'CTCTrainer' object has no attribute 'use_amp'

[its-ogawa]

Use the latest huggingsound.

#!pip list | grep huggingsound
huggingsound 0.1.4

AttributeError occurs when finetune is performed as shown in the sample below.
https://github.com/jonatasgrosman/huggingsound#fine-tuning

/usr/local/lib/python3.7/dist-packages/huggingsound/trainer.py in training_step(self, model, inputs)
432 inputs = self._prepare_inputs(inputs)
433
--> 434 if self.use_amp:
435 with torch.cuda.amp.autocast():
436 loss = self.compute_loss(model, inputs)

AttributeError: 'CTCTrainer' object has no attribute 'use_amp'

Can you find the cause?

[its-ogawa]

It is not a correct solution, but I removed use_amp from the source code. That is, I commented it out and ran it.
https://github.com/jonatasgrosman/huggingsound/blob/main/huggingsound/trainer.py

        #if self.use_amp:
        # with torch.cuda.amp.autocast():
        # loss = self.compute_loss(model, inputs)
        #else:
            loss = self.compute_loss(model, inputs)

        #if self.use_amp:
        #self.scaler.scale(loss).backward()
        #elif self.deepspeed:
        if self.deepspeed:
            self.deepspeed.backward(loss)
        else:
            loss.backward()

This is a completely reckless response, and I hope you can point me to the correct way to handle this.

I don't know if this is the cause or not, but I am having a problem where evaluate fails for a finetuned model in this state.
I hope you can refer to this as well.
cf. #46

[arikhalperin]

It's related to transformers version. I use the latest transformers, the correct fix is to use: use_cuda_amq:

if self.use_cuda_amp:
self.scaler.scale(loss).backward()
elif self.deepspeed:
self.deepspeed.backward(loss)
else:
loss.backward()

[its-ogawa]

Thanks for answering.

What version of Transformer do you have?
And how did you install it and how did you install it?

[arikhalperin]

transformers==4.21.1
pip install transformers

[nkaenzig-aifund]

transformers changed the name of the Trainer property use_amp to use_cuda_amp in this PR:
huggingface/transformers#17138

Use transformers==4.19.2 or rename the variable as suggested by arikhalperin