File size: 13,596 Bytes
e34aada
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
# SPDX-FileCopyrightText: Copyright (c) 2021-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: LicenseRef-NvidiaProprietary
#
# NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
# property and proprietary rights in and to this material, related
# documentation and any modifications thereto. Any use, reproduction,
# disclosure or distribution of this material and related documentation
# without an express license agreement from NVIDIA CORPORATION or
# its affiliates is strictly prohibited.

"""Discriminator architectures from the paper
"Efficient Geometry-aware 3D Generative Adversarial Networks"."""

import numpy as np
import torch
import torch.nn as nn
from modules.eg3ds.torch_utils.ops import upfirdn2d
from modules.eg3ds.models.networks_stylegan2 import DiscriminatorBlock, MappingNetwork, DiscriminatorEpilogue
from modules.eg3ds.models.cond_encoder import LM3D_Win_Encoder

from utils.commons.hparams import hparams


class SingleDiscriminator(torch.nn.Module):
    def __init__(self,
        img_resolution,                 # Input resolution.
        img_channels        =3,         # Number of input color channels.
        architecture        = 'resnet', # Architecture: 'orig', 'skip', 'resnet'.
        channel_base        = 32768,    # Overall multiplier for the number of channels.
        channel_max         = 512,      # Maximum number of channels in any layer.
        num_fp16_res        = 4,        # Use FP16 for the N highest resolutions.
        conv_clamp          = 256,      # Clamp the output of convolution layers to +-X, None = disable clamping.
        cmap_dim            = None,     # Dimensionality of mapped conditioning label, None = default.
        sr_upsample_factor  = 1,        # Ignored for SingleDiscriminator
        block_kwargs        = {},       # Arguments for DiscriminatorBlock.
        mapping_kwargs      = {},       # Arguments for MappingNetwork.
        epilogue_kwargs     = {},       # Arguments for DiscriminatorEpilogue.
    ):
        super().__init__()
        self.camera_dim = 25
        if hparams['cond_type'] == 'idexp_lm3d_normalized':
            self.cond_dim = 204
        else:
            self.cond_dim = 0
        c_dim = self.camera_dim
        if self.cond_dim > 0:
            cond_out_dim = hparams['cond_out_dim']
            c_dim += cond_out_dim
            self.cond_encoder = LM3D_Win_Encoder(self.cond_dim, hid_dim=hparams['cond_hid_dim'], out_dim=cond_out_dim, smo_size=hparams['smo_win_size'])
        self.c_dim = c_dim

        self.img_resolution = img_resolution
        self.img_resolution_log2 = int(np.log2(img_resolution))
        self.img_channels = img_channels
        self.block_resolutions = [2 ** i for i in range(self.img_resolution_log2, 2, -1)]
        channels_dict = {res: min(channel_base // res, channel_max) for res in self.block_resolutions + [4]}
        fp16_resolution = max(2 ** (self.img_resolution_log2 + 1 - num_fp16_res), 8)

        if cmap_dim is None:
            cmap_dim = channels_dict[4]
        if c_dim == 0:
            cmap_dim = 0

        common_kwargs = dict(img_channels=img_channels, architecture=architecture, conv_clamp=conv_clamp)
        cur_layer_idx = 0
        for res in self.block_resolutions:
            in_channels = channels_dict[res] if res < img_resolution else 0
            tmp_channels = channels_dict[res]
            out_channels = channels_dict[res // 2]
            use_fp16 = (res >= fp16_resolution)
            block = DiscriminatorBlock(in_channels, tmp_channels, out_channels, resolution=res,
                first_layer_idx=cur_layer_idx, use_fp16=use_fp16, **block_kwargs, **common_kwargs)
            setattr(self, f'b{res}', block)
            cur_layer_idx += block.num_layers
        if c_dim > 0:
            self.mapping = MappingNetwork(z_dim=0, c_dim=c_dim, w_dim=cmap_dim, num_ws=None, w_avg_beta=None, **mapping_kwargs)
        self.b4 = DiscriminatorEpilogue(channels_dict[4], cmap_dim=cmap_dim, resolution=4, **epilogue_kwargs, **common_kwargs)

    def forward(self, img, camera, cond=None, update_emas=False, **block_kwargs):
        img = img['image']

        _ = update_emas # unused
        x = None
        for res in self.block_resolutions:
            block = getattr(self, f'b{res}')
            x, img = block(x, img, **block_kwargs)

        cmap = None
        c = camera
        if self.cond_dim > 0:
            cond_feat = self.cond_encoder(cond)
            c = torch.cat([c, cond_feat], dim=-1) # [b, 25+8]
        
        cmap = self.mapping(None, c)
        x = self.b4(x, img, cmap)
        return x
    
    def extra_repr(self):
        return f'c_dim={self.c_dim:d}, img_resolution={self.img_resolution:d}, img_channels={self.img_channels:d}'

#----------------------------------------------------------------------------

def filtered_resizing(image_orig_tensor, size, f, filter_mode='antialiased'):
    if filter_mode == 'antialiased':
        ada_filtered_64 = torch.nn.functional.interpolate(image_orig_tensor, size=(size, size), mode='bilinear', align_corners=False, antialias=True)
    elif filter_mode == 'classic':
        ada_filtered_64 = upfirdn2d.upsample2d(image_orig_tensor, f, up=2)
        ada_filtered_64 = torch.nn.functional.interpolate(ada_filtered_64, size=(size * 2 + 2, size * 2 + 2), mode='bilinear', align_corners=False)
        ada_filtered_64 = upfirdn2d.downsample2d(ada_filtered_64, f, down=2, flip_filter=True, padding=-1)
    elif filter_mode == 'none':
        ada_filtered_64 = torch.nn.functional.interpolate(image_orig_tensor, size=(size, size), mode='bilinear', align_corners=False)
    elif type(filter_mode) == float:
        assert 0 < filter_mode < 1

        filtered = torch.nn.functional.interpolate(image_orig_tensor, size=(size, size), mode='bilinear', align_corners=False, antialias=True)
        aliased  = torch.nn.functional.interpolate(image_orig_tensor, size=(size, size), mode='bilinear', align_corners=False, antialias=False)
        ada_filtered_64 = (1 - filter_mode) * aliased + (filter_mode) * filtered
        
    return ada_filtered_64

#----------------------------------------------------------------------------

class DualDiscriminator(torch.nn.Module):
    def __init__(self):
        super().__init__()
        channel_base = hparams['base_channel']
        channel_max = hparams['max_channel']
        conv_clamp = 256
        cmap_dim = None
        disc_c_noise = 0.
        block_kwargs = {'freeze_layers': 0}
        mapping_kwargs = {}
        epilogue_kwargs = {'mbstd_group_size': 4}
        architecture = 'resnet' # Architecture: 'orig', 'skip', 'resnet'.
        
        img_channels = 3
        img_channels *= 2

        self.camera_dim = 25
        if hparams['cond_type'] == 'idexp_lm3d_normalized':
            self.cond_dim = 204
        else:
            self.cond_dim = 0
        c_dim = self.camera_dim

        if self.cond_dim > 0:
            cond_out_dim = hparams['cond_out_dim']
            c_dim += cond_out_dim
            self.cond_encoder = LM3D_Win_Encoder(self.cond_dim, hid_dim=hparams['cond_hid_dim'], out_dim=cond_out_dim, smo_size=hparams['smo_win_size'])

        self.img_resolution = hparams['final_resolution']
        self.img_resolution_log2 = int(np.log2(self.img_resolution))
        self.img_channels = 3

        self.block_resolutions = [2 ** i for i in range(self.img_resolution_log2, 2, -1)]
        channels_dict = {res: min(channel_base // res, channel_max) for res in self.block_resolutions + [4]}
        self.num_fp16_res = hparams['num_fp16_layers_in_discriminator']
        fp16_resolution = max(2 ** (self.img_resolution_log2 + 1 - self.num_fp16_res), 8)
        if cmap_dim is None:
            cmap_dim = channels_dict[4]
        if c_dim == 0:
            cmap_dim = 0

        common_kwargs = dict(img_channels=img_channels, architecture=architecture, conv_clamp=conv_clamp)
        cur_layer_idx = 0
        for res in self.block_resolutions:
            in_channels = channels_dict[res] if res < self.img_resolution else 0
            tmp_channels = channels_dict[res]
            out_channels = channels_dict[res // 2]
            use_fp16 = (res >= fp16_resolution)
            block = DiscriminatorBlock(in_channels, tmp_channels, out_channels, resolution=res,
                first_layer_idx=cur_layer_idx, use_fp16=use_fp16, **block_kwargs, **common_kwargs)
            setattr(self, f'b{res}', block)
            cur_layer_idx += block.num_layers

        self.mapping = MappingNetwork(z_dim=0, c_dim=c_dim, w_dim=cmap_dim, num_ws=None, w_avg_beta=None, **mapping_kwargs)
        self.b4 = DiscriminatorEpilogue(channels_dict[4], cmap_dim=cmap_dim, resolution=4, **epilogue_kwargs, **common_kwargs)
        self.register_buffer('resample_filter', upfirdn2d.setup_filter([1,3,3,1]))
        self.disc_c_noise = disc_c_noise

    def forward(self, img, camera, cond=None, update_emas=False, **block_kwargs):
        image_raw = filtered_resizing(img['image_raw'], size=img['image'].shape[-1], f=self.resample_filter)
        img = torch.cat([img['image'], image_raw], 1)

        _ = update_emas # unused
        x = None
        for res in self.block_resolutions:
            block = getattr(self, f'b{res}')
            x, img = block(x, img, **block_kwargs)

        cmap = None

        c = camera
        if self.cond_dim > 0:
            cond_feat = self.cond_encoder(cond)
            c = torch.cat([c, cond_feat], dim=-1) # [b, 25+8]
        if self.disc_c_noise > 0: 
            c += torch.randn_like(c) * c.std(0) * self.disc_c_noise
        
        cmap = self.mapping(None, c)
        x = self.b4(x, img, cmap)
        return x

    def extra_repr(self):
        return f'c_dim={self.c_dim:d}, img_resolution={self.img_resolution:d}, img_channels={self.img_channels:d}'

#----------------------------------------------------------------------------

class DummyDualDiscriminator(torch.nn.Module):
    def __init__(self,
        c_dim,                          # Conditioning label (C) dimensionality.
        img_resolution,                 # Input resolution.
        img_channels,                   # Number of input color channels.
        architecture        = 'resnet', # Architecture: 'orig', 'skip', 'resnet'.
        channel_base        = 32768,    # Overall multiplier for the number of channels.
        channel_max         = 512,      # Maximum number of channels in any layer.
        num_fp16_res        = 4,        # Use FP16 for the N highest resolutions.
        conv_clamp          = 256,      # Clamp the output of convolution layers to +-X, None = disable clamping.
        cmap_dim            = None,     # Dimensionality of mapped conditioning label, None = default.
        block_kwargs        = {},       # Arguments for DiscriminatorBlock.
        mapping_kwargs      = {},       # Arguments for MappingNetwork.
        epilogue_kwargs     = {},       # Arguments for DiscriminatorEpilogue.
    ):
        super().__init__()
        img_channels *= 2

        self.c_dim = c_dim
        self.img_resolution = img_resolution
        self.img_resolution_log2 = int(np.log2(img_resolution))
        self.img_channels = img_channels
        self.block_resolutions = [2 ** i for i in range(self.img_resolution_log2, 2, -1)]
        channels_dict = {res: min(channel_base // res, channel_max) for res in self.block_resolutions + [4]}
        fp16_resolution = max(2 ** (self.img_resolution_log2 + 1 - num_fp16_res), 8)

        if cmap_dim is None:
            cmap_dim = channels_dict[4]
        if c_dim == 0:
            cmap_dim = 0

        common_kwargs = dict(img_channels=img_channels, architecture=architecture, conv_clamp=conv_clamp)
        cur_layer_idx = 0
        for res in self.block_resolutions:
            in_channels = channels_dict[res] if res < img_resolution else 0
            tmp_channels = channels_dict[res]
            out_channels = channels_dict[res // 2]
            use_fp16 = (res >= fp16_resolution)
            block = DiscriminatorBlock(in_channels, tmp_channels, out_channels, resolution=res,
                first_layer_idx=cur_layer_idx, use_fp16=use_fp16, **block_kwargs, **common_kwargs)
            setattr(self, f'b{res}', block)
            cur_layer_idx += block.num_layers
        if c_dim > 0:
            self.mapping = MappingNetwork(z_dim=0, c_dim=c_dim, w_dim=cmap_dim, num_ws=None, w_avg_beta=None, **mapping_kwargs)
        self.b4 = DiscriminatorEpilogue(channels_dict[4], cmap_dim=cmap_dim, resolution=4, **epilogue_kwargs, **common_kwargs)
        self.register_buffer('resample_filter', upfirdn2d.setup_filter([1,3,3,1]))

        self.raw_fade = 1

    def forward(self, img, c, update_emas=False, **block_kwargs):
        self.raw_fade = max(0, self.raw_fade - 1/(500000/32))

        image_raw = filtered_resizing(img['image_raw'], size=img['image'].shape[-1], f=self.resample_filter) * self.raw_fade
        img = torch.cat([img['image'], image_raw], 1)

        _ = update_emas # unused
        x = None
        for res in self.block_resolutions:
            block = getattr(self, f'b{res}')
            x, img = block(x, img, **block_kwargs)

        cmap = None
        if self.c_dim > 0:
            cmap = self.mapping(None, c)
        x = self.b4(x, img, cmap)
        return x

    def extra_repr(self):
        return f'c_dim={self.c_dim:d}, img_resolution={self.img_resolution:d}, img_channels={self.img_channels:d}'

#----------------------------------------------------------------------------