Target {} is out of bounds

🤗Transformers
1

Hi,
I am following this fantastic notebook to fine-tune a multi classifier.

Context:

  1. I am using my own dataset.
  2. Dataset is a CSV file with two values, text and label.
  3. Labels are all numbers.
  4. I have 7 labels.
  5. When loading the pre-trained model, I am assigning num_labels=7.
from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased",num_labels=7)

When training, I am receiving this error:

/usr/local/lib/python3.7/dist-packages/torch/nn/functional.py in cross_entropy(input, target, weight, size_average, ignore_index, reduce, reduction, label_smoothing)
   2844     if size_average is not None or reduce is not None:
   2845         reduction = _Reduction.legacy_get_string(size_average, reduce)
-> 2846     return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index, label_smoothing)
   2847 
   2848 

IndexError: Target 7 is out of bounds.

I have tried changing the number of labels to 2 and 5 and that didn’t solve the issue. Still getting out of bounds error.

Training arguments:

from transformers import TrainingArguments, Trainer

training_args = TrainingArguments(
    output_dir="./results",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    num_train_epochs=5,
    weight_decay=0.01,
)


trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokinized_jobs["train"],
    eval_dataset=tokinized_jobs["test"],
    tokenizer=tokenizer,
    data_collator=data_collator,
)

trainer.train()

and here is how tokenized data look like

DatasetDict({
    train: Dataset({
        features: ['attention_mask', 'input_ids', 'label', 'text', 'token_type_ids'],
        num_rows: 1598
    })
    test: Dataset({
        features: ['attention_mask', 'input_ids', 'label', 'text', 'token_type_ids'],
        num_rows: 400
    })
})

Sample:

{
'attention_mask': [1, 1, 1, 1, 1, 1, 1],
 'input_ids': [101, 1015, 1011, 2095, 3325, 6871, 102],
 'label': 2,
 'text': '1-year experience preferred',
 'token_type_ids': [0, 0, 0, 0, 0, 0, 0]
}

I tried it on Colab with GPU and TPU.

Any idea what is the issue?

2

I found the solution. It was an indexing issue with my labels.
My labels were starting from 1 to 8, I changed them to 0…7 and that fixed the issue for me.

Credit to this answer on Stackoverflow.

Hope this will help someone in the future.

2 Likes
3

Thanks, this worked for me

4

How do we go about changing the labels for semantic segmentation mask? i’ve attemped np.where() and converting back to a PIL image with no success

5

For Anyone who has already ensured they’re zero indexed, what worked for me was to specify num_labels as a parameter to the model:

model = AutoModelForSequenceClassification.from_pretrained(“FacebookAI/roberta-base”, num_labels = 52)