gentaiscool
The CER of librispeech?
hi @flp1990 that depends to the architecture you use. You may check #34. There are some results reported there. I think, you can get better results by carefully tuning the model (+ deeper + wider models + regularization, and larger beam search).
@gentaiscool thanks to reply,i have check #34 .i took your initial (except i set batch_size = 8,because memory limit )setup to trainning model on 960 hours librispeech datasets.After 3 days trainning(1 rtx titan ) it has obtain Good results. 2020-11-13 09:32:47,406 - (Epoch 20) TRAIN LOSS:0.1903 CER:6.53% LR:0.0000167 2020-11-13 09:32:47,461 - VALID 2020-11-13 09:33:31,062 - VALID SET 0 LOSS:0.1780 CER:5.78% 2020-11-13 09:33:34,481 - SHUFFLE 2020-11-13 09:33:34,487 - TRAIN 2020-11-13 13:55:36,838 - (Epoch 21) TRAIN LOSS:0.1832 CER:6.25% LR:0.0000163 2020-11-13 13:55:36,894 - VALID 2020-11-13 13:56:31,427 - VALID SET 0 LOSS:0.1735 CER:5.59% 2020-11-13 13:56:35,347 - SHUFFLE 2020-11-13 13:56:35,349 - TRAIN I want know your best results cer is (transformer the batch_size bigger the results is better )?and how many epochs usually in train?thanks.
I didn't record the results before. But, you can try 50-100 epochs. Btw, the CER showed in the log is not the autoregressive CER. You need to run the inference/test to check the real CER.
I didn't record the results before. But, you can try 50-100 epochs. Btw, the CER showed in the log is not the autoregressive CER. You need to run the inference/test to check the real CER.
What does "the CER showed in the log is not the autoregressive CER. You need to run the inference/test to check the real CER." mean?
I didn't record the results before. But, you can try 50-100 epochs. Btw, the CER showed in the log is not the autoregressive CER. You need to run the inference/test to check the real CER.
What does "the CER showed in the log is not the autoregressive CER. You need to run the inference/test to check the real CER." mean?
The CER during training is calculated with target-gt as input of decoder, so it must be higher than the true CER. You should run the test.py which gets the output one by one.
How to achieve training on librispeech database? I have been unable to complete the operation of downloading librispeech database. May I ask what modification should be made to the Librispeech.py file if I download the database by myself? In addition, after making changes to the database, what train instruction should I use to complete the training process? Thank you for your answer.
