TODOs as of 11/29
Opened this issue · 2 comments
ZhaofengWu commented
- Seeding (Zhaofeng)
-
Dataset
s should probably sharechar2idx
as we want unseen chars to be generated as well (Joseph) - Sort through the data points in increaseing video length
- Fix any dataset problem (e.g. micro not loading) (Joseph)
- Check [INAUDIBLE] token #13 (Joseph)
- Write something that puts dataloader/model/training/eval/logging/etc. together. Somethine like the current
test_better_model.py
but more sophisticated? Perhaps with tensorboard support (Joseph) - Finish eval function with loss and CER output (Zhaofeng)
- (optional) Plug in a WER implementation into eval function (Yutong)
- Maybe just write one ourselves
- Joint CTC+attention model (Zhaofeng)
- CNN frame processing (Joseph)
- Experiments. See #18. (Yutong & Joseph)
- Inference (beam search) (Yutong)
- Implement exact-foward-approximate-backward?
joseph-zhong commented
Fuck me.....
vdalv commented