delete and rewrite

app.py

@@ -56,9 +56,9 @@ random_seed = 42
 video_length = 201
 W = 1024
 H = W
+device = "cuda" #if torch.cuda.is_available() else "cpu"
 
 def get_pipe_image_and_video_predictor():
-    device="cpu"
     vae = AutoencoderKLWan.from_pretrained("./model/vae", torch_dtype=torch.float16)
     transformer = Transformer3DModel.from_pretrained("./model/transformer", torch_dtype=torch.float16)
     scheduler = UniPCMultistepScheduler.from_pretrained("./model/scheduler")
@@ -177,7 +177,7 @@ def preprocess_for_removal(images, masks):
         out_masks.append(msk_resized)
     arr_images = np.stack(out_images)
     arr_masks = np.stack(out_masks)
-    return torch.from_numpy(arr_images).half(), torch.from_numpy(arr_masks).half()
+    return torch.from_numpy(arr_images).half().to(device), torch.from_numpy(arr_masks).half().to(device)
 
 @spaces.GPU(duration=300)
 def inference_and_return_video(dilation_iterations, num_inference_steps, video_state=None):
@@ -189,13 +189,8 @@ def inference_and_return_video(dilation_iterations, num_inference_steps, video_s
     images = np.array(images)
     masks = np.array(masks)
     img_tensor, mask_tensor = preprocess_for_removal(images, masks)
-    print("mask_tensor shape:",mask_tensor.shape)
-    img_tensor=img_tensor.to("cuda")
-    mask_tensor=mask_tensor.to("cuda")
     mask_tensor = mask_tensor[:,:,:,:1]
 
-    pipe.to("cuda")
-
     if mask_tensor.shape[1] < mask_tensor.shape[2]:
         height = 480
         width = 832
@@ -211,7 +206,7 @@ def inference_and_return_video(dilation_iterations, num_inference_steps, video_s
                 height=height,
                 width=width,
                 num_inference_steps=int(num_inference_steps),
-                generator=torch.Generator(device="cuda").manual_seed(random_seed),
+                generator=torch.Generator(device=device).manual_seed(random_seed),
                 iterations=int(dilation_iterations)
         ).frames[0]
 
@@ -223,7 +218,7 @@ def inference_and_return_video(dilation_iterations, num_inference_steps, video_s
     clip.write_videofile(video_file, codec='libx264', audio=False, verbose=False, logger=None)
     return video_file
 
-@spaces.GPU(duration=150)
+@spaces.GPU(duration=100)
 def track_video(n_frames, video_state):
 
     input_points = video_state["input_points"]
@@ -247,12 +242,7 @@ def track_video(n_frames, video_state):
     images = [cv2.resize(img, (W_, H_)) for img in images]
     video_state["origin_images"] = images
     images = np.array(images)
-
-    sam2_checkpoint = "./SAM2-Video-Predictor/checkpoints/sam2_hiera_large.pt"
-    config = "sam2_hiera_l.yaml"
-    video_predictor = build_sam2_video_predictor(config, sam2_checkpoint, device="cuda")
-    
-    inference_state = video_predictor.init_state(images=images/255, device="cuda")
+    inference_state = video_predictor.init_state(images=images/255, device=device)
     video_state["inference_state"] = inference_state
 
     if len(torch.from_numpy(video_state["masks"][0]).shape) == 3: