Experiment Task Classification
Execution classes
| Class | ETA | Execution | Storage |
|---|---|---|---|
| L0 | <= 2 h | local interactive | data/results under /tmp |
| L1 | 2-12 h | local 1-8 GPU | data/results under /tmp |
| J1 | > 12 h | launchjob, one 8xA100 node | data from /s3-code, results on NFS |
| J2 | multi-stage or many seeds | separate launchjob YAML per run | data from /s3-code, results on NFS |
Plan mapping
Local completion status and claim boundaries are recorded in
docs/LOCAL_EXPERIMENT_CONCLUSIONS.md. The classifications below remain the
execution policy for reruns and long follow-up work.
| ID | Task | Class now | Launch dependency |
|---|---|---|---|
| E0 | Environment, S3 and DDP smoke | L0 | none |
| E0 | ImageNet-v2 S3 ReLU6, 300 epochs, 3 seeds | J2 | staged baseline anchor |
| E1 | Filter reconstruction and frequency metric audit | L0/L1 | trained checkpoint |
| E2 | Sinusoid/random-field/cue-conflict synthetic study | L0/L1 | none |
| E3 | CIFAR-100 gate screening | L1 | E2 implementation |
| E3 | ImageNet-v2 five-gate confirmatory matrix, 3 seeds | J2 | validated E0 anchor |
| E4 | Cheap gate interventions at evaluation | L1 | E0 checkpoint |
| E4 | Causal controls that require retraining | J2 | E4 shortlist |
| E5 | S1/S2/S4 single-seed scale context | J2 | validated E0 anchor |
| E6-E9 | Corruption/OOD/attack evaluation | L1; split if >12 h | frozen checkpoints |
| E10 | Fine-grained transfer | L1/J1 by dataset | frozen ImageNet checkpoint |
| E11 | COCO/ADE20K/Cityscapes training | J2 | detection/segmentation adapters |
| E12 | Local GPU/CPU profiling | L0/L1 | frozen checkpoints |
| E12 | iPhone/Jetson profiling | external device | exported artifact |
ImageNet-v2 staged batch
The revised machine-readable protocol is
configs/imagenet_v2_protocol.yaml. It prepares exactly 21 long-run YAMLs:
15 S3 confirmatory runs across ReLU6, ReLU, corrected learned-smooth,
ReLU6-only, and no-gate; one fixed-smooth-c6 diagnostic; two conditional
recipe/BN audits; and three single-seed ReLU6 scale-context runs for S1/S2/S4.
Only the S3 ReLU6 seed-0 E0 anchor is initially submit-ready. The remaining
confirmatory and scaling runs are held behind explicit dependencies, while the
two recipe audits are conditional on a documented reproduction discrepancy.
See docs/IMAGENET_V2_PROTOCOL.md and depoly/task_matrix.yaml; the existence
of a generated YAML is not submission authorization.
The local controller prerequisite is now passed with 7/7 complete runs. It does not unlock a launch by itself: corrected-smooth seed 0 remains held behind the common S3 ReLU6 anchor. Activation-only is tested for operator-replacement non-inferiority, not used as a material-loss gate that blocks ReLU/smooth tests.
E3 adaptive gating, E4 retraining, E10 transfer, and E11 dense prediction are outside this 21-task matrix and remain blocked until their runners, data contracts, and local smoke tests are complete.
Data registry
| Dataset | Required S3 prefix | Local cache |
|---|---|---|
| CIFAR-10 | s3://snap-research-cv-code/ywang29/datasets/cifar-10/ | /tmp/gmnet_data/cifar-10 |
| ImageNet-1K | s3://snap-research-cv-code/ywang29/datasets/imagenet-1k/ | direct mount for launch; optional /tmp stage |
| ImageNet-C | s3://snap-research-cv-code/ywang29/datasets/imagenet-c/ | /tmp/gmnet_data/imagenet-c |
| OOD/fine-grained/dense | dataset-specific child prefix | /tmp/gmnet_data/ |
Only CIFAR-10 and ImageNet-1K are confirmed present at preparation time.