hyperconformer_8M on librispeech train-clean-100 does not perform as well as expected #2955
Replies: 1 comment 4 replies
-
|
Hello, have you found the problem and the solution?I also used hyperconformer_8M to reproduce Table 2 in the paper.I further modified the attention_type to get conformer_8M, and its WER is 15.57. |
Beta Was this translation helpful? Give feedback.
-
|
I'm afraid not! Good luck with your journey |
Beta Was this translation helpful? Give feedback.
-
|
Make sure that your batch size is large enough. Play with grad accum and nb of GPU. The batch size must be large. |
Beta Was this translation helpful? Give feedback.
-
|
Hello,I have tried setting grad_accumulation_factor to 2 and 4 respectively, but there remains a significant performance gap for conformer_8M.I am currently running experiments with a larger batch size, increasing it from 16 to 32 by using more GPUs. |
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Hi all,
I'm trying to train hyperconformer_8M as a baseline for my experiments, training only on train-clean-100. The architecture certainly takes less memory and is faster to train, but the WER is higher than expected. I followed the train.py script, using hparams/hyperconformer_8M.yaml.
Table 2 from the paper suggests the result should be 6.76% WER on test-other, but I get 16.74% WER.
Does anyone know what might be causing this?
The only things I can think of that can cause this discrepancy are different decoding or LM scoring strategies, or the data got corrupted on my server.
The full metrics for the final checkpoint were
I trained on an RTXA4500 20GB.
Beta Was this translation helpful? Give feedback.
All reactions