openpi / droid /scripts /REPROCESSING_INSTRUCTIONS.md
zhicao's picture
Upload folder using huggingface_hub
b584148 verified
|
Raw
History Blame Contribute Delete
3.55 kB

Reprocess 4,976 DROID Episodes Across 5 Machines

Quick Start

Run this command on each of the 5 machines:

cd /root/workspace/code/wmrl/Dual-Dynamics-Models/DROID-main
bash scripts/reprocess_all_5machines.sh <MACHINE_ID>

Where <MACHINE_ID> is:

  • Machine 1: 0
  • Machine 2: 1
  • Machine 3: 2
  • Machine 4: 3
  • Machine 5: 4

Episode Distribution

Each machine processes a subset of the 4,976 episodes:

Machine Episodes Count
0 0 → 995 995
1 995 → 1990 995
2 1990 → 2985 995
3 2985 → 3980 995
4 3980 → 4976 996

Each machine distributes its episodes across 8 GPUs (~124 episodes per GPU).

Monitoring Progress

Check logs

# Watch GPU 0 log in real-time
tail -f /mnt/kevin/data/droid_processed_1000pts/reprocessing_logs/m0_gpu0.log

# Check all GPU logs on machine 0
ls -lh /mnt/kevin/data/droid_processed_1000pts/reprocessing_logs/m0_*.log

Check completion

# Count completed GPUs on machine 0
grep 'Complete:' /mnt/kevin/data/droid_processed_1000pts/reprocessing_logs/m0_*.log | wc -l
# Should output: 8 (when all GPUs done)

# Check success/error counts
grep -h 'Success:' /mnt/kevin/data/droid_processed_1000pts/reprocessing_logs/m0_*.log

Monitor GPU usage

watch -n 1 nvidia-smi
# Should see all 8 GPUs at 70-90% utilization

Verify Results

After all machines complete, verify track counts:

cd /root/workspace/code/wmrl/Dual-Dynamics-Models/DROID-main

/mnt/kevin/envs/miniconda3/envs/atm_ati_vdm_droid/bin/python -c "
import numpy as np
from pathlib import Path
import random

data_dir = Path('/mnt/kevin/data/droid_processed_1000pts/data')
npz_files = sorted(data_dir.glob('episode_*.npz'))

# Sample 50 random episodes
samples = random.sample(npz_files, 50)

correct = 0
wrong = 0

for npz_path in samples:
    data = np.load(npz_path, allow_pickle=True)
    wrist_count = data['tracks_wrist'].shape[1]

    if wrist_count == 1105:
        correct += 1
    else:
        wrong += 1
        print(f'{npz_path.name}: wrist={wrist_count}')

print(f'\nResults from 50 random samples:')
print(f'  Correct (1105): {correct}')
print(f'  Wrong: {wrong}')
"

Expected output: All 50 samples should have 1105 wrist points.

Troubleshooting

Stop all reprocessing

pkill -f 'python.*episode'

Restart a specific machine

# Stop
pkill -f 'python.*episode'

# Wait a few seconds
sleep 5

# Restart
bash scripts/reprocess_all_5machines.sh <MACHINE_ID>

Check for stuck processes

ps aux | grep python | grep episode

Estimated Time

  • Per episode: ~30-60 seconds (depending on length)
  • Per GPU: ~124 episodes × 45 sec avg = ~93 minutes
  • Total time: ~1.5-2 hours (all machines in parallel)

What Gets Updated

Each .npz file will be updated with:

  • tracks_wrist: 1112 → 1105 points (7 mesh + 98 grid + 1000 random)
  • tracks_exterior: Already correct (1105 points)
  • mesh_vertices_2d_wrist_fixed: Already exists
  • ✓ All other fields: Unchanged

Features

  • GPU distribution: All 40 GPUs (5 machines × 8 GPUs) utilized
  • OOM handling: Automatic retry with batching (batch_size=600)
  • Correct track counts: 1105 points for both views
  • Progress tracking: Individual log files per GPU
  • Resumable: Skips already-processed episodes (checks track count)