search

ictnlp / NMLA-NAT

Code for NeurIPS 2022 Spotlight paper " Non-Monotonic Latent Alignments for CTC-Based Non-Autoregressive Machine Translation"
MIT License
20 stars 2 forks source link

'None' loss without exception #1

Open baoguangsheng opened 1 year ago

baoguangsheng commented 1 year ago

Hi, I run the training on IWSLT17 En-De translation dataset, but got following 'None' loss without raising any exception. Could you help? Thanks!

2023-02-23 11:14:21 | INFO | train | epoch 001 | loss None | nll_loss None | ppl 0 | wps 59093 | ups 4 | wpb None | bsz None | num_updates None | lr None | gnorm None | clip None | oom None | loss_scale None | train_wall 82 | wall 85 2023-02-23 11:14:23 | INFO | valid | epoch 001 | valid on 'valid' subset | loss None | nll_loss None | ppl 0 | wps 143368 | wpb None | bsz None | num_updates 328

shaochenze commented 1 year ago

Thanks for pointing out the problem! It's a logging problem and should not affect the model performance, so you can keep training your model in this way. I will fix the problem later.

TingxunShi commented 1 year ago

NMLA-NAT/fairseq/metrics.py line 123 was commented out. Recover this line you can get the detailed statistical information