negative ctc_loss when training

TensorSpeech / TensorFlowASR

:zap: TensorFlowASR: Almost State-of-the-art Automatic Speech Recognition in Tensorflow 2. Supported languages that can use characters or subwords

https://huylenguyen.com/asr

Apache License 2.0

938 stars 245 forks source link

negative ctc_loss when training #172

Closed TinBarbie closed 2 years ago

TinBarbie commented 3 years ago

Hi, this repo is really awesome. I'm training deepspeech2 model with vietnamese dataset (vivos + viet_tts) and encounting an issue when the ctc_loss is negative after some batches. Is it possible to be negative? if not, any idea to fix this issue? Thanks. Capture

nglehuy commented 3 years ago

@TinBarbie The loss shouldn't be negative, can you train only on vivos to see if the issue happens?

nglehuy commented 2 years ago

Feel free to reopen if the issue still exists. Make sure your dataset contains only characters specified in vocab file.