File size: 4,532 Bytes
00b1139
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import sys
import time
import glob
import random
import skimage
import skimage.io
import numpy as np
from skimage import io, color

import skimage
import skimage.io
from PIL import Image
import cv2

import torch
import torch.nn as nn
from torch.nn import functional as F

import timm
import torchvision
from torchvision.models.feature_extraction import create_feature_extractor

L_range = 100
ab_min = -128
ab_max = 127
ab_range = ab_max - ab_min   
        
def extract_zip(input_zip):
    input_zip=ZipFile(input_zip)
    return {name: input_zip.read(name) for name in input_zip.namelist()}

def normalize_lab_channels(x):
    # Normalize L
    x[:,:,0] = x[:,:,0] / L_range

    # Normalize AB     
    x[:,:,1] = (x[:,:,1]-ab_min) / ab_range        
    x[:,:,2] = (x[:,:,2]-ab_min) / ab_range
    return x

def normalized_lab_to_rgb(lab):
    lab[:,:,0] = (lab[:,:,0] * L_range)
    lab[:,:,1] = (lab[:,:,1] * ab_range) + ab_min 
    lab[:,:,2] = (lab[:,:,2] * ab_range) + ab_min 
    return color.lab2rgb(lab)

def torch_normalized_lab_to_rgb(lab):
    for i in range(lab.shape[0]):
        lab[i,0,:,:] = torch.clip(lab[i,0,:,:] * L_range, 0, L_range)
        lab[i,1,:,:] = torch.clip((lab[i,1,:,:] * ab_range) + ab_min, ab_min, ab_max)
        lab[i,2,:,:] = torch.clip((lab[i,2,:,:] * ab_range) + ab_min, ab_min, ab_max)
        
    for i in range(lab.shape[0]):
        lab[i] = torch.from_numpy( color.lab2rgb(lab[i].permute(1,2,0).detach().cpu().numpy()) ).permute(2,0,1)
        
    return lab

class Encoder(nn.Module):
    def __init__(self):
        super(Encoder, self).__init__()       
        self.backend_model = timm.create_model('efficientnetv2_rw_s', pretrained=True)        
        self.backend = create_feature_extractor(self.backend_model, 
            return_nodes=['blocks.0', 'blocks.1', 'blocks.2', 'blocks.3', 'act2']) 

    def forward(self, x):
        features = self.backend(x)
        return list(features.values())

class UpSample(nn.Sequential):
    def __init__(self, in_channels, out_channels):
        skip_input, output_features = in_channels, out_channels
        
        super(UpSample, self).__init__()        
        self.convA = nn.Conv2d(skip_input, output_features, kernel_size=3, stride=1, padding=1, padding_mode='reflect', bias=False)
        self.leakyreluA = nn.LeakyReLU(0.2)
        self.convB = nn.Conv2d(output_features, output_features, kernel_size=3, stride=1, padding=1, padding_mode='reflect', bias=False)
        self.leakyreluB = nn.LeakyReLU(0.2)
        
        self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False)

    def forward(self, x, concat_with=None):
        up_x = self.upsample(x)
        
        if concat_with is not None:
            up_x = torch.cat([up_x, concat_with], dim=1)
        
        return self.leakyreluB( self.convB( self.leakyreluA( self.convA( up_x ) )  )  )
    
class Decoder(nn.Module):
    def __init__(self, num_features=1792 * 1, decoder_width=None):
        super(Decoder, self).__init__()
        features = int(num_features * decoder_width)
        
        self.conv2 =  nn.Sequential(
            nn.Conv2d(num_features, features, kernel_size=1, stride=1, padding=0, bias=False),
            nn.LeakyReLU(0.2),
        )

        self.up1 = UpSample(in_channels=features//1 + 152 - 24, out_channels=features//2)
        self.up2 = UpSample(in_channels=features//2 + 80 - 16, out_channels=features//4)
        self.up3 = UpSample(in_channels=features//4 + 56 - 8, out_channels=features//8)
        self.up4 = UpSample(in_channels=features//8 + 32 - 8, out_channels=features//16)    
        
        self.up5 = UpSample(in_channels=features//16, out_channels=features//16) 

        self.conv3 = nn.Conv2d(features//16, 2, kernel_size=1, stride=1, padding=0, bias=False)
        
    def forward(self, features):
        blocks0, blocks1, blocks2, blocks3, x = features
        
        x = self.conv2(x)
        
        x = self.up1(x, blocks3)
        x = self.up2(x, blocks2)
        x = self.up3(x, blocks1)
        x = self.up4(x, blocks0)
        
        x = self.up5(x)
        
        x_final = self.conv3(x)
        
        return x_final
    
class ColorizeNet(nn.Module):
    def __init__(self, decoder_width):
        super(ColorizeNet, self).__init__()
        self.encoder = Encoder()
        self.decoder = Decoder(decoder_width=decoder_width)

    def forward(self, x):        
        features_x = self.encoder(x)
        return self.decoder( features_x )