Skip to content

Finetune/v61 v62 training#265

Open
dcaustin33 wants to merge 12 commits into
ByteDance-Seed:mainfrom
dcaustin33:finetune/v61-v62-training
Open

Finetune/v61 v62 training#265
dcaustin33 wants to merge 12 commits into
ByteDance-Seed:mainfrom
dcaustin33:finetune/v61-v62-training

Conversation

@dcaustin33

Copy link
Copy Markdown

No description provided.

dcaustin33 and others added 4 commits April 29, 2026 21:31
Introduces scripts/ with a metric depth finetune pipeline (train loop,
dataset, losses, wandb logging, sanity checks) and adds wandb dep to
pyproject. Ignores training artifacts (checkpoints/, wandb/,
finetune_cache/) and the drone_v59_depth dataset.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- scripts/run_inference.py: one script covering both with-GT (metrics + 6-panel compare) and no-GT (2-panel compare) runs across multiple epochs.
- notebooks/kbz_inference_viewer.ipynb: no-GT viewer for KBZ_4_9 predictions with focal rescale (7200/300), random-pick hover plot, and click-to-pin depth probe.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- config.py: subset_size default 1000 -> None, so omitting --subset-size no
  longer silently caps training to 1000 samples.
- config.py / train.py: add --save-every-epochs (last epoch always saved).
- .gitignore: ignore logs/.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@dcaustin33
dcaustin33 force-pushed the finetune/v61-v62-training branch from 733fe5c to 3c599d6 Compare July 3, 2026 14:40
dcaustin33 and others added 8 commits July 3, 2026 14:43
Dead code superseded by scripts/run_inference.py (the "unified" inference
script from the same commit): run_inference.py handles arbitrary
--images-dir/--ckpt-dir plus optional GT metrics, and the KBZ viewer
notebook already points at it. Nothing referenced run_kbz_inference.py.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Stale planning doc; no longer referenced.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Frame-by-frame viewer for datasets/consecutive_groups: group dropdown +
frame slider/play over 11-frame consecutive groups, three-panel Plotly
FigureWidget (RGB | v62 depth | v63 depth) with hover-to-read depth in
meters, shared colour scale, and an optional signed v63-v62 diff map.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Config-driven MODEL_A/MODEL_B ({label, dir}); all panels, hovers, diff,
  and titles derive from labels. GROUP_RE/GROUP_SIZE (None = variable len)
  and DEPTH_SCALE are configurable too.
- Linked zoom/pan across all three panels via matched axes.
- Apply focal rescale DEPTH_SCALE = 7200/300 = 24 -> real-world meters.
- RGB hover reports both models' depth at the pixel; frame slider adapts
  to each group's frame count.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Relabel the two models as v63_e7 / v63_e4 so the compared epochs are
distinguishable in panels, hovers, and the diff. Clear all cell outputs.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Point run_train.sh at datasets/drone_v64_depth and checkpoints/drone_v64,
tag run name with v64. All other params match the v63 run.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Rename consecutive_v62_v63_viewer.ipynb -> two_model_depth_viewer.ipynb
and add a modebar draw-line tool on the RGB panel: drag across an object
of known real-world size, enter its size (mm) and camera focal (px), and
read the pinhole depth = (size_mm/1000) * focal_px / pixel_span alongside
the median predicted depth of both models under that same line.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…+ standalone pinhole calculator

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant