File size: 19,009 Bytes
991941e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 | #!/usr/bin/env python3
"""
RICL Training Verification Script
Run this BEFORE training to ensure everything is configured correctly.
Usage:
cd /projects/extern/kisski/kisski-spath/dir.project/VLA_Groot/in_context_learning/ricl_openpi
python scripts/verify_ricl_training_setup.py
"""
import os
import sys
import json
import numpy as np
from pathlib import Path
# ============================================================
# CONFIGURATION
# ============================================================
WORK_DIR = "/projects/extern/kisski/kisski-spath/dir.project/VLA_Groot/in_context_learning/ricl_openpi"
DATA_DIR = "/projects/extern/kisski/kisski-spath/dir.project/VLA_Groot/merged_libero_mask_depth_noops_lerobot_10"
CONTEXT_DIR = os.path.join(WORK_DIR, "rag/ricl_training_context_libero_10_test")
CHECKPOINT_PATH = os.path.join(WORK_DIR, "pi0_fast_base_params")
ASSETS_DIR = os.path.join(WORK_DIR, "assets")
sys.path.insert(0, WORK_DIR)
passed = 0
failed = 0
def check_pass(msg):
global passed
passed += 1
print(f" ✓ {msg}")
def check_fail(msg):
global failed
failed += 1
print(f" ❌ {msg}")
# ============================================================
# CHECK 1: Verify all paths exist
# ============================================================
print("=" * 60)
print("CHECK 1: Verify all paths exist")
print("=" * 60)
paths_to_check = {
"WORK_DIR": WORK_DIR,
"DATA_DIR": DATA_DIR,
"CONTEXT_DIR": CONTEXT_DIR,
"CHECKPOINT_PATH (pi0_fast_base_params)": CHECKPOINT_PATH,
"ASSETS_DIR": ASSETS_DIR,
}
for name, path in paths_to_check.items():
if os.path.exists(path):
check_pass(f"{name}: {path}")
else:
check_fail(f"{name} MISSING: {path}")
print()
# ============================================================
# CHECK 2: Verify JAX/Orbax checkpoint structure
# ============================================================
print("=" * 60)
print("CHECK 2: Verify base model checkpoint (JAX/Orbax)")
print("=" * 60)
required_ckpt_files = ["_METADATA", "_sharding", "manifest.ocdbt"]
for f in required_ckpt_files:
fpath = os.path.join(CHECKPOINT_PATH, f)
if os.path.exists(fpath):
size = os.path.getsize(fpath)
check_pass(f"{f}: {size} bytes")
else:
check_fail(f"{f} MISSING")
ocdbt_dir = os.path.join(CHECKPOINT_PATH, "ocdbt.process_0")
if os.path.exists(ocdbt_dir):
num_files = len(os.listdir(ocdbt_dir))
check_pass(f"ocdbt.process_0/: {num_files} shard files")
else:
check_fail("ocdbt.process_0/ directory MISSING")
print()
# ============================================================
# CHECK 3: Verify RICL context directory
# ============================================================
print("=" * 60)
print("CHECK 3: Verify RICL retrieval context")
print("=" * 60)
required_context_files = {
"nn_indices.npy": "Nearest neighbor indices",
"nn_distances.npy": "Nearest neighbor distances",
"actions.npy": "Action chunks",
"states.npy": "State vectors",
"metadata.json": "Frame metadata",
"embeddings.npy": "Visual embeddings",
"index.faiss": "FAISS index",
}
for fname, desc in required_context_files.items():
fpath = os.path.join(CONTEXT_DIR, fname)
if os.path.exists(fpath):
size_mb = os.path.getsize(fpath) / 1e6
check_pass(f"{fname} ({desc}): {size_mb:.1f} MB")
else:
check_fail(f"{fname} ({desc}) MISSING")
print()
# ============================================================
# CHECK 4: Load and validate context data shapes
# ============================================================
print("=" * 60)
print("CHECK 4: Validate context data shapes and contents")
print("=" * 60)
try:
nn_indices = np.load(os.path.join(CONTEXT_DIR, "nn_indices.npy"))
nn_distances = np.load(os.path.join(CONTEXT_DIR, "nn_distances.npy"))
actions = np.load(os.path.join(CONTEXT_DIR, "actions.npy"))
states = np.load(os.path.join(CONTEXT_DIR, "states.npy"))
with open(os.path.join(CONTEXT_DIR, "metadata.json"), "r") as f:
metadata = json.load(f)
num_frames = len(metadata)
print(f" Total frames: {num_frames}")
print(f" nn_indices shape: {nn_indices.shape}")
print(f" nn_distances shape: {nn_distances.shape}")
print(f" actions shape: {actions.shape}")
print(f" states shape: {states.shape}")
# Validate shapes match
if nn_indices.shape[0] == num_frames:
check_pass(f"nn_indices rows ({nn_indices.shape[0]}) match metadata ({num_frames})")
else:
check_fail(f"nn_indices rows ({nn_indices.shape[0]}) != metadata ({num_frames})")
if nn_distances.shape[0] == num_frames:
check_pass(f"nn_distances rows ({nn_distances.shape[0]}) match metadata ({num_frames})")
else:
check_fail(f"nn_distances rows ({nn_distances.shape[0]}) != metadata ({num_frames})")
if actions.shape[0] == num_frames:
check_pass(f"actions rows ({actions.shape[0]}) match metadata ({num_frames})")
else:
check_fail(f"actions rows ({actions.shape[0]}) != metadata ({num_frames})")
# Check action dimensions
action_dim = actions.shape[-1] if len(actions.shape) > 1 else 0
print(f" Action dimension: {action_dim}")
if action_dim == 7:
check_pass(f"Action dimension is 7 (LIBERO standard)")
else:
check_fail(f"Action dimension is {action_dim}, expected 7")
# Action horizon
if len(actions.shape) == 3:
action_horizon = actions.shape[1]
print(f" Action horizon: {action_horizon}")
# Check nn_indices are valid (no out-of-range)
max_idx = nn_indices.max()
if max_idx < num_frames:
check_pass(f"nn_indices max ({max_idx}) < num_frames ({num_frames})")
else:
check_fail(f"nn_indices max ({max_idx}) >= num_frames ({num_frames}) - OUT OF RANGE!")
# Check no self-retrieval (query != demo)
if nn_indices.shape[1] >= 1:
self_retrieval_count = np.sum(nn_indices[:, 0] == np.arange(num_frames))
self_pct = 100 * self_retrieval_count / num_frames
if self_pct < 5:
check_pass(f"Self-retrieval rate: {self_pct:.1f}% (low, good)")
else:
check_fail(f"Self-retrieval rate: {self_pct:.1f}% (HIGH - might be a bug!)")
# Distance statistics
top1_dist = nn_distances[:, 0]
print(f"\n Top-1 distance statistics:")
print(f" Min: {top1_dist.min():.4f}")
print(f" Max: {top1_dist.max():.4f}")
print(f" Mean: {top1_dist.mean():.4f}")
print(f" Median: {np.median(top1_dist):.4f}")
print(f" Std: {top1_dist.std():.4f}")
if top1_dist.mean() > 0:
check_pass(f"Distances are non-zero (mean={top1_dist.mean():.4f})")
else:
check_fail(f"Distances are all zero - retrieval might be broken!")
except Exception as e:
check_fail(f"Error loading context: {e}")
import traceback
traceback.print_exc()
print()
# ============================================================
# CHECK 5: Verify dataset structure
# ============================================================
print("=" * 60)
print("CHECK 5: Verify dataset structure")
print("=" * 60)
meta_path = os.path.join(DATA_DIR, "meta")
if os.path.exists(meta_path):
info_path = os.path.join(meta_path, "info.json")
if os.path.exists(info_path):
with open(info_path, "r") as f:
info = json.load(f)
check_pass(f"Dataset info found")
print(f" Total episodes: {info.get('total_episodes', 'N/A')}")
print(f" Total frames: {info.get('total_frames', 'N/A')}")
features = info.get("features", {})
image_keys = [k for k in features if "image" in k.lower() and "mask" not in k.lower() and "depth" not in k.lower()]
print(f" RGB image keys: {image_keys}")
# Check video files exist
video_dir = os.path.join(DATA_DIR, "videos", "chunk-000")
if os.path.exists(video_dir):
check_pass(f"Video directory exists: {video_dir}")
for img_key in ["observation.images.image", "observation.images.wrist_image"]:
key_dir = os.path.join(video_dir, img_key)
if os.path.exists(key_dir):
num_vids = len([f for f in os.listdir(key_dir) if f.endswith(".mp4")])
check_pass(f"{img_key}: {num_vids} video files")
else:
check_fail(f"{img_key} video directory MISSING")
else:
check_fail(f"Video directory MISSING: {video_dir}")
else:
check_fail(f"info.json MISSING")
else:
check_fail(f"meta directory MISSING")
# Check tasks.jsonl
tasks_file = os.path.join(DATA_DIR, "meta", "tasks.jsonl")
if os.path.exists(tasks_file):
task_count = 0
with open(tasks_file, "r") as f:
for line in f:
task_count += 1
check_pass(f"tasks.jsonl: {task_count} tasks")
else:
check_fail(f"tasks.jsonl MISSING")
print()
# ============================================================
# CHECK 6: Retrieval quality debug (sample pairs)
# ============================================================
print("=" * 60)
print("CHECK 6: Retrieval quality debug (sample pairs)")
print("=" * 60)
try:
import random
random.seed(42)
# Load task mapping
task_mapping = {}
with open(os.path.join(DATA_DIR, "meta", "tasks.jsonl"), "r") as f:
for line in f:
item = json.loads(line)
if "task_index" in item and "task" in item:
task_mapping[item["task_index"]] = item["task"]
# Load parquet for task indices
import pandas as pd
parquet_files = sorted(list(Path(DATA_DIR, "data").rglob("*.parquet")))
dfs = [pd.read_parquet(f) for f in parquet_files[:5]] # Load a subset for speed
df = pd.concat(dfs, ignore_index=True)
# Sample and display pairs
num_samples = 5
sample_indices = random.sample(range(min(len(metadata), len(df))), min(num_samples, len(metadata), len(df)))
same_task_count = 0
for i, query_idx in enumerate(sample_indices):
demo_idx = int(nn_indices[query_idx, 0])
distance = nn_distances[query_idx, 0]
query_meta = metadata[query_idx]
demo_meta = metadata[demo_idx]
# Get tasks
query_global = query_meta.get("global_frame_idx", query_idx)
demo_global = demo_meta.get("global_frame_idx", demo_idx)
query_task_idx = -1
demo_task_idx = -1
if query_global < len(df):
query_row = df.iloc[query_global]
query_task_idx = int(query_row.get("task_index", -1))
if demo_global < len(df):
demo_row = df.iloc[demo_global]
demo_task_idx = int(demo_row.get("task_index", -1))
query_task = task_mapping.get(query_task_idx, "Unknown")
demo_task = task_mapping.get(demo_task_idx, "Unknown")
same_task = query_task_idx == demo_task_idx
if same_task:
same_task_count += 1
# Action similarity
action_mse = np.mean((actions[query_idx] - actions[demo_idx]) ** 2)
print(f"\n Pair {i+1}/{num_samples}:")
print(f" Query: ep={query_meta.get('episode_idx', '?')}, frame={query_meta.get('frame_idx', '?')}")
print(f" Task: '{query_task[:70]}'")
print(f" Demo: ep={demo_meta.get('episode_idx', '?')}, frame={demo_meta.get('frame_idx', '?')}")
print(f" Task: '{demo_task[:70]}'")
print(f" Distance: {distance:.4f} | Action MSE: {action_mse:.4f} | Same task: {same_task}")
same_task_pct = 100 * same_task_count / num_samples
print(f"\n Same-task retrieval rate (in sample): {same_task_pct:.0f}%")
if same_task_pct >= 50:
check_pass(f"Same-task retrieval is reasonable ({same_task_pct:.0f}%)")
else:
print(f" ⚠ Low same-task retrieval - this could indicate cross-task retrieval (may be intended)")
except Exception as e:
check_fail(f"Error in retrieval debug: {e}")
import traceback
traceback.print_exc()
print()
# ============================================================
# CHECK 7: Verify RiclLiberoDataset can load
# ============================================================
print("=" * 60)
print("CHECK 7: Verify RiclLiberoDataset loads correctly")
print("=" * 60)
try:
from openpi.data.ricl_libero_dataset import RiclLiberoDataset
TARGET_ACTION_HORIZON = 10 # Must match pi0fast-LIBERO default
dataset = RiclLiberoDataset(
data_dir=DATA_DIR,
context_dir=CONTEXT_DIR,
action_horizon=TARGET_ACTION_HORIZON, # Truncate from precomputed 50 → 10
use_action_interpolation=True,
lambda_decay=10.0,
num_retrieved_observations=1,
)
check_pass(f"Dataset created successfully with {len(dataset)} samples")
# Try loading one sample
print(" Loading sample [0]...")
sample = dataset[0]
print(f" Sample keys: {sorted(sample.keys())}")
# Check expected keys (matching actual RiclLiberoDataset output format)
expected_keys = [
"query_observation.images.image",
"query_observation.images.wrist_image",
"query_observation.state",
"query_actions",
"query_prompt",
]
for key in expected_keys:
if key in sample:
val = sample[key]
if isinstance(val, np.ndarray):
check_pass(f"{key}: shape={val.shape}, dtype={val.dtype}")
else:
check_pass(f"{key}: type={type(val).__name__}")
else:
check_fail(f"{key} MISSING from sample")
# Check demo keys
demo_keys = [k for k in sample.keys() if "retrieved" in k or "demo" in k]
print(f" Demo keys: {demo_keys}")
for key in demo_keys:
val = sample[key]
if isinstance(val, np.ndarray):
check_pass(f"{key}: shape={val.shape}, dtype={val.dtype}")
else:
check_pass(f"{key}: type={type(val).__name__}")
# Check interpolation weights
interp_keys = [k for k in sample.keys() if "lamda" in k or "lambda" in k or "interp" in k]
if interp_keys:
for key in interp_keys:
val = sample[key]
if isinstance(val, np.ndarray):
print(f" {key}: shape={val.shape}, range=[{val.min():.4f}, {val.max():.4f}]")
else:
print(f" {key}: {val}")
check_pass("Interpolation weights present")
else:
print(f" ⚠ No interpolation weight keys found (checked: lamda, lambda, interp)")
# ============================================================
# CHECK 7b: Verify action truncation (50 → 10)
# ============================================================
print()
print("=" * 60)
print(f"CHECK 7b: Verify action truncation (precomputed=50 → target={TARGET_ACTION_HORIZON})")
print("=" * 60)
# Precomputed actions shape
precomputed_horizon = actions.shape[1]
print(f" Precomputed context action_horizon: {precomputed_horizon}")
print(f" Target action_horizon: {TARGET_ACTION_HORIZON}")
# Check query_actions shape
query_actions = sample["query_actions"]
if query_actions.shape[0] == TARGET_ACTION_HORIZON:
check_pass(f"query_actions truncated correctly: shape={query_actions.shape} (horizon={TARGET_ACTION_HORIZON})")
else:
check_fail(f"query_actions NOT truncated: shape={query_actions.shape} (expected horizon={TARGET_ACTION_HORIZON})")
# Check demo actions shape
demo_actions = sample["retrieved_0_actions"]
if demo_actions.shape[0] == TARGET_ACTION_HORIZON:
check_pass(f"retrieved_0_actions truncated correctly: shape={demo_actions.shape} (horizon={TARGET_ACTION_HORIZON})")
else:
check_fail(f"retrieved_0_actions NOT truncated: shape={demo_actions.shape} (expected horizon={TARGET_ACTION_HORIZON})")
# Verify truncation preserves data (first 10 of 50 should match)
raw_demo_actions_full = actions[int(nn_indices[0, 0])] # Full 50-step from context
raw_demo_truncated = raw_demo_actions_full[:TARGET_ACTION_HORIZON]
if np.allclose(demo_actions, raw_demo_truncated, atol=1e-5):
check_pass(f"Truncated demo actions match first {TARGET_ACTION_HORIZON} steps of precomputed context")
else:
check_fail(f"Truncated demo actions DO NOT match precomputed context first {TARGET_ACTION_HORIZON} steps!")
# Verify multiple samples to ensure consistency
num_verify = 5
all_correct = True
for vi in range(1, min(num_verify + 1, len(dataset))):
s = dataset[vi]
if s["query_actions"].shape[0] != TARGET_ACTION_HORIZON:
all_correct = False
break
if s["retrieved_0_actions"].shape[0] != TARGET_ACTION_HORIZON:
all_correct = False
break
if all_correct:
check_pass(f"Truncation verified across {num_verify} additional samples")
else:
check_fail(f"Truncation inconsistent across samples!")
except Exception as e:
check_fail(f"Error loading dataset: {e}")
import traceback
traceback.print_exc()
print()
# ============================================================
# CHECK 8: Verify config resolution
# ============================================================
print("=" * 60)
print("CHECK 8: Verify training config resolution")
print("=" * 60)
try:
import etils.epath as epath
# Check PI0_FAST_BASE_CHECKPOINT resolution
local_ckpt = epath.Path("pi0_fast_base_params")
if local_ckpt.exists():
check_pass(f"PI0_FAST_BASE_CHECKPOINT resolves to LOCAL: pi0_fast_base_params")
else:
check_fail(f"pi0_fast_base_params NOT found locally - will try S3 (requires network!)")
# Check assets
assets_franka = os.path.join(WORK_DIR, "pi0_fast_base", "assets", "franka")
if os.path.exists(assets_franka):
check_pass(f"Assets dir (franka) exists")
else:
# Check alternative location
assets_alt = os.path.join(WORK_DIR, "assets")
if os.path.exists(assets_alt):
check_pass(f"Assets base dir exists: {assets_alt}")
else:
check_fail(f"No assets directory found")
except Exception as e:
check_fail(f"Config check error: {e}")
print()
# ============================================================
# SUMMARY
# ============================================================
print("=" * 60)
print("VERIFICATION COMPLETE")
print("=" * 60)
total = passed + failed
print(f"\n ✓ Passed: {passed}/{total}")
print(f" ❌ Failed: {failed}/{total}")
if failed == 0:
print("\n 🎉 ALL CHECKS PASSED! Ready to train.")
else:
print(f"\n ⚠ {failed} check(s) failed. Fix before training.")
print(f"""
To train RICL, submit:
cd {WORK_DIR}
sbatch slurm/train_ricl_libero.slurm
""")
|