File size: 2,823 Bytes
95b7c73 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 | # synth_captioning
Captioning pipeline for the synthetic PinPan datasets (`enrue1893/PinPan_Train_Syn`), extracted
from `JointGen-backup`. Perspective-crop rendering happens upstream (external simulator); this
bundle only covers the **captioning stage**: pulling pre-rendered panorama + 5 perspective clips
per sample and generating structured JSON captions with a locally-served Qwen VLM
(`Qwen/Qwen3.6-27B`).
## Files
```
data/dataset_streaming_template/
stream_process_pinpan_train_syn_captions_only.py # entry point — run this
stream_process.py # generic streaming template (reference)
README.md # streaming pattern docs
data/_5traj_4view_analysis/
pano_conditioning.py # provides equi2pers() (ERP -> pinhole)
qwen36_test/
caption_qwen36_8k_5traj.py # caption_9view_qwen() engine
caption_qwen36_pano_perspective_pair.py # low-level ERP/pinhole + Qwen helpers
vllm_serve_wrapper.py # `vllm serve` wrapper (GEMM kernel workaround)
README.md # vLLM server launch instructions
requirements/
qwen36_env.txt # pinned Python deps
```
Import chain: `stream_process_pinpan_train_syn_captions_only.py` -> `caption_qwen36_8k_5traj.py`
(`q5`) -> `caption_qwen36_pano_perspective_pair.py` -> `pano_conditioning.py` (`equi2pers`).
## Setup
1. `python -m venv venv && source venv/bin/activate && pip install -r requirements/qwen36_env.txt`
2. Serve the captioning model: see `qwen36_test/README.md`
(`python qwen36_test/vllm_serve_wrapper.py serve Qwen/Qwen3.6-27B --host 0.0.0.0 --port 8000 ...`).
3. **Fix the hardcoded conditioning path.** `caption_qwen36_pano_perspective_pair.py` defines:
```python
DEFAULT_PANO_CONDITIONING_PATH = (
"/home/work/.local/khs/Projects/PanoWorld/cosmos_predict2/_src/predict2/utils/pano_conditioning.py"
)
```
That absolute path won't exist on a new machine. Point it at the copy included in this bundle
(`data/_5traj_4view_analysis/pano_conditioning.py`), e.g. edit the constant or symlink it into
place before running.
4. Put an HF **write** token at `data/dataset_streaming_template/.hftoken` (gitignored, not
included here) — used to upload results to `OUTPUT_REPO`. Override the input/output repos,
`QWEN36_TEST_DIR`, `VLLM_BASE_URLS`, `VLLM_MODEL`, etc. via env vars (see the top of
`stream_process_pinpan_train_syn_captions_only.py`).
5. Run:
```bash
cd data/dataset_streaming_template
python stream_process_pinpan_train_syn_captions_only.py
```
Source: `JointGen-backup` (khs2001d), `data/dataset_streaming_template/` and `qwen36_test/`.
|