Does it work for you?

README.md

@@ -4,10 +4,10 @@ emoji: 👨‍🎨
 colorFrom: blue
 colorTo: red
 sdk: gradio
-sdk_version: 6.6.0
+sdk_version: 3.50.2
 app_file: web-demos/hugging_face/app.py
 pinned: false
 license: apache-2.0
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

requirements.txt

@@ -6,6 +6,8 @@ numpy
 scipy
 matplotlib
 scikit-image
+torch>=1.7.1
+torchvision>=0.8.2
 imageio-ffmpeg
 pyyaml
 requests
@@ -17,9 +19,14 @@ gitpython
 git+https://github.com/cheind/py-thin-plate-spline
 hickle
 tensorboard
+numpy
 git+https://github.com/facebookresearch/segment-anything.git
+gradio
 opencv-python
+matplotlib
+pyyaml
+av
 openmim
 tqdm
 psutil
-omegaconf
+omegaconf

web-demos/hugging_face/app.py

@@ -12,8 +12,6 @@ import torch
 import torchvision
 import numpy as np
 import gradio as gr
-import spaces
-import av
 
 from tools.painter import mask_painter
 from track_anything import TrackingAnything
@@ -62,7 +60,6 @@ def get_prompt(click_state, click_input):
     return prompt
 
 # extract frames from upload video
-@spaces.GPU
 def get_frames_from_video(video_input, video_state):
     """
     Args:
@@ -139,7 +136,6 @@ def get_frames_from_video(video_input, video_state):
                         gr.update(visible=True, value=operation_log), gr.update(visible=status_ok, value=operation_log)
 
 # get the select frame from gradio slider
-@spaces.GPU
 def select_template(image_selection_slider, video_state, interactive_state, mask_dropdown):
 
     # images = video_state[1]
@@ -163,25 +159,6 @@ def get_end_number(track_pause_number_slider, video_state, interactive_state):
     return video_state["painted_images"][track_pause_number_slider],interactive_state, operation_log, operation_log
 
 # use sam to get the mask
-@spaces.GPU
-def _sam_refine_gpu(video_state, point_prompt, click_state, interactive_state, coordinate):
-    # prompt for sam model
-    model.samcontroler.sam_controler.reset_image()
-    model.samcontroler.sam_controler.set_image(video_state["origin_images"][video_state["select_frame_number"]])
-    prompt = get_prompt(click_state=click_state, click_input=coordinate)
-
-    mask, logit, painted_image = model.first_frame_click(
-                                                      image=video_state["origin_images"][video_state["select_frame_number"]],
-                                                      points=np.array(prompt["input_point"]),
-                                                      labels=np.array(prompt["input_label"]),
-                                                      multimask=prompt["multimask_output"],
-                                                      )
-    video_state["masks"][video_state["select_frame_number"]] = mask
-    video_state["logits"][video_state["select_frame_number"]] = logit
-    video_state["painted_images"][video_state["select_frame_number"]] = painted_image
-    return painted_image, video_state, interactive_state
-
-
 def sam_refine(video_state, point_prompt, click_state, interactive_state, evt:gr.SelectData):
     """
     Args:
@@ -189,18 +166,27 @@ def sam_refine(video_state, point_prompt, click_state, interactive_state, evt:gr
         point_prompt: flag for positive or negative button click
         click_state: [[points], [labels]]
     """
-    # Extract the click coordinate here (outside @spaces.GPU) because
-    # gr.SelectData does not pickle across the ZeroGPU fork boundary.
     if point_prompt == "Positive":
         coordinate = "[[{},{},1]]".format(evt.index[0], evt.index[1])
         interactive_state["positive_click_times"] += 1
     else:
         coordinate = "[[{},{},0]]".format(evt.index[0], evt.index[1])
         interactive_state["negative_click_times"] += 1
+    
+    # prompt for sam model
+    model.samcontroler.sam_controler.reset_image()
+    model.samcontroler.sam_controler.set_image(video_state["origin_images"][video_state["select_frame_number"]])
+    prompt = get_prompt(click_state=click_state, click_input=coordinate)
 
-    painted_image, video_state, interactive_state = _sam_refine_gpu(
-        video_state, point_prompt, click_state, interactive_state, coordinate
-    )
+    mask, logit, painted_image = model.first_frame_click( 
+                                                      image=video_state["origin_images"][video_state["select_frame_number"]], 
+                                                      points=np.array(prompt["input_point"]),
+                                                      labels=np.array(prompt["input_label"]),
+                                                      multimask=prompt["multimask_output"],
+                                                      )
+    video_state["masks"][video_state["select_frame_number"]] = mask
+    video_state["logits"][video_state["select_frame_number"]] = logit
+    video_state["painted_images"][video_state["select_frame_number"]] = painted_image
 
     operation_log = [("[Must Do]", "Add mask"), (": add the current displayed mask for video segmentation.\n", None),
                      ("[Optional]", "Remove mask"), (": remove all added masks.\n", None),
@@ -245,20 +231,6 @@ def show_mask(video_state, interactive_state, mask_dropdown):
     return select_frame, operation_log, operation_log
 
 # tracking vos
-def _tracking_duration(video_state, interactive_state, mask_dropdown):
-    try:
-        origin = video_state.get("origin_images") if isinstance(video_state, dict) else None
-        start = video_state.get("select_frame_number", 0) if isinstance(video_state, dict) else 0
-        end = interactive_state.get("track_end_number") if isinstance(interactive_state, dict) else None
-        total = len(origin) if origin is not None else 60
-        n = (end if end else total) - start
-        n = max(1, int(n))
-    except Exception:
-        n = 60
-    return min(300, max(60, int(0.4 * n + 30)))
-
-
-@spaces.GPU(duration=_tracking_duration)
 def vos_tracking_video(video_state, interactive_state, mask_dropdown):
     operation_log = [("",""), ("Tracking finished! Try to click the Inpainting button to get the inpainting result.","Normal")]
     model.cutie.clear_memory()
@@ -319,18 +291,7 @@ def vos_tracking_video(video_state, interactive_state, mask_dropdown):
     #### shanggao code for mask save
     return video_output, video_state, interactive_state, operation_log, operation_log
 
-# inpaint
-def _inpaint_duration(video_state, resize_ratio_number, dilate_radius_number, raft_iter_number, subvideo_length_number, neighbor_length_number, ref_stride_number, mask_dropdown):
-    try:
-        origin = video_state.get("origin_images") if isinstance(video_state, dict) else None
-        n = len(origin) if origin is not None else 60
-        ratio = float(resize_ratio_number) if resize_ratio_number else 1.0
-    except Exception:
-        n, ratio = 60, 1.0
-    return min(420, max(90, int(n * ratio * 1.3 + 45)))
-
-
-@spaces.GPU(duration=_inpaint_duration)
+# inpaint 
 def inpaint_video(video_state, resize_ratio_number, dilate_radius_number, raft_iter_number, subvideo_length_number, neighbor_length_number, ref_stride_number, mask_dropdown):
     operation_log = [("",""), ("Inpainting finished!","Normal")]
 
@@ -368,42 +329,16 @@ def inpaint_video(video_state, resize_ratio_number, dilate_radius_number, raft_i
 def generate_video_from_frames(frames, output_path, fps=30):
     """
     Generates a video from a list of frames.
-
+    
     Args:
         frames (list of numpy arrays): The frames to include in the video.
         output_path (str): The path to save the generated video.
         fps (int, optional): The frame rate of the output video. Defaults to 30.
     """
-    frames = np.asarray(frames)
-    if frames.ndim != 4:
-        raise ValueError(f"Expected frames of shape (T, H, W, 3); got {frames.shape}")
+    frames = torch.from_numpy(np.asarray(frames))
     if not os.path.exists(os.path.dirname(output_path)):
         os.makedirs(os.path.dirname(output_path))
-
-    # torchvision.io.write_video was removed in newer torchvision; encode via PyAV.
-    height, width = frames.shape[1], frames.shape[2]
-    # libx264 requires even dimensions.
-    pad_h = height % 2
-    pad_w = width % 2
-    if pad_h or pad_w:
-        frames = np.pad(frames, ((0, 0), (0, pad_h), (0, pad_w), (0, 0)), mode="edge")
-        height += pad_h
-        width += pad_w
-
-    container = av.open(output_path, mode="w")
-    try:
-        stream = container.add_stream("libx264", rate=int(round(fps)) or 1)
-        stream.width = width
-        stream.height = height
-        stream.pix_fmt = "yuv420p"
-        for img in frames:
-            frame = av.VideoFrame.from_ndarray(np.ascontiguousarray(img.astype(np.uint8)), format="rgb24")
-            for packet in stream.encode(frame):
-                container.mux(packet)
-        for packet in stream.encode():
-            container.mux(packet)
-    finally:
-        container.close()
+    torchvision.io.write_video(output_path, frames, fps=fps, video_codec="libx264")
     return output_path
 
 def restart():
@@ -619,7 +554,7 @@ with gr.Blocks(theme=gr.themes.Monochrome(), css=css) as iface:
                         visible=False,
                         min_width=100,
                         scale=1,)
-                    with gr.Row(elem_classes="mask_button_group"):
+                    with gr.Row(scale=2, elem_classes="mask_button_group"):
                         Add_mask_button = gr.Button(value="Add mask", interactive=True, visible=False, elem_classes="add_button")
                         remove_mask_button = gr.Button(value="Remove mask", interactive=True, visible=False, elem_classes="remove_button")
                         clear_button_click = gr.Button(value="Clear clicks", interactive=True, visible=False, elem_classes="clear_button")
@@ -741,5 +676,5 @@ with gr.Blocks(theme=gr.themes.Monochrome(), css=css) as iface:
     )
     gr.Markdown(article)
 
-iface.queue()
+iface.queue(concurrency_count=1)
 iface.launch(debug=True)

web-demos/hugging_face/requirements.txt

@@ -6,6 +6,7 @@ hickle
 tensorboard
 numpy
 git+https://github.com/facebookresearch/segment-anything.git
+gradio
 opencv-python
 matplotlib
 pyyaml