Fix doc examples: cannot import name

[ydshieh]

What does this PR do?

In doc examples in some PT model files, there are cannot import name error. This PR fixes this issue.

Who can review?

@patrickvonplaten

Question:

@patrickvonplaten , in modeling_unispeech.py, there are examples using

from transformers import UniSpeechSatFeatureExtractor

And in some docstring, there are

:class:`~transformers.UniSpeechSatProcessor`

However, these can't be imported

ImportError: cannot import name 'UniSpeechSatFeatureExtractor' from 'transformers' (/home/ydshieh/Desktop/ydshieh/transformers/src/transformers/__init__.py)

Is there any reason to make these objects invisible to transformers?

[ydshieh]

can't keep copy here, because there is no BigBirdPegasusTokenizer

[patrickvonplaten]

could we try to keep the copy and change Bart->BigBirdPegasus to BartDecoderWrapper->BigBirdPegasusDecoderWrapper and BartForCausalLM->BigBirdPegasusForCausalLM? This way the tokenizer should stay the same :-)

[ydshieh]

Let me try it.

[ydshieh]

The following does the trick:

• BartDecoderWrapper->BigBirdPegasusDecoderWrapper
• BartForCausalLM->BigBirdPegasusForCausalLM
• BartPreTrainedModel->BigBirdPegasusPreTrainedModel
• BartTokenizer->PegasusTokenizer
• 'facebook/bart-large'->"google/bigbird-pegasus-large-arxiv"

[patrickvonplaten]

perfect!

[patrickvonplaten]

Great catch - thanks a lot for fixing it! I think we should do the exact same thing for the speech models (UniSpeech, SEW, UniSpeechSAT) actually. Feel free to open a PR if you want - I'll put it on my ToDo list as well though

[ydshieh]

@patrickvonplaten

So UniSpeechSatFeatureExtractor shouldn't be used by UniSpeechSatModel , and it is supposed to use Wav2Vec2FeatureExtractor?

I have seen

_PROCESSOR_FOR_DOC = "Wav2Vec2Processor"
_SEQ_CLASS_PROCESSOR_FOR_DOC = "Wav2Vec2FeatureExtractor"

and UniSpeechSatFeatureExtractor is not visible to transformers. But I don't have the full context.

[patrickvonplaten]

@patrickvonplaten

So UniSpeechSatFeatureExtractor shouldn't be used by UniSpeechSatModel , and it is supposed to use Wav2Vec2FeatureExtractor?

I have seen

_PROCESSOR_FOR_DOC = "Wav2Vec2Processor"
_SEQ_CLASS_PROCESSOR_FOR_DOC = "Wav2Vec2FeatureExtractor"

and UniSpeechSatFeatureExtractor is not visible to transformers. But I don't have the full context.

Yeah there is no UniSpeechFeatureExtractor since it's the same as Wav2Vec2FeatureExtractor so we should instead just use Wav2Vec2FeatureExtractor. Happy to resolve this is another PR though!

[patrickvonplaten]

@ydshieh - if ok for you, I would like to merge this one now and then if you want we could open a new one for UniSpeechFeatureExtractor

[patrickvonplaten]

Thanks a lot!

[ydshieh]

@ydshieh - if ok for you, I would like to merge this one now and then if you want we could open a new one for UniSpeechFeatureExtractor

Sure, ok for me.

[sgugger]

Thanks for fixing!