mcleanie/UCPE-useful-checkpoints

UCPE Useful Checkpoints

Public backup for selected UCPE checkpoints from:

/mnt/dataset/zt8kgo/ucpe_ckpt/logs/

This repo intentionally does not mirror the whole logs directory. The source tree was about 1.3T and contained many large prediction and optimizer artifacts. The files here are the small set most useful for continuing the PanShot nodistort and soft camera RoPE work after remote shutdown.

Related Code

Clone the codebase:

git clone https://github.com/jinhong-ni/CCVG.git UCPE
cd UCPE
git checkout soft-camera-rope-f1

Primary recovery instructions are also copied in:

context/REMOTE_SHUTDOWN_RECOVERY.md
context/fix_base_rope.md

Dataset

Use the non-distorted PanShot dataset for all future experiments:

https://huggingface.co/datasets/mcleanie/PanShot-nodistort

On the source machine the dataset path was:

/mnt/dataset/zt8kgo/ucpe_ckpt/data/UCPE/PanShot-nodistort

Uploaded Checkpoints

HF path	Source run	Why kept
checkpoints/muc9h5wj/epoch=9-step=10000.ckpt	muc9h5wj	PanShot-nodistort, soft_camera_rope: true, replace_soft_camera, alpha: 0.05, 10k steps.
checkpoints/16zjeilr/epoch=9-step=10000.ckpt	16zjeilr	Soft camera RoPE gated blend run, gated_blend_soft_camera, alpha: 0.0, 10k steps. Config notes RealEstate10k data path, so check before reuse.
checkpoints/bv3dnw7a/epoch=9-step=10000.ckpt	bv3dnw7a	Soft camera RoPE residual output run, residual_camera_rope_output, 10k steps.
checkpoints/wan22ucpe-20260526-005237/epoch=0-step=1000.ckpt	wan22ucpe-20260526-005237	Wan UCPE baseline-ish run using Wan-AI/Wan2.1-T2V-1.3B, relray_absmap, 1k steps.
checkpoints/soft-camera-rope-replace/20260603_143149_hard_replace_fast/epoch=0-step=100/mp_rank_00_model_states.pt	soft-camera-rope-replace/20260603_143149_hard_replace_fast	DeepSpeed model-state-only artifact for the hard-replace fast soft-camera experiment. Optimizer shards omitted to avoid uploading training-state bulk.

Each run also includes its config.yaml when available. wan22ucpe-20260526-005237 also includes train.log.

Omitted Files

last.ckpt siblings were not uploaded for the selected Lightning runs because they are the same training endpoint as the explicit epoch=... checkpoint and would cost another 3G each. Prediction folders, W&B binary logs, and DeepSpeed optimizer shards were also omitted.

Restore Notes

The Lightning checkpoints were saved with PyTorch Lightning 2.6.1 in the UCPE conda environment. Restore the environment from the code repo files:

conda create -n UCPE --file environment-ucpe-explicit.txt
conda activate UCPE
python -m pip install -r requirements-ucpe-pip-freeze.txt
python -m pip install -e .

Then point configs at the restored nodistort dataset path. If a checkpoint path is hard-coded in a config, update it to the local downloaded file.