File size: 11,014 Bytes
7a87926 879189e 91de477 9e9ce7f 91de477 9e9ce7f 879189e 7a87926 |
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 |
"""
Model loading utilities for DA3 and other models.
"""
import logging
import os
from pathlib import Path
from typing import Dict, Optional
import torch # type: ignore
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# HuggingFace cache location (RunPod optimization)
# ---------------------------------------------------------------------------
def _ensure_workspace_cache_env() -> None:
"""
Ensure HF/torch caches live under /workspace when available.
RunPod pods typically mount a volume at /workspace; placing caches there reduces
repeated downloads across restarts/redeploys.
"""
workspace = Path(os.environ.get("YLFF_WORKSPACE_DIR", "/workspace"))
try:
workspace.mkdir(parents=True, exist_ok=True)
except Exception:
return
cache_root = workspace / ".cache"
hf_root = cache_root / "huggingface"
try:
hf_root.mkdir(parents=True, exist_ok=True)
(hf_root / "hub").mkdir(parents=True, exist_ok=True)
(hf_root / "transformers").mkdir(parents=True, exist_ok=True)
(cache_root / "torch").mkdir(parents=True, exist_ok=True)
except Exception:
# If we can't create directories, still set env defaults (caller may have perms)
pass
os.environ.setdefault("XDG_CACHE_HOME", str(cache_root))
os.environ.setdefault("HF_HOME", str(hf_root))
os.environ.setdefault("HUGGINGFACE_HUB_CACHE", str(hf_root / "hub"))
os.environ.setdefault("TRANSFORMERS_CACHE", str(hf_root / "transformers"))
os.environ.setdefault("TORCH_HOME", str(cache_root / "torch"))
_ensure_workspace_cache_env()
# Optimize cuDNN for consistent input sizes (faster convolutions)
if torch.backends.cudnn.is_available():
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.deterministic = False # Allow non-deterministic for speed
logger.debug("cuDNN benchmark mode enabled for faster training")
# Available DA3 models and their characteristics
DA3_MODELS = {
# Main Series - Unified depth-ray representation
"depth-anything/DA3-GIANT": {
"series": "main",
"size": "giant",
"capabilities": [
"mono_depth",
"multi_view_depth",
"pose_conditioned_depth",
"pose_estimation",
"3d_gaussians",
],
"metric": False,
"description": "Largest model, best quality, all capabilities",
},
"depth-anything/DA3-LARGE": {
"series": "main",
"size": "large",
"capabilities": [
"mono_depth",
"multi_view_depth",
"pose_conditioned_depth",
"pose_estimation",
"3d_gaussians",
],
"metric": False,
"description": "Large model, good quality, all capabilities",
},
"depth-anything/DA3-BASE": {
"series": "main",
"size": "base",
"capabilities": [
"mono_depth",
"multi_view_depth",
"pose_conditioned_depth",
"pose_estimation",
"3d_gaussians",
],
"metric": False,
"description": "Base model, balanced quality/speed, all capabilities",
},
"depth-anything/DA3-SMALL": {
"series": "main",
"size": "small",
"capabilities": [
"mono_depth",
"multi_view_depth",
"pose_conditioned_depth",
"pose_estimation",
"3d_gaussians",
],
"metric": False,
"description": "Small model, fastest, all capabilities",
},
# Metric Series - Real-world scale depth
"depth-anything/DA3Metric-LARGE": {
"series": "metric",
"size": "large",
"capabilities": ["mono_depth", "metric_depth"],
"metric": True,
"description": "Specialized for metric depth estimation (real-world scale)",
},
# Monocular Series - High-quality relative depth
"depth-anything/DA3Mono-LARGE": {
"series": "mono",
"size": "large",
"capabilities": ["mono_depth"],
"metric": False,
"description": "High-quality relative monocular depth",
},
# Nested Series - Best for metric reconstruction
"depth-anything/DA3NESTED-GIANT-LARGE": {
"series": "nested",
"size": "giant-large",
"capabilities": [
"mono_depth",
"multi_view_depth",
"pose_conditioned_depth",
"pose_estimation",
"metric_depth",
],
"metric": True,
"description": "Combines giant model with metric model for real-world metric scale",
"recommended_for": ["ba_validation", "fine_tuning", "metric_reconstruction"],
},
}
def get_recommended_model(use_case: str = "ba_validation") -> str:
"""
Get recommended model for a specific use case.
Args:
use_case: One of:
- "ba_validation": BA validation and fine-tuning (needs pose + metric depth)
- "pose_estimation": Camera pose estimation
- "metric_depth": Metric depth estimation
- "mono_depth": Monocular depth estimation
- "fast": Fast inference (smaller model)
Returns:
Recommended model name
"""
recommendations = {
"ba_validation": "depth-anything/DA3NESTED-GIANT-LARGE", # Best: metric + pose
"fine_tuning": "depth-anything/DA3NESTED-GIANT-LARGE", # Best: metric + pose
"pose_estimation": "depth-anything/DA3-LARGE", # Good balance
"metric_depth": "depth-anything/DA3Metric-LARGE", # Specialized
"mono_depth": "depth-anything/DA3Mono-LARGE", # Specialized
"fast": "depth-anything/DA3-SMALL", # Fastest
"best_quality": "depth-anything/DA3-GIANT", # Highest quality
}
model = recommendations.get(use_case, "depth-anything/DA3-LARGE")
logger.info(f"Recommended model for '{use_case}': {model}")
return model
def list_available_models() -> Dict[str, Dict]:
"""List all available DA3 models with their characteristics."""
return DA3_MODELS.copy()
def get_model_info(model_name: str) -> Optional[Dict]:
"""Get information about a specific model."""
return DA3_MODELS.get(model_name)
def load_da3_model(
model_name: Optional[str] = None,
device: str = "cuda",
use_case: Optional[str] = None,
compile_model: bool = True,
compile_mode: str = "reduce-overhead",
) -> torch.nn.Module:
"""
Load pretrained DA3 model with optional compilation optimizations.
Args:
model_name: HuggingFace model name or local path.
If None and use_case is provided, uses recommended model.
device: Device to load model on
use_case: Optional use case to get recommended model if model_name not provided
compile_model: Whether to compile model with torch.compile (PyTorch 2.0+)
compile_mode: Compilation mode: "default", "reduce-overhead", "max-autotune"
Returns:
Loaded DA3 model
"""
# Auto-select model if not provided
if model_name is None:
if use_case:
model_name = get_recommended_model(use_case)
logger.info(f"Auto-selected model for '{use_case}': {model_name}")
else:
model_name = "depth-anything/DA3-LARGE" # Default fallback
logger.info(f"Using default model: {model_name}")
# Get model info
model_info = get_model_info(model_name)
if model_info:
logger.info(f"Loading {model_info['series']} series model: {model_name}")
logger.info(f" Description: {model_info['description']}")
if model_info.get("recommended_for"):
logger.info(f" Recommended for: {', '.join(model_info['recommended_for'])}")
try:
# Try to import DA3 API
from depth_anything_3.api import DepthAnything3 # type: ignore
# Robust device selection: Fallback from MPS if not available (e.g. running on Linux)
if device == "mps" and not torch.backends.mps.is_available():
if torch.cuda.is_available():
logger.warning("MPS requested but not available. Falling back to CUDA.")
device = "cuda"
else:
logger.warning("MPS requested but not available. Falling back to CPU.")
device = "cpu"
# Monkeypatch torch.cuda.is_bf16_supported to avoid initializing CUDA on CPU machines
# AND return True to force usage of bfloat16, which is required for CPU autocast
# (PyTorch CPU AMP does not support float16, only bfloat16)
if hasattr(torch.cuda, "is_bf16_supported"):
logger.info("Patching torch.cuda.is_bf16_supported=True to enable CPU bfloat16 autocast")
torch.cuda.is_bf16_supported = lambda: True
logger.info(f"Loading DA3 model: {model_name} on {device}")
model = DepthAnything3.from_pretrained(model_name)
model = model.to(device)
# Compile model for faster inference/training (PyTorch 2.0+)
# Disable compilation on MPS (often unstable or unsupported)
if device == "mps":
compile_model = False
logger.info("Disabling torch.compile on MPS device")
if compile_model and hasattr(torch, "compile"):
try:
logger.info(f"Compiling model with torch.compile (mode={compile_mode})...")
model = torch.compile(model, mode=compile_mode, fullgraph=False)
logger.info("Model compilation successful")
except Exception as e:
logger.warning(f"Model compilation failed: {e}. Continuing without compilation.")
elif compile_model:
logger.warning(
"torch.compile not available (requires PyTorch 2.0+). Skipping compilation."
)
model.eval()
return model
except ImportError:
logger.error(
"DA3 not found. Install with:\n"
" git clone https://github.com/ByteDance-Seed/Depth-Anything-3.git\n"
" cd Depth-Anything-3\n"
" pip install -e ."
)
raise
except Exception as e:
logger.error(f"Failed to load DA3 model: {e}")
raise
def load_model_from_checkpoint(
model: torch.nn.Module,
checkpoint_path: Path,
device: str = "cuda",
) -> torch.nn.Module:
"""
Load model weights from checkpoint.
Args:
model: Model architecture
checkpoint_path: Path to checkpoint
device: Device to load on
Returns:
Model with loaded weights
"""
checkpoint = torch.load(checkpoint_path, map_location=device)
if "model_state_dict" in checkpoint:
model.load_state_dict(checkpoint["model_state_dict"])
else:
model.load_state_dict(checkpoint)
logger.info(f"Loaded model from {checkpoint_path}")
return model
|