Error in RAG finetuning script

[shamanez]

Environment info

ut the missing fields in that output! -->

• transformers version: 3.4.0
• Platform: Linux-4.18.0-147.5.1.el8_1.x86_64-x86_64-with-centos-8.1.1911-Core
• Python version: 3.6.8
• PyTorch version (GPU?): 1.7.0 (True)
• Tensorflow version (GPU?): not installed (NA)
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help

@patrickvonplaten, @lhoestq

Information

I am using RAG fine-tuning script. During the fine-tuning process, it says

torch.nn.modules.module.ModuleAttributeError: 'GenerativeQAModule' object has no attribute 'opt'

The bug exactly appears in [line 332 in finetune.py (https://github.com/huggingface/transformers/blob/master/examples/rag/finetune.py#L332)

To reproduce

I have installed the transformers library from the source. Not the pip.

[shamanez]

It is related with the optimizer initialization in the finetune.py script. Seems like even in the lightning_base.py there is no initialization for the optimizer.

[shamanez]

@lhoestq

any idea on this?

I managed to work around by calling the optimizer initialization inside the train_dataloader function in finetune.py.

[lhoestq]

Well the optimizer/scheduler is already defined in examples/lightningbase.py in BaseTransoformer.configure_optimizers. Not sure why the train_dataloader() function in finetune.py tries to define the scheduler. This must have been a bad copy paste...

I think we should remove those lines

transformers/examples/rag/finetune.py

Lines 326 to 336 in 17b1fd8

	t_total = (
	(len(dataloader.dataset) // (self.hparams.train_batch_size * max(1, self.hparams.gpus)))
	// self.hparams.accumulate_grad_batches
	* float(self.hparams.max_epochs)
	)
	scheduler = get_linear_schedule_with_warmup(
	self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=t_total
	)
	if max(scheduler.get_last_lr()) > 0:
	warnings.warn("All learning rates are 0")
	self.lr_scheduler = scheduler

I just tried to remove them and now I'm getting this other issue #7816
I'll fix this one as well and make a PR

[shamanez]

That what I was thinking since there is a specific def in lightningbase.py.