, tokenizer, data_args.block_size, training_args.device, None, interaction_matrix) # feature interaction
results[example['pid']] = generated_tags
grouping_results[example['pid']] = {}
grouping_results[example['pid']]['all'] = generated_tags
# print('all:', str(generated_tags))
# diverse generation (according to context)
if data_args.tag2contexts:
for context in context_list:
generated_tags = beam_decode(beam_width, top_k, model,
example, tokenizer,
data_args.block_size,
training_args.device,
context2ids[context])
grouping_results[example['pid']][context] = generated_tags
# print(context, ':', str(generated_tags))
# with np.printoptions(precision=2, suppress=True): # feature interaction
# print(interaction_matrix)
# print(interaction_matrix.sum(1))
# print(interaction_matrix / interaction_matrix.sum(1))
results_save_path = os.path.join(training_args.output_dir,
'results.json')
with open(results_save_path, 'w') as f:
logger.info("saved results.json into %s", training_args.output_dir)
json.dump(results, f)
grouping_results_save_path = os.path.join(training_args.output_dir,
'grouping_results.json')
with open(grouping_results_save_path, 'w') as f:
logger.info("saved grouping_results.json into %s",
training_args.output_dir)
json.dump(grouping_results, f)
def prediction_loop(
self,
dataloader: DataLoader,
description: str,
prediction_loss_only: Optional[bool] = None,
ignore_keys: Optional[List[str]] = None,
metric_key_prefix: str = "eval",
) -> PredictionOutput:
"""
Prediction/evaluation loop, shared by :obj:`Trainer.evaluate()` and :obj:`Trainer.predict()`.
Works both with or without labels.
"""
if not isinstance(dataloader.dataset, collections.abc.Sized):
raise ValueError("dataset must implement __len__")
prediction_loss_only = (
prediction_loss_only if prediction_loss_only is not None else self.args.prediction_loss_only
)
if self.args.deepspeed and not self.args.do_train:
# no harm, but flagging to the user that deepspeed config is ignored for eval
# flagging only for when --do_train wasn't passed as only then it's redundant
logger.info("Detected the deepspeed argument but it will not be used for evaluation")
model = self.model
# multi-gpu eval
if self.args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Note: in torch.distributed mode, there's no point in wrapping the model
# inside a DistributedDataParallel as we'll be under `no_grad` anyways
# if full fp16 is wanted on eval and this ``evaluation`` or ``predict`` isn't called while
# ``train`` is running, half it first and then put on device
batch_size = dataloader.batch_size
num_examples = self.num_examples(dataloader)
logger.info("***** Running %s *****", description)
logger.info(" Num examples = %d", num_examples)
logger.info(" Batch size = %d", batch_size)
losses_host: torch.Tensor = None
preds_host: Union[torch.Tensor, List[torch.Tensor]] = None
labels_host: Union[torch.Tensor, List[torch.Tensor]] = None
gumbel_host: Union[torch.Tensor, List[torch.Tensor]] = None
sentence_labels_host: Union[torch.Tensor, List[torch.Tensor]] = None
sentence_indicator_host: Union[torch.Tensor, List[torch.Tensor]] = None
world_size = 1
if is_torch_tpu_available():
world_size = xm.xrt_world_size()
elif self.args.local_rank != -1:
world_size = dist.get_world_size()
world_size = max(1, world_size)
eval_losses_gatherer = DistributedTensorGatherer(world_size, num_examples, make_multiple_of=batch_size)
if not prediction_loss_only:
preds_gatherer = DistributedTensorGatherer(world_size, num_examples)
labels_gatherer = DistributedTensorGatherer(world_size, num_examples)
gumbel_gatherer = DistributedTensorGatherer(world_size, num_examples)
sentence_labels_gatherer = DistributedTensorGatherer(world_size, num_examples)
sentence_indicator_gatherer = DistributedTensorGatherer(world_size, num_examples)
model.eval()
if self.args.past_index >= 0:
self._past = None
self.callback_handler.eval_dataloader = dataloader
for step, inputs in enumerate(dataloader):
loss, logits, labels, gumbel_output, sentence_labels, sentence_indicator = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
if loss is not None:
losses = loss.repeat(batch_size)
losses_host = losses if losses_host is None else torch.cat((losses_host, losses), dim=0)
if logits is not None:
preds_host = logits if preds_host is None else nested_concat(preds_host, logits, padding_index=-100)
if labels is not None:
labels_host = labels if labels_host is None else nested_concat(labels_host, labels, padding_index=-100)
if gumbel_output is not None:
gumbel_host = gumbel_output if gumbel_host is None else nested_concat(gumbel_host, gumbel_output, padding_index=-1)
if sentence_labels is not None:
sentence_labels_host = sentence_labels if sentence_labels_host is None else nested_concat(sentence_labels_host, sentence_labels, padding_index=-1)
if sentence_indicator is not None:
sentence_indicator_host = sentence_indicator if sentence_indicator_host is None else nested_concat(sentence_indicator_host, sentence_indicator, padding_index=-100)
self.control = self.callback_handler.on_prediction_step(self.args, self.state, self.control)
# Gather all tensors and put them back on the CPU if we have done enough accumulation steps.
if self.args.eval_accumulation_steps is not None and (step + 1) % self.args.eval_accumulation_steps == 0:
eval_losses_gatherer.add_arrays(self._gather_and_numpify(losses_host, "eval_losses"))
if not prediction_loss_only:
preds_gatherer.add_arrays(self._gather_and_numpify(preds_host, "eval_preds"))
labels_gatherer.add_arrays(self._gather_and_numpify(labels_host, "eval_label_ids"))
gumbel_gatherer.add_arrays(self._gather_and_numpify(gumbel_host, "eval_gumbel_output"))
sentence_labels_gatherer.add_arrays(self._gather_and_numpify(sentence_labels_host, "eval_sentence_idxs"))
sentence_indicator_gatherer.add_arrays(self._gather_and_numpify(sentence_indicator_host, "eval_sentence_indicator"))
# Set back to None to begin a new accumulation
losses_host, preds_host, labels_host, gumbel_host, sentence_labels_host, sentence_indicator_host = None, None, None, None, None, None
if self.args.past_index and hasattr(self, "_past"):
# Clean the state at the end of the evaluation loop
delattr(self, "_past")
# Gather all remaining tensors and put them back on the CPU
eval_losses_gatherer.add_arrays(self._gather_and_numpify(losses_host, "eval_losses"))
if not prediction_loss_only:
preds_gatherer.add_arrays(self._gather_and_numpify(preds_host, "eval_preds"))
labels_gatherer.add_arrays(self._gather_and_numpify(labels_host, "eval_label_ids"))
gumbel_gatherer.add_arrays(self._gather_and_numpify(gumbel_host, "eval_gumbel_output"))
sentence_labels_gatherer.add_arrays(self._gather_and_numpify(sentence_labels_host, "eval_sentence_idxs"))
sentence_indicator_gatherer.add_arrays(self._gather_and_numpify(sentence_indicator_host, "eval_sentence_indicator"))
eval_loss = eval_losses_gatherer.finalize()
preds = preds_gatherer.finalize() if not prediction_loss_only else None
label_ids = labels_gatherer.finalize() if not prediction_loss_only else None
gumbel_outputs = gumbel_gatherer.finalize() if not prediction_loss_only else None
sentence_idxs = sentence_labels_gatherer.finalize() if not prediction_loss_only else None
sentence_indicators = sentence_indicator_gatherer.finalize() if not prediction_loss_only else None
print(sentence_idxs, 'test')
if self.compute_metrics is not None and preds is not None and label_ids is not None:
metrics = self.compute_metrics(preds, label_ids, gumbel_outputs, sentence_idxs, sentence_indicators)
else:
metrics = {}
if eval_loss is not None:
metrics[f"{metric_key_prefix}_loss"] = eval_loss.mean().item()
# Prefix all keys with metric_key_prefix + '_'
for key in list(metrics.keys()):
if not key.startswith(f"{metric_key_prefix}_"):
metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)
return PredictionOutput(predictions=preds, label_ids=label_ids, metrics=metrics)
def prediction_loop(
self, dataloader: DataLoader, description: str, prediction_loss_only: Optional[bool] = None, use_tqdm: Optional[bool] = True,
reduce_other_outputs:Callable[[Tuple[torch.tensor]], Any] = None,
) -> PredictionOutput:
"""
Prediction/evaluation loop, shared by :obj:`Trainer.evaluate()` and :obj:`Trainer.predict()`.
Works both with or without labels.
"""
if hasattr(self, "_prediction_loop"):
warnings.warn(
"The `_prediction_loop` method is deprecated and won't be called in a future version, define `prediction_loop` in your subclass.",
FutureWarning,
)
return self._prediction_loop(dataloader, description, prediction_loss_only=prediction_loss_only)
prediction_loss_only = prediction_loss_only if prediction_loss_only is not None else self.prediction_loss_only
model = self.model
# multi-gpu eval
if self.args.n_gpu > 1:
model = torch.nn.DataParallel(model)
else:
model = self.model
# Note: in torch.distributed mode, there's no point in wrapping the model
# inside a DistributedDataParallel as we'll be under `no_grad` anyways.
batch_size = dataloader.batch_size
logger.info("***** Running %s *****", description)
logger.info(" Num examples = %d", self.num_examples(dataloader))
logger.info(" Batch size = %d", batch_size)
eval_losses: List[float] = []
preds: torch.Tensor = None
label_ids: torch.Tensor = None
other_outputs: Tuple[torch.Tensor] = None
model.eval()
if is_torch_tpu_available():
dataloader = pl.ParallelLoader(dataloader, [self.args.device]).per_device_loader(self.args.device)
if self.args.past_index >= 0:
self._past = None
data_iterators = tqdm(dataloader, desc=description) if use_tqdm else dataloader
reduce_other_outputs = reduce_other_outputs if reduce_other_outputs is not None else self.reduce_other_outputs
for inputs in data_iterators:
loss, logits, labels, other_outputs_ = self.prediction_step(model, inputs, prediction_loss_only)
if loss is not None:
eval_losses.append(loss)
if logits is not None:
preds = logits if preds is None else torch.cat((preds, logits), dim=0)
if labels is not None:
label_ids = labels if label_ids is None else torch.cat((label_ids, labels), dim=0)
if other_outputs_ is not None:
#print(list(o.size() for o in other_outputs_))
#if other_outputs is not None:
# print(list(o.size() for o in others))
if reduce_other_outputs is not None:
other_outputs = other_outputs_ if other_outputs is None else tuple(
reduce_other_outputs(output, output_) for output, output_ in zip(other_outputs, other_outputs_)
) # 不加tuple只用()就是generator
if self.args.past_index and hasattr(self, "_past"):
# Clean the state at the end of the evaluation loop
delattr(self, "_past")
if self.args.local_rank != -1:
# In distributed mode, concatenate all results from all nodes:
if preds is not None:
preds = self.distributed_concat(preds, num_total_examples=self.num_examples(dataloader))
if label_ids is not None:
label_ids = self.distributed_concat(label_ids, num_total_examples=self.num_examples(dataloader))
#if other_outputs is not None: # [TODO] maybe error!!! 不熟悉distributed训练
# other_outputs = tuple(self.distributed_concat(o, num_total_examples=self.num_examples(dataloader)) for o in other_outputs)
elif is_torch_tpu_available():
# tpu-comment: Get all predictions and labels from all worker shards of eval dataset
if preds is not None:
preds = xm.mesh_reduce("eval_preds", preds, torch.cat)
if label_ids is not None:
label_ids = xm.mesh_reduce("eval_label_ids", label_ids, torch.cat)
#if others is not None: # [TODO] maybe error!!! 不熟悉tpu训练, 这里就不考虑了
#others = tuple(xm.mesh_reduce("eval_label_ids", label_ids, torch.cat)
# Finally, turn the aggregated tensors into numpy arrays.
if preds is not None:
preds = preds.cpu().numpy()
if label_ids is not None:
label_ids = label_ids.cpu().numpy()
if other_outputs is not None:
other_outputs = other_outputs # 假设一切都在self.reduce_other_outputs中处理好了
if self.compute_metrics is not None and preds is not None and label_ids is not None:
metrics = self.compute_metrics(EvalPrediction(predictions=preds, label_ids=label_ids))
else:
metrics = {}
if len(eval_losses) > 0:
metrics[f"{description}_loss"] = np.mean(eval_losses)
# Prefix all keys with eval_
#for key in list(metrics.keys()):
# if not key.startswith("eval_"):
# metrics[f"eval_{key}"] = metrics.pop(key)
for key in list(metrics.keys()):
if not key.startswith(description):
tqdm_prefix = ""
new_key = key
if key[0] == "_":
tqdm_prefix = "_"
new_key = key[1:]
metrics[tqdm_prefix+description+"_"+new_key] = metrics.pop(key)
return PredictionOutput(predictions=preds, label_ids=label_ids, metrics=metrics, other_outputs=other_outputs)
def train(self, model_path: Optional[str] = None):
"""
Main training entry point.
Args:
model_path (:obj:`str`, `optional`):
Local path to the model if the model to train has been instantiated from a local path. If present,
training will resume from the optimizer/scheduler states loaded here.
"""
if not self.call_once_called and self.auto_tb_writer:
self.call_once()
self.call_once_called = True
train_dataloader = self.get_train_dataloader()
if self.args.max_steps > 0:
t_total = self.args.max_steps
num_train_epochs = (
self.args.max_steps // (len(train_dataloader) // self.args.gradient_accumulation_steps) + 1
)
else:
t_total = int(len(train_dataloader) // self.args.gradient_accumulation_steps * self.args.num_train_epochs)
num_train_epochs = self.args.num_train_epochs
optimizer, scheduler = self.get_optimizers(num_training_steps=t_total)
# Check if saved optimizer or scheduler states exist
if (
model_path is not None
and os.path.isfile(os.path.join(model_path, "optimizer.pt"))
and os.path.isfile(os.path.join(model_path, "scheduler.pt"))
):
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(model_path, "optimizer.pt"), map_location=self.args.device)
)
scheduler.load_state_dict(torch.load(os.path.join(model_path, "scheduler.pt")))
model = self.model
if self.args.fp16:
if not is_apex_available():
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=self.args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if self.args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if self.args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[self.args.local_rank],
output_device=self.args.local_rank,
find_unused_parameters=True,
)
if self.tb_writer is not None:
self.tb_writer.add_text("args", self.args.to_json_string())
self.tb_writer.add_hparams(self.args.to_sanitized_dict(), metric_dict={})
# Train!
if is_torch_tpu_available():
total_train_batch_size = self.args.train_batch_size * xm.xrt_world_size()
else:
total_train_batch_size = (
self.args.train_batch_size
* self.args.gradient_accumulation_steps
* (torch.distributed.get_world_size() if self.args.local_rank != -1 else 1)
)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", self.num_examples(train_dataloader))
logger.info(" Num Epochs = %d", num_train_epochs)
logger.info(" Instantaneous batch size per device = %d", self.args.per_device_train_batch_size)
logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d", total_train_batch_size)
logger.info(" Gradient Accumulation steps = %d", self.args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
self.global_step = 0
self.epoch = 0
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if model_path is not None:
# set global_step to global_step of last saved checkpoint from model path
try:
self.global_step = int(model_path.split("-")[-1].split("/")[0])
epochs_trained = self.global_step // (len(train_dataloader) // self.args.gradient_accumulation_steps)
steps_trained_in_current_epoch = self.global_step % (
len(train_dataloader) // self.args.gradient_accumulation_steps
)
logger.info(" Continuing training from checkpoint, will skip to saved global_step")
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", self.global_step)
logger.info(" Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
except ValueError:
self.global_step = 0
logger.info(" Starting fine-tuning.")
tr_loss = 0.0
logging_loss = 0.0
model.zero_grad()
train_iterator = trange(
epochs_trained, int(num_train_epochs), desc="Epoch", disable=not self.is_local_master()
)
for epoch in train_iterator:
if isinstance(train_dataloader, DataLoader) and isinstance(train_dataloader.sampler, DistributedSampler):
train_dataloader.sampler.set_epoch(epoch)
if is_torch_tpu_available():
parallel_loader = pl.ParallelLoader(train_dataloader, [self.args.device]).per_device_loader(
self.args.device
)
epoch_iterator = tqdm(parallel_loader, desc="Iteration", disable=not self.is_local_master())
else:
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=not self.is_local_master())
# Reset the past mems state at the beginning of each epoch if necessary.
if self.args.past_index >= 0:
self._past = None
tqdm_dict = {}
for step, inputs in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
tr_loss += self.training_step(model, inputs, optimizer)
if (step + 1) % self.args.gradient_accumulation_steps == 0 or (
# last step in epoch but step is always smaller than gradient_accumulation_steps
len(epoch_iterator) <= self.args.gradient_accumulation_steps
and (step + 1) == len(epoch_iterator)
):
if self.args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), self.args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), self.args.max_grad_norm)
if is_torch_tpu_available():
xm.optimizer_step(optimizer)
else:
optimizer.step()
scheduler.step()
model.zero_grad()
self.global_step += 1
self.epoch = epoch + (step + 1) / len(epoch_iterator)
if (self.args.logging_steps > 0 and self.global_step % self.args.logging_steps == 0) or (
self.global_step == 1 and self.args.logging_first_step
):
logs: Dict[str, float] = {}
logs["loss"] = (tr_loss - logging_loss) / self.args.logging_steps
# backward compatibility for pytorch schedulers
logs["learning_rate"] = (
scheduler.get_last_lr()[0]
if version.parse(torch.__version__) >= version.parse("1.4")
else scheduler.get_lr()[0]
)
logging_loss = tr_loss
tqdm_dict['train_loss'] = logs["loss"]
self.log(logs)
if self.args.evaluate_during_training and self.global_step % self.args.eval_steps == 0:
eval_metrics = self.evaluate(use_tqdm=self.use_tqdm_in_eval, log=True, description="eval") # also save the metrics
for k,v in eval_metrics.items():
if k[0] == '_':
tqdm_dict[k[1:]] = v
if self.args.evaluate_trainset_during_training and self.global_step % self.args.eval_trainset_steps == 0:
train_metrics = self.evaluate(eval_dataloader=self.eval_train_dataloader,use_tqdm=self.use_tqdm_in_eval, log=True, description="eval_train")
for k,v in train_metrics.items():
if k[0] == '_':
tqdm_dict[k[1:]] = v
if self.args.save_steps > 0 and self.global_step % self.args.save_steps == 0:
# In all cases (even distributed/parallel), self.model is always a reference
# to the model we want to save.
if hasattr(model, "module"):
assert model.module is self.model
else:
assert model is self.model
# Save model checkpoint
output_dir = os.path.join(self.args.output_dir, self.exp_dir, f"{PREFIX_CHECKPOINT_DIR}-{self.global_step}")
self.save_model(output_dir)
if self.is_world_master():
self._rotate_checkpoints()
if is_torch_tpu_available():
xm.rendezvous("saving_optimizer_states")
xm.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
xm.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
elif self.is_world_master():
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
epoch_iterator.set_postfix(**tqdm_dict)
train_iterator.set_postfix(**tqdm_dict)
if self.args.max_steps > 0 and self.global_step > self.args.max_steps:
epoch_iterator.close()
break
if self.args.max_steps > 0 and self.global_step > self.args.max_steps:
train_iterator.close()
break
if self.args.tpu_metrics_debug or self.args.debug:
# tpu-comment: Logging debug metrics for PyTorch/XLA (compile, execute times, ops, etc.)
xm.master_print(met.metrics_report())
if self.tb_writer:
self.tb_writer.close()
log_dirs = os.listdir(self.log_dir)
for name in log_dirs:
if re.match("\d{10}.*", name): # unknown dir generated by tensorboardX
shutil.rmtree(os.path.join(self.log_dir, name))
if self.compare_datasets:
self.tb_writer_for_train.close()
self.tb_writer_for_eval.close()
if self.args.past_index and hasattr(self, "_past"):
# Clean the state at the end of training
delattr(self, "_past")
logger.info("\n\nTraining completed. Do not forget to share your model on huggingface.co/models =)\n\n")
return TrainOutput(self.global_step, tr_loss / self.global_step)
# 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.
"""
A subclass of `Trainer` specific to Question-Answering tasks
"""
from transformers import Trainer, is_datasets_available, is_torch_tpu_available
from transformers.trainer_utils import PredictionOutput
if is_datasets_available():
import datasets
if is_torch_tpu_available():
import torch_xla.core.xla_model as xm
import torch_xla.debug.metrics as met
class QuestionAnsweringTrainer(Trainer):
def __init__(self,
*args,
eval_examples=None,
post_process_function=None,
**kwargs):
super().__init__(*args, **kwargs)
self.eval_examples = eval_examples
self.post_process_function = post_process_function
def evaluate(self,
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,
use_auth_token=True if model_args.use_auth_token else None,
)
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
data_files = {}
dataset_args = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = (
data_args.train_file.split(".")[-1]
if data_args.train_file is not None
else data_args.validation_file.split(".")[-1]
)
if extension == "txt":
extension = "text"
dataset_args["keep_linebreaks"] = data_args.keep_linebreaks
raw_datasets = load_dataset(
extension,
data_files=data_files,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
**dataset_args,
)
# If no validation data is there, validation_split_percentage will be used to divide the dataset.
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
extension,
data_files=data_files,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
**dataset_args,
)
raw_datasets["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
**dataset_args,
)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# 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:
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
train_dataset = train_dataset.select(range(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:
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(max_eval_samples))
def preprocess_logits_for_metrics(logits, labels):
if isinstance(logits, tuple):
# Depending on the model and config, logits may contain extra tensors,
# like past_key_values, but logits always come first
logits = logits[0]
return logits.argmax(dim=-1)
metric = load_metric("accuracy")
def compute_metrics(eval_preds):
preds, labels = eval_preds
# preds have the same shape as the labels, after the argmax(-1) has been calculated
# by preprocess_logits_for_metrics but we need to shift the labels
labels = labels[:, 1:].reshape(-1)
preds = preds[:, :-1].reshape(-1)
return metric.compute(predictions=preds, references=labels)
# 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,
compute_metrics=compute_metrics if training_args.do_eval and not is_torch_tpu_available() else None,
preprocess_logits_for_metrics=preprocess_logits_for_metrics
if training_args.do_eval and not is_torch_tpu_available()
else None,
)
# 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 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. 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 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.config_name:
config = BertConfig.from_pretrained(model_args.config_name,
cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = BertConfig.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 = BertTokenizer.from_pretrained(
model_args.tokenizer_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
tokenizer = BertTokenizer.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported, but 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 = BertForTagRankingLate.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 = BertForTagRanking.from_config(config)
# add vocab for special tokens and hashtags
special_tokens = ['<img>', '<loc>', '<time>']
num_added_special_toks = tokenizer.add_tokens(special_tokens)
print('We have added', num_added_special_toks, 'special tokens')
tokenizer.img_token = '<img>'
tokenizer.loc_token = '<loc>'
tokenizer.time_token = '<time>'
print(tokenizer.convert_tokens_to_ids(special_tokens))
assert tokenizer.img_token == '<img>'
assert tokenizer.loc_token == '<loc>'
assert tokenizer.time_token == '<time>'
with open(data_args.tag_list) as f:
tag_list = f.readlines()
tag_list = ' '.join(tag_list).replace('\n', '').split()
num_added_toks = tokenizer.add_tokens(tag_list)
print('tag_list:', data_args.tag_list)
print('We have added', num_added_toks, 'tokens for hashtags')
print('total vocab_size:', len(tokenizer))
model.resize_token_embeddings(len(tokenizer))
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
data_collator = DataCollatorForTagGeneration(config.vocab_size)
training_args.per_device_eval_batch_size = 1 # force eval_batch as 1
# Initialize our Trainer
trainer = Trainer(model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
data_collator=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)
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
if training_args.do_eval:
logger.info("*** Evaluate ***")
dataloader = trainer.get_eval_dataloader(eval_dataset)
# multi-gpu eval
if training_args.n_gpu > 1:
model = torch.nn.DataParallel(model)
description = "Evaluation"
batch_size = dataloader.batch_size
logger.info("***** Running %s *****", description)
logger.info(" Num examples = %d", len(dataloader.dataset))
logger.info(" Batch size = %d", batch_size)
model.eval()
if is_torch_tpu_available():
dataloader = pl.ParallelLoader(
dataloader,
[training_args.device]).per_device_loader(training_args.device)
results = {}
for eid, example in enumerate(tqdm(dataloader, desc=description)):
feature = convert_example_to_feature(example, tokenizer,
data_args.block_size)
image_ids = torch.tensor([feature['image_ids']],
dtype=torch.long).to(training_args.device)
location_ids = torch.tensor([feature['location_ids']],
dtype=torch.long).to(
training_args.device)
time_ids = torch.tensor([feature['time_ids']],
dtype=torch.long).to(training_args.device)
text_ids = torch.tensor([feature['text_ids']],
dtype=torch.long).to(training_args.device)
pid = feature['pid']
inputs = {
'image_ids': image_ids,
'location_ids': location_ids,
'time_ids': time_ids,
'text_ids': text_ids
}
with torch.no_grad():
outputs = model(**inputs)
logits = outputs[0]
logit_for_cls = logits[0]
orig_vocab_size = 30522
added_special_toks_size = 3 # <img>, <loc>, <time>
logit_for_cls[:orig_vocab_size +
added_special_toks_size] = -float('inf')
probabilities = F.softmax(logit_for_cls, 0).detach().cpu()
probs, predicted_indices = torch.topk(probabilities, k=10)
predicted_tokens = tokenizer.convert_ids_to_tokens(
predicted_indices)
while pid in results:
pid = pid + '_'
results[pid] = predicted_tokens
results_save_path = os.path.join(training_args.output_dir,
'results.json')
with open(results_save_path, 'w') as f:
logger.info("saved results.json into %s", training_args.output_dir)
json.dump(results, f)
