Training:
- finetune.py: optional track/base param groups (WANTRACK_TRACK_GROUP) so the
track pathway can take a different LR (WANTRACK_TRACK_LR_MULT); patch_embedding
gradient masked to the track slot [:, 36:]. Raises if nothing matches.
- track_warmup.py callback: hold the pretrained DiT at lr 0 while the track
pathway warms up, then ramp it back in.
14B configs + held-alloc launchers:
- finetune_wantrack_{synth_sparse_fixed,synth_sparse_random,openvid_sparse_merged_bias,
synth_stage2}_14b + openvid_14b_{warmup,lwlr} + 1.3B stage2 configs.
- run_wan14b_held.sh (held alloc + auto-restart + rank cleanup, per-stage WANTRACK
env), acquire_nodes.sh, chain_stepAB.sh, run_stepB_seed.sh, run_stepC_merge.sh,
watch_gate.sh, run_synth_stage2_slurm.sh.
Init / diagnostics:
- convert_trackwan_init_v2.py: --pe-src to lift patch_embedding[:, 36:] and
--track-src to lift track_encoder together (co-adapted merge); sharded-source support.
- check_gate_growth.py, gradient_flow_analysis.py, gradient_flow_summarize.py.
Data pipeline: segment_tracks.py sharding; decode/segment/tracks slurm scripts;
precompute_null_text, two_track_val, cfg_ablation, lwlr_experiment.
Inference: app_multi_trace.py updates.
examples/inference/gradio/trackwan/app_action.py:
Bring the interactive drawing app in line with the sparse training
regime our overfit runs used (WANTRACK_SPARSE=1). New 'Track budget'
radio: 'dense' keeps the old full 50x50 grid; 'sparse' (new default)
builds a much smaller [Tpx, num_moving + num_bg, 2] tensor -- the
moving handle points from the drawn path + num_bg random background
points from the training grid (seeded by the seed slider). New
'num_bg' slider (0-60, default 20) matches WANTRACK_EXTRA_RANDOM=20.
Overlay logic updated to draw every point in sparse mode instead of
grid-strided subsampling that assumed N==2500.
examples/train/scenario/worldmodel/finetune_wantrack_synth_sparse_random_warmstart.yaml:
Test config for the curriculum idea: start from sparse-fixed@1000
and let random-sampling sparse training take over. Compared cold-
start sparse-random (~0.8%) vs warmstart (~3.2%) vs fixed (~8.8%)
divergence.
Add the training configs used for the golf-overfit and 20-clip synthetic
overfit series exercising every combination of the recent changes:
golf_nonzero / golf_overfit_nonzero -- non-zero-init proj ablation
synth_generic / synth_varied -- prompt-conditioning ablation
(generic vs per-clip caption)
synth_varied_d64 -- id_dim=64 (paper spec)
synth_varied_d64_nobias -- + bias=False on track head
synth_varied_d64_nobias_fixed -- + WANTRACK_FIXED_SAMPLE=1
synth_curriculum -- random-sample warmstart from
a fixed-sample checkpoint
synth_varied_full_val -- max_train_steps=1 harness that
regenerates all 20 val clips
in 3 modes from a checkpoint
synth_sparse_fixed / synth_sparse_random -- 1-per-object + 20 background
extras (Yongqi's recipe)
All resume from the trackwan_1.3b_i2v_control_init family (or the
d64_nobias variant) and enable the new joint text+motion CFG (w_t=3,
w_m=1.5) in track_validation.
_augment_tracks gated the object-coverage + low-rank-weighted + uniform draw on
`0 < max_points < N`, so with the default WANTRACK_MAX_POINTS=2500 and the full
2500-track grid (max_points == N) the entire subsample block was skipped -- the
model saw all 2500 tracks every step and only chunk-masking ran. Gate on
`min_points < N` instead so K~U[1000,2500] actually samples (verified: kept
1179/1437/1573 across steps, was a constant 2500).
Also: TrackValidation now applies the same sampler (loads object_ids/track_weights,
runs _augment_tracks deterministically) so validation is conditioned on -- and
overlays -- the sampled subset, not the full grid. Env-gated WANTRACK_DEBUG=1 logs
per-step sampling stats (kept count, object coverage, coord ranges). Adds a
single-video (golf) overfit config for debugging.
Sampler-on finetune configs (coverage + low-rank-weighted + uniform track
sampling, p_mask chunk masking) across three data regimes: unfiltered
egocentric EPIC-200, camera-static-filtered EPIC-200 (single kitchen), and the
Wan-200 synthetic set re-run with the sampler on (vs the AUG=0 overfit baseline).
- _augment_tracks: default random 1000-2500 track sampling (was 1-200), contiguous
chunk masking (was per-frame), diversity weighting (favor moving/unique tracks,
keep some static), and an object-coverage hook (>=1 track per SAM segment via
optional object_ids). Sampling ON by default; CFG dropout now default off.
- trackwan_infer.generate: motion classifier-free guidance (guidance_scale) using
the base no-tracks I2V branch as the unconditional.
- test_motion_cfg.py: quick guidance-scale EPE sweep.
- finetune_wantrack_wan200.yaml: 50->200 data-scale run (same params).
convert_droid.py builds a set where the first frame is similar but motion is
diverse, the design that forces trace reliance: it scans DROID (LeRobot mirror),
clusters frame-0 embeddings, and writes the tightest cluster as 121-frame libx264
mp4s + manifest with ONE generic caption, dropping straight into extract_tracks ->
i2v_track preprocess. Adds finetune_wantrack_droid_overfit.yaml (Fun-Control init,
WANTRACK_AUG=0 clean overfit, fresh output_dir) and gitignores .gradio/.
The overfit teacher ignored the first frame (validation looked T2V): the
init came from FastWan2.1-T2V-1.3B, whose image-conditioning channels are
zero/untrained and which has no CLIP pathway, so first-frame and track
conditioning received ~no gradient.
Re-base on Wan2.1-Fun-1.3B-InP (pretrained concat-I2V + CLIP):
- convert_trackwan_init: widen patch-embed base_in(36)->52, keep the 36
pretrained image-cond channels, zero-init only the 16 track channels;
inherit image_dim from the base config; merge sharded safetensors.
- trackwan config: image_dim=1280 builds the DiT image_embedder.
- preprocess: compute + store clip_feature (CLIP frame-0 embedding) and
add image_encoder/image_processor to required modules.
- schema: clip_feature_{bytes,shape,dtype} in pyarrow_schema_i2v_track.
- wantrack / track_validation / trackwan_infer: feed
encoder_hidden_states_image through train, validation, and inference.
- yaml: point at the Fun-InP init + re-preprocessed data.
Verified: step-0 validation reproduces the GT first frame (PSNR ~24-26 dB
vs ~9-11 before), i.e. I2V is enforced before any training.
shaoxiongduan