Spaces:
Runtime error
Runtime error
File size: 12,337 Bytes
df4a21a | 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 | """
Wrapper for DeiT Distilled submodel.
"""
import json
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
from PIL import Image
from torchvision import transforms
try:
import timm
TIMM_AVAILABLE = True
except ImportError:
TIMM_AVAILABLE = False
from app.core.errors import InferenceError, ConfigurationError
from app.core.logging import get_logger
from app.models.wrappers.base_wrapper import BaseSubmodelWrapper
from app.services.explainability import attention_rollout, heatmap_to_base64, compute_focus_summary
logger = get_logger(__name__)
def create_custom_mlp_head(in_features: int = 768, num_classes: int = 2) -> nn.Sequential:
"""
Create custom MLP head for DeiT model matching training configuration.
Returns nn.Sequential to match saved state dict keys (0, 1, 4 indices).
"""
return nn.Sequential(
nn.LayerNorm(in_features), # 0
nn.Linear(in_features, 512), # 1
nn.GELU(), # 2 (no params)
nn.Dropout(p=0.2), # 3 (no params)
nn.Linear(512, num_classes) # 4
)
class DeiTDistilledWrapper(BaseSubmodelWrapper):
"""
Wrapper for DeiT Distilled model.
Model expects 224x224 RGB images with ImageNet normalization.
Uses a custom MLP head for classification.
"""
def __init__(
self,
repo_id: str,
config: Dict[str, Any],
local_path: str
):
super().__init__(repo_id, config, local_path)
self._model: Optional[nn.Module] = None
self._transform: Optional[transforms.Compose] = None
self._device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self._threshold = config.get("threshold", 0.5)
logger.info(f"Initialized DeiTDistilledWrapper for {repo_id}")
def load(self) -> None:
"""Load the DeiT model with custom head and trained weights."""
if not TIMM_AVAILABLE:
raise ConfigurationError(
message="timm package not installed. Run: pip install timm",
details={"repo_id": self.repo_id}
)
weights_path = Path(self.local_path) / "deit_distilled_final.pt"
preprocess_path = Path(self.local_path) / "preprocess.json"
if not weights_path.exists():
raise ConfigurationError(
message=f"deit_distilled_final.pt not found in {self.local_path}",
details={"repo_id": self.repo_id, "expected_path": str(weights_path)}
)
try:
# Load preprocessing config
preprocess_config = {}
if preprocess_path.exists():
with open(preprocess_path, "r") as f:
preprocess_config = json.load(f)
# Build transform pipeline
input_size = preprocess_config.get("input_size", 224)
if isinstance(input_size, list):
input_size = input_size[0]
normalize_config = preprocess_config.get("normalize", {})
mean = normalize_config.get("mean", [0.485, 0.456, 0.406])
std = normalize_config.get("std", [0.229, 0.224, 0.225])
# Use bicubic interpolation as specified
interpolation = preprocess_config.get("interpolation", "bicubic")
interp_mode = transforms.InterpolationMode.BICUBIC if interpolation == "bicubic" else transforms.InterpolationMode.BILINEAR
self._transform = transforms.Compose([
transforms.Resize((input_size, input_size), interpolation=interp_mode),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
# Create model architecture
model_name = self.config.get("model_name", "deit_base_distilled_patch16_224")
num_classes = self.config.get("num_classes", 2)
# Create base model without pretrained weights
self._model = timm.create_model(model_name, pretrained=False, num_classes=0)
# Replace heads with custom MLP heads (Sequential assigned directly)
# Note: state dict has separate keys for head and head_dist, so don't share
hidden_dim = 768 # DeiT base hidden dimension
self._model.head = create_custom_mlp_head(hidden_dim, num_classes)
self._model.head_dist = create_custom_mlp_head(hidden_dim, num_classes)
# Load trained weights
state_dict = torch.load(weights_path, map_location=self._device, weights_only=True)
self._model.load_state_dict(state_dict)
self._model.to(self._device)
self._model.eval()
# Mark as loaded
self._predict_fn = self._run_inference
logger.info(f"Loaded DeiT Distilled model from {self.repo_id}")
except ConfigurationError:
raise
except Exception as e:
logger.error(f"Failed to load DeiT Distilled model: {e}")
raise ConfigurationError(
message=f"Failed to load model: {e}",
details={"repo_id": self.repo_id, "error": str(e)}
)
def _run_inference(
self,
image_tensor: torch.Tensor,
explain: bool = False
) -> Dict[str, Any]:
"""Run model inference on preprocessed tensor."""
heatmap = None
if explain:
# Collect attention weights from all blocks
attentions: List[torch.Tensor] = []
handles = []
# Hook into attention modules to capture weights
# DeiT blocks structure: blocks[i].attn
def create_attn_hook():
stored_attn = []
def hook(module, inputs, outputs):
# Get q, k from the module's forward computation
# inputs[0] is x of shape [B, N, C]
x = inputs[0]
B, N, C = x.shape
# Access the attention module's parameters
qkv = module.qkv(x) # [B, N, 3*dim]
qkv = qkv.reshape(B, N, 3, module.num_heads, C // module.num_heads)
qkv = qkv.permute(2, 0, 3, 1, 4) # [3, B, heads, N, dim_head]
q, k, v = qkv[0], qkv[1], qkv[2]
# Compute attention weights
scale = (C // module.num_heads) ** -0.5
attn = (q @ k.transpose(-2, -1)) * scale
attn = attn.softmax(dim=-1) # [B, heads, N, N]
# Average over heads
attn_avg = attn.mean(dim=1) # [B, N, N]
stored_attn.append(attn_avg.detach())
return hook, stored_attn
all_stored_attns = []
for block in self._model.blocks:
hook_fn, stored = create_attn_hook()
all_stored_attns.append(stored)
handle = block.attn.register_forward_hook(hook_fn)
handles.append(handle)
try:
with torch.no_grad():
logits = self._model(image_tensor)
probs = F.softmax(logits, dim=1)
prob_fake = probs[0, 1].item()
pred_int = 1 if prob_fake >= self._threshold else 0
# Get attention from hooks
attention_list = [stored[0] for stored in all_stored_attns if len(stored) > 0]
if attention_list:
# Stack: [num_layers, B, N, N]
attention_stack = torch.stack(attention_list, dim=0)
# Compute rollout - returns (grid_size, grid_size) heatmap
attention_map = attention_rollout(
attention_stack[:, 0], # [num_layers, N, N]
head_fusion="mean", # Already averaged
discard_ratio=0.0,
num_prefix_tokens=2 # DeiT has CLS + distillation token
) # Returns (14, 14) for DeiT-Base
# Resize to image size
from PIL import Image as PILImage
heatmap_img = PILImage.fromarray(
(attention_map * 255).astype(np.uint8)
).resize((224, 224), PILImage.BILINEAR)
heatmap = np.array(heatmap_img).astype(np.float32) / 255.0
finally:
for handle in handles:
handle.remove()
else:
with torch.no_grad():
# In eval mode, DeiT returns single tensor
logits = self._model(image_tensor)
probs = F.softmax(logits, dim=1)
prob_fake = probs[0, 1].item()
pred_int = 1 if prob_fake >= self._threshold else 0
result = {
"logits": logits[0].cpu().numpy().tolist(),
"prob_fake": prob_fake,
"pred_int": pred_int
}
if heatmap is not None:
result["heatmap"] = heatmap
return result
def predict(
self,
image: Optional[Image.Image] = None,
image_bytes: Optional[bytes] = None,
explain: bool = False,
**kwargs
) -> Dict[str, Any]:
"""
Run prediction on an image.
Args:
image: PIL Image object
image_bytes: Raw image bytes (will be converted to PIL Image)
explain: If True, compute attention rollout heatmap
Returns:
Standardized prediction dictionary with optional heatmap
"""
if self._model is None or self._transform is None:
raise InferenceError(
message="Model not loaded",
details={"repo_id": self.repo_id}
)
try:
# Convert bytes to PIL Image if needed
if image is None and image_bytes is not None:
import io
image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
elif image is not None:
image = image.convert("RGB")
else:
raise InferenceError(
message="No image provided",
details={"repo_id": self.repo_id}
)
# Preprocess
image_tensor = self._transform(image).unsqueeze(0).to(self._device)
# Run inference
result = self._run_inference(image_tensor, explain=explain)
# Standardize output
class_mapping = self.config.get("class_mapping", {"0": "real", "1": "fake"})
pred_int = result["pred_int"]
output = {
"pred_int": pred_int,
"pred": class_mapping.get(str(pred_int), "unknown"),
"prob_fake": result["prob_fake"],
"meta": {
"model": self.name,
"threshold": self._threshold,
"logits": result["logits"]
}
}
# Add heatmap if requested
if explain and "heatmap" in result:
heatmap = result["heatmap"]
output["heatmap_base64"] = heatmap_to_base64(heatmap)
output["explainability_type"] = "attention_rollout"
output["focus_summary"] = compute_focus_summary(heatmap)
return output
except InferenceError:
raise
except Exception as e:
logger.error(f"Prediction failed for {self.repo_id}: {e}")
raise InferenceError(
message=f"Prediction failed: {e}",
details={"repo_id": self.repo_id, "error": str(e)}
)
|