wav2vec pre-trained models require inconsistent versions of fairseq

The pre-trained models available for the wav2vec flavors use outdated, inconsistent, and undocumented versions of fairseq.

First of all, which version of fairseq is required for each of the pre-trained models is not documented on this README. Users may expect that this would be (a) current master branch on the GitHub, or (b) whatever version of fairseq was automatically chosen by the user’s Python package manager (e.g. pip install fairseq, which in my case gave me 0.10.2). Unfortunately, neither of these assumptions is correct for all of the pre-trained models.

The models listed under wav2vec 2.0 Base and Large are compatible with 0.10.2. (I have not tried all of them.) They fail to initialize with a GitHub based fairseq installation.

The XLSR-53 model, which is the one I actually wanted to use, is compatible with neither 0.10.2 nor top-of-tree at time of writing.

I have not tried wav2vec 1.0 or vq-wav2vec, but the required versions for those is also not documented.

The easiest way to solve this would be to document the required versions for each of these pre-trained downloads, preferably with a link to the GitHub tag for the needed repo state. Better yet would be an update to the pre-trained models every time the repo is updated in such a way that they break, and links to all the historical versions of the models so users stuck on a particular version for some reason can download the model version compatible with their fairseq version.

[GreatDarrenSun]

I think I have found the answer，I downloaded version 7061a0f of master，reinstall it，and the error is gone

[stale[bot]]

Closing this issue after a prolonged period of inactivity. If this issue is still present in the latest release, please create a new issue with up-to-date information. Thank you!