aditya

This would be a great feature to add (I've got a couple of projects that use `hmm_marginal` and would really benefit from time varying transition probabilities). Would I be able...

It seems like this wasn't the case in the oldest commit of the model: https://github.com/facebookresearch/fairseq/blob/392fce8a9873e54eca71cfca9d98f2685fdf6238/fairseq/models/wav2vec.py#L405 and was introduced around the time of vq-wav2vec?

Also shouldn't `wav2vec_predictions.infonce` be set to `False` for wav2vec large?