Wav2vec 2.0 doing inference on newly trained transformer lm model gives missing key error

[amant555]

🐛 Bug

Hi, I trained a transformer LM model for inferencing in Wav2vec 2.0. It gave a similar error of getting None in args in model dict. Like in previous similar issue, I changed the value of state["args"] to state["cfg"]["model"] but the data key was missing in that. So again changed it to state["cfg"]["tasks"] where it gave a new error of missing key language_modeling. It's probably a issue because of moving repo to hydra config. I tried to find a solution for this, but there is distribution of keys in different namespaces that are needed.

Can you look into it. Mainly the error is arising from init method of fairseqlm in w2l_decoder.py.

[alexeib]

is this still an issue? plz reopen if so

[max-15s]

I faced the same issue, even with 'transformer_lm.wmt19.en'. My setup: torch: 1.9.0 from pip fairseq: from master CUDA Version: 11.2 (Although it's irrelevant, I've tested this on GPU too) The problem arises from checkpoint = torch.load(args.kenlm_model, map_location="cpu"). Here the checkpoint is a dict. However, it is treated as a OmegaConf. If we check the type of checkpoint in line 383 and cast it, we won't face this problem:

if "cfg" in checkpoint and checkpoint["cfg"] is not None:
    if isinstance(checkpoint["cfg"], dict):
        lm_args = OmegaConf.create(checkpoint["cfg"])
    else:
        lm_args = checkpoint["cfg"]

@alexeib, it is a very minor error. But, I would like to make a pull request.

[alexeib]

please make a pull request @max-15s, thanks!