File size: 10,820 Bytes
1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 0f94e43 1aea709 |
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 |
#!/usr/bin/env python3
"""
MatAnyone Model Loader
Handles MatAnyone loading with proper device initialization
"""
import os
import time
import logging
import traceback
from pathlib import Path
from typing import Optional, Dict, Any
import torch
import numpy as np
logger = logging.getLogger(__name__)
class MatAnyoneLoader:
"""Dedicated loader for MatAnyone models"""
def __init__(self, device: str = "cuda", cache_dir: str = "./checkpoints/matanyone_cache"):
self.device = device
self.cache_dir = cache_dir
os.makedirs(self.cache_dir, exist_ok=True)
self.model = None
self.model_id = "PeiqingYang/MatAnyone"
self.load_time = 0.0
def load(self) -> Optional[Any]:
"""
Load MatAnyone model
Returns:
Loaded model or None
"""
logger.info(f"Loading MatAnyone model: {self.model_id}")
# Try loading strategies in order
strategies = [
("official", self._load_official),
("fallback", self._load_fallback)
]
for strategy_name, strategy_func in strategies:
try:
logger.info(f"Trying MatAnyone loading strategy: {strategy_name}")
start_time = time.time()
model = strategy_func()
if model:
self.load_time = time.time() - start_time
self.model = model
logger.info(f"MatAnyone loaded successfully via {strategy_name} in {self.load_time:.2f}s")
return model
except Exception as e:
logger.error(f"MatAnyone {strategy_name} strategy failed: {e}")
logger.debug(traceback.format_exc())
continue
logger.error("All MatAnyone loading strategies failed")
return None
def _load_official(self) -> Optional[Any]:
"""Load using official MatAnyone API"""
from matanyone import InferenceCore
# Create processor - pass model ID as positional argument
processor = InferenceCore(self.model_id)
# Ensure processor is properly initialized for the device
if hasattr(processor, 'device'):
processor.device = self.device
# Move model components to device if they exist
if hasattr(processor, 'model'):
if hasattr(processor.model, 'to'):
processor.model = processor.model.to(self.device)
processor.model.eval()
# Patch the processor to handle inputs properly
self._patch_processor(processor)
return processor
def _patch_processor(self, processor):
"""
Patch the MatAnyone processor to handle device placement and tensor formats correctly
"""
original_step = getattr(processor, 'step', None)
original_process = getattr(processor, 'process', None)
device = self.device
def safe_wrapper(*args, **kwargs):
"""Universal wrapper that handles both step and process calls"""
try:
# Handle different calling patterns
# Pattern 1: step(image, mask, idx_mask=False)
# Pattern 2: process(image, mask)
# Pattern 3: Called with just args
# Pattern 4: Called with kwargs
image = None
mask = None
idx_mask = kwargs.get('idx_mask', False)
# Extract image and mask
if 'image' in kwargs and 'mask' in kwargs:
image = kwargs['image']
mask = kwargs['mask']
elif len(args) >= 2:
image = args[0]
mask = args[1]
if len(args) > 2:
idx_mask = args[2]
elif len(args) == 1:
# Might be called with just mask for refinement
mask = args[0]
# Create dummy image if needed
if isinstance(mask, np.ndarray):
h, w = mask.shape[:2] if mask.ndim >= 2 else (512, 512)
image = np.zeros((h, w, 3), dtype=np.uint8)
elif isinstance(mask, torch.Tensor):
h, w = mask.shape[-2:] if mask.dim() >= 2 else (512, 512)
image = torch.zeros((h, w, 3), dtype=torch.uint8)
if image is None or mask is None:
logger.error(f"MatAnyone called with invalid args: {len(args)} args, kwargs: {kwargs.keys()}")
# Return something safe
if mask is not None:
return mask
return np.ones((512, 512), dtype=np.float32) * 0.5
# Convert to tensors on correct device
if isinstance(image, np.ndarray):
image = torch.from_numpy(image).to(device)
elif isinstance(image, torch.Tensor):
image = image.to(device)
if isinstance(mask, np.ndarray):
mask = torch.from_numpy(mask).to(device)
elif isinstance(mask, torch.Tensor):
mask = mask.to(device)
# Fix image format (ensure CHW or NCHW)
if image.dim() == 2: # Grayscale HW
image = image.unsqueeze(0) # CHW
elif image.dim() == 3:
# Check if HWC or CHW
if image.shape[-1] in [1, 3, 4]: # HWC
image = image.permute(2, 0, 1) # CHW
# Add batch if needed
if image.shape[0] in [1, 3, 4]: # CHW
image = image.unsqueeze(0) # NCHW
elif image.dim() == 4:
# Already NCHW, ensure correct channel position
if image.shape[-1] in [1, 3, 4]: # NHWC
image = image.permute(0, 3, 1, 2) # NCHW
# Fix mask format
if mask.dim() == 2:
mask = mask.unsqueeze(0) # Add channel: CHW
elif mask.dim() == 3:
if mask.shape[0] > 4: # Likely HWC
mask = mask.permute(2, 0, 1) # CHW
# Ensure float and normalized
if image.dtype != torch.float32:
image = image.float()
if not idx_mask and mask.dtype != torch.float32:
mask = mask.float()
if image.max() > 1.0:
image = image / 255.0
if not idx_mask and mask.max() > 1.0:
mask = mask / 255.0
# Call original method if it exists
if original_step:
try:
result = original_step(image, mask, idx_mask=idx_mask)
# Convert result back to numpy if needed
if isinstance(result, torch.Tensor):
result = result.cpu().numpy()
return result
except Exception as e:
logger.error(f"MatAnyone original step failed: {e}")
# Fallback: return slightly processed mask
if isinstance(mask, torch.Tensor):
# Apply slight smoothing
import torch.nn.functional as F
mask = F.avg_pool2d(mask.unsqueeze(0), 3, stride=1, padding=1)
mask = mask.squeeze(0).cpu().numpy()
return mask
except Exception as e:
logger.error(f"MatAnyone safe_wrapper failed: {e}")
import traceback
logger.debug(traceback.format_exc())
# Return safe fallback
if 'mask' in locals() and mask is not None:
if isinstance(mask, torch.Tensor):
return mask.cpu().numpy()
return mask
return np.ones((512, 512), dtype=np.float32) * 0.5
# Apply patches to both methods
if hasattr(processor, 'step'):
processor.step = safe_wrapper
logger.info("Patched MatAnyone step method")
if hasattr(processor, 'process'):
processor.process = safe_wrapper
logger.info("Patched MatAnyone process method")
# Also add a direct call method
processor.__call__ = safe_wrapper
def _load_fallback(self) -> Optional[Any]:
"""Create fallback processor for testing"""
class FallbackMatAnyone:
def __init__(self, device):
self.device = device
def step(self, image, mask, idx_mask=False, **kwargs):
"""Pass through mask with minor smoothing"""
if isinstance(mask, np.ndarray):
# Apply slight Gaussian blur for edge smoothing
import cv2
if mask.ndim == 2:
smoothed = cv2.GaussianBlur(mask, (5, 5), 1.0)
return smoothed
elif mask.ndim == 3:
smoothed = np.zeros_like(mask)
for i in range(mask.shape[0]):
smoothed[i] = cv2.GaussianBlur(mask[i], (5, 5), 1.0)
return smoothed
return mask
def process(self, image, mask, **kwargs):
"""Alias for step"""
return self.step(image, mask, **kwargs)
logger.warning("Using fallback MatAnyone (limited refinement)")
return FallbackMatAnyone(self.device)
def cleanup(self):
"""Clean up resources"""
if self.model:
del self.model
self.model = None
if torch.cuda.is_available():
torch.cuda.empty_cache()
def get_info(self) -> Dict[str, Any]:
"""Get loader information"""
return {
"loaded": self.model is not None,
"model_id": self.model_id,
"device": self.device,
"load_time": self.load_time,
"model_type": type(self.model).__name__ if self.model else None
} |