Tatiana Likhomanenko

@samin9796 1) run saving pytorch model into txt file ``` python3 wav2letter/recipes/models/utilities/convlm_serializer/save_pytorch_model.py /data/ahnaf/fairseq_folder/models/decoder/convlm_models/word_14B/checkpoint_best.pt /data/ahnaf/fairseq_folder/models/decoder/convlm_models/word_14B/checkpoint_best.weights ``` 2) then convert to bin format with ``` wav2letter/build/recipes/models/utilities/convlm_serializer/SerializeConvLM \ model.arch \ /fairseq_folder/models/decoder/convlm_models/word_14B/checkpoint_best.weights \ /fairseq_folder/models/decoder/convlm_models/word_14B/checkpoint_best.bin...

Sorry, a bit wrong arch, could you try to use this one? ``` # input in format (t, b, 1, 1) V -1 0 1 1 # after emb (c,...

What is the output for 70 samples (I mean there should be time for decoding printed on the screen)? How many tokens do you have in the tokens file? What...

Looks fine. You can try to set beamsize=10, beamsizetoken=10 and check on the full dataset. There could be some delay in adding to the log. When you said it gets...

Still strange, you can try to run with `beamthreashold=5` or even less to see what will happen. Do you have very large input? I would like to see the hist...

@samin9796 > Is this the same arch according to your above comment? I am still facing difficulty while writing the model.arch file. Could you please modify the previous arch file?...

Several questions: - what acoustic model did you train? s2s or CTC? - what is validation perplexity of 4gram compared to the convlm? - advice: try to set 200k vocab,...

And yes, all decoder parameters are sensitive for switching from ngram to convlm, like lm weight, wordscore, eosscore (for s2s), silweight (for asg). Possibly the problem with decoding exactly in...

Then set `--silweight=0` and optimise the `wordscore` in range (-3, 3) and `lmweight`. They will be totally different from ngram model decoding.

For me this is really very suspicious. There could be some corner cases, like for Librivox models ngram and convlm gives close results. Could you send what is the perplexity...