File size: 4,408 Bytes
929f451
 
 
 
 
 
 
 
 
 
 
 
ee8e678
929f451
774a99f
 
 
 
 
 
 
 
 
 
929f451
 
def8bfd
929f451
 
 
 
 
 
 
 
 
 
 
 
 
 
ac5c2a4
 
 
 
929f451
774a99f
929f451
 
 
 
774a99f
929f451
 
 
 
 
 
 
def8bfd
 
 
 
 
 
929f451
def8bfd
84aa40a
ac5c2a4
 
def8bfd
929f451
ac5c2a4
def8bfd
929f451
ac5c2a4
def8bfd
929f451
ac5c2a4
def8bfd
929f451
ac5c2a4
929f451
ac5c2a4
 
929f451
ac5c2a4
929f451
774a99f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
929f451
 
 
 
 
af03c71
929f451
 
774a99f
 
929f451
 
af03c71
774a99f
929f451
 
774a99f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import gradio as gr

import torch
from PIL import Image
import numpy as np
import random
from einops import rearrange
import matplotlib.pyplot as plt
from torchvision.transforms import v2

from model import MAE_ViT, MAE_Encoder, MAE_Decoder, MAE_Encoder_FeatureExtractor

path  = [['images/cat.jpg'], ['images/dog.jpg'], ['images/horse.jpg'], ['images/airplane.jpg'], ['images/truck.jpg']]
device = torch.device("cpu")

model_name = "model/no_mode/vit-t-mae-pretrain.pt"
model_no_mode = torch.load(model_name, map_location='cpu')
model_no_mode.eval()
model_no_mode.to(device)

model_name = "model/bottom_256/vit-t-mae-pretrain.pt"
model_pca_mode = torch.load(model_name, map_location='cpu')
model_pca_mode.eval()
model_pca_mode.to(device)

transform = v2.Compose([
        v2.Resize((32, 32)),
        v2.ToTensor(), 
        v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
    ])

# Load and Preprocess the Image
def load_image(image_path, transform):
    img = Image.open(image_path).convert('RGB')
    img = transform(img).unsqueeze(0)  # Add batch dimension
    return img

def show_image(img, title):
    img = rearrange(img, "c h w -> h w c")
    img = (img.cpu().detach().numpy() + 1) / 2  # Normalize to [0, 1]

    plt.imshow(img)
    plt.axis('off')
    plt.title(title)

# Visualize a Single Image
def visualize_single_image_no_mode(image_path):
    img = load_image(image_path, transform).to(device)
    
    # Run inference
    with torch.no_grad():
        predicted_img, mask = model_no_mode(img)
    
    # Convert the tensor back to a displayable image
    # masked image
    im_masked = img * (1 - mask)

    # MAE reconstruction pasted with visible patches
    im_paste = img * (1 - mask) + predicted_img * mask

    # resize the image to 96 x 96
    img = v2.functional.resize(img[0], (96, 96))
    im_masked = v2.functional.resize(im_masked[0], (96, 96))
    predicted_img = v2.functional.resize(predicted_img[0], (96, 96))
    im_paste = v2.functional.resize(im_paste[0], (96, 96))
    
    # make the plt figure larger
    plt.figure(figsize=(18, 8))

    plt.subplot(1, 4, 1)
    show_image(img, "original")

    plt.subplot(1, 4, 2)
    show_image(im_masked, "masked")

    plt.subplot(1, 4, 3)
    show_image(predicted_img, "reconstruction")

    plt.subplot(1, 4, 4)
    show_image(im_paste, "reconstruction + visible")

    plt.tight_layout()

    # convert the plt figure to a numpy array
    plt.savefig("output.png")

    return np.array(plt.imread("output.png"))

def visualize_single_image_pca_mode(image_path):
    img = load_image(image_path, transform).to(device)
    
    # Run inference
    with torch.no_grad():
        predicted_img, mask = model_pca_mode(img)
    
    # Convert the tensor back to a displayable image
    # masked image
    im_masked = img * (1 - mask)

    # MAE reconstruction pasted with visible patches
    im_paste = img * (1 - mask) + predicted_img * mask

    # remove the batch dimension
    im_masked = im_masked[0]
    predicted_img = predicted_img[0]
    im_paste = im_paste[0]
    
    # make the plt figure larger
    plt.figure(figsize=(18, 8))

    plt.subplot(1, 4, 1)
    show_image(img, "original")

    plt.subplot(1, 4, 2)
    show_image(im_masked, "masked")

    plt.subplot(1, 4, 3)
    show_image(predicted_img, "reconstruction")

    plt.subplot(1, 4, 4)
    show_image(im_paste, "reconstruction + visible")

    plt.tight_layout()

    # convert the plt figure to a numpy array
    plt.savefig("output.png")

    return np.array(plt.imread("output.png"))

inputs_image = [
    gr.components.Image(type="filepath", label="Input Image"),
]

outputs_image = [
    gr.components.Image(type="numpy", label="Output Image"),
]

inference_no_mode = gr.Interface(
    fn=visualize_single_image_no_mode,
    inputs=inputs_image,
    outputs=outputs_image,
    examples=path,
    cache_examples = False,
    title="MAE-ViT Image Reconstruction",
    description="This is a demo of the MAE-ViT model for image reconstruction.",
)

inference_pca_mode = gr.Interface(
    fn=visualize_single_image_pca_mode,
    inputs=inputs_image,
    outputs=outputs_image,
    examples=path,
    title="MAE-ViT Image Reconstruction",
    description="This is a demo of the MAE-ViT model for image reconstruction.",
)

gr.TabbedInterface(
    [inference_no_mode, inference_pca_mode],
    tab_names=['Normal Mode', 'PCA Mode']
).queue().launch()