mirror of
https://github.com/huggingface/lerobot.git
synced 2026-05-25 13:40:00 +00:00
Jade Choghari
51 lines
1.6 KiB
Bash
51 lines
1.6 KiB
Bash
#!/bin/bash
|
|
|
|
# Example script to run synthetic data generation with Qwen VLM
|
|
# This generates user prompts and robot utterances for hierarchical policy training
|
|
|
|
# Configuration
|
|
REPO_ID="lerobot/libero_video"
|
|
MODEL="Qwen/Qwen3-VL-30B-A3B-Instruct"
|
|
# or: MODEL="Qwen/Qwen2-VL-7B-Instruct"
|
|
|
|
|
|
OUTPUT_DIR="/fsx/jade_choghari/outputs/libero-annotate"
|
|
|
|
BATCH_SIZE=16
|
|
TEMPERATURE=0.9
|
|
SAMPLE_INTERVAL=5.0 # generate dialogue every 1 second (all episodes processed)
|
|
|
|
# Run subtask annotation
|
|
python /admin/home/jade_choghari/lerobot/src/lerobot/policies/pi05_full/annotate/subtask_annotate.py \
|
|
--repo-id "$REPO_ID" \
|
|
--video-key observation.images.image \
|
|
--output-dir "$OUTPUT_DIR" \
|
|
--output-repo-id "jadechoghari/libero-annotate" \
|
|
--batch-size "$BATCH_SIZE" \
|
|
# run synthetic data generation (all episodes processed)
|
|
# python examples/dataset/annotate_pgen.py \
|
|
# --repo-id "$REPO_ID" \
|
|
# --model "$MODEL" \
|
|
# --output-dir "$OUTPUT_DIR" \
|
|
# --temperature "$TEMPERATURE" \
|
|
# --batch-size "$BATCH_SIZE" \
|
|
# --sample-interval "$SAMPLE_INTERVAL" \
|
|
# --image-key observation.images.base \
|
|
# --num-image-views-per-sample 1
|
|
|
|
# for faster testing, increase sample interval:
|
|
# --sample-interval 5.0 # Samples every 5 seconds (much faster)
|
|
|
|
# to push to hub after generation:
|
|
# add --push-to-hub flag
|
|
|
|
# efficient batch processing: 4 episodes at once
|
|
# python examples/dataset/annotate_pgen.py \
|
|
# --repo-id "$REPO_ID" \
|
|
# --model "$MODEL" \
|
|
# --output-dir "$OUTPUT_DIR" \
|
|
# --video-mode \
|
|
# --video-key observation.images.up \
|
|
# --video-batch-size "$BATCH_SIZE" \
|
|
# --sample-interval 1.0
|