Ahnaf Mozib Samin

@wolfassi123 Did you able to fix it? I am also getting beeping sound with training a model on LJSpeech dataset.

Hi @tlikhomanenko This is the command I used for training: > **python3 train.py /data/ahnaf/fairseq_folder/models/first_try --save-dir /data/ahnaf/fairseq_folder/models/decoder/convlm_models/word_14B --task=language_modeling --arch=fconv_lm --fp16 --max-epoch=48 --optimizer=nag --lr=0.5 --lr-scheduler=fixed --decoder-embed-dim=128 --clip-norm=0.1 --decoder-layers='[(512, 5)] + [(128, 1,...

@tlikhomanenko I created a model.arch file as you mentioned and then ran the command. It shows following error: > F0605 06:28:50.317502 32772 Utils.cpp:291] mismatch between the number of parameters in...

@tlikhomanenko Thanks a lot. It worked. However, when I try to decode 6500 samples (20 hours), it gets stuck. This is from the terminal: > Skipping unknown entry: 'কম্পিটিশনে' Skipping...

@tlikhomanenko For 70 samples, actual decoding time is 270.871sec and it took 3.8 sec per sample with nthread decoder=1. I have 178 tokens in the tokens file. I used ctc...

I waited for more or less an hour but found nothing at terminal. I just checked one or two times in the log file. One more thing, I built a...

@tlikhomanenko I have only 10-15 sec inputs. I am going to test with `beamthreashold=5` and then let you know. About LM: I trained a new LM using the following command:...

@tlikhomanenko Using Convlm I get WER 12.3%, 4 gram LM results in WER 6.3 %. Without applying any LM, I get WER 20%. My training ppl is around 70 and...

@tlikhomanenko Thank you for your reply. I trained AM with CTC and trained word-based LM. I don't have answers to the other questions yet but I will let you know....

@tlikhomanenko Hi! I trained a convLM with 273k vocab and found out that n-gram is still better by a large margin. Then I experimented with a smaller text corpus (only...