added whisper

[huseinzol05]

quick example,

from transformers import AutoProcessor, WhisperForConditionalGeneration
from cut_cross_entropy.transformers import cce_patch
import requests
from datasets import Audio

sr = 16000
audio = Audio(sampling_rate=sr)
processor = AutoProcessor.from_pretrained('openai/whisper-small')
model = WhisperForConditionalGeneration.from_pretrained('openai/whisper-small')
model = cce_patch(model)
_ = model.cuda()
r = requests.get('https://github.com/mesolitica/malaya-speech/raw/master/speech/7021-79759-0004.wav')
input_str = '<|en|><|0.0|> Without going to any such extreme as this, we can easily see on reflection how vast an influence on the ideas and conceptions,<|10.6|><|11.04|> as well as on the principles of action in mature years, must be exerted by the nature and character of the images which<|19.84|><|19.84|>the period of infancy and childhood impress upon mind.<|24.14|>'
y = audio.decode_example(audio.encode_example(r.content))['array']
labels = processor.tokenizer(input_str, add_special_tokens=False, return_tensors = 'pt').to('cuda').input_ids
decoder_input_ids = labels[:, :-1]
labels = labels[:, 1:]
batch = processor(
    [y],
    return_tensors='pt',
    sampling_rate=processor.feature_extractor.sampling_rate,
    device = 'cuda'
).to('cuda')
batch["labels"] = labels
batch["decoder_input_ids"] = decoder_input_ids
model(**batch)