Spaces:
Sleeping
Sleeping
File size: 10,532 Bytes
ef791b1 | 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 | # utils.py
from ultralytics import YOLO
import os
import cv2
import numpy as np
from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import torch
from dotenv import load_dotenv
from pathlib import Path
load_dotenv()
# β
Load .env from this backend folder
load_dotenv(Path(__file__).resolve().parent / ".env")
# β
Now fetch your Gemini key
GEMINI_API_KEY = os.getenv("GOOGLE_API_KEY")
if not GEMINI_API_KEY:
print("β GOOGLE_API_KEY not found! Check your .env file path.")
else:
print("β
GOOGLE_API_KEY loaded successfully.")
# ========================================
# YOLO Object Detection
# ========================================
yolo_model = YOLO("yolov8n.pt")
def detect_objects(image_path):
"""Detect objects in image using YOLO"""
results = yolo_model(image_path)
detected_objects = []
for result in results:
for cls in result.boxes.cls:
class_idx = int(cls)
class_name = result.names[class_idx]
detected_objects.append(class_name)
# Remove duplicates
return list(set(detected_objects))
# ========================================
# ViT Model for AI Detection (used in main.py)
# ========================================
MODEL_PATH = os.path.join(os.path.dirname(__file__), "vit-ai-vs-real-model")
processor = AutoImageProcessor.from_pretrained(MODEL_PATH)
model = AutoModelForImageClassification.from_pretrained(MODEL_PATH)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()
# ========================================
# Grad-CAM Visualization
# ========================================
def generate_gradcam(image_path, output_path="gradcam_result.jpg"):
"""
Generates Grad-CAM heatmap for the ViT model and saves it as an image file.
"""
image = Image.open(image_path).convert("RGB")
inputs = processor(images=image, return_tensors="pt").to(device)
# Enable gradient computation
inputs["pixel_values"].requires_grad_(True)
# Lists to capture attention weights and gradients
attention_weights = []
attention_grads = []
def forward_hook(module, input, output):
"""Capture attention weights during forward pass"""
if isinstance(output, tuple) and len(output) > 1:
attention_weights.append(output[1])
else:
attention_weights.append(None)
def backward_hook(module, grad_input, grad_output):
"""Capture gradients during backward pass"""
if grad_output[0] is not None:
attention_grads.append(grad_output[0])
# Register hooks on the last attention layer
last_attn = model.vit.encoder.layer[-1].attention.attention
fwd_handle = last_attn.register_forward_hook(forward_hook)
bwd_handle = last_attn.register_full_backward_hook(backward_hook)
try:
print("π Running forward pass...")
# Forward pass
outputs = model(**inputs)
pred_class = torch.argmax(outputs.logits, dim=-1).item()
score = outputs.logits[:, pred_class]
print(f"π― Predicted class: {pred_class}, Score: {score.item():.4f}")
# Backward pass
model.zero_grad()
score.backward()
print(f"π Captured {len(attention_weights)} attention weights")
print(f"π Captured {len(attention_grads)} gradients")
# Check if we captured valid attention data
if len(attention_weights) == 0 or attention_weights[0] is None:
print("β οΈ No attention weights captured - using alternative method")
# Alternative: Use the last hidden state's gradients
patch_grad = inputs["pixel_values"].grad
if patch_grad is not None:
# Average over channels and batch
cam = patch_grad.abs().mean(dim=[0, 1]).detach().cpu().numpy()
# Resize to reasonable spatial dimensions
import torch.nn.functional as F
h, w = cam.shape
target_size = 14 # ViT typically uses 14x14 patches
cam_tensor = torch.from_numpy(cam).unsqueeze(0).unsqueeze(0)
cam = F.interpolate(cam_tensor, size=(target_size, target_size),
mode='bilinear', align_corners=False)
cam = cam.squeeze().numpy()
else:
print("β οΈ No gradients found - using center-focused fallback")
# Create a simple center-focused heatmap as fallback
cam = np.ones((14, 14))
center = 7
for i in range(14):
for j in range(14):
dist = np.sqrt((i - center)**2 + (j - center)**2)
cam[i, j] = max(0, 1 - dist / 10)
else:
print("β
Processing attention weights...")
# Get attention weights and gradients
attn = attention_weights[0] # [batch, heads, seq_len, seq_len]
if len(attention_grads) > 0:
grad = attention_grads[0]
# Average over heads
weights = grad.mean(dim=1).squeeze(0) # [seq_len, seq_len]
attn_map = attn.mean(dim=1).squeeze(0) # [seq_len, seq_len]
# Weight the attention map by gradients
cam = (weights * attn_map).sum(dim=0).detach().cpu().numpy()
else:
# Just use attention without gradients
attn_map = attn.mean(dim=1).squeeze(0) # [seq_len, seq_len]
cam = attn_map.mean(dim=0).detach().cpu().numpy()
# Remove CLS token (first position) and reshape to spatial dimensions
cam = cam[1:] # Remove CLS token
size = int(np.sqrt(len(cam)))
cam = cam[:size*size].reshape(size, size)
# Normalize
cam = np.maximum(cam, 0)
if cam.max() > 0:
cam = cam / cam.max()
else:
cam = np.ones_like(cam) * 0.5
print(f"π CAM shape: {cam.shape}, min: {cam.min():.4f}, max: {cam.max():.4f}")
except Exception as e:
print(f"β Error during Grad-CAM computation: {e}")
import traceback
traceback.print_exc()
# Fallback: create simple heatmap
print("π Using fallback heatmap")
cam = np.ones((14, 14)) * 0.5
finally:
# Remove hooks
fwd_handle.remove()
bwd_handle.remove()
# Resize CAM to match input image size
cam = cv2.resize(cam, (image.size[0], image.size[1]))
# Create heatmap
heatmap = cv2.applyColorMap(np.uint8(255 * cam), cv2.COLORMAP_JET)
heatmap = cv2.cvtColor(heatmap, cv2.COLOR_BGR2RGB)
# Overlay on original image
overlay = np.array(image) * 0.6 + heatmap * 0.4
overlay = np.uint8(overlay)
# Save Grad-CAM image
cv2.imwrite(output_path, cv2.cvtColor(overlay, cv2.COLOR_RGB2BGR))
print("π’ Grad-CAM generated!")
print(f"πΎ Saved at: {os.path.abspath(output_path)}")
return output_path
def generate_gemini_summary(
original_image_path: str = None,
grad_cam_image_path: str = None,
classification: str = None,
probability_percent: float = None,
objects: list = None
) -> str:
"""
Unified Gemini summary generator for both vision and text-only models.
Works for Gemini 1.5 (vision) and text-only Gemini models.
"""
try:
if not GEMINI_MODEL:
raise ValueError("No Gemini model available")
print(f"π€ Using Gemini model: {GEMINI_MODEL}")
gemini_model_instance = genai.GenerativeModel(GEMINI_MODEL)
# Detect if model supports vision
is_vision_model = any(x in GEMINI_MODEL.lower() for x in ["1.5", "2.0", "2.5", "vision"])
# Vision model β send both images
if original_image_path and grad_cam_image_path and is_vision_model:
with open(original_image_path, "rb") as f1, open(grad_cam_image_path, "rb") as f2:
img_original = {"mime_type": "image/jpeg", "data": f1.read()}
img_gradcam = {"mime_type": "image/jpeg", "data": f2.read()}
prompt = f"""
You are an expert in AI image forensics and explainable AI (XAI).
Two images are provided:
1. The original image.
2. The Grad-CAM heatmap (red/yellow = strong focus regions).
The detection model's result:
β’ Classification: {classification or "Unknown"}
β’ Probability: {probability_percent or 0:.2f}%
Write a concise 2β3 sentence explanation describing:
- Which regions were highlighted by Grad-CAM.
- Why those regions indicate the modelβs decision.
- How the confidence reflects these cues.
"""
response = gemini_model_instance.generate_content([prompt, img_original, img_gradcam])
return response.text.strip()
# Text-only fallback
else:
prompt = f"""
The model detected these objects: {', '.join(objects or [])}.
Classification result: {classification or "Unknown"} ({probability_percent or 0:.2f}% confidence).
Write a short 2β3 sentence explanation of why the image might appear {classification.lower() if classification else 'AI-generated or real'}.
"""
response = gemini_model_instance.generate_content(prompt)
return response.text.strip()
except Exception as e:
print(f"β οΈ Gemini summary error: {e}")
err = str(e)
if "429" in err or "quota" in err.lower():
print("π‘ Quota exceeded β fallback summary used.")
if not objects:
return "This image likely contains a few simple visual features, but detailed analysis is unavailable."
elif len(objects) == 1:
return f"This image features a {objects[0]}."
elif len(objects) == 2:
return f"This image shows a {objects[0]} and a {objects[1]}."
else:
return f"This image contains {', '.join(objects[:-1])}, and {objects[-1]}."
else:
return f"Unable to generate Gemini summary. Possible reason: {err}"
|