File size: 11,421 Bytes
f4c3c2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ff50bb1
f4c3c2b
 
 
 
e90f2c5
f4c3c2b
1c472db
f4c3c2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ff50bb1
e90f2c5
f4c3c2b
ff50bb1
 
 
 
f4c3c2b
e90f2c5
ff50bb1
f4c3c2b
ff50bb1
 
f4c3c2b
 
ff50bb1
f4c3c2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ff50bb1
f4c3c2b
 
 
 
e90f2c5
 
 
f4c3c2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e90f2c5
f4c3c2b
 
 
 
 
 
 
 
 
 
 
 
ff50bb1
f4c3c2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e90f2c5
f4c3c2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ff50bb1
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
# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto.  Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.

"""Generate images using pretrained network pickle."""

import math
import legacy
import clip
import dnnlib
import numpy as np
import torch
import torch.nn.functional as F
from torchvision.transforms import Compose, Resize, CenterCrop
from PIL import Image
from torch_utils import misc
from torch_utils.ops import upfirdn2d
import id_loss
from copy import deepcopy


def block_forward(self, x, img, ws, shapes, force_fp32=False, fused_modconv=None, **layer_kwargs):
        misc.assert_shape(ws, [None, self.num_conv + self.num_torgb, self.w_dim])
        w_iter = iter(ws.unbind(dim=1))
        dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32
        memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format
        if fused_modconv is None:
            with misc.suppress_tracer_warnings(): # this value will be treated as a constant
                fused_modconv = (not self.training) and (dtype == torch.float32 or int(x.shape[0]) == 1)

        # Input.
        if self.in_channels == 0:
            x = self.const.to(dtype=dtype, memory_format=memory_format)
            x = x.unsqueeze(0).repeat([ws.shape[0], 1, 1, 1])
        else:
            misc.assert_shape(x, [None, self.in_channels, self.resolution // 2, self.resolution // 2])
            x = x.to(dtype=dtype, memory_format=memory_format)

        # Main layers.
        if self.in_channels == 0:
            x = self.conv1(x, next(w_iter)[...,:shapes[0]], fused_modconv=fused_modconv, **layer_kwargs)
        elif self.architecture == 'resnet':
            y = self.skip(x, gain=np.sqrt(0.5))
            x = self.conv0(x, next(w_iter), fused_modconv=fused_modconv, **layer_kwargs)
            x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv, gain=np.sqrt(0.5), **layer_kwargs)
            x = y.add_(x)
        else:
            x = self.conv0(x, next(w_iter)[...,:shapes[0]], fused_modconv=fused_modconv, **layer_kwargs)
            x = self.conv1(x, next(w_iter)[...,:shapes[1]], fused_modconv=fused_modconv, **layer_kwargs)

        # ToRGB.
        if img is not None:
            misc.assert_shape(img, [None, self.img_channels, self.resolution // 2, self.resolution // 2])
            img = upfirdn2d.upsample2d(img, self.resample_filter)
        if self.is_last or self.architecture == 'skip':
            y = self.torgb(x, next(w_iter)[...,:shapes[2]], fused_modconv=fused_modconv)
            y = y.to(dtype=torch.float32, memory_format=torch.contiguous_format)
            img = img.add_(y) if img is not None else y

        assert x.dtype == dtype
        assert img is None or img.dtype == torch.float32
        return x, img

def unravel_index(index, shape):
    out = []
    for dim in reversed(shape):
        out.append(index % dim)
        index = index // dim
    return tuple(reversed(out))

def find_direction(
    GIn,
    text_prompt: str,
    truncation_psi: float = 0.7,
    noise_mode: str = "const",
    resolution: int = 256,
    identity_power: float = 0.5,
):
    G = deepcopy(GIn)
    seeds=np.random.randint(0, 1000, 128)

    batch_size=1
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

    # Labels
    class_idx=None
    label = torch.zeros([1, G.c_dim], device=device).requires_grad_()
    if G.c_dim != 0:
        label[:, class_idx] = 1

    model, preprocess = clip.load("ViT-B/32", device=device)
    text = clip.tokenize([text_prompt]).to(device)
    text_features = model.encode_text(text)

    # Generate images
    for i in G.parameters():
        i.requires_grad = True

    mean = torch.as_tensor((0.48145466, 0.4578275, 0.40821073), dtype=torch.float, device=device)
    std = torch.as_tensor((0.26862954, 0.26130258, 0.27577711), dtype=torch.float, device=device)
    if mean.ndim == 1:
        mean = mean.view(-1, 1, 1)
    if std.ndim == 1:
        std = std.view(-1, 1, 1)

    transf = Compose([Resize(224, interpolation=Image.BICUBIC), CenterCrop(224)])

    styles_array = []
    for seed_idx, seed in enumerate(seeds):
        if seed == seeds[-1]:
            print('Generating image for seed %d (%d/%d) ...' % (seed, seed_idx, len(seeds)))
        z = torch.from_numpy(np.random.RandomState(seed).randn(1, G.z_dim)).to(device)
        ws = G.mapping(z, label, truncation_psi=truncation_psi)

        block_ws = []
        with torch.autograd.profiler.record_function('split_ws'):
            misc.assert_shape(ws, [None, G.synthesis.num_ws, G.synthesis.w_dim])
            ws = ws.to(torch.float32)

            w_idx = 0
            for res in G.synthesis.block_resolutions:
                block = getattr(G.synthesis, f'b{res}')
                block_ws.append(ws.narrow(1, w_idx, block.num_conv + block.num_torgb))
                w_idx += block.num_conv

        styles = torch.zeros(1, 26, 512, device=device)
        styles_idx = 0
        temp_shapes = []
        for res, cur_ws in zip(G.synthesis.block_resolutions, block_ws):
            block = getattr(G.synthesis, f'b{res}')

            if res == 4:
                temp_shape = (block.conv1.affine.weight.shape[0], block.conv1.affine.weight.shape[0], block.torgb.affine.weight.shape[0])
                styles[0, :1, :] = block.conv1.affine(cur_ws[0, :1, :])
                styles[0, 1:2, :] = block.torgb.affine(cur_ws[0, 1:2, :])
                if seed_idx == (len(seeds) - 1):
                    block.conv1.affine = torch.nn.Identity()
                    block.torgb.affine = torch.nn.Identity()
                styles_idx += 2
            else:
                temp_shape = (block.conv0.affine.weight.shape[0], block.conv1.affine.weight.shape[0], block.torgb.affine.weight.shape[0])
                styles[0,styles_idx:styles_idx+1,:temp_shape[0]] = block.conv0.affine(cur_ws[0,:1,:])
                styles[0,styles_idx+1:styles_idx+2,:temp_shape[1]] = block.conv1.affine(cur_ws[0,1:2,:])
                styles[0,styles_idx+2:styles_idx+3,:temp_shape[2]] = block.torgb.affine(cur_ws[0,2:3,:])
                if seed_idx == (len(seeds) - 1):
                    block.conv0.affine = torch.nn.Identity()
                    block.conv1.affine = torch.nn.Identity()
                    block.torgb.affine = torch.nn.Identity()
                styles_idx += 3
            temp_shapes.append(temp_shape)

        styles = styles.detach()
        styles_array.append(styles)

    resolution_dict = {256: 6, 512: 7, 1024: 8}
    id_coeff = identity_power
    styles_direction = torch.zeros(1, 26, 512, device=device)
    styles_direction_grad_el2 = torch.zeros(1, 26, 512, device=device)
    styles_direction.requires_grad_()

    global id_loss2
    #id_loss = id_loss.IDLoss("a").to(device).eval()
    id_loss2 = id_loss.IDLoss("a").to(device).eval()

    temp_photos = []
    grads = []
    for i in range(math.ceil(len(seeds) / batch_size)):

        styles = torch.vstack(styles_array[i*batch_size:(i+1)*batch_size]).to(device)
        seed = seeds[i]

        styles_idx = 0
        x2 = img2 = None

        for k, (res, cur_ws) in enumerate(zip(G.synthesis.block_resolutions, block_ws)):
            block = getattr(G.synthesis, f'b{res}')
            if k > resolution_dict[resolution]:
                continue

            if res == 4:
                x2, img2 = block_forward(block, x2, img2, styles[:, styles_idx:styles_idx+2, :], temp_shapes[k], noise_mode=noise_mode, force_fp32=True)
                styles_idx += 2
            else:
                x2, img2 = block_forward(block, x2, img2, styles[:, styles_idx:styles_idx+3, :], temp_shapes[k], noise_mode=noise_mode, force_fp32=True)
                styles_idx += 3

        img2_cpu = img2.detach().cpu().numpy()
        temp_photos.append(img2_cpu)
        if i > 3:
            continue

        styles2 = styles + styles_direction

        styles_idx = 0
        x = img = None
        for k, (res, cur_ws) in enumerate(zip(G.synthesis.block_resolutions, block_ws)):
            block = getattr(G.synthesis, f'b{res}')
            if k > resolution_dict[resolution]:
                continue
            if res == 4:
                x, img = block_forward(block, x, img, styles2[:, styles_idx:styles_idx+2, :], temp_shapes[k], noise_mode=noise_mode, force_fp32=True)
                styles_idx += 2
            else:
                x, img = block_forward(block, x, img, styles2[:, styles_idx:styles_idx+3, :], temp_shapes[k], noise_mode=noise_mode, force_fp32=True)
                styles_idx += 3

        identity_loss, _ = id_loss2(img, img2)
        identity_loss *= id_coeff
        img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255)
        img = (transf(img.permute(0, 3, 1, 2)) / 255).sub_(mean).div_(std)
        image_features = model.encode_image(img)
        cos_sim = -1*F.cosine_similarity(image_features, (text_features[0]).unsqueeze(0))
        (identity_loss + cos_sim.sum()).backward(retain_graph=True)

    styles_direction.grad[:, list(range(26)), :] = 0
    with torch.no_grad():
        styles_direction *= 0

    for i in range(math.ceil(len(seeds) / batch_size)):

        seed = seeds[i]
        styles = torch.vstack(styles_array[i*batch_size:(i+1)*batch_size]).to(device)
        img2 = torch.tensor(temp_photos[i]).to(device)
        styles2 = styles + styles_direction

        styles_idx = 0
        x = img = None
        for k, (res, cur_ws) in enumerate(zip(G.synthesis.block_resolutions, block_ws)):
            block = getattr(G.synthesis, f'b{res}')
            if k > resolution_dict[resolution]:
                continue

            if res == 4:
                x, img = block_forward(block, x, img, styles2[:, styles_idx:styles_idx+2, :], temp_shapes[k], noise_mode=noise_mode, force_fp32=True)
                styles_idx += 2
            else:
                x, img = block_forward(block, x, img, styles2[:, styles_idx:styles_idx+3, :], temp_shapes[k], noise_mode=noise_mode, force_fp32=True)
                styles_idx += 3

        identity_loss, _ = id_loss2(img, img2)
        identity_loss *= id_coeff
        img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255)
        img = (transf(img.permute(0, 3, 1, 2)) / 255).sub_(mean).div_(std)
        image_features = model.encode_image(img)
        cos_sim = -1*F.cosine_similarity(image_features, (text_features[0]).unsqueeze(0))
        (identity_loss + cos_sim.sum()).backward(retain_graph=True)

        styles_direction.grad[:, [0, 1, 4, 7, 10, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25], :] = 0

        if i % 2 == 1:
            styles_direction.data = (styles_direction - styles_direction.grad * 5)
            grads.append(styles_direction.grad.clone())
            styles_direction.grad.data.zero_()
            if i > 3:
                styles_direction_grad_el2[grads[-2] * grads[-1] < 0] += 1

    styles_direction = styles_direction.detach()
    styles_direction[styles_direction_grad_el2 > (len(seeds) / batch_size) / 4] = 0

    return styles_direction.cpu().numpy()