reduce time

app.py

@@ -89,8 +89,8 @@ def zeroshot_classifier(model, classnames, templates, device):
         zeroshot_weights = torch.stack(zeroshot_weights).cuda()
     return zeroshot_weights*100
 
-# device = "cuda" if torch.cuda.is_available() else "cpu"
-device = "cpu"
+device = "cuda" if torch.cuda.is_available() else "cpu"
+# device = "cpu"
 # Instantiate model
 vis_model = CLIPModel_Super("ViT-B/16", device=device, download_root="./ckpt")
 vis_model.eval()
@@ -232,6 +232,7 @@ def visualization(image, submodular_image_set, saved_json_file, index=None, comp
     
     ax3.set_title('Insertion Curve', fontsize=54)
     
+    fig.tight_layout()
     fig.canvas.draw()
     img_curve = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
     img_curve = img_curve.reshape(fig.canvas.get_width_height()[::-1] + (3,))
@@ -331,6 +332,8 @@ def update_image(thumbnail_name):
 # 创建 Gradio 界面
 with gr.Blocks() as demo:
     gr.Markdown("# Semantic Region Attribution via Submodular Subset Selection")  # 使用Markdown添加标题
+    
+    gr.Markdown("Since huggingface only has ordinary CPUs available, our sub-region division is relatively coarse-grained, which may affect the model performance. The inference time is about 5 minutes. If you are interested, you can try our source code. We have written many scripts to facilitate visualization.")
     with gr.Row():
         with gr.Column():
             # 第一排：上传图像输入框和一个缩略图