# H3D One-Sample Overfit Tmux Run

Started from Cursor on GPU 2 and GPU 3.

## Sample

- `edit_id`: `del_0008dc75fb3648f2af4ca8c4d711e53e_000`
- `obj_id`: `0008dc75fb3648f2af4ca8c4d711e53e`
- `edit_type`: `deletion`
- prompt: `Remove the detailed hand views from the character model.`

## Hyperparameters

- `max_samples`: `1`
- `max_steps`: `5000`
- `batch_size_per_gpu`: `1`
- `batch_split`: `1`
- `lr`: `0.0001`
- `i_print`: `1`
- `i_log`: `1`
- `i_save`: `100`
- `no_sample_images`: `true`
- trainable params: `--train_only_editing_weights`

## Configs

- Stage 1: `configs/editing/ss_flow_img_dit_L_16l8_fp16_h3d_overfit1.json`
- Stage 2: `configs/editing/slat_flow_img_dit_L_64l8p2_fp16_h3d_overfit1.json`

## Commands

Stage 1 on GPU 2:

```bash
CUDA_VISIBLE_DEVICES=2 PYTHONUNBUFFERED=1 torchrun --nproc_per_node=1 --master_port=12381 train_torchrun.py \
  --config configs/editing/ss_flow_img_dit_L_16l8_fp16_h3d_overfit1.json \
  --output_dir ./work_dirs/Editing_Training/h3d_overfit1_img_to_voxel_5k \
  --train_only_editing_weights
```

Stage 2 on GPU 3:

```bash
CUDA_VISIBLE_DEVICES=3 PYTHONUNBUFFERED=1 torchrun --nproc_per_node=1 --master_port=12382 train_torchrun.py \
  --config configs/editing/slat_flow_img_dit_L_64l8p2_fp16_h3d_overfit1.json \
  --output_dir ./work_dirs/Editing_Training/h3d_overfit1_voxel_to_texture_5k \
  --train_only_editing_weights
```

## Logs

- Stage 1: `logs/h3d_overfit1_s1_gpu2.log`
- Stage 2: `logs/h3d_overfit1_s2_gpu3.log`

## Tmux

- session: `h3d_overfit1`
- windows: `s1_gpu2`, `s2_gpu3`