search

NVIDIA / OpenSeq2Seq

Toolkit for efficient experimentation with Speech Recognition, Text2Speech and NLP
https://nvidia.github.io/OpenSeq2Seq
Apache License 2.0
1.54k stars 369 forks source link

DS 2 pre-trained model epoch numbers? #402

Closed inchpunch closed 5 years ago

inchpunch commented 5 years ago

On page: https://nvidia.github.io/OpenSeq2Seq/html/speech-recognition/deepspeech2.html

For WER=6.71%, it says "a pre-trained model which was trained for 200 epochs". However, in the given configuration file "ds2_large_8gpus_mp", it uses 8 GPUs but "num_epochs": 50.

I tried 50 epochs with 4 GPUs and got WER=8.5%. Wondering if the reference WER 6.71% was from 200 epochs or 50 epochs?

borisgin commented 5 years ago

We trained the model for 200 epochs using 8 GPUs.

inchpunch commented 5 years ago

Thanks for the information.