app.py

import spaces
import gradio as gr

import torch
from PIL import Image
import numpy as np
import time

import gradio as gr

from draft_gradio_backbone import extract_features
from ncut_pytorch import NCUT, rgb_from_tsne_3d, rgb_from_umap_3d


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,
):        
    logging_str = ""
    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,
    ).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()
    if embedding_method == "UMAP":
        X_3d, rgb = rgb_from_umap_3d(
            eigvecs,
            n_neighbors=n_neighbors,
            min_dist=min_dist,
            device="cuda" if torch.cuda.is_available() else "cpu",
        )
        # print(f"UMAP time: {time.time() - start:.2f}s")
        logging_str += f"UMAP time: {time.time() - start:.2f}s\n"
    elif embedding_method == "t-SNE":    
        X_3d, rgb = rgb_from_tsne_3d(
            eigvecs,
            num_sample=num_sample_tsne,
            perplexity=perplexity,
            knn=knn_tsne,
            device="cuda" if torch.cuda.is_available() else "cpu",
        )
        # print(f"t-SNE time: {time.time() - start:.2f}s")
        logging_str += f"t-SNE time: {time.time() - start:.2f}s\n"
    else:
        raise ValueError(f"Embedding method {embedding_method} not supported.")
    
    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.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']

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,
):
    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 to {num_sample_tsne-1}.")
        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"
    
    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,
    )
    logging_str += _logging_str
    rgb = dont_use_too_much_green(rgb)
    return to_pil_images(rgb), logging_str

@spaces.GPU(duration=15)
def quick_run(images, **kwargs):
    return ncut_run(images, **kwargs)

@spaces.GPU(duration=30)
def long_run(images, **kwargs):
    return ncut_run(images, **kwargs)

@spaces.GPU(duration=60)
def longer_run(images, **kwargs):
    return ncut_run(images, **kwargs)

@spaces.GPU(duration=120)
def super_duper_long_run(images, **kwargs):
    return ncut_run(images, **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,
):
    if images is None:
        gr.Warning("No images selected.")
        return [], "No images selected."
    
    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,
    }
    num_images = len(images)
    if num_images > 100:
        return super_duper_long_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)
    
with gr.Blocks() as demo:

    with gr.Row():
        with gr.Column(scale=5, min_width=200):
            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')
        
            gr.Markdown('### Load Examples 👇')
            load_images_button = gr.Button("Load", elem_id="load-images-button")
            hide_button = gr.Button("Hide", elem_id="hide-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.click(
                fn=lambda: gr.update(visible=False),
                outputs=example_gallery
            )

        with gr.Column(scale=5, min_width=200):
            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")
            model_dropdown = gr.Dropdown(["SAM(sam_vit_b)", "MobileSAM", "DiNO(dinov2_vitb14_reg)", "CLIP(openai/clip-vit-base-patch16)"], label="Model", value="SAM(sam_vit_b)", elem_id="model_name")
            layer_slider = gr.Slider(0, 11, step=1, label="Layer", value=11, elem_id="layer")
            num_eig_slider = gr.Slider(1, 1000, step=1, label="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="Affinity focal gamma", value=0.3, elem_id="affinity_focal_gamma", info="decrease for shaper NCUT")
            
            with gr.Accordion("Additional Parameters", open=False):
                node_type_dropdown = gr.Dropdown(["attn: attention output", "mlp: mlp output", "block: sum of residual"], label="Node type", value="block: sum of residual", elem_id="node_type", info="which feature to take from each layer?")
                num_sample_ncut_slider = gr.Slider(100, 50000, step=100, label="num_sample (NCUT)", value=10000, elem_id="num_sample_ncut", info="Nyström approximation")
                knn_ncut_slider = gr.Slider(1, 100, step=1, label="KNN (NCUT)", value=10, elem_id="knn_ncut", info="Nyström approximation")
                embedding_method_dropdown = gr.Dropdown(["t-SNE", "UMAP"], label="Embedding method", value="t-SNE", elem_id="embedding_method")
                num_sample_tsne_slider = gr.Slider(100, 1000, step=100, label="num_sample (t-SNE/UMAP)", value=300, elem_id="num_sample_tsne", info="Nyström approximation")
                knn_tsne_slider = gr.Slider(1, 100, step=1, label="KNN (t-SNE/UMAP)", value=10, elem_id="knn_tsne", info="Nyström approximation")
                perplexity_slider = gr.Slider(10, 500, step=10, label="Perplexity (t-SNE)", value=150, elem_id="perplexity")
                n_neighbors_slider = gr.Slider(10, 500, step=10, label="n_neighbors (UMAP)", value=150, elem_id="n_neighbors")
                min_dist_slider = gr.Slider(0.1, 1, step=0.1, label="min_dist (UMAP)", value=0.1, elem_id="min_dist")

            # logging text box
            logging_text = gr.Textbox("Logging information", label="Logging", elem_id="logging", type="text", placeholder="Logging information")
            
    def load_default_images():
        return default_images, default_outputs

    def empty_input_and_output():
        return [], []
    
    load_images_button.click(load_default_images, outputs=[input_gallery, output_gallery])
    clear_images_button.click(empty_input_and_output, 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
        ],
        outputs=[output_gallery, logging_text]
    )


demo.launch()