fridge_m_v2: gen SETUP.md
Browse files- fridge_m_v2/GEN_SETUP.md +89 -0
fridge_m_v2/GEN_SETUP.md
ADDED
|
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# fridge_m_v2 video generation — setup record
|
| 2 |
+
|
| 3 |
+
Auto-launched 2026-05-14 by the same agent that ran sim datagen. The
|
| 4 |
+
gen phase starts per-node automatically once that node's sim
|
| 5 |
+
chunked_runner_v2 exits (no concurrent sim+gen — that's the
|
| 6 |
+
lock-contention bug from `datagen_pipeline_fridge_m.md` gotcha #6).
|
| 7 |
+
|
| 8 |
+
## Pipeline
|
| 9 |
+
|
| 10 |
+
Per sim chunk:
|
| 11 |
+
1. `depth_methods.py` (Sim2RealGen_Video_Model/scripts) generates depth
|
| 12 |
+
+ robot-mask mp4s for the chunk's h5 trajs.
|
| 13 |
+
2. `inference_long_video.py` (Sim2RealGen_Video_Model/sim2realgen)
|
| 14 |
+
conditions Wan2.1-Fun-V1.1-1.3B + LoRA on the depth + mask + prompt
|
| 15 |
+
for ext and wrist views.
|
| 16 |
+
3. `push_gen_chunk.py` uploads the chunk dir to HF as one commit
|
| 17 |
+
(mp4s + prompts/ + _logs/, but not _depth/_inf_tmp/_shard).
|
| 18 |
+
|
| 19 |
+
Frozen scripts: `/scr/ravenh/fridge_m_v2/scripts/` on each node.
|
| 20 |
+
|
| 21 |
+
## Depth settings (per user spec)
|
| 22 |
+
|
| 23 |
+
```
|
| 24 |
+
--use_raw_pct
|
| 25 |
+
--raw_pct 1,99
|
| 26 |
+
--inv_clip 1,20
|
| 27 |
+
--inv_map log
|
| 28 |
+
--depth_size 416,240
|
| 29 |
+
```
|
| 30 |
+
|
| 31 |
+
## Inference settings
|
| 32 |
+
|
| 33 |
+
```
|
| 34 |
+
LoRA: lora_ckpts/Wan2.1-Fun-V1.1-1.3B-DualControl-3View-PrevFrame_minimal_lora/step-4740.safetensors
|
| 35 |
+
height: 240
|
| 36 |
+
width: 416
|
| 37 |
+
clip_length: 81
|
| 38 |
+
num_inference_steps: 30
|
| 39 |
+
cfg_scale: 5.0
|
| 40 |
+
seed: 1000 + seq_idx * 13 (seq_offset varies per chunk)
|
| 41 |
+
```
|
| 42 |
+
|
| 43 |
+
## Prompts
|
| 44 |
+
|
| 45 |
+
`/scr/ravenh/fridge_m_v2/scripts/prompts.json` is the source of truth.
|
| 46 |
+
Every prompt is `"Video of a real {scene}. {fixed_prefix}"`.
|
| 47 |
+
|
| 48 |
+
Fixed prefix is a strict scene constraint: **"A Franka Panda robot arm
|
| 49 |
+
is mounted on a black mobile platform in front of a stainless steel
|
| 50 |
+
single door refrigerator with a stainless steel handle.…"**
|
| 51 |
+
|
| 52 |
+
The 10 scenes vary kitchen background but all explicitly include
|
| 53 |
+
*"stainless steel single door fridge with a stainless steel handle"*
|
| 54 |
+
plus wall colors (gray/white) and floor (light yellow/white/gray) per
|
| 55 |
+
the original `datagen_pipeline_fridge_m.md` spec.
|
| 56 |
+
|
| 57 |
+
## HF target
|
| 58 |
+
|
| 59 |
+
```
|
| 60 |
+
Ravenh97/generated_video:fridge_m_v2/
|
| 61 |
+
GEN_SETUP.md
|
| 62 |
+
gen_chunks/
|
| 63 |
+
chunk_001/ (viscam6)
|
| 64 |
+
chunk_002/
|
| 65 |
+
...
|
| 66 |
+
chunk_101/ (viscam7)
|
| 67 |
+
...
|
| 68 |
+
chunk_201/ (viscam3 slurm job 15387074, 4 GPUs, ROOT=/scr/ravenh/fridge_m_v2)
|
| 69 |
+
...
|
| 70 |
+
chunk_301/ (viscam3 slurm job 15438619, 3 GPUs, ROOT=/scr/ravenh/fridge_m_v2_3gpu)
|
| 71 |
+
...
|
| 72 |
+
```
|
| 73 |
+
|
| 74 |
+
One commit per chunk (mp4s + prompts + per-traj log).
|
| 75 |
+
|
| 76 |
+
## Wall estimate
|
| 77 |
+
|
| 78 |
+
Per memory `agent_3drawers_gen_viscam7`, A5000 ≈ ~3 h/chunk for Wan2.1
|
| 79 |
+
1.3B at 240x416x81. L40S is ~6× faster.
|
| 80 |
+
|
| 81 |
+
- viscam6 (4× A5000), ~8-10 chunks: ~24-30 h
|
| 82 |
+
- viscam7 (4× A5000), ~8-10 chunks: ~24-30 h
|
| 83 |
+
- viscam3 (4× 3090), ~6-7 chunks: 3090 ≈ A5000 speed for inference; ~18-21 h
|
| 84 |
+
- viscam3 (3× 3090), ~7 chunks: 3 GPUs → ~21 h
|
| 85 |
+
|
| 86 |
+
All four run in parallel; bottleneck ≈ 24-30 h from gen start.
|
| 87 |
+
|
| 88 |
+
Each chunk is pushed atomically when it completes, so HF builds up
|
| 89 |
+
incrementally — the dataset is usable as soon as the first chunk lands.
|