Backup / gmnet /deploy /README.md
YFanwang's picture
Add files using upload-large-folder tool
318674b verified
|
Raw History Blame Contribute Delete
3.5 kB

GmNet ImageNet-v2 Long-Run Configurations

This directory keeps the requested depoly spelling. The active matrix is the staged ImageNet-v2 protocol in journal_exp/configs/imagenet_v2_protocol.yaml; generation never invokes launchjob.

The post-hoc, single-seed mechanism follow-up requested after the legacy jobs completed is isolated in single_seed_followup_20260715/. Its task names, outputs, approvals, and protocol are separate from the formal ImageNet-v2 matrix; see that directory's README before submitting any YAML.

Current Protocol State

Exactly 21 YAMLs are prepared below imagenet_v2/:

Role Runs Initial status
S3 confirmatory: ReLU6/ReLU/corrected learned-smooth/ReLU6-only/no-gate, 3 seeds each 15 1 ready, 14 held
S3 fixed-smooth-c6 diagnostic, seed 0 1 held
Release-hyperparameter recipe audits with paper BN and historical full BN 2 conditional
S1/S2/S4 ReLU6 scale context, seed 0 3 held
Total 21 1 ready, 18 held, 2 conditional

Only imv2_e0_s3_relu6_seed0 has submission_allowed: true. Every other YAML is intentionally prepared but locked. A YAML file existing is not permission to submit it; task_matrix.yaml is authoritative for current status, dependencies, role, and phase.

All run names contain imv2 and outputs are isolated below /nfs/ywang29/GmNet/runs/imagenet_v2. Jobs read ImageNet-1K from /s3-code/ywang29/datasets/imagenet-1k, whose canonical source is s3://snap-research-cv-code/ywang29/datasets/imagenet-1k/.

Every generated command supplies CODE_MANIFEST_PATH=configs/imagenet_v2_code_manifest.json. The launcher refuses code/config drift before training and before official evaluation. imagenet_v2/data_manifest_canonical.json freezes exact sample counts, the full relative-path/target index, and one deterministic raw-file content sample per class.

Staged Unlock

  1. Submit only the phase-0 ReLU6 S3 seed-0 anchor after manual review.
  2. Validate its fixed-epoch metrics, logs, checkpoint, and recipe before unlocking phase-1 seed-0 gate comparisons.
  3. Unlock each gate's seeds 1 and 2 only after that gate's seed-0 run is valid.
  4. Run fixed-c6 and S1/S2/S4 only as diagnostics after their dependencies pass.
  5. Keep recipe audits conditional. The paper-BN audit requires a documented anchor discrepancy; historical full-BN additionally requires an unresolved topology discrepancy. These are approximate recipe-bundle diagnostics using this repository's runner, not release-exact timm reproductions.

Held and conditional YAMLs have a default-deny shell guard in both setup and training commands. They require the exact task-specific GMNET_PROTOCOL_UNLOCK_TASK value after documented prerequisite review.

The primary confirmatory result is the three-seed fixed-epoch comparison. Do not select checkpoints or gates by the ImageNet validation maximum.

Generate And Validate

cd /nfs/ywang29/GmNet/journal_exp
python scripts/generate_deploy.py
python scripts/generate_deploy.py --check

Generation copies resources, image, mounts, ownership, and cost fields from /nfs/ywang29/LongLive/deploy/jul11_vgp/j11_vgp01_base.yaml. It also removes unexpected stale YAMLs only when they carry the generator header; hand-written YAMLs are never deleted. --check reports both missing/drifted expected files and stale generated files without changing the filesystem.

No launchjob has been submitted by this preparation step.