File size: 10,992 Bytes
75e0942 | 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 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 | """
Preprocess real-world Franka data for Ctrl-World training.
- Videos: from /mnt/filesystem-g0/task_data/ (raw 672x376 MP4)
→ center-crop to 5:3 aspect (627x376) → resize to 320x192 → SVD VAE encode
- Actions: from /mnt/filesystem-g0/task_data_320_square_tracks/ (10D, gripper binarized)
→ convert rot6d to axis-angle → save 7D + 10D
Output: Ctrl-World dataset format under dataset_example/realworld/
Usage:
cd /mnt/filesystem-g0/Dual-Dynamics-Models/Ctrl-World
conda activate atm_ati_vdm
# Single GPU:
python scripts/preprocess_realworld.py \
--svd_path checkpoints/svd
# Multi-GPU:
accelerate launch scripts/preprocess_realworld.py \
--svd_path checkpoints/svd
"""
import argparse
import glob
import json
import os
import sys
import cv2
import h5py
import numpy as np
import torch
from tqdm import tqdm
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from models.action_conversion import actions_10d_to_7d
# Camera serial → view name
SERIAL_TO_VIEW = {
"14203948": "agentview",
"18204585": "eye_in_hand",
}
TASKS = ["task_1", "task_2", "task_3", "task_4"]
TASK_INSTRUCTIONS = {
"task_1": "put the pink noodle in the bowl",
"task_2": "put the bread in the bowl",
"task_3": "pour the pasta into the pan",
"task_4": "close the right cabinet door",
}
# Raw: 672x376. Center-crop width to 627 (5:3 ratio), then resize to 320x192.
RAW_W, RAW_H = 672, 376
TARGET_W, TARGET_H = 320, 192
# 5:3 crop: height stays 376, width = 376 * 5/3 = 626.67 → 627 (round up, will be resized anyway)
CROP_W = round(RAW_H * TARGET_W / TARGET_H) # 627
CROP_LEFT = (RAW_W - CROP_W) // 2 # 22
CROP_RIGHT = CROP_LEFT + CROP_W # 649
def read_mp4_frames(mp4_path):
"""Read all frames from MP4. Returns (T, H, W, 3) uint8 RGB."""
cap = cv2.VideoCapture(mp4_path)
frames = []
while True:
ret, frame = cap.read()
if not ret:
break
frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
cap.release()
return np.array(frames)
def crop_and_resize(frames):
"""672x376 → center-crop to 627x376 → resize to 320x192.
Input/Output: (T, H, W, 3) uint8 RGB.
"""
cropped = frames[:, :, CROP_LEFT:CROP_RIGHT, :] # (T, 376, 627, 3)
resized = np.zeros((len(cropped), TARGET_H, TARGET_W, 3), dtype=np.uint8)
for i in range(len(cropped)):
resized[i] = cv2.resize(cropped[i], (TARGET_W, TARGET_H), interpolation=cv2.INTER_CUBIC)
return resized
def convert_actions_7d(actions_10d):
"""Convert 10D → 7D using shared utility (models/action_conversion.py)."""
return actions_10d_to_7d(actions_10d)
def get_split_demos(tracks_dir):
"""Read train/val split from 320_square_tracks directory.
Returns {task: {demo_name: split}}.
"""
splits = {}
for task in TASKS:
splits[task] = {}
for split in ["train", "val"]:
split_dir = os.path.join(tracks_dir, task, split)
if not os.path.exists(split_dir):
continue
for f in glob.glob(os.path.join(split_dir, "*.hdf5")):
demo_name = os.path.splitext(os.path.basename(f))[0]
splits[task][demo_name] = split
return splits
def process_demo(
raw_dir, tracks_dir, output_dir, task, demo_name, split, vae, device
):
"""Process a single demo: videos from raw MP4, actions from tracks HDF5."""
demo_dir = os.path.join(raw_dir, task, demo_name)
tracks_path = os.path.join(tracks_dir, task, split, f"{demo_name}.hdf5")
episode_id = f"{task}_{demo_name}"
# Skip if already done
latent_check = os.path.join(output_dir, "latent_videos", split, episode_id, "0.pt")
if os.path.exists(latent_check):
return "skip"
# Load actions from tracks HDF5
with h5py.File(tracks_path, "r") as f:
actions_10d = np.array(f["root/actions"]) # (T, 10)
T = actions_10d.shape[0]
# Convert actions
actions_7d = convert_actions_7d(actions_10d)
# Separate fields for annotation (DROID-compatible format)
# cartesian_position: (T, 6) list of lists — [pos(3), axis_angle(3)]
# gripper_position: (T,) flat list of scalars — matches DROID format
cartesian_position = actions_7d[:, 0:6].tolist()
gripper_position = actions_7d[:, 6].tolist()
# Process videos from raw MP4
# DROID view order: 0=exterior_1 (third-person), 1=exterior_2, 2=wrist
# Our mapping: agentview→slot 0, zeros→slot 1, eye_in_hand→slot 2
VIEW_SLOT = {
"agentview": 0, # exterior_1 equivalent
"eye_in_hand": 2, # wrist equivalent
}
video_dir = os.path.join(output_dir, "videos", split, episode_id)
latent_dir = os.path.join(output_dir, "latent_videos", split, episode_id)
os.makedirs(video_dir, exist_ok=True)
os.makedirs(latent_dir, exist_ok=True)
view_latents = {}
for serial, view_name in SERIAL_TO_VIEW.items():
slot = VIEW_SLOT[view_name]
mp4_path = os.path.join(demo_dir, f"serial_{serial}_left.mp4")
if not os.path.exists(mp4_path):
return f"missing mp4: {mp4_path}"
frames = read_mp4_frames(mp4_path) # (T_raw, 376, 672, 3)
if len(frames) > T:
frames = frames[:T]
elif len(frames) < T:
return f"frame mismatch: raw={len(frames)}, actions={T}"
# Crop and resize to 320x192
frames_resized = crop_and_resize(frames) # (T, 192, 320, 3) uint8
# Save MP4
save_mp4(frames_resized, os.path.join(video_dir, f"{slot}.mp4"), fps=10)
# VAE encode
x = torch.from_numpy(frames_resized).float().permute(0, 3, 1, 2) / 255.0 * 2 - 1
x = x.to(device)
with torch.no_grad():
latents = []
for i in range(0, len(x), 32):
batch = x[i:i+32]
latent = vae.encode(batch).latent_dist.sample().mul_(vae.config.scaling_factor).cpu()
latents.append(latent)
latent_tensor = torch.cat(latents, dim=0) # (T, 4, 24, 40)
view_latents[slot] = latent_tensor
torch.save(latent_tensor, os.path.join(latent_dir, f"{slot}.pt"))
# Slot 1: zeros (no second exterior camera)
zero_latent = torch.zeros_like(list(view_latents.values())[0])
torch.save(zero_latent, os.path.join(latent_dir, "1.pt"))
black_frames = np.zeros((T, TARGET_H, TARGET_W, 3), dtype=np.uint8)
save_mp4(black_frames, os.path.join(video_dir, "1.mp4"), fps=10)
# Write annotation JSON
instruction = TASK_INSTRUCTIONS[task]
annotation = {
"texts": [instruction],
"episode_id": episode_id,
"video_length": T,
"videos": [
{"video_path": f"videos/{split}/{episode_id}/0.mp4"},
{"video_path": f"videos/{split}/{episode_id}/1.mp4"},
{"video_path": f"videos/{split}/{episode_id}/2.mp4"},
],
"latent_videos": [
{"latent_video_path": f"latent_videos/{split}/{episode_id}/0.pt"},
{"latent_video_path": f"latent_videos/{split}/{episode_id}/1.pt"},
{"latent_video_path": f"latent_videos/{split}/{episode_id}/2.pt"},
],
"states": actions_7d.tolist(),
"states_10d": actions_10d.tolist(),
"observation.state.cartesian_position": cartesian_position,
"observation.state.gripper_position": gripper_position,
}
ann_dir = os.path.join(output_dir, "annotation", split)
os.makedirs(ann_dir, exist_ok=True)
with open(os.path.join(ann_dir, f"{episode_id}.json"), "w") as f:
json.dump(annotation, f, indent=2)
return "ok"
def save_mp4(frames_rgb, path, fps=10):
"""Save (T, H, W, 3) uint8 RGB array as MP4."""
T, H, W, _ = frames_rgb.shape
fourcc = cv2.VideoWriter_fourcc(*"mp4v")
writer = cv2.VideoWriter(path, fourcc, fps, (W, H))
for i in range(T):
writer.write(cv2.cvtColor(frames_rgb[i], cv2.COLOR_RGB2BGR))
writer.release()
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--raw_dir", default="/mnt/filesystem-g0/task_data",
help="Raw data with MP4s")
parser.add_argument("--tracks_dir", default="/mnt/filesystem-g0/task_data_320_square_tracks",
help="320 square tracks HDF5 (for actions, split)")
parser.add_argument("--output_dir", default="dataset_example/realworld",
help="Output in Ctrl-World dataset format")
parser.add_argument("--svd_path", default="checkpoints/svd",
help="Path to SVD model (for VAE)")
args = parser.parse_args()
print(f"Raw videos: {args.raw_dir}")
print(f"Actions: {args.tracks_dir}")
print(f"Output: {args.output_dir}")
print(f"SVD VAE: {args.svd_path}")
# Load VAE
from diffusers.models import AutoencoderKLTemporalDecoder
try:
from accelerate import Accelerator
accelerator = Accelerator()
device = accelerator.device
is_main = accelerator.is_main_process
except Exception:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
is_main = True
vae = AutoencoderKLTemporalDecoder.from_pretrained(args.svd_path, subfolder="vae").to(device)
vae.eval()
vae.requires_grad_(False)
print(f"VAE loaded on {device}")
# Get splits
splits = get_split_demos(args.tracks_dir)
# Build work list
work = []
for task in TASKS:
for demo_name, split in sorted(splits[task].items()):
work.append((task, demo_name, split))
print(f"Total demos: {len(work)}")
# Process
ok, skip, err = 0, 0, 0
for task, demo_name, split in tqdm(work, desc="Processing", disable=not is_main):
result = process_demo(
args.raw_dir, args.tracks_dir, args.output_dir,
task, demo_name, split, vae, device
)
if result == "ok":
ok += 1
elif result == "skip":
skip += 1
else:
err += 1
if is_main:
print(f" ERROR {task}/{demo_name}: {result}")
if is_main:
print(f"\nDone: {ok} processed, {skip} skipped, {err} errors")
# Quick verification
sample_latent = glob.glob(os.path.join(args.output_dir, "latent_videos", "train", "*", "0.pt"))
if sample_latent:
t = torch.load(sorted(sample_latent)[0], map_location="cpu")
print(f"Sample latent shape: {t.shape}") # expect (T, 4, 24, 40)
sample_ann = glob.glob(os.path.join(args.output_dir, "annotation", "train", "*.json"))
if sample_ann:
with open(sorted(sample_ann)[0]) as f:
ann = json.load(f)
print(f"Sample annotation: episode={ann['episode_id']}, T={ann['video_length']}, "
f"7D action shape=({len(ann['states'])}, {len(ann['states'][0])})")
if __name__ == "__main__":
main()
|