Repository navigation
Checkpoints provided over huggingface low performance #19
Description
Hi, congratulations on this work.
I tried to load in pretrained ckpts 5s db5, epn612, 500ms db8.
Created the appropriate datsaets with same window len, but I get very low scores 32%-ish when I perform trainer.test
I also found a channel_embed key in the statedict that is not used by current model in repo.
Should I retrain on 200 ms data to reproduce paper numbers, is this expected with provided checkpoints?
Activity
Thoriri commentedon Sep 3, 2026
@MatteoFasulo Maybe you can give some pointers.
MatteoFasulo commentedon Sep 3, 2026
Hello @bamfpga,
If you're seeing an unused key in the state dict, it's likely because you're still on the old TinyMyo code. Could you double-check that you're using the updated version from PR #8?
bamfpga commentedon Sep 8, 2026
Yes, I am using latest version of the code.
I'd like to reproduce the paper results on finetuning for 200ms EPN-612, can I use the pretrained checkpoint provided? In config it says img_size=1000, but I think for 200 ms I'd need img_size=40 checkpoint.
Are released pretrined and downstream task checkpoints all for 5s, is their performance below paper's 200ms checkpoints?
MatteoFasulo commentedon Sep 8, 2026
The released pretrained checkpoint is compatible with all downstream tasks. The individual task-specific checkpoints are only provided for the 5s setup. For your 200ms experiment on EPN-612, you can safely use the shared pretrained checkpoint, but please make sure to set
img_size=40in your configuration.Thoriri commentedon Oct 12, 2026
@bamfpga I'll close this for now, just re-open if you have more questions.