Hi, I am training Mbart50 for conditional text generation. I prepared the dataset in the required format using datasets.load_dataset. Later, I loaded the model and tokenizer given as
from transformers import MBartForConditionalGeneration, MBart50Tokenizer
from transformers import Trainer, TrainingArguments
tokenizer = MBart50Tokenizer.from_pretrained("facebook/mbart-large-50")
model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50")
I have specified the training arguments as
training_args = TrainingArguments(
output_dir=output_dir,
overwrite_output_dir = True,
#do_train = True,
num_train_epochs=1,
per_device_train_batch_size=4,
per_device_eval_batch_size=4,
save_steps=100,
save_total_limit=3,
evaluation_strategy='steps',
save_strategy = 'steps',
learning_rate = 8e-5,
lr_scheduler_type = 'linear',
weight_decay=0.01,
logging_dir='./logs',
logging_steps=10,
)
and trainer as
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=val_dataset,
tokenizer=tokenizer,
)
I am training using trainer.train() and training is done fine.
But later I wanted to include compute_metrics in the tarining. I have defined the compute_metrics as
from datasets import load_metric
from transformers import EvalPrediction
rouge = load_metric('rouge')
def compute_metrics(p: EvalPrediction):
predictions = p.predictions
labels= p.label_ids
decoded_preds = tokenizer.batch_decode(predictions, skip_special_tokens=True)
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
decoded_preds = ["\n".join(nltk.sent_tokenize(pred.strip())) for pred in decoded_preds]
decoded_labels = ["\n".join(nltk.sent_tokenize(label.strip())) for label in decoded_labels]
result = rouge.compute(decoded_preds, decoded_labels)
result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in predictions]
results["gen_len"] = np.mean(prediction_lens)
results = {k: round(v, 4) for k, v in result.items()}
return results
and then modified the trainer as
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=val_dataset,
tokenizer=tokenizer,
compute_metrics = compute_metrics
)
When I run the training using trainer.train(), it trains till first logging step and starts doing evaluation but in between the evaluation it stops.
I am not able to understand what is going wrong here. Can somebody help me understand and fix it