deep-learningpytorchhuggingface-tokenizershuggingfacectc

TypeError: max() received an invalid combination of arguments when trying to use beam search decoding

I'm trying to run simple example of decode WAV2VEC2 output with beam search (without LM):

from pyctcdecode       import build_ctcdecoder
from transformers      import Wav2Vec2ForCTC, Wav2Vec2Processor
from torchaudio.utils  import download_asset

import torch
import librosa

processor        = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
model            = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")

FILE_NAME        = "tutorial-assets/Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.wav"
SPEECH_FILE      = download_asset(FILE_NAME)

speech, sr       = librosa.load(SPEECH_FILE, sr=16000)
input_values     = processor(speech, sampling_rate=16000, return_tensors="pt").input_values

logits           = model(input_values).logits
vocabulary       = list(processor.tokenizer.get_vocab().keys())
log_probs        = torch.nn.functional.log_softmax(logits[0])

decoder          = build_ctcdecoder(vocabulary)
text             = decoder.decode(log_probs)

I'm getting the error:

TypeError: max() received an invalid combination of arguments - got (keepdims=bool, out=NoneType, axis=int, ), but expected one of:
 * ()
 * (Tensor other)
 * (int dim, bool keepdim)
 * (name dim, bool keepdim)

As you can see I'm using pyctcdecode.

How can I decode the output of the wav2vec2 model with a beam search algorithm?

Solution

  • According to this issue on Github this error should be resolved by converting your logits to a numpy array:

    logits = model(input_values).logits.numpy()