feat(2026): distill to cnn-medium-attn+RoPE, add inference CLI, restructure dist/

[AmitMY]

Summary

Distills the 2026 experimental work down to a clean, minimal production path. Removes all architecture branches we didn't end up using and all loss/augmentation variants that didn't improve results.

Changes

• model.py (805→294 lines): cnn-medium-attn + RoPE only — removed bilstm, bigru, tcn, cnn-fast-slow, cnn-local-attn, cnn-lstm, cnn-large, cnn; removed focal loss, label smoothing, b-dice, per-head weighted loss, legacy flags
• train.py: removed curriculum callbacks (SegmentationDataModule, CurriculumCallback, FrameCurriculumCallback); simplified to direct get_dataloader() path
• args.py: removed ~30 unused args; defaults reflect proven recipe (hidden_dim=384, depth=6, nhead=8, dice_loss_weight=1.0, epochs=200)
• dataset.py: removed acceleration and speed_aug branches
• evaluate.py: added as tracked file; removed windowed/LSTM eval path
• bin.py (new): 2026 inference CLI — loads .ckpt, preprocesses pose, runs chunked RoPE inference, writes ELAN
• old/bin.py: updated dist path to dist/2023/
• dist/2023/: moved 2023 TorchScript .pth models from old/dist/; added README
• dist/2026/: added EXPERIMENTS.md + findings README.md

Key findings (dist/2026/README.md)

Helped: RoPE transformer, Dice loss (weight=1.0), fps_aug (essential — disabling costs −9pp Sign IoU), body_part_dropout=0.1 (+10.5pp Phrase25), frame_dropout=0.15 (essential regularisation), velocity features, no_face, hidden_dim=384, HM(sign,phrase) validation metric, inference chunk_size=num_frames (bug fix: +12.8pp Phrase IoU)

Didn't help: attention padding mask (−7pp S25 consistently), B-frame Dice loss, focal loss, label smoothing, speed augmentation, frame curriculum, removing frame_dropout (catastrophic phrase overfitting after ~50 epochs)

Best result so far (E145, 1024 frames)

	Sign IoU	Phrase IoU
50fps	0.595	0.907
25fps	0.569	0.880
HM		0.705

🤖 Generated with Claude Code

[AmitMY]

where is the dockerfile? update to the main README that says how to train? how to run?

[AmitMY]

did we check which is better? likeliest or probs_to_segments? from thinking that we might not need thresholds anymore? that would simplify the code a lot

[AmitMY]

Haven't done a systematic comparison, but from threshold-sweep experiments (--tune_threshold) likeliest (argmax) has generally matched or beaten threshold-based on IoU — and it's simpler. Made likeliest the default. Threshold-based still available via --threshold + --tune_threshold if needed. Happy to add a explicit comparison to the eval script if you'd like.

[AmitMY]

please make an evaluation of likeliest vs threshold and report here. if likeliest is the same/better, we can remove the threshold code, and all threshold tuning code. simplifies the repo

[AmitMY]

Running E165 (1536fr, new code) to get a clean comparison. Will report here once done. Previous models (E162, E163) were trained with probs_to_segments as the validation metric, so comparing on those is skewed. E165 is the first model trained with the new fps-normalised velocity and ClassifierHead, so results will be directly comparable.

[AmitMY]

Ablation results (E165-1536-batch8-drop01-fixchunk-hmval-3h, dev, 50fps):

Decoding	Sign IoU	Phrase IoU	HM
likeliest (argmax)	0.5598	0.8892	0.6871
threshold (default 0.5)	0.5519	0.8885	0.6809
tuned threshold (swept)	0.5706	0.8937	0.6965

Tuned threshold beats likeliest by +0.011 Sign IoU, so we keep threshold-based decoding for final reporting. likeliest is the fast-path default (no tuning needed); --tune_threshold is the recommended path for final numbers.

Note: training validation used likeliest; the eval gap vs tuned threshold is ~1pp.

[AmitMY]

Updated main README.md with Docker-based training/evaluation instructions and a pointer to Dockerfile.train. The Dockerfile was already in the repo root but not documented in the README.