Fluctuating training loss

ming024 / FastSpeech2

An implementation of Microsoft's "FastSpeech 2: Fast and High-Quality End-to-End Text to Speech"

MIT License

1.69k stars 515 forks source link

Open 299792459b opened 10 months ago

299792459b commented 10 months ago

Red is training loss Orange is valid loss

Other datasets show same behavior

Training loss fluctuates up and down drastically. Anyone know why is this happening? Thank you

melodyze-ai commented 9 months ago

Try lowering learning rate.

asarsembayev commented 2 months ago

Try lowering learning rate.

which parameter in config is actually LR?