W&B image uploads.

§ 00Real samples from your last run

Pulled from wandb/run-20260424_225626-izsiodv7/ — what actually shipped to wandb.ai during your most recent gate-detector training. Same files, byte-for-byte.

Training mosaic — train_batch0.jpg

16 augmented training images stitched into a 4×4 grid, with ground-truth boxes drawn. This is the exact view of your dataset post-augmentation that the model saw on the first batch.

Validation: ground truth vs. predictions

Every validation epoch, the same val images are uploaded twice — once with GT boxes (left) and once with the model's predictions at that epoch (right). Scrubbing the W&B media slider shows predictions sharpening as training progresses.

Dataset statistics — labels.jpg

Histograms of class balance, box-size distribution, and x/y centroid heatmap. Gives you a one-glance read on whether your dataset has sneaky biases (all gates clustered top-center, etc.).

Curves and confusion — training trajectory

Uploaded once at training end. The PR curve and confusion matrix tell you "is the model good?"; results.png tells you "did it converge cleanly?"

§ 01The exact files

Ultralytics writes these files to runs/detect/<name>/ (or runs/pose/<name>/ for keypoints), and the W&B callback uploads each one as a media artifact attached to your run. Open any of them inside the W&B UI under Files or in the Media panel.

Training batch mosaics — your training data

These are the most "your-data-shaped" uploads. Each mosaic embeds 16 training images at native resolution, post-augmentation (mosaic, flip, HSV, mixup if enabled). If you trained on dataset_gates_dcl_web/, the frames you scraped end up here.

Validation batch images — your val set + predictions

These get re-uploaded each validation epoch (every epoch by default) so you can scrub the slider in the W&B media viewer and watch predictions sharpen over training. Ultralytics overwrites the same filename — W&B versions them server-side.

Dataset summary plots — label statistics

These describe your dataset shape, not individual images. Useful for spotting label imbalance (e.g. all gates clustered in image center → policy will overfit to that prior).

Metric curves — training trajectory

§ 02When uploads happen

The Ultralytics callback fires on these events:

For our typical config (200 epochs, validation every epoch), one detector run uploads:

• 3 training mosaics (one-time, ~ 200 KB each)
• 2 dataset plots (one-time)
• 6 validation mosaics × 200 epochs = 1200 validation image uploads
• 7 final curve plots

Roughly 200–400 MB of media per run end-to-end.

§ 03How the upload happens (code path)

Our code does not call wandb.log() directly. Instead, we pre-init a W&B run before model.train() so Ultralytics' built-in callback picks up the active run and routes everything through it.

# train_apex.py:238 — pre-init pattern
def _wandb_begin(project: str, run_name: str):
    if _env_str("AIGP_WANDB", "1") != "1":
        return None
    try:
        import wandb
        return wandb.init(project=project, name=run_name, reinit=True)
    except Exception as e:
        print(f"  [W&B] init skipped ({e})")
        return None

Inside YOLO.train(), Ultralytics' callbacks/wb.py hooks fire. Each callback grabs the file path written by plot_images(), wraps it in wandb.Image(), and logs it via wandb.log({"train_batch": wandb.Image(path)}, step=epoch).

The pre-init dance exists because of an old bug: Ultralytics 8.4.32 on Windows crashed when its W&B callback tried to parse paths containing the : in drive letters (C:\Users\...). Pre-initing means the callback finds an existing run instead of creating one with the buggy path-parsing code path.

§ 04Where they live on wandb.ai

For a run named apex_yolo11n in project aigp-gate-detector, the URL is:

https://wandb.ai/<your-username>/aigp-gate-detector/runs/<auto-id>

§ 05Privacy & access

• Default visibility: W&B projects are private by default — only your account sees them. Verify on the project's Settings → Privacy page.
• Team accounts: if your account is part of a W&B team, all team members see all team runs.
• Public sharing: changing a project to public exposes all images uploaded across all historic runs. There is no per-run public toggle.
• Deletion: deleting a run removes its media. The W&B server retains backups for ~30 days.
• Logs. stdout often contains your dataset paths, file names, and shell environment. Treat the Logs tab as semi-public if your project is public.

§ 06How to control or disable

Disable W&B entirely for one run

set AIGP_WANDB=0
python train_apex.py detector --dataset dataset_gates_mega --epochs 200

Reads AIGP_WANDB at train_apex.py:239. 0 skips wandb.init(); nothing leaves the box.

Disable just media (keep scalar metrics)

Set WANDB_LOG_MODEL=false and edit your run's settings. Or set Ultralytics' plots=False training arg — this prevents the JPGs from being written, so the callback has nothing to upload.

# Edit train_apex.py call site:
results = m.train(..., plots=False, ...)

Run W&B in offline mode (upload later or never)

set WANDB_MODE=offline
python train_apex.py detector ...

Run is recorded to ./wandb/offline-run-<ts>/. To upload later: wandb sync wandb/offline-run-*. To never upload: just leave it on disk.

Switch project (for sensitive datasets)

python train_apex.py detector \
  --dataset my_private_dataset \
  --wandb-project aigp-private-research

Use a separate, locked-down project for runs you want isolated from your main dashboard.

Audit what already uploaded

The wandb/ directory in this repo (gitignored, but on disk) contains every run's local copy. Check wandb/run-<ts>/files/media/images/ to see exactly what shipped. Same content as the W&B server, byte-for-byte.

§ 07FAQ

Are full training images uploaded, or only thumbnails?

Mosaics are at imgsz (default 640 px) per cell — so a 4×4 mosaic is 2560×2560 px. That's the same fidelity as your training pipeline saw. They're saved at JPEG quality 95 by Ultralytics' plot_images(). Not thumbnails.

Are model weights uploaded?

By default, no. Set WANDB_LOG_MODEL=true if you want them uploaded as W&B artifacts. We don't, because models live in models/latest/ with our own backup/promote rotation.

Does the policy phase upload anything?

No. Phase 3 (PPO) doesn't use W&B at all in our wiring. Metrics there go to TensorBoard (output/apex_policy/tb_apex/) and the new live Rich dashboard. Stay-local-only.

What about RF-DETR runs?

RF-DETR uses its own roboflow.deploy() path and is not currently wired to W&B in train_apex.py. If you add it, it would log via the same Ultralytics-style callback pattern.

What if I want every val image, not just three mosaics?

Edit Ultralytics' default val_batches=3. There's no clean kwarg — patch the source or set cfg.val.plots hooks. Most teams find 3 mosaics × 16 images = 48 val images per epoch is plenty for visual eyeballing.

Cross-refs: training runbook · model eval dashboard · related code: train_apex.py:236-260, train_apex.py:413, train_apex.py:620.