app.py

from typing import Optional, Tuple
from einops import rearrange
import torch
import torch.nn.functional as F
from PIL import Image
from torch import nn
import numpy as np
import os
import time

import gradio as gr

import spaces

USE_CUDA = torch.cuda.is_available()

print("CUDA is available:", USE_CUDA)

def transform_images(images, resolution=(1024, 1024)):
    images = [image.convert("RGB").resize(resolution) for image in images]
    # Convert to torch tensor
    images = [torch.tensor(np.array(image).transpose(2, 0, 1)).float() / 255 for image in images]
    # Normalize
    images = [(image - 0.5) / 0.5 for image in images]
    images = torch.stack(images)
    return images

class MobileSAM(nn.Module):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)

        from mobile_sam import sam_model_registry

        url = 'https://raw.githubusercontent.com/ChaoningZhang/MobileSAM/master/weights/mobile_sam.pt'
        model_type = "vit_t"
        sam_checkpoint = "mobile_sam.pt"
        if not os.path.exists(sam_checkpoint):
            import requests
            r = requests.get(url)
            with open(sam_checkpoint, 'wb') as f:
                f.write(r.content)

        device = 'cuda' if USE_CUDA else 'cpu'

        mobile_sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)

        def new_forward_fn(self, x):
            shortcut = x

            x = self.conv1(x)
            x = self.act1(x)

            x = self.conv2(x)
            x = self.act2(x)
            
            self.attn_output = rearrange(x.clone(), "b c h w -> b h w c")

            x = self.conv3(x)
            
            self.mlp_output = rearrange(x.clone(), "b c h w -> b h w c")

            x = self.drop_path(x)

            x += shortcut
            x = self.act3(x)
            
            self.block_output = rearrange(x.clone(), "b c h w -> b h w c")

            return x

        setattr(mobile_sam.image_encoder.layers[0].blocks[0].__class__, "forward", new_forward_fn)

        def new_forward_fn2(self, x):
            H, W = self.input_resolution
            B, L, C = x.shape
            assert L == H * W, "input feature has wrong size"
            res_x = x
            if H == self.window_size and W == self.window_size:
                x = self.attn(x)
            else:
                x = x.view(B, H, W, C)
                pad_b = (self.window_size - H %
                            self.window_size) % self.window_size
                pad_r = (self.window_size - W %
                            self.window_size) % self.window_size
                padding = pad_b > 0 or pad_r > 0

                if padding:
                    x = F.pad(x, (0, 0, 0, pad_r, 0, pad_b))

                pH, pW = H + pad_b, W + pad_r
                nH = pH // self.window_size
                nW = pW // self.window_size
                # window partition
                x = x.view(B, nH, self.window_size, nW, self.window_size, C).transpose(2, 3).reshape(
                    B * nH * nW, self.window_size * self.window_size, C)
                x = self.attn(x)
                # window reverse
                x = x.view(B, nH, nW, self.window_size, self.window_size,
                            C).transpose(2, 3).reshape(B, pH, pW, C)

                if padding:
                    x = x[:, :H, :W].contiguous()

                x = x.view(B, L, C)

            hw = np.sqrt(x.shape[1]).astype(int)
            self.attn_output = rearrange(x.clone(), "b (h w) c -> b h w c", h=hw)
            
            x = res_x + self.drop_path(x)

            x = x.transpose(1, 2).reshape(B, C, H, W)
            x = self.local_conv(x)
            x = x.view(B, C, L).transpose(1, 2)

            mlp_output = self.mlp(x)
            self.mlp_output = rearrange(mlp_output.clone(), "b (h w) c -> b h w c", h=hw)
            
            x = x + self.drop_path(mlp_output)
            self.block_output = rearrange(x.clone(), "b (h w) c -> b h w c", h=hw)
            return x

        setattr(mobile_sam.image_encoder.layers[1].blocks[0].__class__, "forward", new_forward_fn2)


        mobile_sam.to(device=device)
        mobile_sam.eval()
        self.image_encoder = mobile_sam.image_encoder
        
    
    @torch.no_grad()
    def forward(self, x):
        with torch.no_grad():
            x = torch.nn.functional.interpolate(x, size=(1024, 1024), mode="bilinear")
        out = self.image_encoder(x)

        attn_outputs, mlp_outputs, block_outputs = [], [], []
        for i_layer in range(len(self.image_encoder.layers)):
            for i_block in range(len(self.image_encoder.layers[i_layer].blocks)):
                blk = self.image_encoder.layers[i_layer].blocks[i_block]
                attn_outputs.append(blk.attn_output)
                mlp_outputs.append(blk.mlp_output)
                block_outputs.append(blk.block_output)
        return attn_outputs, mlp_outputs, block_outputs

mobilesam = MobileSAM()

def image_mobilesam_feature(
    images,
    node_type="block",
    layer=-1,
):

    if USE_CUDA:
        images = images.cuda()

    feat_extractor = mobilesam
    if USE_CUDA:
        feat_extractor = feat_extractor.cuda()

    # attn_outputs, mlp_outputs, block_outputs = [], [], []
    outputs = []
    for i in range(images.shape[0]):
        attn_output, mlp_output, block_output = feat_extractor(
            images[i].unsqueeze(0)
        )
        out_dict = {
            "attn": attn_output,
            "mlp": mlp_output,
            "block": block_output,
        }
        out = out_dict[node_type]
        out = out[layer]
        outputs.append(out.cpu())
    outputs = torch.cat(outputs, dim=0)
    
    mobilesam = mobilesam.cpu()
    return outputs



class SAM(torch.nn.Module):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        from segment_anything import sam_model_registry, SamPredictor
        from segment_anything.modeling.sam import Sam
        
        checkpoint = "sam_vit_b_01ec64.pth"
        if not os.path.exists(checkpoint):
            checkpoint_url = 'https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth'
            import requests
            r = requests.get(checkpoint_url)
            with open(checkpoint, 'wb') as f:
                f.write(r.content)

        sam: Sam = sam_model_registry["vit_b"](checkpoint=checkpoint)

        from segment_anything.modeling.image_encoder import (
            window_partition,
            window_unpartition,
        )

        def new_block_forward(self, x: torch.Tensor) -> torch.Tensor:
            shortcut = x
            x = self.norm1(x)
            # Window partition
            if self.window_size > 0:
                H, W = x.shape[1], x.shape[2]
                x, pad_hw = window_partition(x, self.window_size)

            x = self.attn(x)
            # Reverse window partition
            if self.window_size > 0:
                x = window_unpartition(x, self.window_size, pad_hw, (H, W))
            self.attn_output = x.clone()

            x = shortcut + x
            mlp_outout = self.mlp(self.norm2(x))
            self.mlp_output = mlp_outout.clone()
            x = x + mlp_outout
            self.block_output = x.clone()

            return x

        setattr(sam.image_encoder.blocks[0].__class__, "forward", new_block_forward)

        self.image_encoder = sam.image_encoder
        self.image_encoder.eval()
        if USE_CUDA:
            self.image_encoder = self.image_encoder.cuda()

    @torch.no_grad()
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        with torch.no_grad():
            x = torch.nn.functional.interpolate(x, size=(1024, 1024), mode="bilinear")
        out = self.image_encoder(x)

        attn_outputs, mlp_outputs, block_outputs = [], [], []
        for i, blk in enumerate(self.image_encoder.blocks):
            attn_outputs.append(blk.attn_output)
            mlp_outputs.append(blk.mlp_output)
            block_outputs.append(blk.block_output)
        attn_outputs = torch.stack(attn_outputs)
        mlp_outputs = torch.stack(mlp_outputs)
        block_outputs = torch.stack(block_outputs)
        return attn_outputs, mlp_outputs, block_outputs

sam = SAM()

def image_sam_feature(
    images,
    node_type="block",
    layer=-1,
):

    if USE_CUDA:
        images = images.cuda()

    feat_extractor = sam
    if USE_CUDA:
        feat_extractor = feat_extractor.cuda()

    # attn_outputs, mlp_outputs, block_outputs = [], [], []
    outputs = []
    for i in range(images.shape[0]):
        attn_output, mlp_output, block_output = feat_extractor(
            images[i].unsqueeze(0)
        )
        out_dict = {
            "attn": attn_output,
            "mlp": mlp_output,
            "block": block_output,
        }
        out = out_dict[node_type]
        out = out[layer]
        outputs.append(out.cpu())
    outputs = torch.cat(outputs, dim=0)
    
    sam = sam.cpu()
    
    return outputs


class DiNOv2(torch.nn.Module):
    def __init__(self, ver="dinov2_vitb14_reg"):
        super().__init__()
        self.dinov2 = torch.hub.load("facebookresearch/dinov2", ver)
        self.dinov2.requires_grad_(False)
        self.dinov2.eval()
        if USE_CUDA:
            self.dinov2 = self.dinov2.cuda()

        def new_block_forward(self, x: torch.Tensor) -> torch.Tensor:
            def attn_residual_func(x):
                return self.ls1(self.attn(self.norm1(x)))

            def ffn_residual_func(x):
                return self.ls2(self.mlp(self.norm2(x)))

            attn_output = attn_residual_func(x)
            self.attn_output = attn_output.clone()
            x = x + attn_output
            mlp_output = ffn_residual_func(x)
            self.mlp_output = mlp_output.clone()
            x = x + mlp_output
            block_output = x
            self.block_output = block_output.clone()
            return x

        setattr(self.dinov2.blocks[0].__class__, "forward", new_block_forward)

    @torch.no_grad()
    def forward(self, x):

        out = self.dinov2(x)

        attn_outputs, mlp_outputs, block_outputs = [], [], []
        for i, blk in enumerate(self.dinov2.blocks):
            attn_outputs.append(blk.attn_output)
            mlp_outputs.append(blk.mlp_output)
            block_outputs.append(blk.block_output)

        attn_outputs = torch.stack(attn_outputs)
        mlp_outputs = torch.stack(mlp_outputs)
        block_outputs = torch.stack(block_outputs)
        return attn_outputs, mlp_outputs, block_outputs

dinov2 = DiNOv2()

def image_dino_feature(images, node_type="block", layer=-1):

    if USE_CUDA:
        images = images.cuda()

    feat_extractor = dinov2
    if USE_CUDA:
        feat_extractor = feat_extractor.cuda()

    # attn_outputs, mlp_outputs, block_outputs = [], [], []
    outputs = []
    for i in range(images.shape[0]):
        attn_output, mlp_output, block_output = feat_extractor(
            images[i].unsqueeze(0)
        )
        out_dict = {
            "attn": attn_output,
            "mlp": mlp_output,
            "block": block_output,
        }
        out = out_dict[node_type]
        out = out[layer]
        outputs.append(out.cpu())
    outputs = torch.cat(outputs, dim=0)
    outputs = rearrange(outputs[:, 5:, :], "b (h w) c -> b h w c", h=32, w=32)
    
    dinov2 = dinov2.cpu()
    return outputs


class CLIP(torch.nn.Module):
    def __init__(self):
        super().__init__()

        from transformers import CLIPProcessor, CLIPModel

        model = CLIPModel.from_pretrained("openai/clip-vit-base-patch16")
        # processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch16")
        self.model = model.eval()
        if USE_CUDA:
            self.model = self.model.cuda()

        def new_forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: torch.Tensor,
            causal_attention_mask: torch.Tensor,
            output_attentions: Optional[bool] = False,
        ) -> Tuple[torch.FloatTensor]:

            residual = hidden_states

            hidden_states = self.layer_norm1(hidden_states)
            hidden_states, attn_weights = self.self_attn(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                causal_attention_mask=causal_attention_mask,
                output_attentions=output_attentions,
            )
            self.attn_output = hidden_states.clone()
            hidden_states = residual + hidden_states

            residual = hidden_states
            hidden_states = self.layer_norm2(hidden_states)
            hidden_states = self.mlp(hidden_states)
            self.mlp_output = hidden_states.clone()
            
            hidden_states = residual + hidden_states

            outputs = (hidden_states,)

            if output_attentions:
                outputs += (attn_weights,)

            self.block_output = hidden_states.clone()
            return outputs
        
        setattr(self.model.vision_model.encoder.layers[0].__class__, "forward", new_forward)

    @torch.no_grad()
    def forward(self, x): 

        out = self.model.vision_model(x)

        attn_outputs, mlp_outputs, block_outputs = [], [], []
        for i, blk in enumerate(self.model.vision_model.encoder.layers):
            attn_outputs.append(blk.attn_output)
            mlp_outputs.append(blk.mlp_output)
            block_outputs.append(blk.block_output)

        attn_outputs = torch.stack(attn_outputs)
        mlp_outputs = torch.stack(mlp_outputs)
        block_outputs = torch.stack(block_outputs)
        return attn_outputs, mlp_outputs, block_outputs

clip = CLIP()

def image_clip_feature(
    images, node_type="block", layer=-1
):
    if USE_CUDA:
        images = images.cuda()

    feat_extractor = clip
    if USE_CUDA:
        feat_extractor = feat_extractor.cuda()

    # attn_outputs, mlp_outputs, block_outputs = [], [], []
    outputs = []
    for i in range(images.shape[0]):
        attn_output, mlp_output, block_output = feat_extractor(
            images[i].unsqueeze(0)
        )
        out_dict = {
            "attn": attn_output,
            "mlp": mlp_output,
            "block": block_output,
        }
        out = out_dict[node_type]
        out = out[layer]
        outputs.append(out.cpu())
    outputs = torch.cat(outputs, dim=0)
    
    clip = clip.cpu()
    return outputs



import hashlib
import pickle
import sys
from collections import OrderedDict

# Cache dictionary with limited size
class LimitedSizeCache(OrderedDict):
    def __init__(self, max_size_bytes):
        self.max_size_bytes = max_size_bytes
        self.current_size_bytes = 0
        super().__init__()

    def __setitem__(self, key, value):
        item_size = self.get_item_size(value)
        # Evict items until there is enough space
        while self.current_size_bytes + item_size > self.max_size_bytes:
            self.popitem(last=False)
        super().__setitem__(key, value)
        self.current_size_bytes += item_size

    def __delitem__(self, key):
        value = self[key]
        super().__delitem__(key)
        self.current_size_bytes -= self.get_item_size(value)

    def get_item_size(self, value):
        """Estimate the size of the value in bytes."""
        return sys.getsizeof(value)

# Initialize the cache with a 4GB limit
cache = LimitedSizeCache(max_size_bytes=4 * 1024 * 1024 * 1024)  # 4GB

def compute_hash(*args, **kwargs):
    """Compute a unique hash based on the function arguments."""
    hasher = hashlib.sha256()
    pickled_args = pickle.dumps((args, kwargs))
    hasher.update(pickled_args)
    return hasher.hexdigest()


@spaces.GPU(duration=30)
def run_model_on_image(images, model_name="sam", node_type="block", layer=-1):
    global USE_CUDA
    USE_CUDA = True
    
    if model_name == "SAM(sam_vit_b)":
        if not USE_CUDA:
            gr.warning("GPU not detected. Running SAM on CPU, ~30s/image.")
        result = image_sam_feature(images, node_type=node_type, layer=layer)
    elif model_name == 'MobileSAM':
        result = image_mobilesam_feature(images, node_type=node_type, layer=layer)
    elif model_name == "DiNO(dinov2_vitb14_reg)":
        result = image_dino_feature(images, node_type=node_type, layer=layer)
    elif model_name == "CLIP(openai/clip-vit-base-patch16)":
        result = image_clip_feature(images, node_type=node_type, layer=layer)
    else:
        raise ValueError(f"Model {model_name} not supported.")
    
    USE_CUDA = False
    return result

def extract_features(images, model_name="mobilesam", node_type="block", layer=-1):
    resolution_dict = {
        "mobilesam": (1024, 1024),
        "sam(sam_vit_b)": (1024, 1024),
        "dinov2(dinov2_vitb14_reg)": (448, 448),
        "clip(openai/clip-vit-base-patch16)": (224, 224),
    }
    images = transform_images(images, resolution=resolution_dict[model_name])
    
    # Compute the cache key
    cache_key = compute_hash(images, model_name, node_type, layer)

    # Check if the result is already in the cache
    if cache_key in cache:
        print("Cache hit!")
        return cache[cache_key]
    
    result = run_model_on_image(images, model_name=model_name, node_type=node_type, layer=layer)

    # Store the result in the cache
    cache[cache_key] = result
    
    return result

def compute_ncut(
    features,
    num_eig=100,
    num_sample_ncut=10000,
    affinity_focal_gamma=0.3,
    knn_ncut=10,
    knn_tsne=10,
    num_sample_tsne=1000,
    perplexity=500,
):
    from ncut_pytorch import NCUT, rgb_from_tsne_3d

    start = time.time()
    eigvecs, eigvals = NCUT(
        num_eig=num_eig,
        num_sample=num_sample_ncut,
        device="cpu",
        affinity_focal_gamma=affinity_focal_gamma,
        knn=knn_ncut,
    ).fit_transform(features.reshape(-1, features.shape[-1]))
    print(f"NCUT time (cpu): {time.time() - start:.2f}s")
    
    start = time.time()
    X_3d, rgb = rgb_from_tsne_3d(
        eigvecs,
        num_sample=num_sample_tsne,
        perplexity=perplexity,
        knn=knn_tsne,
    )
    print(f"t-SNE time (cpu): {time.time() - start:.2f}s")
    
    # print("input shape:", features.shape)
    # print("output shape:", rgb.shape)
    
    rgb = rgb.reshape(features.shape[:3] + (3,))
    return rgb


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
    ]


def main_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,
    num_sample_tsne=1000,
    knn_tsne=10,
    perplexity=500,
):
    if perplexity >= num_sample_tsne:
        # raise gr.Error("Perplexity must be less than the number of samples for t-SNE.")
        gr.Warning("Perplexity must be less than the number of samples for t-SNE.\n" f"Setting perplexity to {num_sample_tsne-1}.")
        perplexity = num_sample_tsne - 1
    
    images = [image[0] for image in images]
    
    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")
    
    rgb = 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,
        perplexity=perplexity,
    )
    rgb = dont_use_too_much_green(rgb)
    return to_pil_images(rgb)

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']

demo = gr.Interface(
    main_fn,
    [
        gr.Gallery(value=default_images, label="Select images", show_label=False, elem_id="images", columns=[3], rows=[1], object_fit="contain", height="auto", type="pil"),
        gr.Dropdown(["MobileSAM", "SAM(sam_vit_b)", "DiNO(dinov2_vitb14_reg)", "CLIP(openai/clip-vit-base-patch16)"], label="Model", value="MobileSAM", elem_id="model_name"),
        gr.Slider(0, 11, step=1, label="Layer", value=11, elem_id="layer", info="which layer of the image backbone features"),
        gr.Slider(1, 1000, step=1, label="Number of eigenvectors", value=100, elem_id="num_eig", info='increase for more object parts, decrease for whole object'),
    ],
    gr.Gallery(value=default_outputs, label="NCUT Embedding", show_label=False, elem_id="ncut", columns=[3], rows=[1], object_fit="contain", height="auto"),
    additional_inputs=[
        gr.Dropdown(["attn", "mlp", "block"], label="Node type", value="block", elem_id="node_type", info="attn: attention output, mlp: mlp output, block: sum of residual stream"),
        gr.Slider(0.01, 1, step=0.01, label="Affinity focal gamma", value=0.3, elem_id="affinity_focal_gamma", info="decrease for more aggressive cleaning on the affinity matrix"),
        gr.Slider(100, 10000, step=100, label="num_sample (NCUT)", value=5000, elem_id="num_sample_ncut", info="for Nyström approximation"),
        gr.Slider(1, 100, step=1, label="KNN (NCUT)", value=10, elem_id="knn_ncut", info="for Nyström approximation"),
        gr.Slider(100, 1000, step=100, label="num_sample (t-SNE)", value=500, elem_id="num_sample_tsne", info="for Nyström approximation. Adding will slow down t-SNE quite a lot"),
        gr.Slider(1, 100, step=1, label="KNN (t-SNE)", value=10, elem_id="knn_tsne", info="for Nyström approximation"),
        gr.Slider(10, 500, step=10, label="Perplexity (t-SNE)", value=250, elem_id="perplexity", info="for t-SNE"),
        
    ]
)

demo.launch()