Add a new eighth training session, with 2000 steps.
FangSen9000
a9be817
Start the more advanced training again. Aim for 800 steps. For now, evaluate every 100 steps.
FangSen9000
df420df
Start the more advanced training again. Aim for 800 steps. For now, evaluate every 100 steps.
FangSen9000
36f5bb6
Track SMKD .h5 artifacts via LFS
FangSen9000
b4b4729
Then, after the initial version training, dropout was added, and the training was continued until 38 epochs.
FangSen9000
f2932c3
fix typo
FangSen9000
556aa9b
Add frame-annotated gloss files and documentation
FangSen9000+1: Claude
875e074
- Generated ASLLRP_utterances_with_frames.json with precise frame numbers for each gloss
- Added comprehensive documentation for ASLLRP dataset structure and usage
- Updated .gitignore to exclude xlsx files (using CSV instead)
- Moved analysis scripts to SignX/doc/ directory
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Remove "Position" to make the display more tidy.
FangSen9000
3a4b7bf
A more perfect version requires checking whether the segmentation is correct.
FangSen9000
193d58a
Increase the threshold corresponding to the feature-frame; add the corresponding original image of smkd
FangSen9000
741864d
Add a progress bar corresponding to the gloss that can control the threshold.
FangSen9000+1: Claude
cfee709
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Relative time and the original frame can now be displayed.
FangSen9000
7162aa8
Successfully completed the attention-based segmentation inference
FangSen9000
9803b71
The reasoning script no longer deteriorates. Although sltunet requires more training to remove unknown tokens; the @ symbol has been removed.
FangSen9000
65e4828
Restore SignX files from pre-reset snapshot
FangSen9000
7393a38
Attempted to submit 4 changes, although the reasoning degraded, the reasoning could still run.
FangSen9000
1eb306c
A little normalization for Dino and some run-time optimizations were added, but they didn't make much difference. Then it was reverted to the Resnet version.
FangSen9000
0e1a71d
The dino3 model can be submitted for normal training, but the loss and wer training results are not obvious.
FangSen9000
de1bf29
The dino3 model can be submitted for normal training, but the loss and wer training results are not obvious.
FangSen9000
6cfcd99
It seems that Dino Base can now be trained normally.
FangSen9000
2402b3f
Delete the videos related to dwpose on the online platform
FangSen9000
080b426
Clean up the old training logs
FangSen9000
a2734a8
Optimize the name of the training folder
FangSen9000
37a8276
Delete the original version of CSLR plus
FangSen9000
35db2f5
add some part mp4
FangSen9000
181b71e
Add ASLLRP model-2040 checkpoint
FangSen9000
7293c60
Track ASLLRP checkpoints via git-lfs
FangSen9000
7139eb5
Add SLTUNET training code and configs (no large checkpoints)
