gmnet/code/journal_exp/docs/OFFICIAL_LOCAL_COMMANDS.md

Official Local E4 and E12 Commands

Scope and scheduling

These jobs are local and expected to finish well below 12 hours, so they do not need additional launchjob YAML files in depoly/. The shared shell entry point is scripts/run_e4_e12_official.sh. It defaults to DRY_RUN=1, resumes by skipping result files whose status starts with completed, and only overwrites completed work when FORCE=1 is explicitly supplied.

The shell contains the export lines from /nfs/ywang29/LongLive/scripts/init_run.sh verbatim. Do not edit those lines.

Inspect the exact commands without running anything:

cd /nfs/ywang29/GmNet/journal_exp
DRY_RUN=1 bash scripts/run_e4_e12_official.sh plan

E4 frozen interventions

Wait until GPUs 0-2 are available, then evaluate all five frozen-weight modes on the full 10,000-image CIFAR-100 test set for each of the three independently trained ReLU6 checkpoints. The three seeds run concurrently, one process on each of GPUs 0, 1, and 2. The shell waits for every process, propagates any failure, and aggregates only after all seeds succeed:

cd /nfs/ywang29/GmNet/journal_exp
DRY_RUN=0 bash scripts/run_e4_e12_official.sh e4

Inputs are the three /tmp/gmnet_runs/e3_cifar100/e3_c100_s1_relu6_seed{0,1,2}/checkpoint_last.pt files. Per-seed results and the cross-seed aggregate.json/AGGREGATE.md are written under /tmp/gmnet_runs/e4_cifar100_relu6_official/.

Estimated wall time: 2-6 minutes, depending on concurrent GPU and filesystem load. This is a frozen-checkpoint diagnostic, not matched intervention retraining.

Legacy ImageNet clean validation

After GPU 0 is free, evaluate the historical public checkpoint on all 50,000 ImageNet-1K validation images using the verified local staging cache at /tmp/gmnet_data/imagenet-1k-val:

cd /nfs/ywang29/GmNet/journal_exp
DRY_RUN=0 bash scripts/run_e4_e12_official.sh legacy-clean

The result is written to /tmp/gmnet_runs/e4_legacy_imagenet_clean_official/. Report Top-1, Top-5, NLL, and 15-bin ECE only as a historical full-BN compatibility-topology result. It is not an E0 reproduction of the paper-default topology. Each result records the canonical s3://snap-research-cv-code/ywang29/datasets/imagenet-1k/val source, the local cache access path, and the contents and SHA256 of the .READY staging manifest.

Estimated wall time: 5-15 minutes from the local staged data.

E12 profiling

Run E12 last, with no other CPU or GPU workload on the host. The shell requires an explicit confirmation and also refuses to continue if nvidia-smi reports any active compute PID:

cd /nfs/ywang29/GmNet/journal_exp
CONFIRM_IDLE=1 DRY_RUN=0 bash scripts/run_e4_e12_official.sh e12

Five independent Python processes run sequentially. Each measures CUDA FP32 and BF16 at batch 1/32, CPU FP32 at batch 1/32, ONNX validation, and ONNX Runtime CPU latency. Results and the five-process aggregate are written under /tmp/gmnet_runs/e12_official/. INT8 remains blocked until a validated whole-model calibration and quantization pipeline exists.

Estimated wall time: 15-30 minutes. Measurements are reportable only from an otherwise idle host.

After all phases finish, aggregation can be regenerated idempotently with:

cd /nfs/ywang29/GmNet/journal_exp
DRY_RUN=0 bash scripts/run_e4_e12_official.sh aggregate