File size: 5,405 Bytes
0145b71
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4a4961c
0145b71
 
 
 
 
 
 
 
 
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
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from torch.nn import Conv2d, BatchNorm2d, PReLU, Sequential, Module
import math

from .helpers import get_blocks, bottleneck_IR, bottleneck_IR_SE

import sys, os
sys.path.append(os.path.dirname(__file__) + os.sep + '../')
from model import EqualLinear


"""
Modified from [pSp](https://github.com/eladrich/pixel2style2pixel)
"""


class GradualStyleBlock(Module):
    def __init__(self, in_c, out_c, spatial):
        super(GradualStyleBlock, self).__init__()
        self.out_c = out_c
        self.spatial = spatial
        num_pools = int(np.log2(spatial))
        modules = []
        modules += [Conv2d(in_c, out_c, kernel_size=3, stride=2, padding=1),
                    nn.LeakyReLU()]
        for i in range(num_pools - 1):
            modules += [
                Conv2d(out_c, out_c, kernel_size=3, stride=2, padding=1),
                nn.LeakyReLU()
            ]
        self.convs = nn.Sequential(*modules)
        self.linear = EqualLinear(out_c, out_c, lr_mul=1)

    def forward(self, x):
        x = self.convs(x)
        x = x.view(-1, self.out_c)
        x = self.linear(x)
        return x


class GradualStyleEncoder(Module):
    def __init__(self, num_layers, mode='ir', n_styles=18):
        super(GradualStyleEncoder, self).__init__()
        assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152'
        assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se'
        blocks = get_blocks(num_layers)
        if mode == 'ir':
            unit_module = bottleneck_IR
        elif mode == 'ir_se':
            unit_module = bottleneck_IR_SE
        self.input_layer = Sequential(Conv2d(3, 64, (3, 3), 1, 1, bias=False),
                                      BatchNorm2d(64),
                                      PReLU(64))
        modules = []
        for block in blocks:
            for bottleneck in block:
                modules.append(unit_module(bottleneck.in_channel,
                                           bottleneck.depth,
                                           bottleneck.stride))
        self.body = Sequential(*modules)

        self.styles = nn.ModuleList()
        self.style_count = n_styles  # opts.n_styles
        self.coarse_ind = 3
        self.middle_ind = 7
        for i in range(self.style_count):
            if i < self.coarse_ind:
                style = GradualStyleBlock(512, 512, 16)
            elif i < self.middle_ind:
                style = GradualStyleBlock(512, 512, 32)
            else:
                style = GradualStyleBlock(512, 512, 64)
            self.styles.append(style)
        self.latlayer1 = nn.Conv2d(256, 512, kernel_size=1, stride=1, padding=0)
        self.latlayer2 = nn.Conv2d(128, 512, kernel_size=1, stride=1, padding=0)

    def _upsample_add(self, x, y):
        '''Upsample and add two feature maps.
        Args:
          x: (Variable) top feature map to be upsampled.
          y: (Variable) lateral feature map.
        Returns:
          (Variable) added feature map.
        Note in PyTorch, when input size is odd, the upsampled feature map
        with `F.upsample(..., scale_factor=2, mode='nearest')`
        maybe not equal to the lateral feature map size.
        e.g.
        original input size: [N,_,15,15] ->
        conv2d feature map size: [N,_,8,8] ->
        upsampled feature map size: [N,_,16,16]
        So we choose bilinear upsample which supports arbitrary output sizes.
        '''
        _, _, H, W = y.size()
        return F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True) + y

    def forward(self, x):
        x = self.input_layer(x)

        latents = []
        modulelist = list(self.body._modules.values())
        for i, l in enumerate(modulelist):
            x = l(x)
            if i == 6:
                c1 = x
            elif i == 20:
                c2 = x
            elif i == 23:
                c3 = x

        for j in range(self.coarse_ind):
            latents.append(self.styles[j](c3))

        p2 = self._upsample_add(c3, self.latlayer1(c2))
        for j in range(self.coarse_ind, self.middle_ind):
            latents.append(self.styles[j](p2))

        p1 = self._upsample_add(p2, self.latlayer2(c1))
        for j in range(self.middle_ind, self.style_count):
            latents.append(self.styles[j](p1))

        out = torch.stack(latents, dim=1)
        return out


def get_keys(d, name):
    if 'state_dict' in d:
        d = d['state_dict']
    d_filt = {k[len(name) + 1:]: v for k, v in d.items() if k[:len(name)] == name}
    return d_filt


class PSPEncoder(Module):
    def __init__(self, encoder_ckpt_path, output_size=1024):
        super(PSPEncoder, self).__init__()
        n_styles = int(math.log(output_size, 2)) * 2 - 2
        self.encoder = GradualStyleEncoder(50, 'ir_se', n_styles)

        print('Loading psp encoders weights from irse50!')
        encoder_ckpt = torch.load(encoder_ckpt_path, map_location='cpu')
        self.encoder.load_state_dict(get_keys(encoder_ckpt, 'encoder'), strict=True)
        self.latent_avg = encoder_ckpt['latent_avg']

        self.face_pool = torch.nn.AdaptiveAvgPool2d((256, 256))

    def forward(self, x):
        x = self.face_pool(x)
        codes = self.encoder(x)
        codes = codes + self.latent_avg.repeat(codes.shape[0], 1, 1)
        return codes