Training result looks abnormal

Dear everyone,

I followed guide(https://github.com/mozilla/DeepSpeech/blob/master/TRAINING.rst) and trained with CommonVoice English.

Training is done. But result looks abnormal. I don’t know why.

Would you plz help me?

— Training --------------------------------------------------------------------------------------

I Saved new best validating model with loss 127.796996 to: /home/sangwon.lee/.local/share/deepspeech/checkpoints/best_dev-61796
Epoch 4 | Training | Elapsed Time: 2:24:50 | Steps: 15449 | Loss: 141.306873
Epoch 4 | Validation | Elapsed Time: 0:09:15 | Steps: 3151 | Loss: 128.151905 | Dataset: data/CommonVoice/clips/dev.csv
Epoch 5 | Training | Elapsed Time: 2:21:31 | Steps: 15449 | Loss: 141.785106
Epoch 5 | Validation | Elapsed Time: 0:09:15 | Steps: 3151 | Loss: 127.252613 | Dataset: data/CommonVoice/clips/dev.csv
I Saved new best validating model with loss 127.252613 to: /home/sangwon.lee/.local/share/deepspeech/checkpoints/best_dev-92694
Epoch 6 | Training | Elapsed Time: 2:23:25 | Steps: 15449 | Loss: 141.677562
Epoch 6 | Validation | Elapsed Time: 0:09:15 | Steps: 3151 | Loss: 128.376973 | Dataset: data/CommonVoice/clips/dev.csv
I Early stop triggered as (for last 4 steps) validation loss: 128.376973 with standard deviation: 0.369841 and mean: 127.733838
I FINISHED optimization in 18:21:15.930897

----Result------------------------------------------------------------------------------------------------

INFO:tensorflow:Restoring parameters from /home/sangwon.lee/.local/share/deepspeech/checkpoints/best_dev-92694
I1209 17:22:46.776838 139829627459328 saver.py:1280] Restoring parameters from /home/sangwon.lee/.local/share/deepspeech/checkpoints/best_dev-92694
I Restored variables from best validation checkpoint at /home/sangwon.lee/.local/share/deepspeech/checkpoints/best_dev-92694, step 92694
Testing model on data/CommonVoice/clips/test.csv
Test epoch | Steps: 12580 | Elapsed Time: 1:51:24
Test on data/CommonVoice/clips/test.csv - WER: 0.985325, CER: 0.863086, loss: 117.808365

WER: 2.000000, CER: 0.888889, loss: 54.817043

  • wav: file:///home/sangwon.lee/deepSpeech/DeepSpeech/data/CommonVoice/clips/common_voice_en_54384.wav
  • src: “undefined”
  • res: “i entreat”

WER: 2.000000, CER: 0.794118, loss: 104.489899

  • wav: file:///home/sangwon.lee/deepSpeech/DeepSpeech/data/CommonVoice/clips/common_voice_en_574469.wav
  • src: “aesthetically speaking it was ugly”
  • res: “i i i i i i i i i i”

WER: 1.857143, CER: 0.750000, loss: 108.312141

  • wav: file:///home/sangwon.lee/deepSpeech/DeepSpeech/data/CommonVoice/clips/common_voice_en_49828.wav
  • src: “the men fell into an animated discussion”
  • res: “i i i i i i i i i i i i i”

WER: 1.750000, CER: 0.857143, loss: 60.598770

  • wav: file:///home/sangwon.lee/deepSpeech/DeepSpeech/data/CommonVoice/clips/common_voice_en_147994.wav
  • src: “then a man approached”
  • res: "i i i i i i i "

WER: 1.727273, CER: 0.787500, loss: 193.794495

  • wav: file:///home/sangwon.lee/deepSpeech/DeepSpeech/data/CommonVoice/clips/common_voice_en_87016.wav
  • src: “epilepsy is a condition which sends the sufferer into uncontrollable convulsions”
  • res: “i i i i i i i i i i i i i i i i i i i”

I don’t either, and since you shared none of the command line you ran to perform this training, we can’t help you.

Could you please use proper code formatting as well? Console output is hard to read without.