Create app.py

app.py

@@ -0,0 +1,77 @@
+import torch
+import timm
+import numpy as np
+from PIL import Image
+import gradio as gr
+import json
+from torchvision import transforms
+from pytorch_grad_cam import GradCAMPlusPlus
+from pytorch_grad_cam.utils.image import show_cam_on_image
+from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
+
+# --- Config ---
+MODEL_WEIGHTS = "efficientnetv2_best_model.pth"
+CLASS_MAPPING = "class_mapping.json"
+
+# --- Device ---
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+# --- Load class names ---
+with open(CLASS_MAPPING, "r") as f:
+    class_names = json.load(f)
+    # handle both list and dict style
+    if isinstance(class_names, dict):
+        class_names = [class_names[str(i)] for i in range(len(class_names))]
+
+NUM_CLASSES = len(class_names)
+
+# --- Model ---
+model = timm.create_model('efficientnetv2_rw_s', pretrained=False, num_classes=NUM_CLASSES, drop_rate=0.3)
+model.load_state_dict(torch.load(MODEL_WEIGHTS, map_location=device))
+model.to(device)
+model.eval()
+
+# --- Preprocessing ---
+imagenet_mean = [0.485, 0.456, 0.406]
+imagenet_std = [0.229, 0.224, 0.225]
+val_transform = transforms.Compose([
+    transforms.Resize(256),
+    transforms.CenterCrop(224),
+    transforms.ToTensor(),
+    transforms.Normalize(mean=imagenet_mean, std=imagenet_std)
+])
+
+# --- Grad-CAM setup (for EfficientNetV2, last block) ---
+target_layer = model.blocks[-1] if hasattr(model, "blocks") else model.layer4[-1]
+cam = GradCAMPlusPlus(model=model, target_layers=[target_layer], use_cuda=(device.type=='cuda'))
+
+# --- Gradio Inference + Explainability ---
+def predict_and_explain(img: Image.Image):
+    # Preprocess
+    image_pil = img.convert("RGB").resize((224, 224))
+    input_tensor = val_transform(image_pil).unsqueeze(0).to(device)
+    with torch.no_grad():
+        output = model(input_tensor)
+        pred_idx = output.argmax().item()
+    pred_name = class_names[pred_idx]
+    # Grad-CAM
+    targets = [ClassifierOutputTarget(pred_idx)]
+    grayscale_cam = cam(input_tensor=input_tensor, targets=targets)[0]
+    image_np = np.array(image_pil).astype(np.float32) / 255.0
+    cam_image = show_cam_on_image(image_np, grayscale_cam, use_rgb=True)
+    return Image.fromarray(cam_image), f"Prediction: {pred_name} (class index {pred_idx})"
+
+demo = gr.Interface(
+    fn=predict_and_explain,
+    inputs=gr.Image(type="pil", label="Upload Car Image"),
+    outputs=[
+        gr.Image(label="Grad-CAM++ Output"),
+        gr.Text(label="Prediction")
+    ],
+    title="🚗 EfficientNetV2 Car Classifier + Grad-CAM Demo",
+    description="Upload a car photo to classify its make/model/year and visualize the model's attention with Grad-CAM.",
+    allow_flagging='never'
+)
+
+if __name__ == "__main__":
+    demo.launch()