From other posts, I know that WER calculation and report preparation takes a very long time (when display_step > 0) as its implemented in CPU. So, I run my training with display_step = 0. This speeds up the training speed A LOT (training time for one epoch reduced to 34 seconds from 30 minutes on Google Colab) on both version 0.2 and 0.3 of DeepSpeech. But, the downside is, I don’t get to see any WER result and can’t see how the model is performing on word level.
Is there any way to show WER for some epoch without lengthy operation? I only require an average WER value for some epochs and nothing more.
As you said, WER computation is CPU intensive, there’s no magic here. But you could just report once every ten epochs, for example, by adjusting display_step ?