Update app.py

app.py

@@ -31,29 +31,32 @@ model = torch.quantization.quantize_dynamic(
 
 processor = DPTImageProcessor.from_pretrained("Intel/dpt-swinv2-tiny-256")
 
-color_map = cv2.applyColorMap(np.arange(256, dtype=np.uint8), cv2.COLORMAP_INFERNO)
-color_map_gpu = torch.from_numpy(color_map).to(device)
+color_map = torch.from_numpy(cv2.applyColorMap(np.arange(256, dtype=np.uint8), cv2.COLORMAP_INFERNO)).to(device)
+
+input_tensor = torch.zeros((1, 3, 128, 128), dtype=torch.float32, device=device)
 
 def preprocess_image(image):
-    resized = cv2.resize(image, (128, 72), interpolation=cv2.INTER_AREA)
-    return torch.from_numpy(resized.transpose(2, 0, 1)).float().div(255).unsqueeze(0).to(device)
+    return cv2.resize(image, (128, 72), interpolation=cv2.INTER_AREA).transpose(2, 0, 1).astype(np.float32) / 255.0
 
 @torch.inference_mode()
 def process_frame(image):
     if image is None:
         return None
-    input_tensor = preprocess_image(image)
+    preprocessed = preprocess_image(image)
+    input_tensor = torch.from_numpy(preprocessed).unsqueeze(0).to(device)
     
     predicted_depth = model(input_tensor).predicted_depth
     depth_map = predicted_depth.squeeze()
     
-    # Rounding instead of discretization
-    depth_map = torch.round(depth_map * 100) / 100
+    # Discretization on GPU
+    num_bins = 1000
+    depth_min, depth_max = depth_map.min(), depth_map.max()
+    bins = torch.linspace(depth_min, depth_max, num_bins, device=device)
+    depth_map = torch.bucketize(depth_map, bins)
     
     depth_map = (depth_map - depth_map.min()) / (depth_map.max() - depth_map.min())
     depth_map = (depth_map * 255).byte()
-    
-    depth_map_colored = color_map_gpu[depth_map]
+    depth_map_colored = color_map[depth_map]
     
     return cv2.cvtColor(depth_map_colored.cpu().numpy(), cv2.COLOR_BGR2RGB)