---
tags:
  - smooth-hierarchy
  - canary
  - ucf101
  - video
size_categories:
  - n<1K
---

# smooth-hierarchy-ucf101 — Canary v1

Canary run artifacts for the `smooth-hierarchy-ucf101` experiment. This is a **pipeline-verification run**, not the main training run. It establishes that:

- FFmpeg 7.1 + torchcodec 0.11.0 + torch 2.11.0+cu130 work on NYU torch H200 nodes
- The 4-loss FiLM U-Net (L1 recon @ β=1, L2 collapse @ β=0, L3 abstraction-margin, L4 temporal-margin) trains end-to-end
- Checkpoints save every 100 train steps + every epoch
- W&B logs all 4 loss components and reconstruction panels at β ∈ {0, 0.1, 0.5, 0.9, 1.0}

## Provenance
- **Job**: torch:6876770 on gh116 (NVIDIA H200, 143 GB VRAM)
- **Elapsed**: 6m28s, exit 0
- **Global steps**: 400 (2 epochs × 200 `--limit_train_batches`)
- **W&B**: [canary-6876770](https://wandb.ai/aditijc5-new-york-university/smooth-hierarchy/runs/9yngjsbe)
- **Repo head**: `348c9cf`

## Files
| File | Purpose |
|------|---------|
| `last.ckpt` | Most recent checkpoint (47 MB) — rolling |
| `smooth-hierarchy-epoch=01-step=00000400.ckpt` | Final step-indexed checkpoint |
| `smooth-hierarchy-epoch=epoch=01.ckpt` | Epoch-1 end checkpoint |
| `canary.sbatch` | The sbatch script used for this run |
| `canary-6876770.out` / `canary-6876770.err` | Full job logs |
| `canary_metadata.parquet` | Run config + hyperparams for reproducibility |
| `uv.lock` | Exact Python dep resolution for bit-exact env reconstruction |

## Hyperparameters (canary)
- Dataset: UCF101 (101 classes, 13,320 videos) pulled from `quchenyuan/UCF101-ZIP` HF mirror (CRCV source was flaky)
- Image size: 64 × 64
- Batch size: 128
- Max epochs: 2
- `--limit_train_batches`: 200
- LR: 1e-4
- λ₁ = λ₂ = λ₃ = λ₄ = 1.0
- `--skip_beta_power`: 1.0
- `--recon_all_betas`: False
- `--frames_per_clip`: 32, `--step_between_clips`: 8, `--delta_t`: 2

## How to reproduce
1. Obtain the smooth-hierarchy training repo at head `348c9cf` from its source of record (private repo; see experiment owner).
2. On an HPC with an H100/H200/A100 GPU, run `setup.sbatch` (fetches UCF101 via HF mirror, builds venv via uv using the `uv.lock` in this dataset).
3. `sbatch canary.sbatch` with `WANDB_API_KEY` + `HF_TOKEN` available in `~/.raca/keys.env`.

Exact dependency set is pinned in `uv.lock` in this dataset.