File size: 6,168 Bytes
6df18f5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
"""
"Edited Alpha-CLIP", as proposed in the CLick2Mask paper https://arxiv.org/abs/2409.08272.
Evaluates the similarity between the masked edited region, and the un-localized prompt
(prompt without the word indicating addition ('add', 'insert', etc.),
and without the location to be edited.
A mask indicating the edit made is extracted automatically,
and a similarity is calculated between the masked output and the un-localized prompt, using Alpha-CLIP.

Can optionally output the image with the extracted masks overlayed.
"""
import os

import torch
from torchvision import transforms
from PIL import Image
import torch.nn as nn
import numpy as np
import cv2
from einops import rearrange
import warnings
warnings.filterwarnings("ignore", message="PyTorch version 1.7.1 or higher is recommended")
import alpha_clip

DEST_SIZE = (512, 512)


def make_overlay(im, mask, alpha=1.0, beta=0.5):
    mask[:, :, 0] = 0
    mask[:, :, 2] = 0
    ret = cv2.addWeighted(im, alpha, mask, beta, 0)
    ret = np.clip(ret, a_min=0, a_max=1)
    return ret


class EditedAlphaCLip:
    def __init__(self, ac_scale=336, device="cuda:0"):
        assert ac_scale in (224, 336)
        self.device = device
        self.ac_size = (ac_scale, ac_scale)
        if self.ac_size == (336, 336):
            self.ac_model, self.ac_preprocess = alpha_clip.load(
                "ViT-L/14@336px",
                alpha_vision_ckpt_pth="./checkpoints/clip_l14_336_grit1m_fultune_8xe.pth",
                device=self.device,
            )
        else:
            self.ac_model, self.ac_preprocess = alpha_clip.load(
                "ViT-L/14",
                alpha_vision_ckpt_pth="./checkpoints/clip_l14_grit20m_fultune_2xe.pth",
                device=self.device,
            )

        self.im_to_sqz32 = lambda x: rearrange(
            x.cpu().numpy().squeeze().astype(np.float32), "c h w -> h w c"
        )
        self.im_to_cat_32 = lambda x: rearrange(
            torch.stack([x.squeeze(0).squeeze(0)] * 3, dim=0)
            .cpu()
            .numpy()
            .astype(np.float32),
            "c h w -> h w c",
        )

    def save_im(self, im, path):
        os.makedirs(os.path.dirname(path), exist_ok=True)
        Image.fromarray((im * 255).round().astype("uint8")).save(path, quality=95)

    def read_image(self, img_path):
        image = Image.open(img_path).convert("RGB")
        if image.size != DEST_SIZE:
            image = image.resize(DEST_SIZE, Image.LANCZOS)
        image = np.array(image)
        image = image.astype(np.float32) / 255.0
        image = image[None].transpose(0, 3, 1, 2)
        image = torch.from_numpy(image).to(self.device)
        return image

    # Edited Alpha-CLIP (higher is better)
    @torch.no_grad()
    def edited_alpha_clip_sim(self, image_in_p, image_out_p, prompt, save_outs=None):
        """
        Args:
            image_in_p: The input image path
            image_out_p: The output image path
            prompt: The un-localized prompt (as explained above)
            save_outs: If given, will save:
                * The output image with extracted mask overlayed to <save_outs>_out_masked.jpg,
                * The output image to <save_outs>_out.jpg.
                * The input image to <save_outs>_in.jpg.
            All in size (512, 512).
        """
        assert type(prompt) is str
        prompt = [prompt]
        image_in = self.read_image(image_in_p)
        image_out = self.read_image(image_out_p)

        mask_transform = transforms.Compose(
            [nn.AdaptiveAvgPool2d(self.ac_size), transforms.Normalize(0.5, 0.26)]
        )
        image_transform = transforms.Compose(
            [
                transforms.Resize(self.ac_size, interpolation=Image.BICUBIC),
                transforms.Normalize(
                    (0.48145466, 0.4578275, 0.40821073),
                    (0.26862954, 0.26130258, 0.27577711),
                ),
            ]
        )

        mask = self.extract_mask(image_in=image_in, image_out=image_out)
        alpha = mask_transform(mask).half()
        image = image_transform(image_out).half()
        image_features = self.ac_model.visual(image, alpha)
        image_features = image_features / image_features.norm(dim=-1, keepdim=True)

        text = alpha_clip.tokenize(prompt).to(self.device)
        text_features = self.ac_model.encode_text(text)
        text_features = text_features / text_features.norm(dim=-1, keepdim=True)

        alpha_loss = image_features @ text_features.T

        alpha_loss = alpha_loss.mean(dim=0)

        if save_outs:
            self.save_im(
                make_overlay(self.im_to_sqz32(image_out), self.im_to_cat_32(mask)),
                f"{save_outs}_out_masked.jpg",
            )
            self.save_im(self.im_to_sqz32(image_out), f"{save_outs}_out.jpg")
            self.save_im(self.im_to_sqz32(image_in), f"{save_outs}_in.jpg")

        return alpha_loss

    def create_multiple_convex_hulls(self, binary_mask, min_hull_area=100):
        if binary_mask.is_cuda:
            binary_mask = binary_mask.cpu()
        np_mask = binary_mask.squeeze().numpy().astype(np.uint8)

        contours, _ = cv2.findContours(
            np_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE
        )
        if not contours:
            return binary_mask

        all_hulls_mask = np.zeros_like(np_mask)
        for contour in contours:
            hull = cv2.convexHull(contour)
            hull_area = cv2.contourArea(hull)
            if hull_area >= min_hull_area:
                cv2.drawContours(all_hulls_mask, [hull], 0, 1, -1)

        hull_tensor = torch.from_numpy(all_hulls_mask).unsqueeze(0).unsqueeze(0).half()
        hull_tensor = hull_tensor.to(self.device)

        return hull_tensor

    def extract_mask(self, image_in, image_out):
        mask = (torch.mean(torch.abs(image_in - image_out), dim=1) > 0.1).half()
        pool_for_min = nn.MaxPool2d(3, stride=1, padding=1)
        mask = -pool_for_min(-mask)
        pool_for_max = nn.MaxPool2d(5, stride=1, padding=2)
        mask = pool_for_max(mask)
        mask = self.create_multiple_convex_hulls(mask)

        return mask