# Author: Huzheng Yang # %% USE_SPACES = True if USE_SPACES: # huggingface ZeroGPU try: import spaces except ImportError: USE_SPACES = False # run on standard GPU import gradio as gr import torch from PIL import Image import numpy as np import time import gradio as gr from backbone import extract_features from ncut_pytorch import NCUT, eigenvector_to_rgb def compute_ncut( features, num_eig=100, num_sample_ncut=10000, affinity_focal_gamma=0.3, knn_ncut=10, knn_tsne=10, embedding_method="UMAP", num_sample_tsne=300, perplexity=150, n_neighbors=150, min_dist=0.1, sampling_method="fps", ): logging_str = "" num_nodes = np.prod(features.shape[:3]) if num_nodes / 2 < num_eig: # raise gr.Error("Number of eigenvectors should be less than half the number of nodes.") gr.Warning("Number of eigenvectors should be less than half the number of nodes.\n" f"Setting num_eig to {num_nodes // 2 - 1}.") num_eig = num_nodes // 2 - 1 logging_str += f"Number of eigenvectors should be less than half the number of nodes.\n" f"Setting num_eig to {num_nodes // 2 - 1}.\n" start = time.time() eigvecs, eigvals = NCUT( num_eig=num_eig, num_sample=num_sample_ncut, device="cuda" if torch.cuda.is_available() else "cpu", affinity_focal_gamma=affinity_focal_gamma, knn=knn_ncut, sample_method=sampling_method, ).fit_transform(features.reshape(-1, features.shape[-1])) # print(f"NCUT time: {time.time() - start:.2f}s") logging_str += f"NCUT time: {time.time() - start:.2f}s\n" start = time.time() _, rgb = eigenvector_to_rgb( eigvecs, method=embedding_method, num_sample=num_sample_tsne, perplexity=perplexity, n_neighbors=n_neighbors, min_distance=min_dist, knn=knn_tsne, device="cuda" if torch.cuda.is_available() else "cpu", ) logging_str += f"{embedding_method} time: {time.time() - start:.2f}s\n" rgb = rgb.reshape(features.shape[:3] + (3,)) return rgb, logging_str def dont_use_too_much_green(image_rgb): # make sure the foval 40% of the image is red leading x1, x2 = int(image_rgb.shape[1] * 0.3), int(image_rgb.shape[1] * 0.7) y1, y2 = int(image_rgb.shape[2] * 0.3), int(image_rgb.shape[2] * 0.7) sum_values = image_rgb[:, x1:x2, y1:y2].mean((0, 1, 2)) sorted_indices = sum_values.argsort(descending=True) image_rgb = image_rgb[:, :, :, sorted_indices] return image_rgb def to_pil_images(images): return [ Image.fromarray((image * 255).cpu().numpy().astype(np.uint8)).resize((256, 256), Image.Resampling.NEAREST) for image in images ] default_images = ['./images/image_0.jpg', './images/image_1.jpg', './images/image_2.jpg', './images/image_3.jpg', './images/image_5.jpg'] default_outputs = ['./images/ncut_0.jpg', './images/ncut_1.jpg', './images/ncut_2.jpg', './images/ncut_3.jpg', './images/ncut_5.jpg'] default_outputs_independent = ['./images/ncut_0_independent.jpg', './images/ncut_1_independent.jpg', './images/ncut_2_independent.jpg', './images/ncut_3_independent.jpg', './images/ncut_5_independent.jpg'] downscaled_images = ['./images/image_0_small.jpg', './images/image_1_small.jpg', './images/image_2_small.jpg', './images/image_3_small.jpg', './images/image_5_small.jpg'] downscaled_outputs = ['./images/ncut_0_small.jpg', './images/ncut_1_small.jpg', './images/ncut_2_small.jpg', './images/ncut_3_small.jpg', './images/ncut_5_small.jpg'] example_items = downscaled_images[:3] + downscaled_outputs[:3] def ncut_run( images, model_name="SAM(sam_vit_b)", layer=-1, num_eig=100, node_type="block", affinity_focal_gamma=0.3, num_sample_ncut=10000, knn_ncut=10, embedding_method="UMAP", num_sample_tsne=1000, knn_tsne=10, perplexity=500, n_neighbors=500, min_dist=0.1, sampling_method="fps", old_school_ncut=False, ): logging_str = "" if perplexity >= num_sample_tsne or n_neighbors >= num_sample_tsne: # raise gr.Error("Perplexity must be less than the number of samples for t-SNE.") gr.Warning("Perplexity/n_neighbors must be less than the number of samples.\n" f"Setting Perplexity to {num_sample_tsne-1}.") logging_str += f"Perplexity/n_neighbors must be less than the number of samples.\n" f"Setting Perplexity to {num_sample_tsne-1}.\n" perplexity = num_sample_tsne - 1 n_neighbors = num_sample_tsne - 1 node_type = node_type.split(":")[0].strip() images = [image[0] for image in images] # remove the label start = time.time() features = extract_features( images, model_name=model_name, node_type=node_type, layer=layer ) # print(f"Feature extraction time (gpu): {time.time() - start:.2f}s") logging_str += f"Backbone time: {time.time() - start:.2f}s\n" if not old_school_ncut: # joint across all images rgb, _logging_str = compute_ncut( features, num_eig=num_eig, num_sample_ncut=num_sample_ncut, affinity_focal_gamma=affinity_focal_gamma, knn_ncut=knn_ncut, knn_tsne=knn_tsne, num_sample_tsne=num_sample_tsne, embedding_method=embedding_method, perplexity=perplexity, n_neighbors=n_neighbors, min_dist=min_dist, sampling_method=sampling_method, ) logging_str += _logging_str rgb = dont_use_too_much_green(rgb) if old_school_ncut: # individual images logging_str += "Running NCut for each image independently\n" rgb = [] for i_image in range(features.shape[0]): feature = features[i_image] _rgb, _logging_str = compute_ncut( feature[None], num_eig=num_eig, num_sample_ncut=num_sample_ncut, affinity_focal_gamma=affinity_focal_gamma, knn_ncut=knn_ncut, knn_tsne=knn_tsne, num_sample_tsne=num_sample_tsne, embedding_method=embedding_method, perplexity=perplexity, n_neighbors=n_neighbors, min_dist=min_dist, sampling_method=sampling_method, ) logging_str += _logging_str rgb.append(_rgb[0]) return to_pil_images(rgb), logging_str def _ncut_run(*args, **kwargs): try: return ncut_run(*args, **kwargs) except Exception as e: gr.Error(str(e)) return [], "Error: " + str(e) if USE_SPACES: @spaces.GPU(duration=13) def quick_run(*args, **kwargs): return _ncut_run(*args, **kwargs) @spaces.GPU(duration=30) def long_run(*args, **kwargs): return _ncut_run(*args, **kwargs) @spaces.GPU(duration=60) def longer_run(*args, **kwargs): return _ncut_run(*args, **kwargs) @spaces.GPU(duration=120) def super_duper_long_run(*args, **kwargs): return _ncut_run(*args, **kwargs) if not USE_SPACES: def quick_run(*args, **kwargs): return _ncut_run(*args, **kwargs) def long_run(*args, **kwargs): return _ncut_run(*args, **kwargs) def longer_run(*args, **kwargs): return _ncut_run(*args, **kwargs) def super_duper_long_run(*args, **kwargs): return _ncut_run(*args, **kwargs) def run_fn( images, model_name="SAM(sam_vit_b)", layer=-1, num_eig=100, node_type="block", affinity_focal_gamma=0.3, num_sample_ncut=10000, knn_ncut=10, embedding_method="UMAP", num_sample_tsne=1000, knn_tsne=10, perplexity=500, n_neighbors=500, min_dist=0.1, sampling_method="fps", old_school_ncut=False, ): if images is None: gr.Warning("No images selected.") return [], "No images selected." if sampling_method == "fps": sampling_method = "farthest" kwargs = { "model_name": model_name, "layer": layer, "num_eig": num_eig, "node_type": node_type, "affinity_focal_gamma": affinity_focal_gamma, "num_sample_ncut": num_sample_ncut, "knn_ncut": knn_ncut, "embedding_method": embedding_method, "num_sample_tsne": num_sample_tsne, "knn_tsne": knn_tsne, "perplexity": perplexity, "n_neighbors": n_neighbors, "min_dist": min_dist, "sampling_method": sampling_method, "old_school_ncut": old_school_ncut, } num_images = len(images) if num_images > 100: return super_duper_long_run(images, **kwargs) if num_images > 50: return longer_run(images, **kwargs) if num_images > 10: return long_run(images, **kwargs) if embedding_method == "UMAP": if perplexity >= 250 or num_sample_tsne >= 500: return longer_run(images, **kwargs) return long_run(images, **kwargs) if embedding_method == "t-SNE": if perplexity >= 250 or num_sample_tsne >= 500: return long_run(images, **kwargs) return quick_run(images, **kwargs) return quick_run(images, **kwargs) def make_input_images_section(): gr.Markdown('### Input Images') input_gallery = gr.Gallery(value=[], label="Select images", show_label=False, elem_id="images", columns=[3], rows=[1], object_fit="contain", height="auto", type="pil", show_share_button=False) submit_button = gr.Button("🔴RUN", elem_id="submit_button") clear_images_button = gr.Button("🗑️Clear", elem_id='clear_button') return input_gallery, submit_button, clear_images_button def make_example_images_section(): gr.Markdown('### Load from Cloud Dataset 👇') load_images_button = gr.Button("Load Example", elem_id="load-images-button") example_gallery = gr.Gallery(value=example_items, label="Example Set A", show_label=False, columns=[3], rows=[2], object_fit="scale-down", height="200px", show_share_button=False, elem_id="example-gallery") hide_button = gr.Button("Hide Example", elem_id="hide-button") hide_button.click( fn=lambda: gr.update(visible=False), outputs=example_gallery ) return load_images_button, example_gallery, hide_button def make_dataset_images_section(): with gr.Accordion("➜ Load from dataset", open=True): dataset_names = [ 'UCSC-VLAA/Recap-COCO-30K', 'nateraw/pascal-voc-2012', 'johnowhitaker/imagenette2-320', 'jainr3/diffusiondb-pixelart', 'JapanDegitalMaterial/Places_in_Japan', 'Borismile/Anime-dataset', ] dataset_dropdown = gr.Dropdown(dataset_names, label="Dataset name", value="UCSC-VLAA/Recap-COCO-30K", elem_id="dataset") num_images_slider = gr.Slider(1, 200, step=1, label="Number of images", value=9, elem_id="num_images") random_seed_slider = gr.Number(0, label="Random seed", value=42, elem_id="random_seed") load_dataset_button = gr.Button("Load Dataset", elem_id="load-dataset-button") def load_dataset_images(dataset_name, num_images=10, random_seed=42): from datasets import load_dataset try: dataset = load_dataset(dataset_name) key = list(dataset.keys())[0] dataset = dataset[key] except Exception as e: gr.Error(f"Error loading dataset {dataset_name}: {e}") return None if num_images > len(dataset): num_images = len(dataset) image_idx = np.random.RandomState(random_seed).choice(len(dataset), num_images, replace=False) image_idx = image_idx.tolist() images = [dataset[i]['image'] for i in image_idx] return images load_dataset_button.click(load_dataset_images, inputs=[dataset_dropdown, num_images_slider, random_seed_slider], outputs=[input_gallery]) return dataset_dropdown, num_images_slider, random_seed_slider, load_dataset_button def make_output_images_section(): gr.Markdown('### Output Images') output_gallery = gr.Gallery(value=[], label="NCUT Embedding", show_label=False, elem_id="ncut", columns=[3], rows=[1], object_fit="contain", height="auto") return output_gallery def make_parameters_section(): gr.Markdown('### Parameters') model_dropdown = gr.Dropdown(["SAM(sam_vit_b)", "MobileSAM", "DiNO(dinov2_vitb14_reg)", "CLIP(openai/clip-vit-base-patch16)", "MAE(vit_base)"], label="Backbone", value="SAM(sam_vit_b)", elem_id="model_name") layer_slider = gr.Slider(0, 11, step=1, label="Backbone: Layer index", value=11, elem_id="layer") node_type_dropdown = gr.Dropdown(["attn: attention output", "mlp: mlp output", "block: sum of residual"], label="Backbone: Layer type", value="block: sum of residual", elem_id="node_type", info="which feature to take from each layer?") num_eig_slider = gr.Slider(1, 1000, step=1, label="NCUT: Number of eigenvectors", value=100, elem_id="num_eig", info='increase for more clusters') affinity_focal_gamma_slider = gr.Slider(0.01, 1, step=0.01, label="NCUT: Affinity focal gamma", value=0.5, elem_id="affinity_focal_gamma", info="decrease for shaper segmentation") with gr.Accordion("➜ Click to expand: more parameters", open=False): num_sample_ncut_slider = gr.Slider(100, 50000, step=100, label="NCUT: num_sample", value=10000, elem_id="num_sample_ncut", info="Nyström approximation") sampling_method_dropdown = gr.Dropdown(["fps", "random"], label="NCUT: Sampling method", value="fps", elem_id="sampling_method") knn_ncut_slider = gr.Slider(1, 100, step=1, label="NCUT: KNN", value=10, elem_id="knn_ncut", info="Nyström approximation") embedding_method_dropdown = gr.Dropdown(["tsne_3d", "umap_3d", "umap_shpere", "tsne_2d", "umap_2d"], label="Coloring method", value="tsne_3d", elem_id="embedding_method") num_sample_tsne_slider = gr.Slider(100, 1000, step=100, label="t-SNE/UMAP: num_sample", value=300, elem_id="num_sample_tsne", info="Nyström approximation") knn_tsne_slider = gr.Slider(1, 100, step=1, label="t-SNE/UMAP: KNN", value=10, elem_id="knn_tsne", info="Nyström approximation") perplexity_slider = gr.Slider(10, 500, step=10, label="t-SNE: Perplexity", value=150, elem_id="perplexity") n_neighbors_slider = gr.Slider(10, 500, step=10, label="UMAP: n_neighbors", value=150, elem_id="n_neighbors") min_dist_slider = gr.Slider(0.1, 1, step=0.1, label="UMAP: min_dist", value=0.1, elem_id="min_dist") return [model_dropdown, layer_slider, node_type_dropdown, num_eig_slider, affinity_focal_gamma_slider, num_sample_ncut_slider, knn_ncut_slider, embedding_method_dropdown, num_sample_tsne_slider, knn_tsne_slider, perplexity_slider, n_neighbors_slider, min_dist_slider, sampling_method_dropdown] with gr.Blocks() as demo: with gr.Tab('AlignedCut (Images)'): with gr.Row(): with gr.Column(scale=5, min_width=200): input_gallery, submit_button, clear_images_button = make_input_images_section() load_images_button, example_gallery, hide_button = make_example_images_section() dataset_dropdown, num_images_slider, random_seed_slider, load_dataset_button = make_dataset_images_section() with gr.Column(scale=5, min_width=200): output_gallery = make_output_images_section() [ model_dropdown, layer_slider, node_type_dropdown, num_eig_slider, affinity_focal_gamma_slider, num_sample_ncut_slider, knn_ncut_slider, embedding_method_dropdown, num_sample_tsne_slider, knn_tsne_slider, perplexity_slider, n_neighbors_slider, min_dist_slider, sampling_method_dropdown ] = make_parameters_section() # logging text box logging_text = gr.Textbox("Logging information", label="Logging", elem_id="logging", type="text", placeholder="Logging information") load_images_button.click(lambda x: (default_images, default_outputs), outputs=[input_gallery, output_gallery]) clear_images_button.click(lambda x: ([], []), outputs=[input_gallery, output_gallery]) submit_button.click( run_fn, inputs=[ input_gallery, model_dropdown, layer_slider, num_eig_slider, node_type_dropdown, affinity_focal_gamma_slider, num_sample_ncut_slider, knn_ncut_slider, embedding_method_dropdown, num_sample_tsne_slider, knn_tsne_slider, perplexity_slider, n_neighbors_slider, min_dist_slider, sampling_method_dropdown ], outputs=[output_gallery, logging_text] ) with gr.Tab('NCut (Legacy)'): gr.Markdown('#### Ncut, not aligned, no Nyström approximation') gr.Markdown('1. Each image is solved independently, _color is not aligned across images_') gr.Markdown('2. No Nyström approximation') gr.Markdown('### NCut (Legacy) vs. AlignedCut:') with gr.Row(): with gr.Column(scale=5, min_width=200): gr.Markdown('#### Pros') gr.Markdown('- Easy Solution. Use less eigenvectors.') gr.Markdown('- Exact solution. No Nyström approximation.') with gr.Column(scale=5, min_width=200): gr.Markdown('#### Cons') gr.Markdown('- Not aligned. Distance is not preserved across images. No pseudo-labeling or correspondence.') gr.Markdown('- Poor complexity scaling. Unable to handle large number of pixels.') gr.Markdown('---') with gr.Row(): with gr.Column(scale=5, min_width=200): gr.Markdown(' ') with gr.Column(scale=5, min_width=200): gr.Markdown('_color is not aligned across images_ 👇') with gr.Row(): with gr.Column(scale=5, min_width=200): input_gallery, submit_button, clear_images_button = make_input_images_section() load_images_button, example_gallery, hide_button = make_example_images_section() dataset_dropdown, num_images_slider, random_seed_slider, load_dataset_button = make_dataset_images_section() example_gallery.visible = False hide_button.visible = False with gr.Column(scale=5, min_width=200): output_gallery = make_output_images_section() [ model_dropdown, layer_slider, node_type_dropdown, num_eig_slider, affinity_focal_gamma_slider, num_sample_ncut_slider, knn_ncut_slider, embedding_method_dropdown, num_sample_tsne_slider, knn_tsne_slider, perplexity_slider, n_neighbors_slider, min_dist_slider, sampling_method_dropdown ] = make_parameters_section() old_school_ncut_checkbox = gr.Checkbox(label="Old school NCut", value=True, elem_id="old_school_ncut") invisible_list = [old_school_ncut_checkbox, num_sample_ncut_slider, knn_ncut_slider, num_sample_tsne_slider, knn_tsne_slider, sampling_method_dropdown] for item in invisible_list: item.visible = False # logging text box logging_text = gr.Textbox("Logging information", label="Logging", elem_id="logging", type="text", placeholder="Logging information") load_images_button.click(lambda x: (default_images, default_outputs_independent), outputs=[input_gallery, output_gallery]) clear_images_button.click(lambda x: ([], []), outputs=[input_gallery, output_gallery]) submit_button.click( run_fn, inputs=[ input_gallery, model_dropdown, layer_slider, num_eig_slider, node_type_dropdown, affinity_focal_gamma_slider, num_sample_ncut_slider, knn_ncut_slider, embedding_method_dropdown, num_sample_tsne_slider, knn_tsne_slider, perplexity_slider, n_neighbors_slider, min_dist_slider, sampling_method_dropdown, old_school_ncut_checkbox ], outputs=[output_gallery, logging_text] ) demo.launch(share=True) # %%