update hyper

app.py

@@ -29,14 +29,15 @@ DEFAULT_K = 400
 DEFAULT_WSIZE = 224
 DEFAULT_GAMMA = 3.0
 DEFAULT_TAU = 0.01
+DEFAULT_R = 13
 
 # Function to reset hyperparameters to default values
 def reset_hyperparams():
-    return DEFAULT_WSIZE, DEFAULT_K, DEFAULT_GAMMA, DEFAULT_ALPHA, DEFAULT_SIGMA, DEFAULT_TAU
+    return DEFAULT_WSIZE, DEFAULT_K, DEFAULT_GAMMA, DEFAULT_ALPHA, DEFAULT_SIGMA, DEFAULT_TAU, DEFAULT_R
 
 @spaces.GPU
 def segment_image(img: PIL.Image.Image, classnames: str, use_lposs_plus: bool | None, 
-                  winodw_size:int, k:int, gamma:float, alpha:float, sigma: float, tau:float) -> tuple[np.ndarray | PIL.Image.Image | str, list[tuple[np.ndarray | tuple[int, int, int, int], str]]]:
+                  winodw_size:int, k:int, gamma:float, alpha:float, sigma: float, tau:float, r:int) -> tuple[np.ndarray | PIL.Image.Image | str, list[tuple[np.ndarray | tuple[int, int, int, int], str]]]:
     img_tensor = to_torch_tensor(PIL.Image.fromarray(img)).unsqueeze(0).to(device)
     classnames = [c.strip() for c in classnames.split(",")]
     num_classes = len(classnames)
@@ -46,7 +47,7 @@ def segment_image(img: PIL.Image.Image, classnames: str, use_lposs_plus: bool |
     
     preds = lposs(maskclip, dino, img_tensor, classnames, window_size=winodw_size, window_stride=stride, sigma=1/sigma, lp_k_image=k, lp_gamma=gamma, lp_alpha=alpha)
     if use_lposs_plus:
-        preds = lposs_plus(img_tensor, preds, tau=tau, alpha=alpha)
+        preds = lposs_plus(img_tensor, preds, tau=tau, alpha=alpha, r=r)
     preds = F.interpolate(preds, size=img.shape[:-1], mode="bilinear", align_corners=False)
     preds = F.softmax(preds * 100, dim=1).cpu().numpy()
     return (img, [(preds[0, i, :, :], classnames[i]) for i in range(num_classes)])
@@ -69,11 +70,12 @@ with gr.Blocks() as demo:
                 gr.Markdown("Hyper-parameters")
             with gr.Row():
                 window_size = gr.Slider(minimum=112, maximum=448, value=DEFAULT_WSIZE, step=16, label="Window Size")
-                k = gr.Slider(minimum=50, maximum=800, value=DEFAULT_K, step=50, label="k")
-                gamma = gr.Slider(minimum=0.0, maximum=10.0, value=DEFAULT_GAMMA, step=0.5, label="Gamma")
-                alpha = gr.Slider(minimum=0.0, maximum=1.0, value=DEFAULT_ALPHA, step=0.05, label="Alpha")
-                sigma = gr.Slider(minimum=50, maximum=400, value=DEFAULT_SIGMA, step=10, label="Sigma")
-                tau = gr.Slider(minimum=0.0, maximum=1.0, value=DEFAULT_TAU, step=0.01, label="Tau")
+                k = gr.Slider(minimum=50, maximum=800, value=DEFAULT_K, step=50, label="k (LPOSS number of graph neighbors)")
+                gamma = gr.Slider(minimum=0.0, maximum=10.0, value=DEFAULT_GAMMA, step=0.5, label="γ (LPOSS graph edge weight)")
+                sigma = gr.Slider(minimum=50, maximum=400, value=DEFAULT_SIGMA, step=10, label="σ (LPOSS spatial affinity weight)")
+                tau = gr.Slider(minimum=0.0, maximum=1.0, value=DEFAULT_TAU, step=0.01, label="τ (LPOSS+ appearance affinity weight)")
+                r = gr.Slider(minimum=3, maximum=15, value=DEFAULT_R, step=2, label="r (LPOSS+ kernel size)")
+                alpha = gr.Slider(minimum=0.0, maximum=1.0, value=DEFAULT_ALPHA, step=0.05, label="α (amount of propagation)")
             with gr.Row():
                 reset_btn = gr.Button("Reset to Default Values")
     
@@ -89,11 +91,11 @@ with gr.Blocks() as demo:
     with gr.Row():
         segment_btn = gr.Button("Segment Image")
     
-    reset_btn.click(fn=reset_hyperparams, outputs=[window_size, k, gamma, alpha, sigma, tau])
+    reset_btn.click(fn=reset_hyperparams, outputs=[window_size, k, gamma, alpha, sigma, tau, r])
     
     segment_btn.click(
         fn=segment_image,
-        inputs=[input_image, class_names, use_lposs_plus, window_size, k, gamma, alpha, sigma, tau],
+        inputs=[input_image, class_names, use_lposs_plus, window_size, k, gamma, alpha, sigma, tau, r],
         outputs=[output_image]
     )
 

lposs.py

@@ -277,13 +277,13 @@ def get_laplacian(rows, cols, data, N, alpha=0.99):
     return L
 
 
-def lposs_plus(img, preds, tau=0.01, alpha=0.95):
+def lposs_plus(img, preds, tau=0.01, alpha=0.95, r=13):
     preds = preds[0, ...]
     num_classes, h_img, w_img = preds.shape
     preds = preds.permute((1, 2, 0))
     preds = preds.reshape((h_img*w_img, -1))
 
-    rows, cols, pixel_pixel_data, locs = get_pixel_connections(img, neigh=6)
+    rows, cols, pixel_pixel_data, locs = get_pixel_connections(img, neigh=r//2)
     pixel_pixel_data = torch.sqrt(pixel_pixel_data)
     pixel_pixel_data = torch.exp(-pixel_pixel_data / tau)
     L = get_laplacian(rows, cols, pixel_pixel_data, preds.shape[0], alpha=alpha)