Suggestion for introducing "shift_labels" argument for Trainer

[seungeunrho]

Feature request

Add an argument to determine shifting the labels or not.

In TrainingArguments class, an argument named shift_labels should be added.

During training, at here and here, model must check both labels is not None and self.shift_labels is True

e.g.

if labels is not None and self.shift_labels:          # changed
            # Shift so that tokens < n predict n
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()

Default values for shift_labels is False, except for causal language models such as GPT2PreTrainedModel

Related to gpt2 : @patil-suraj and trainer @sgugger

Motivation

In the current state of the code, the shifting of labels for training GPT2LMHeadModel is changing under the use of label_smoothing, which I assume is unintended.

Specifically, training a GPT2LMHeadModel with args.label_smoothing_factor==0 (which is default), the code shifts the labels and computes the loss inside the model.forward(). This assumes that labels have not been shifted to be properly aligned with corresponding input_ids.

However, if I train GPT2LMHeadModel with args.label_smoothing_factor > 0, then the loss is computed here, inside the compute_loss() function of the Trainer. This part assumes labels are already shifted, and does not proceed to shift the labels.

I believe whether to shift labels or not should be explicitly determined by its own argument, not by another argument like label_smoothing_factor. In my case, our team was very frustrated that our training results were totally different by only changing the label_smoothing with same given labels and input_ids. The reason was due to the misalignment of labels and input_ids when turning on the label_smoothing.

Your contribution

I'm willing to make PR after your confirmation.

[sgugger]

I don't think a new TrainingArgument is the right answer here. Some models shift the labels internally, I think it's all the models for causal LM (not jsut GPT-2), so I think instead of a flag, there should be a check when the loss is computed by the Trainer for label smoothing to see if the model class name is inside the MODEL_FOR_CAUSAL_LM_MAPPING_NAMES (to import from the auto module) and then shift the labels.

Let me know if you'd like to proceed with a PR for this fix!

[seungeunrho]

Thanks for quick reply. Your approach seems plausible and I'd like to proceed it. I've read the document for contribution guide thoroughly. Can I just start now? or is there anything I should know before begin?

[sgugger]

You can start, good luck! :-)

[Junpliu]

Feature request

Add an argument to determine shifting the labels or not.

In TrainingArguments class, an argument named shift_labels should be added.

During training, at here and here, model must check both labels is not None and self.shift_labels is True

e.g.

if labels is not None and self.shift_labels:          # changed
            # Shift so that tokens < n predict n
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()

Default values for shift_labels is False, except for causal language models such as GPT2PreTrainedModel

Related to gpt2 : @patil-suraj and trainer @sgugger

Motivation

In the current state of the code, the shifting of labels for training GPT2LMHeadModel is changing under the use of label_smoothing, which I assume is unintended.

Specifically, training a GPT2LMHeadModel with args.label_smoothing_factor==0 (which is default), the code shifts the labels and computes the loss inside the model.forward(). This assumes that labels have not been shifted to be properly aligned with corresponding input_ids.

However, if I train GPT2LMHeadModel with args.label_smoothing_factor > 0, then the loss is computed here, inside the compute_loss() function of the Trainer. This part assumes labels are already shifted, and does not proceed to shift the labels.

I believe whether to shift labels or not should be explicitly determined by its own argument, not by another argument like label_smoothing_factor. In my case, our team was very frustrated that our training results were totally different by only changing the label_smoothing with same given labels and input_ids. The reason was due to the misalignment of labels and input_ids when turning on the label_smoothing.

Your contribution

I'm willing to make PR after your confirmation.

I want to know more about how the prediction text looks like under the label-smoothing case before the bug-fix. Does the model learn an indentity transformation and always predict the last input token repeatedly? I am curious about this.

[zzaebok]

@sgugger May I ask how to shift labels for the custom CausalLM model? Let's assume I made a CustomCausalLM model, which is not mapped in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES. As other models, my CausalLM model (which is modeled based on modeling_llama.py) shifts labels in forward function. When I use hf Trainer, my CaausalLM, and label smoother, how to shift labels in this case?

The current Trainer code is

if labels is not None:
    if unwrap_model(model)._get_name() in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES.values():
        loss = self.label_smoother(outputs, labels, shift_labels=True)
    else:
        loss = self.label_smoother(outputs, labels)

I think the workaround is to add shift_labels=True in else statement. Is there a correct, or better way to shift labels for my custom causal lm model when using label smoother??

[ArthurZucker]

By default, if you copy the end of the forward pass of the LlamaForConditionalGeneration you will see that there is a part where the labels are shifted!