Experiment Preparation Status
Updated on 2026-07-13 UTC. No launchjob has been submitted.
Local execution update
All currently runnable local experiments completed on 2026-07-12. This includes E0 execution smoke, E1 operator/trained-feature audits, E2 synthetic mechanisms, 18 fixed-epoch CIFAR-100 E3 runs, three-seed E4 frozen interventions, historical ImageNet clean validation, partial E6/E8, and five-process E12 profiling. The revised ImageNet pre-gate also completed seven additional 100-epoch runs: three corrected zero-WD learned-smooth, three ReLU6 activation-only, and one fixed-smooth-c6 controller run, all with full clean plus five-corruption evaluation.
The consolidated report is
docs/LOCAL_EXPERIMENT_CONCLUSIONS.md; persistent result tables are under
/nfs/ywang29/GmNet/local_results/20260712/; pre-gate evidence is under
/nfs/ywang29/GmNet/local_results/imagenet_v2_pregate/. The revised 21-task
ImageNet-v2 matrix is prepared but no launchjob has been submitted. Only its
phase-0 ReLU6 seed-0 anchor, staged from S3 to local scratch, is initially
authorized.
Verified inputs
| Input | Location | Verification |
|---|---|---|
| CIFAR-10 | s3://snap-research-cv-code/ywang29/datasets/cifar-10/cifar-10-python.tar.gz |
170,498,071 bytes; SHA256 6d958be074577803d12ecdefd02955f39262c83c16fe9348329d7fe0b5c001ce; 50k/10k examples |
| ImageNet-1K archive | s3://snap-research-cv-code/ywang29/datasets/imagenet-1k/imagenet-1k.tar |
head-object passed; 161,381,969,920 bytes |
| ImageNet launch cache | /tmp/gmnet_data/imagenet-1k |
staged per launch pod from the canonical archive; full count and sample decoding required before training |
| Tiny ImageNet smoke | /tmp/gmnet_data/imagenet-1k-tiny |
5 classes, 10 train, 10 val; all selected samples staged and sample decoding passed |
| CIFAR-100 | s3://snap-research-cv-code/ywang29/datasets/cifar-100/cifar-100-python.tar.gz |
official MD5 eb9058c3a382ffc7106e4002c42a8d85; 50k/10k examples |
| ImageNet-1K validation cache | /tmp/gmnet_data/imagenet-1k-val |
staged from canonical S3 source; 1,000 classes and 50,000 decodable images |
Launch jobs download the complete ImageNet tar to node-local scratch and train
from the extracted /tmp tree. Its frozen manifest verifies 1,281,167/50,000
samples, the complete relative-path/target index, and one raw-file content
sample per class. The archive is removed after successful extraction.
Frozen E0 recipe
The primary recipe uses supplementary Table 8 values: 224 input, AdamW,
learning rate 0.003, weight decay 0.03, global batch 2,048, cosine decay,
5 warmup epochs, 300 epochs, rand-m1-mstd0.5-inc1, label smoothing 0.1,
CutMix 0.4, color jitter 0, and scale-specific DropPath 0/0/0.02/0.02.
Implicit defaults inherited by the public training command are now explicit: Mixup 0.8, Random Erasing 0.25, no AMP, no channels-last, no EMA, and per-epoch distributed BatchNorm-statistics reduction. The conflicting public README recipe (weight decay 0.05, CutMix 0.2, DropPath 0) remains as an audit-only config and is not used for primary experiments.
Completed checks
- 87 unit/integration tests: config inheritance, frozen/audit recipes, staged deploy protocol, gate semantics, S1-S4 topology/parameter counts, forward shape, smooth-gate gradients, RNG-exact resume streams, data/code manifests, formal analysis, evaluation, and robustness helpers.
- S1-S4 parameters: 3,661,600; 6,206,296; 7,791,544; 17,061,460.
- Single-GPU CIFAR-10 real optimizer step and evaluation: passed.
- Single-GPU five-class ImageNet using the frozen S3 recipe: passed.
- Checkpoint
resume=auto: epoch 0/global step 1 resumed to epoch 1/global step 2 with optimizer, scheduler, scaler, every rank's Python/NumPy/Torch/ DataLoader RNG streams, and best metric restored. - Eight-GPU DDP train/eval and NCCL all-reduce/broadcast: passed, exit 0.
- Environment, shell syntax, Python compile, recursive YAML parse, launch config references, and generator drift check: passed.
- Partial epochs cannot overwrite
checkpoint_last.pt; official evaluation requires a complete fixed-final checkpoint plus topology/config/data hashes. - Held/conditional commands default-deny without the exact task-specific unlock token. All generated commands verify the frozen code manifest twice.
The required TORCH_DISTRIBUTED_DEBUG=DETAIL setting emits a PyTorch 2.9
process-group shutdown warning even though all ranks reach the final barrier,
destroy_process_group() is called, collectives pass, and torchrun exits 0.
The earlier device-selection and channels-last stride warnings were removed.
ImageNet-v2 staged queue
| Role | Runs | Protocol state |
|---|---|---|
| S3 five-gate confirmatory comparison, 3 seeds | 15 | 1 ready, 14 held |
| Fixed-smooth-c6 seed-0 diagnostic | 1 | held |
| Release recipe plus paper/full-BN audits | 2 | conditional |
| S1/S2/S4 ReLU6 seed-0 scale context | 3 | held |
| Total | 21 | not submitted |
Run/output names use the imv2 namespace under
/nfs/ywang29/GmNet/runs/imagenet_v2. The staged dependencies and conditional
triggers are defined in configs/imagenet_v2_protocol.yaml; only
imv2_e0_s3_relu6_seed0 currently has submission_allowed: true.
This replaces the earlier all-ready E0/E3/E5 matrix. E3 adaptive gating, E4 causal retraining, E10 transfer, and E11 dense prediction remain outside the matrix until their dedicated runners and datasets pass smoke tests.