Model not getting inputs from custom dataset

🤗Datasets
1

I’m getting the error “IndexError: tuple index out of range” when trying to run the LongformerforQuestionAnswering model on a custom dataset.

I’ve migrated everything over to Huggingface and am loading my dataset to pandas, then converting it to an HF Dataset using Dataset.from_pandas(). The columns are being tokenized using a PretrainedTokenizerFast in this function:

def prepare_train_features(examples, tokenizer):
    tokenized_examples = tokenizer(
        examples["question"],
        examples["context"],
        truncation="only_second",
        max_length=2048,
        return_overflowing_tokens=True,
        return_offsets_mapping=True,
        padding="max_length",
        return_tensors="pt")

    tokenized_examples["start_positions"] = [example["start_idx"] for example in examples]
    tokenized_examples["end_positions"] = [example["end_idx"] for example in examples]

    return tokenized_examples

This returns the dataset which looks like the following image:

Screen Shot 2021-10-15 at 12.07.57 PM
Screen Shot 2021-10-15 at 12.07.57 PM1198Ă—1200 65.7 KB

So everything appears to be correct. The problem occurs in the call to the model itself, the compute_loss step of the Trainer. When I print the value of inputs, it returns an empty tuple.

My training arguments are fairly standard:

training_args = TrainingArguments(output_dir="./models/trained_models",
                                      evaluation_strategy="epoch",
                                      save_strategy="epoch",
                                      fp16=True,
                                      label_names=["start_positions", "end_positions"],
                                      ddp_find_unused_parameters=False,
                                      per_device_train_batch_size=12,
                                      per_device_eval_batch_size=12,
                                      dataloader_num_workers=0,
                                      weight_decay=0.01,
                                      num_train_epochs=EPOCHS,
                                      learning_rate=3e-5,
                                      load_best_model_at_end=True,
                                      metric_for_best_model='rouge',
                                      do_train=True,
                                      do_eval=True)

and I’m using the QuestionAnsweringTrainer provided in this repo:

All of this is in service of trying to use HF’s rouge metric (which didn’t seem to like my torch dataset)

Any help would be greatly appreciated!