forked from NVlabs/stylegan2-ada-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
interpolation.py
362 lines (307 loc) · 15.1 KB
/
interpolation.py
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
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
"""Produce target image from hair and identity images."""
import copy
import glob
import os
from time import perf_counter
from pathlib import Path
import click
import imageio
import numpy as np
import PIL.Image
import torch
import torch.nn.functional as F
import dnnlib
import legacy
from projector_18 import project as project_18
def interpolation(
G,
identity: torch.Tensor, # [C,H,W] and dynamic range [0,255], W & H must match G output resolution
hair: torch.Tensor, # [C,H,W] and dynamic range [0,255], W & H must match G output resolution
*,
alpha1 = 1.0,
alpha2 = 1.0,
num_steps = 500,
w_avg_samples = 10000,
initial_learning_rate = 0.1,
initial_noise_factor = 5e-1,
lr_rampdown_length = 0.25,
lr_rampup_length = 0.05,
noise_ramp_length = 0.75,
regularize_noise_weight = 1e5,
verbose = False,
stop,
device: torch.device,
hair_img_filename,
identity_img_filename
):
assert identity.shape == (G.img_channels, G.img_resolution, G.img_resolution)
def logprint(*args):
if verbose:
print(*args)
G = copy.deepcopy(G).eval().requires_grad_(False).to(device) # type: ignore
# Compute q stats.
logprint(f'Computing W midpoint and stddev using {w_avg_samples} samples...')
w_samples = np.random.RandomState(123).randn(w_avg_samples, G.num_ws)
w_avg = np.mean(w_samples, axis=0, keepdims=True) # [G.w_dim]
w_std = (np.sum((w_samples - w_avg) ** 2) / w_avg_samples) ** 0.5
# Setup noise inputs.
noise_bufs = {name: buf for (name, buf) in G.synthesis.named_buffers() if 'noise_const' in name}
seg_channel_dict = {
'h': 0,
'i': 1,
'b': 2
}
vgg16 = torch.jit.load('vgg16.pt').eval().to(device)
segNet = torch.load('segNet.pt').eval().to('cuda')
segNet_mean = torch.from_numpy(np.load('train-mean.npy')).float().to(device)
segNet_std = torch.from_numpy(np.load('train-std.npy')).float().to(device)
def apply_seg_mask(
x: torch.Tensor,
channel: int
):
x_norm = (x - segNet_mean) / segNet_std
segmentation = segNet(x_norm)['out']
mask = torch.argmax(segmentation, dim=1)
mask = mask.squeeze().to('cuda')
x = x.squeeze()
mask[mask != channel] = 1000
x = torch.where(mask != 1000, x, torch.tensor(255.0).to('cuda'))
return x.unsqueeze(0)
# Features for identity image.
if identity.shape[2] > 256:
masked_identity_img = F.interpolate(identity.unsqueeze(0).to(torch.float32), size=(256, 256), mode='area')
masked_identity_img = apply_seg_mask(masked_identity_img, seg_channel_dict['i']).to('cuda').to(torch.float32)
masked_identity_img = masked_identity_img.to(torch.uint8)
identity_features = vgg16(masked_identity_img, resize_images=False, return_lpips=True)
# Features for hair image.
if hair.shape[2] > 256:
masked_hair_img = F.interpolate(hair.unsqueeze(0).to(torch.float32), size=(256, 256), mode='area')
masked_hair_img = apply_seg_mask(masked_hair_img, seg_channel_dict['h']).to('cuda').to(torch.float32)
masked_hair_img = masked_hair_img.to(torch.uint8)
hair_features = vgg16(masked_hair_img, resize_images=False, return_lpips=True)
# Loading the projection of generated images to save time when debugging
w_h_path = Path("generated-male/{}.npy".format(hair_img_filename))
w_p_path = Path("generated-male/{}.npy".format(identity_img_filename))
# Loading the projection of real images
# w_h_path = Path("male/{}/18x512/{}-projected_w.npz".format(hair_img_filename, hair_img_filename))
# w_p_path = Path("male/{}/18x512/{}-projected_w.npz".format(identity_img_filename, identity_img_filename))
if w_h_path.exists():
# w_h = torch.from_numpy(np.load(w_h_path)['w'][0]).to('cuda')
# w_h = torch.from_numpy(np.load(w_h_path)['w']).to('cuda')
w_h = torch.from_numpy(np.load(w_h_path)[0]).to('cuda')
else:
w_h = project_18(G, hair, device=torch.device('cuda'))[-1]
np.savez(w_h_path, w=w_h.cpu().numpy())
if w_p_path.exists():
# w_p = torch.from_numpy(np.load(w_p_path)['w'][0]).to('cuda')
# w_p = torch.from_numpy(np.load(w_p_path)['w']).to('cuda')
w_p = torch.from_numpy(np.load(w_p_path)[0]).to('cuda')
else:
w_p = project_18(G, identity, device=torch.device('cuda'))[-1]
np.savez(w_p_path, w=w_p.cpu().numpy())
q_opt = torch.nn.Parameter(torch.randn(size=w_p.shape, dtype=torch.float32, requires_grad=True, device=device))
hair_distances = []
identity_distances = []
# list of all target ws through optimization
w_out = torch.zeros([stop] + list(w_h.shape), dtype=torch.float32, device=device)
optimizer = torch.optim.Adam([q_opt] + list(noise_bufs.values()), betas=(0.9, 0.999), lr=initial_learning_rate)
# Init noise.
for buf in noise_bufs.values():
buf[:] = torch.randn_like(buf)
buf.requires_grad = True
print("starting the iterations..")
for step in range(num_steps):
print("iteration ", step)
if step==stop: break
# Learning rate schedule.
t = step / num_steps
w_noise_scale = w_std * initial_noise_factor * max(0.0, 1.0 - t / noise_ramp_length) ** 2
lr_ramp = min(1.0, (1.0 - t) / lr_rampdown_length)
lr_ramp = 0.5 - 0.5 * np.cos(lr_ramp * np.pi)
lr_ramp = lr_ramp * min(1.0, t / lr_rampup_length)
lr = initial_learning_rate * lr_ramp
for param_group in optimizer.param_groups:
param_group['lr'] = lr
# Synth images from opt_w.
w_t = w_p + q_opt.sigmoid() * (w_h - w_p)
ws = w_t.unsqueeze(0)
# Downsample image to 256x256 if it's larger than that. VGG was built for 224x224 images.
target_image = G.synthesis(ws, noise_mode='const')
target_image = (target_image + 1) * (255/2)
if target_image.shape[2] > 256:
target_image = F.interpolate(target_image, size=(256, 256), mode='area')
hair_target_image = apply_seg_mask(target_image, seg_channel_dict['h'])
identity_target_image = apply_seg_mask(target_image, seg_channel_dict['i'])
# Features for synth images.
target_hair_features = vgg16(hair_target_image, resize_images=False, return_lpips=True)
target_identity_features = vgg16(identity_target_image, resize_images=False, return_lpips=True)
# Compute loss
hair_dist = (target_hair_features - hair_features).square().sum()
identity_dist = (target_identity_features - identity_features).square().sum()
hair_distances.append(hair_dist.item())
identity_distances.append(identity_dist.item())
# loss function
dist = alpha1 * hair_dist + alpha2 * identity_dist
# Noise regularization.
reg_loss = 0.0
for v in noise_bufs.values():
noise = v[None,None,:,:] # must be [1,1,H,W] for F.avg_pool2d()
while True:
reg_loss += (noise*torch.roll(noise, shifts=1, dims=3)).mean()**2
reg_loss += (noise*torch.roll(noise, shifts=1, dims=2)).mean()**2
if noise.shape[2] <= 8:
break
noise = F.avg_pool2d(noise, kernel_size=2)
loss = dist + reg_loss * regularize_noise_weight
# Step
print("optimizer steps")
optimizer.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
logprint(f'step {step+1:>4d}/{num_steps}: dist {dist:<4.2f} loss {float(loss):<5.2f}')
# Save projected W for each optimization step.
w_out[step] = w_t.detach()
# Normalize noise.
with torch.no_grad():
for buf in noise_bufs.values():
buf -= buf.mean()
buf *= buf.square().mean().rsqrt()
return w_out, hair_distances, identity_distances
#----------------------------------------------------------------------------
# @click.command()
# @click.option('--network', 'network_pkl', help='Network pickle filename', required=True)
# @click.option('--hair', 'hair_fname', required=True, metavar='FILE')
# @click.option('--identity', 'identity_fname', required=True, metavar='FILE')
# @click.option('--num-steps', help='Number of optimization steps', type=int, default=500, show_default=True)
# @click.option('--seed', help='Random seed', type=int, default=303, show_default=True)
# @click.option('--save-video', help='Save an mp4 video of optimization progress', type=bool, default=True, show_default=True)
# @click.option('--outdir', help='Where to save the output images', required=True, metavar='DIR')
def run_interpolation(
network_pkl: str,
hair_fname: str,
identity_fname: str,
outdir: str,
save_video: bool,
seed: int,
stop: int,
num_steps: int
):
np.random.seed(seed)
torch.manual_seed(seed)
# Load networks.
print('Loading networks from "%s"...' % network_pkl)
device = torch.device('cuda')
with dnnlib.util.open_url(network_pkl) as fp:
G = legacy.load_network_pkl(fp)['G_ema'].requires_grad_(False).to(device) # type: ignore
# Load hair image.
hair_pil = PIL.Image.open(hair_fname).convert('RGB')
w, h = hair_pil.size
s = min(w, h)
hair_pil = hair_pil.crop(((w - s) // 2, (h - s) // 2, (w + s) // 2, (h + s) // 2))
hair_pil = hair_pil.resize((G.img_resolution, G.img_resolution), PIL.Image.LANCZOS)
hair_uint8 = np.array(hair_pil, dtype=np.uint8)
hair_img_filename = hair_fname.split('/')[1].split('.')[0]
# Load identity image.
identity_pil = PIL.Image.open(identity_fname).convert('RGB')
w, h = identity_pil.size
s = min(w, h)
identity_pil = identity_pil.crop(((w - s) // 2, (h - s) // 2, (w + s) // 2, (h + s) // 2))
identity_pil = identity_pil.resize((G.img_resolution, G.img_resolution), PIL.Image.LANCZOS)
identity_uint8 = np.array(identity_pil, dtype=np.uint8)
identity_img_filename = identity_fname.split('/')[1].split('.')[0]
# Optimize interpolation.
start_time = perf_counter()
projected_w_steps, hair_distances, identity_distances = interpolation(
G,
hair=torch.tensor(hair_uint8.transpose([2, 0, 1]), device=device), # pylint: disable=not-callable
identity=torch.tensor(identity_uint8.transpose([2, 0, 1]), device=device),
num_steps=num_steps,
device=device,
verbose=True,
stop=stop,
hair_img_filename=hair_img_filename,
identity_img_filename=identity_img_filename
)
print (f'Elapsed: {(perf_counter()-start_time):.1f} s')
# Create video from results
os.makedirs(outdir, exist_ok=True)
# if save_video:
# video = imageio.get_writer('{}/proj_i{}_h{}.mp4'.format(outdir, identity_img_filename, hair_img_filename), mode='I', fps=10, codec='libx264', bitrate='16M')
# print (f'Saving optimization progress video "{outdir}/proj.mp4"')
# for projected_w in projected_w_steps:
# synth_image = G.synthesis(projected_w.unsqueeze(0), noise_mode='const')
# synth_image = (synth_image + 1) * (255/2)
# synth_image = synth_image.permute(0, 2, 3, 1).clamp(0, 255).to(torch.uint8)[0].cpu().numpy()
# video.append_data(np.concatenate([identity_uint8, hair_uint8, synth_image], axis=1))
# video.close()
# Save final projected frame and W vector.
# hair_pil.save('{}/hair_{}.png'.format(outdir, hair_img_filename))
# identity_pil.save('{}/identity_{}.png'.format(outdir, identity_img_filename))
np.save('{}/synth_i{}_h{}-distances.npy'.format(outdir, identity_img_filename, hair_img_filename), np.concatenate((np.expand_dims(hair_distances, -1), np.expand_dims(identity_distances, -1)), axis=-1))
projected_w = projected_w_steps[-1]
# create image from last style embedding q
synth_image = G.synthesis(projected_w.unsqueeze(0), noise_mode='const')
synth_image = (synth_image + 1) * (255/2)
synth_image = synth_image.permute(0, 2, 3, 1).clamp(0, 255).to(torch.uint8)[0].cpu().numpy()
PIL.Image.fromarray(synth_image, 'RGB').save('{}/synth_i{}_h{}.png'.format(outdir, identity_img_filename, hair_img_filename))
np.savez('{}/synth_i{}_h{}.npz'.format(outdir, identity_img_filename, hair_img_filename), w=projected_w.unsqueeze(0).cpu().numpy())
#----------------------------------------------------------------------------
if __name__ == "__main__":
# running interpolation for all combinations of images
prefix = 'generated-male/'
# imgNumbers = os.listdir(prefix)
imgNames = list(filter(lambda x: '.png' in x, os.listdir(prefix)))
for img1 in imgNames:
for img2 in imgNames:
if img1 == img2: continue
run_interpolation(hair_fname=prefix + img1,
identity_fname=prefix + img2,
network_pkl="ffhq.pkl",
outdir="generated-male-output-dir",
save_video=False,
num_steps=300,
stop=100,
seed=303)
# below is some code used to generate grids
# max_horizontal_imgs = 6
# img_paths = sorted(glob.glob("male_q512/*.png"))
# order = sorted([img_p.split('.')[0] for img_p in glob.glob("male_6/*.jpg")])
# first_imgs = [PIL.Image.open(path+".jpg").convert('RGB') for path in order]
# white_img = PIL.Image.new("RGB", first_imgs[0].size, (255,255,255))
# imgs=[]
#
# for i, img_p in enumerate(img_paths):
# if i % max_horizontal_imgs == 0:
# imgs.append(white_img)
# img = PIL.Image.open(img_p).convert('RGB')
# img_id = img_p.split('/')[-1].split('.')[0].split('synth_')[-1]
# imgs.append(img)
# imgs.append(white_img)
#
# def pil_grid(images, max_horiz=np.iinfo(int).max):
# n_images = len(images)
# n_horiz = min(n_images, max_horiz)
# h_sizes, v_sizes = [0] * n_horiz, [0] * (n_images // n_horiz)
# for i, im in enumerate(images):
# h, v = i % n_horiz, i // n_horiz
# h_sizes[h] = max(h_sizes[h], im.size[0])
# v_sizes[v] = max(v_sizes[v], im.size[1])
# h_sizes, v_sizes = np.cumsum([0] + h_sizes), np.cumsum([0] + v_sizes)
# im_grid = PIL.Image.new('RGB', (h_sizes[-1], v_sizes[-1]), color='white')
# for i, im in enumerate(images):
# im_grid.paste(im, (h_sizes[i % n_horiz], v_sizes[i // n_horiz]))
# return im_grid
#
# grid = pil_grid(imgs, max_horiz=max_horizontal_imgs)
# # create first row
# first_row = pil_grid(first_imgs, max_horiz=len(first_imgs))
# # create first column
# first_imgs.insert(0, white_img)
# first_column = pil_grid(first_imgs, max_horiz=1)
# # append first row to grid
# grid = pil_grid([first_row, grid], max_horiz=1)
# # append first column
# grid = pil_grid([first_column, grid], max_horiz=2)
#
# grid.save("male_q512.png")
#----------------------------------------------------------------------------