def init_training_args(self, model_path: str) -> TrainingArguments:
r"""
构造训练参数.
"""
training_args = TrainingArguments(output_dir=model_path)
training_args.logging_steps = 5000
training_args.save_steps = 5000
training_args.learning_rate = 2e-5
training_args.num_train_epochs = 3
training_args.per_device_train_batch_size = 32
training_args.fp16 = self.fp16
training_args.fp16_opt_level = "O1"
return training_args
def fit(self, train_df, dev_df):
"""
fitting the model based on the train set. validation is done using the dev set
Parameters
----------
:param train_df: dataframe
a pandas dataframe containing data to be trained on
:param dev_df: dataframe
a pandas dataframe containing data to validate on
:return: None
all relevant results are saved under the the location provided to save the model in.
Next a prediction can be done
"""
train_labels = Counter(train_df[self.label_col_name]).keys()
num_labels = len(train_labels)
dev_labels = Counter(train_df[self.label_col_name]).keys()
if num_labels != len(dev_labels):
raise IOError("train and dev datasets contain different number of labels")
# creating a DF for train/test with relevant columns.
# Not clear why the 'alpha' column is needed, but as written here
# (https://towardsdatascience.com/https-medium-com-chaturangarajapakshe-text-classification-with-transformer-models-d370944b50ca) - it is required
train_df = pd.DataFrame({
'id': range(len(train_df)),
'label': train_df[self.label_col_name],
'alpha': ['a'] * train_df.shape[0],
'text': train_df["text"].replace(r'\n', ' ', regex=True)
})
dev_df = pd.DataFrame({
'id': range(len(dev_df)),
'label': dev_df[self.label_col_name],
'alpha': ['a'] * dev_df.shape[0],
'text': dev_df["text"].replace(r'\n', ' ', regex=True)
})
# saving the DF to the new/old folder
train_df.to_csv(os.path.join(self.saving_data_folder, "train.tsv"),
index=False, columns=train_df.columns, sep='\t', header=False)
dev_df.to_csv(os.path.join(self.saving_data_folder, "dev.tsv"),
index=False, columns=dev_df.columns, sep='\t', header=False)
config = AutoConfig.from_pretrained(self.model_name, num_labels=num_labels,
output_attentions=True) ##needed for the visualizations
# loading the actual model to memory
model = BertForSequenceClassification.from_pretrained(self.model_name, config=config)
# Now we need to convert the examples in the dataset to features that the model can understand
# this is a ready made class, provided by HuggingFace
train_dataset = SingleSentenceClassificationProcessor(mode='classification')
dev_dataset = SingleSentenceClassificationProcessor(mode='classification')
# now adding examples (from the DF we created earlier) to the objects we created in the cell above)
_ = train_dataset.add_examples(texts_or_text_and_labels=train_df['text'], labels=train_df[self.label_col_name],
overwrite_examples=True)
_ = dev_dataset.add_examples(texts_or_text_and_labels=dev_df['text'], labels=dev_df[self.label_col_name],
overwrite_examples=True)
train_features = train_dataset.get_features(tokenizer=self.tokenizer, max_length=self.max_length)
test_features = dev_dataset.get_features(tokenizer=self.tokenizer, max_length=self.max_length)
training_args = TrainingArguments("./train")
training_args.do_train = True
# setting the params of the BERT classifier
for cur_param in self.bert_model_params.keys():
try:
training_args.__dict__[cur_param] = eval(self.bert_model_params[cur_param])
except TypeError:
training_args.__dict__[cur_param] = self.bert_model_params[cur_param]
training_args.logging_steps = (len(train_features) - 1) // training_args.per_gpu_train_batch_size + 1
training_args.save_steps = training_args.logging_steps
training_args.output_dir = self.saving_model_folder
training_args.eval_steps = 100
# training_args.logging_dir = "gs://" from torch.utils.tensorboard import SummaryWriter supports google cloud storage
trainer = Trainer(model=model,
args=training_args,
train_dataset=train_features,
eval_dataset=test_features,
compute_metrics=self.compute_metrics)
trainer.train()
# saving the model
self.save_model(model=trainer.model, folder_name='bert_based_model')
def generate_training_args(args, inoculation_step):
training_args = TrainingArguments("tmp_trainer")
training_args.no_cuda = args.no_cuda
training_args.seed = args.seed
training_args.do_train = args.do_train
training_args.do_eval = args.do_eval
training_args.output_dir = os.path.join(args.output_dir, str(inoculation_step)+"-sample")
training_args.evaluation_strategy = args.evaluation_strategy # evaluation is done after each epoch
training_args.metric_for_best_model = args.metric_for_best_model
training_args.greater_is_better = args.greater_is_better
training_args.logging_dir = args.logging_dir
training_args.task_name = args.task_name
training_args.learning_rate = args.learning_rate
training_args.per_device_train_batch_size = args.per_device_train_batch_size
training_args.per_device_eval_batch_size = args.per_device_eval_batch_size
training_args.num_train_epochs = args.num_train_epochs # this is the maximum num_train_epochs, we set this to be 100.
training_args.eval_steps = args.eval_steps
training_args.logging_steps = args.logging_steps
training_args.load_best_model_at_end = args.load_best_model_at_end
if args.save_total_limit != -1:
# only set if it is specified
training_args.save_total_limit = args.save_total_limit
import datetime
date_time = "{}-{}".format(datetime.datetime.now().month, datetime.datetime.now().day)
run_name = "{0}_{1}_{2}_{3}_mlen_{4}_lr_{5}_seed_{6}_metrics_{7}".format(
args.run_name,
args.task_name,
args.model_type,
date_time,
args.max_seq_length,
args.learning_rate,
args.seed,
args.metric_for_best_model
)
training_args.run_name = run_name
training_args_dict = training_args.to_dict()
# for PR
_n_gpu = training_args_dict["_n_gpu"]
del training_args_dict["_n_gpu"]
training_args_dict["n_gpu"] = _n_gpu
HfParser = HfArgumentParser((TrainingArguments))
training_args = HfParser.parse_dict(training_args_dict)[0]
if args.model_path == "":
args.model_path = args.model_type
if args.model_type == "":
assert False # you have to provide one of them.
# Set seed before initializing model.
set_seed(training_args.seed)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info(f"Training/evaluation parameters {training_args}")
return training_args
def fit(self, train_df, dev_df):
"""
fitting the model based on the train set. validation is done using the dev set
Parameters
----------
:param train_df: dataframe
a pandas dataframe containing data to be trained on
:param dev_df: dataframe
a pandas dataframe containing data to validate on
:return: None
all relevant results are saved under the the location provided to save the model in.
Next a prediction can be done
"""
train_labels = Counter(train_df[self.label_col_name]).keys()
num_labels = len(train_labels)
dev_labels = Counter(dev_df[self.label_col_name]).keys()
if num_labels != len(dev_labels):
raise IOError(
"train and dev datasets contain different number of labels")
# creating a DF for train/test with relevant columns.
# Not clear why the 'alpha' column is needed, but as written here
# (https://towardsdatascience.com/https-medium-com-chaturangarajapakshe-text-classification-with-transformer-models-d370944b50ca) - it is required
train_df = pd.DataFrame({
'id':
range(len(train_df)),
'label':
train_df[self.label_col_name],
'alpha': ['a'] * train_df.shape[0],
'text':
train_df["text"].replace(r'\n', ' ', regex=True)
})
dev_df = pd.DataFrame({
'id':
range(len(dev_df)),
'label':
dev_df[self.label_col_name],
'alpha': ['a'] * dev_df.shape[0],
'text':
dev_df["text"].replace(r'\n', ' ', regex=True)
})
# saving the DF to the new/old folder
train_df.to_csv(os.path.join(self.saving_data_folder, "train.tsv"),
index=False,
columns=train_df.columns,
sep='\t',
header=False)
dev_df.to_csv(os.path.join(self.saving_data_folder, "dev.tsv"),
index=False,
columns=dev_df.columns,
sep='\t',
header=False)
config = AutoConfig.from_pretrained(
self.model_name, num_labels=num_labels,
output_attentions=True) ##needed for the visualizations
# loading the actual model to memory
model = BertForSequenceClassification.from_pretrained(self.model_name,
config=config)
# Now we need to convert the examples in the dataset to features that the model can understand
# this is a ready made class, provided by HuggingFace
train_dataset = SingleSentenceClassificationProcessor(
mode='classification')
dev_dataset = SingleSentenceClassificationProcessor(
mode='classification')
# now adding examples (from the DF we created earlier) to the objects we created in the cell above)
_ = train_dataset.add_examples(
texts_or_text_and_labels=train_df['text'],
labels=train_df[self.label_col_name],
overwrite_examples=True)
_ = dev_dataset.add_examples(texts_or_text_and_labels=dev_df['text'],
labels=dev_df[self.label_col_name],
overwrite_examples=True)
train_features = train_dataset.get_features(tokenizer=self.tokenizer,
max_length=self.max_length)
dev_features = dev_dataset.get_features(tokenizer=self.tokenizer,
max_length=self.max_length)
# idea about a self-trainer is taken from here - https://huggingface.co/transformers/main_classes/trainer.html
class MyTrainer(Trainer):
def __init__(self, loss_func=torch.nn.CrossEntropyLoss(),
**kwargs):
self.loss_func = loss_func
super().__init__(**kwargs)
def compute_loss(self, model, inputs):
labels = inputs.pop("labels")
outputs = model(**inputs)
logits = outputs[0]
return self.loss_func(logits, labels)
class FocalLoss(nn.modules.loss._WeightedLoss):
def __init__(self, weight=None, gamma=2, reduction='mean'):
super(FocalLoss, self).__init__(weight, reduction=reduction)
self.gamma = gamma
self.weight = weight # weight parameter will act as the alpha parameter to balance class weights
def forward(self, input, target):
ce_loss = F.cross_entropy(input,
target,
reduction=self.reduction,
weight=self.weight)
pt = torch.exp(-ce_loss)
focal_loss = ((1 - pt)**self.gamma * ce_loss).mean()
return focal_loss
class_weights = compute_class_weight(class_weight='balanced',
classes=np.unique(
list(train_labels)),
y=train_df['label'])
#my_loss_func = torch.nn.CrossEntropyLoss(weight=torch.tensor(class_weights, dtype=torch.float))
my_loss_func = FocalLoss(
weight=torch.tensor(class_weights, dtype=torch.float))
# how to define a trainer and all its arguments is taken from here - https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb
args = TrainingArguments(
"arabic_nlp_model",
evaluation_strategy="epoch",
#learning_rate=1e-5,
learning_rate=1e-4,
per_device_train_batch_size=16,
per_device_eval_batch_size=8,
num_train_epochs=5,
weight_decay=0.01,
load_best_model_at_end=True,
#metric_for_best_model="macro_f1_PN",
)
# setting the params of the BERT classifier
for cur_param in self.bert_model_params.keys():
try:
args.__dict__[cur_param] = eval(
self.bert_model_params[cur_param])
except TypeError:
args.__dict__[cur_param] = self.bert_model_params[cur_param]
args.logging_steps = (len(train_features) -
1) // args.per_device_train_batch_size + 1
args.save_steps = args.logging_steps
args.output_dir = self.saving_model_folder
#training_args.compute_metrics = f1_score
#training_args.compute_metrics = self.compute_metrics
# training_args.logging_dir = "gs://" from torch.utils.tensorboard import SummaryWriter supports google cloud storage
trainer = MyTrainer(loss_func=my_loss_func,
model=model,
args=args,
train_dataset=train_features,
eval_dataset=dev_features,
compute_metrics=self.compute_metrics)
#trainer = Trainer(model=model,
# args=args,
# train_dataset=train_features,
# eval_dataset=dev_features,
# #compute_metrics = compute_metrics)
# compute_metrics=self.compute_metrics)
trainer.train()
# saving the model
self.save_model(model=trainer.model)
