Use pre-trained CTC model for conformer training. #174
Conversation
It improves converging speed.
|
The pre-trained CTC models can be found at They can be used for reproducibility. |
|
LGTM! |
|
How does it affect the speed of convergence and WER? |
I am running the experiments w/ and w/o ali_model. Will post results here once they are available. |
|
Here is the result with ali_model. $ ./mmi_att_transformer_train.py --full-libri=0 --max-duration=300 --ali-model-epoch=7 --world-size=2$ ./mmi_att_transformer_decode.py
(which uses the whole lattice for LM rescoring with output beam size 8, and averaging over the last 5 epochs)
2021-04-23 15:27:14,472 INFO [common.py:373] [test-clean] %WER 5.78% [3038 / 52576, 542 ins, 179 del, 2317 sub ]
2021-04-23 15:33:31,694 INFO [common.py:373] [test-other] %WER 15.88% [8310 / 52343, 1285 ins, 579 del, 6446 sub ]The tensorboard log is available at You can compare its global objf with the one from the third try of #145 (which also used 2GPUs, but without ali_model) this pull-request(Note it uses a lower max-duration, so its number of batches is larger) You can see that it converges faster.
from #145
The valid objf of this pull-request is, however, somewhat higher. this pull-request
from #145
|
It improves converging speed.
This pull-request is just a refactoring of danpovey's work.
The following is a screenshot comparing the obj with and without ali_model.
You can see that the objf with ali_model (using this pull-request) is lower.