Update app.py
Browse files
app.py
CHANGED
|
@@ -11,28 +11,18 @@ import urllib.request
|
|
| 11 |
from torch.nn.functional import interpolate
|
| 12 |
import warnings
|
| 13 |
warnings.filterwarnings('ignore')
|
| 14 |
-
|
| 15 |
-
# Performance optimizations
|
| 16 |
-
torch.backends.cudnn.benchmark = True # Auto-tune for best performance
|
| 17 |
-
torch.set_float32_matmul_precision('high') # Use TF32 on Ampere GPUs
|
| 18 |
-
|
| 19 |
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 20 |
|
| 21 |
@torch.no_grad()
|
| 22 |
def load_model_and_labels():
|
| 23 |
-
"""Load ResNet152 model
|
| 24 |
print("π Loading ResNet152 model...")
|
| 25 |
|
| 26 |
-
#
|
| 27 |
model = models.resnet152(weights='IMAGENET1K_V2')
|
| 28 |
model.eval().to(DEVICE)
|
| 29 |
|
| 30 |
-
#
|
| 31 |
-
if DEVICE.type == 'cuda':
|
| 32 |
-
model = model.half() # FP16 precision for 2x speedup on GPU
|
| 33 |
-
print("β
Model optimized with FP16 precision")
|
| 34 |
-
|
| 35 |
-
# Load ImageNet labels (cached)
|
| 36 |
url = "https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt"
|
| 37 |
response = urllib.request.urlopen(url)
|
| 38 |
labels = [line.decode('utf-8').strip() for line in response.readlines()]
|
|
@@ -46,9 +36,9 @@ model, IMAGENET_LABELS = load_model_and_labels()
|
|
| 46 |
target_layer = model.layer4[-1]
|
| 47 |
gradcam = LayerGradCam(model, target_layer)
|
| 48 |
|
| 49 |
-
#
|
| 50 |
transform = transforms.Compose([
|
| 51 |
-
transforms.Resize((224, 224)
|
| 52 |
transforms.ToTensor(),
|
| 53 |
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
|
| 54 |
])
|
|
@@ -59,34 +49,25 @@ def predict_and_explain(image):
|
|
| 59 |
return "Please upload an image", None, None
|
| 60 |
|
| 61 |
try:
|
| 62 |
-
# Prepare input
|
| 63 |
img_tensor = transform(image).unsqueeze(0).to(DEVICE)
|
| 64 |
-
if DEVICE.type == 'cuda':
|
| 65 |
-
img_tensor = img_tensor.half() # Match model precision
|
| 66 |
|
| 67 |
with torch.no_grad():
|
| 68 |
-
#
|
| 69 |
output = model(img_tensor)
|
| 70 |
-
if DEVICE.type == 'cuda':
|
| 71 |
-
output = output.float() # Convert back for softmax
|
| 72 |
probabilities = torch.softmax(output, dim=1)
|
| 73 |
|
| 74 |
top10_prob, top10_idx = torch.topk(probabilities, 10)
|
| 75 |
pred_class = top10_idx[0][0].item()
|
| 76 |
confidence = top10_prob[0][0].item()
|
| 77 |
|
| 78 |
-
# Generate Grad-CAM
|
| 79 |
-
|
| 80 |
-
grad_tensor = img_tensor.float() # Grad-CAM needs FP32
|
| 81 |
-
else:
|
| 82 |
-
grad_tensor = img_tensor
|
| 83 |
-
|
| 84 |
-
attributions = gradcam.attribute(grad_tensor, target=pred_class)
|
| 85 |
attr_resized = interpolate(attributions, size=(224, 224), mode='bilinear', align_corners=False)
|
| 86 |
attr_np = attr_resized.squeeze().cpu().detach().numpy()
|
| 87 |
attr_np = (attr_np - attr_np.min()) / (attr_np.max() - attr_np.min() + 1e-8)
|
| 88 |
|
| 89 |
-
#
|
| 90 |
fig = plt.figure(figsize=(24, 14))
|
| 91 |
fig.patch.set_facecolor('#0a0a0a')
|
| 92 |
|
|
@@ -194,9 +175,6 @@ def predict_and_explain(image):
|
|
| 194 |
</div>"""
|
| 195 |
top5_html += "</div>"
|
| 196 |
|
| 197 |
-
# Performance info
|
| 198 |
-
precision_info = "FP16 (GPU Accelerated)" if DEVICE.type == 'cuda' else "FP32 (CPU)"
|
| 199 |
-
|
| 200 |
prediction_text = f"""
|
| 201 |
<div class="result-card">
|
| 202 |
<div class="pred-header">
|
|
@@ -204,7 +182,7 @@ def predict_and_explain(image):
|
|
| 204 |
<div class="badge badge-{badge}">{badge_icon} {badge_text}</div>
|
| 205 |
</div>
|
| 206 |
<div class="conf-score">{confidence*100:.2f}%</div>
|
| 207 |
-
<div class="model-tag">π¬ ResNet152
|
| 208 |
<div class="divider"></div>
|
| 209 |
{top5_html}
|
| 210 |
</div>"""
|
|
@@ -288,7 +266,7 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Base(), title="XAI Image Classifi
|
|
| 288 |
<div class="hero-header">
|
| 289 |
<h1>XAI Image Classifier</h1>
|
| 290 |
<p class="subtitle">ResNet152 with Grad-CAM Explainability</p>
|
| 291 |
-
<div class="model-tag">β‘
|
| 292 |
</div>
|
| 293 |
''')
|
| 294 |
|
|
@@ -301,7 +279,7 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Base(), title="XAI Image Classifi
|
|
| 301 |
clear_btn = gr.ClearButton([input_image], value="ποΈ Clear", size="lg", scale=1)
|
| 302 |
|
| 303 |
with gr.Column(scale=1, elem_classes="results-panel"):
|
| 304 |
-
output_text = gr.HTML('<div class="placeholder"><strong>π Welcome to XAI Classifier!</strong><br><br>This classifier uses ResNet152:<br>β’ 82.3% ImageNet Top-1 Accuracy<br>β’
|
| 305 |
|
| 306 |
with gr.Column(elem_classes="viz-section"):
|
| 307 |
gr.HTML("<div class='section-label'>π― Visual Explainability Analysis</div>")
|
|
@@ -314,5 +292,4 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Base(), title="XAI Image Classifi
|
|
| 314 |
predict_btn.click(fn=predict_and_explain, inputs=[input_image], outputs=[output_text, output_image, detailed_heatmap])
|
| 315 |
|
| 316 |
if __name__ == "__main__":
|
| 317 |
-
demo.
|
| 318 |
-
demo.launch(share=True)
|
|
|
|
| 11 |
from torch.nn.functional import interpolate
|
| 12 |
import warnings
|
| 13 |
warnings.filterwarnings('ignore')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 15 |
|
| 16 |
@torch.no_grad()
|
| 17 |
def load_model_and_labels():
|
| 18 |
+
"""Load ResNet152 model for maximum accuracy"""
|
| 19 |
print("π Loading ResNet152 model...")
|
| 20 |
|
| 21 |
+
# ResNet152 (Best accuracy in ResNet family)
|
| 22 |
model = models.resnet152(weights='IMAGENET1K_V2')
|
| 23 |
model.eval().to(DEVICE)
|
| 24 |
|
| 25 |
+
# Load ImageNet labels
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
url = "https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt"
|
| 27 |
response = urllib.request.urlopen(url)
|
| 28 |
labels = [line.decode('utf-8').strip() for line in response.readlines()]
|
|
|
|
| 36 |
target_layer = model.layer4[-1]
|
| 37 |
gradcam = LayerGradCam(model, target_layer)
|
| 38 |
|
| 39 |
+
# Transform for ResNet152
|
| 40 |
transform = transforms.Compose([
|
| 41 |
+
transforms.Resize((224, 224)),
|
| 42 |
transforms.ToTensor(),
|
| 43 |
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
|
| 44 |
])
|
|
|
|
| 49 |
return "Please upload an image", None, None
|
| 50 |
|
| 51 |
try:
|
| 52 |
+
# Prepare input
|
| 53 |
img_tensor = transform(image).unsqueeze(0).to(DEVICE)
|
|
|
|
|
|
|
| 54 |
|
| 55 |
with torch.no_grad():
|
| 56 |
+
# ResNet152 prediction
|
| 57 |
output = model(img_tensor)
|
|
|
|
|
|
|
| 58 |
probabilities = torch.softmax(output, dim=1)
|
| 59 |
|
| 60 |
top10_prob, top10_idx = torch.topk(probabilities, 10)
|
| 61 |
pred_class = top10_idx[0][0].item()
|
| 62 |
confidence = top10_prob[0][0].item()
|
| 63 |
|
| 64 |
+
# Generate Grad-CAM
|
| 65 |
+
attributions = gradcam.attribute(img_tensor, target=pred_class)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
attr_resized = interpolate(attributions, size=(224, 224), mode='bilinear', align_corners=False)
|
| 67 |
attr_np = attr_resized.squeeze().cpu().detach().numpy()
|
| 68 |
attr_np = (attr_np - attr_np.min()) / (attr_np.max() - attr_np.min() + 1e-8)
|
| 69 |
|
| 70 |
+
# Main visualization
|
| 71 |
fig = plt.figure(figsize=(24, 14))
|
| 72 |
fig.patch.set_facecolor('#0a0a0a')
|
| 73 |
|
|
|
|
| 175 |
</div>"""
|
| 176 |
top5_html += "</div>"
|
| 177 |
|
|
|
|
|
|
|
|
|
|
| 178 |
prediction_text = f"""
|
| 179 |
<div class="result-card">
|
| 180 |
<div class="pred-header">
|
|
|
|
| 182 |
<div class="badge badge-{badge}">{badge_icon} {badge_text}</div>
|
| 183 |
</div>
|
| 184 |
<div class="conf-score">{confidence*100:.2f}%</div>
|
| 185 |
+
<div class="model-tag">π¬ ResNet152 Architecture (82.3% ImageNet Accuracy)</div>
|
| 186 |
<div class="divider"></div>
|
| 187 |
{top5_html}
|
| 188 |
</div>"""
|
|
|
|
| 266 |
<div class="hero-header">
|
| 267 |
<h1>XAI Image Classifier</h1>
|
| 268 |
<p class="subtitle">ResNet152 with Grad-CAM Explainability</p>
|
| 269 |
+
<div class="model-tag">β‘ Maximum Accuracy Production Version</div>
|
| 270 |
</div>
|
| 271 |
''')
|
| 272 |
|
|
|
|
| 279 |
clear_btn = gr.ClearButton([input_image], value="ποΈ Clear", size="lg", scale=1)
|
| 280 |
|
| 281 |
with gr.Column(scale=1, elem_classes="results-panel"):
|
| 282 |
+
output_text = gr.HTML('<div class="placeholder"><strong>π Welcome to XAI Classifier!</strong><br><br>This classifier uses ResNet152:<br>β’ 82.3% ImageNet Top-1 Accuracy<br>β’ Grad-CAM Visual Explainability<br>β’ 1000 Object Categories<br><br>Upload an image to see the magic! β¨</div>')
|
| 283 |
|
| 284 |
with gr.Column(elem_classes="viz-section"):
|
| 285 |
gr.HTML("<div class='section-label'>π― Visual Explainability Analysis</div>")
|
|
|
|
| 292 |
predict_btn.click(fn=predict_and_explain, inputs=[input_image], outputs=[output_text, output_image, detailed_heatmap])
|
| 293 |
|
| 294 |
if __name__ == "__main__":
|
| 295 |
+
demo.launch(share=False, show_error=True)
|
|
|