| |
| """Upload all reproduction artifacts to Hugging Face. |
| |
| Model repos (MediumCNN, eta=0.4 lineage): |
| t=0 teacher -> evalstate/cifar10-selftrain-teacher |
| t=2 best iterate -> evalstate/cifar10-selftrain-best-iterate |
| t=9 degraded iterate -> evalstate/cifar10-selftrain-degraded-iterate |
| Dataset repo (all linear + CIFAR artifacts): |
| -> evalstate/denoising-forgetting-linear-repro |
| """ |
| import os, json, glob |
| from huggingface_hub import HfApi |
|
|
| BASE = "/workspace/scratch/research" |
| MDL = f"{BASE}/models" |
| TRAJ = f"{BASE}/trajectories" |
| CFG = f"{BASE}/configs" |
| FIG = f"{BASE}/figures" |
| api = HfApi() |
|
|
| with open(f"{TRAJ}/cifar_trajectory.json") as f: |
| traj = json.load(f) |
| records = traj["records"] |
| primary = [r for r in records if abs(r["eta"] - 0.4) < 1e-9] |
| K = max(r["t"] for r in primary) |
| tstar = min(primary, key=lambda r: r["test_error"])["t"] |
| print(f"primary eta=0.4: K={K}, t*={tstar}") |
|
|
| repos = { |
| "teacher": ("evalstate/cifar10-selftrain-teacher", 0), |
| "best": ("evalstate/cifar10-selftrain-best-iterate", tstar), |
| "degraded": ("evalstate/cifar10-selftrain-degraded-iterate", K), |
| } |
|
|
| cards = { |
| "teacher": """--- |
| license: apache-2.0 |
| tags: [self-training, self-distillation, cifar-10, denoising-forgetting] |
| library_name: pytorch |
| --- |
| # CIFAR-10 Self-Training Teacher (iterate t=0) |
| |
| Small CNN (MediumCNN, ~1.56M params; **not** ResNet-50) trained as the **initial teacher** |
| in a self-distillation trajectory reproducing the denoising-vs-forgetting trade-off from |
| Wu, Yang & Sun, *Why Self-Training Helps and Hurts* (arXiv:2602.14029). |
| |
| ## Training config |
| - Dataset: CIFAR-10 (via HF `uoft-cs/cifar10`), subset n=5000 |
| - Label noise: symmetric, fraction eta=0.4 corrupted (injects initial stochastic error) |
| - Optimizer: SGD (lr=0.1, momentum=0.9, wd=5e-4, nesterov) + cosine, 50 epochs, batch 256 |
| - Augmentation: random crop (pad4, reflect) + horizontal flip + normalize |
| |
| ## Role |
| This is the t=0 iterate (the weak teacher). Students t>=1 are trained from scratch on |
| hard pseudo-labels from the previous iterate on disjoint subsets. The trajectory exhibits |
| a U-shaped test-error curve (denoise then forget). See the report for full details. |
| """, |
| "best": """--- |
| license: apache-2.0 |
| tags: [self-training, self-distillation, cifar-10, denoising-forgetting] |
| library_name: pytorch |
| --- |
| # CIFAR-10 Self-Training Best Iterate (generation t={t}) |
| |
| MediumCNN (~1.56M params): the **best-performing self-trained iterate** in a self-distillation |
| trajectory reproducing the denoising-vs-forgetting trade-off (Wu, Yang & Sun, arXiv:2602.14029). |
| |
| This iterate is the minimum of the U-shaped test-error curve (the denoising-dominant optimum |
| before signal forgetting dominates). Trained from scratch on hard pseudo-labels from iterate |
| {tprev} on a disjoint CIFAR-10 subset (n=5000, eta=0.4 lineage). |
| |
| ## Metrics (test set) |
| - test accuracy: {acc:.2f}% (teacher t=0: {acc0:.2f}%) |
| - test error: {err:.2f}% |
| - mean top-1 confidence: {conf:.3f} |
| """, |
| "degraded": """--- |
| license: apache-2.0 |
| tags: [self-training, self-distillation, cifar-10, denoising-forgetting] |
| library_name: pytorch |
| --- |
| # CIFAR-10 Self-Training Final Degraded Iterate (generation t={t}) |
| |
| MediumCNN (~1.56M params): the **final degraded iterate** (last generation t=K={t}) in a |
| self-distillation trajectory reproducing the denoising-vs-forgetting trade-off |
| (Wu, Yang & Sun, arXiv:2602.14029). |
| |
| Illustrates the signal-forgetting regime: after the optimal stopping point (t*={tstar}), |
| continued self-training on pseudo-labels progressively erodes signal and degrades test |
| accuracy, while confidence keeps rising (miscalibration). |
| |
| ## Metrics (test set) |
| - test accuracy: {acc:.2f}% (best t={tstar}: {accbest:.2f}%) |
| - test error: {err:.2f}% |
| - mean top-1 confidence: {conf:.3f} |
| """, |
| } |
|
|
| rec0 = next(r for r in primary if r["t"] == 0) |
| recbest = next(r for r in primary if r["t"] == tstar) |
| for role, (repo, t) in repos.items(): |
| src = f"{MDL}/iter_medium_eta0p4_t{t}.pth" |
| assert os.path.exists(src), f"missing {src}" |
| rec = next(r for r in primary if r["t"] == t) |
| cfg = {"role": role, "iterate": t, "eta": rec["eta"], "n": rec["n"], |
| "K": rec["K"], "epochs": rec["epochs"], "arch": rec["arch"], |
| "seed": rec["seed"], "label_source": rec["label_source"], |
| "test_acc": rec["test_acc"], "test_error": rec["test_error"], |
| "mean_conf": rec["mean_conf"], |
| "agreement_with_prev": rec["agreement_with_prev"], |
| "architecture": "MediumCNN(4 conv blocks 64/128/256/512 + GAP, ~1.56M params)", |
| "is_best_iterate": role == "best", "is_final_iterate": role == "degraded", |
| "lineage": "eta=0.4 weak-teacher trajectory", |
| "best_iterate_t": tstar} |
| with open(f"/tmp/{role}_config.json", "w") as f: |
| json.dump(cfg, f, indent=2) |
| card_text = cards[role].format( |
| t=t, tprev=t - 1, tstar=tstar, acc=rec["test_acc"], err=rec["test_error"], |
| conf=rec["mean_conf"], acc0=rec0["test_acc"], accbest=recbest["test_acc"]) |
| with open(f"/tmp/{role}_card.md", "w") as f: |
| f.write(card_text) |
| api.upload_file(path_or_fileobj=src, path_in_repo="model.pth", repo_id=repo, repo_type="model") |
| api.upload_file(path_or_fileobj=f"/tmp/{role}_config.json", path_in_repo="config.json", |
| repo_id=repo, repo_type="model") |
| api.upload_file(path_or_fileobj=f"/tmp/{role}_card.md", path_in_repo="README.md", |
| repo_id=repo, repo_type="model") |
| print(f" uploaded {role} (t={t}) -> {repo} acc={rec['test_acc']:.2f}%") |
|
|
| |
| DS = "evalstate/denoising-forgetting-linear-repro" |
| upload_files = [] |
| upload_files += [(p, p.replace(BASE + "/", "")) for p in glob.glob(f"{TRAJ}/*.csv")] |
| upload_files += [(p, p.replace(BASE + "/", "")) for p in glob.glob(f"{TRAJ}/*.json")] |
| upload_files += [(p, p.replace(BASE + "/", "")) for p in glob.glob(f"{CFG}/*.json")] |
| upload_files += [(p, p.replace(BASE + "/", "")) for p in glob.glob(f"{BASE}/*.py")] |
| upload_files += [(p, p.replace(BASE + "/", "")) for p in glob.glob(f"{FIG}/*.png")] |
| upload_files += [(f"{BASE}/part1_log.txt", "part1_log.txt")] |
| upload_files += [(f"{BASE}/part1_summary.txt", "part1_summary.txt")] |
| upload_files += [(f"{TRAJ}/cifar_trajectory_ep15.csv", "trajectories/cifar_trajectory_ep15.csv")] |
| upload_files += [(f"{TRAJ}/cifar_trajectory_ep15.json", "trajectories/cifar_trajectory_ep15.json")] |
| upload_files += [(f"{BASE}/cifar_ep15_log.txt", "cifar_ep15_log.txt")] |
| for fpath, arc in upload_files: |
| if os.path.exists(fpath): |
| api.upload_file(path_or_fileobj=fpath, path_in_repo=arc, repo_id=DS, repo_type="dataset") |
| print(f"uploaded {len(upload_files)} files to dataset {DS}") |
| print("UPLOAD_DONE") |
|
|
