-
Notifications
You must be signed in to change notification settings - Fork 0
/
visualization.py
603 lines (512 loc) · 21.7 KB
/
visualization.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
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
"""
Functions used in various visualization operations, demonstrating
the edits produced by the discovered semantics.
"""
import os
import torch
import numpy as np
from utils import (select_bases, semantic_edit, parse_indices,
key_to_title)
from PIL import Image
from tqdm import tqdm
from torchvision.utils import make_grid
from matplotlib import pyplot as plt
def postprocess(images, min_val=-1.0, max_val=1.0):
"""
Post-process images from torch.Tensor to numpy.ndarray.
Parameters
----------
images : torch.Tensor
A tensor (BxCxHxW) to process.
min_val : float
The minimum value of the input tensor (the default is -1.0).
max_val : float
The maximum value of the input tensor (the default is 1.0).
Returns
-------
numpy.ndarray
A tensor with shape (B, C, H, W) and pixel range [0, 255].
"""
assert isinstance(images, torch.Tensor)
images = images.detach().cpu().numpy()
images = (images - min_val) * 255 / (max_val - min_val)
images = np.clip(images + 0.5, 0, 255).astype(np.uint8)
images = images.transpose(1, 2, 0)
return images
def draw_chart(fig):
"""
Convert a matplotlib.figure.Figure to numpy.ndarray.
Parameters
----------
fig : matplotlib.figure.Figure
Figure to convert to numpy array.
Returns
-------
numpy.ndarray
The numpy array equivalent of the input figure.
"""
fig.canvas.draw()
chart = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
chart = chart.reshape(fig.canvas.get_width_height()[::-1] + (3,))
# crop borders
nonzero_columns = np.count_nonzero(chart != 255, axis=0)[:, 0] > 0
chart = chart.transpose(1, 0, 2)[nonzero_columns].transpose(1, 0, 2)
return chart
def interpolation(G, layers, gan_type, proj_code, direction, magnitudes):
"""
Create a sequence of N synthetic images.
Each image in the sequence is generated by:
I_j = G(z + ε_j*n),
where I_j is a synthetic image (CxHxW), `G` is the generator function, z
corresponds to `proj_code`, ε_j is the magnitude of the edit and n
corresponds to `direction`. Index j \in [1, N].
Parameters
----------
G : torch.nn.module
Generator network that synthesizes images.
layers : list of int
Subset of layers to apply the semantic's effect.
Used only in the case of StyleGAN/StyleGAN2.
gan_type : {'pggan', 'stylegan', 'stylegan2'}
GAN model type.
proj_code : torch.Tensor
desc.
direction : numpy.ndarray
desc.
magnitudes : numpy.ndarray
Returns
-------
list of torch.Tensor
Sequence of edited synthetic images, where each image
I_j is generated using the equation above.
"""
images_per_direction = []
for m in magnitudes:
temp_proj_code = torch.clone(proj_code).detach()
image = semantic_edit(G, layers, gan_type, temp_proj_code, direction, m)
images_per_direction.append(image.squeeze(0))
return images_per_direction
def interpolation_chart(G,
layers,
gan_type,
basis,
proj_code,
magnitudes,
n_directions,
title=None,
begin=None,
end=None,
show_original=True,
**kwargs):
"""
Create a figure with `n_directions` interpolation rows, showcasing how
each direction alters the initial synthesized image.
Parameters
----------
G : torch.nn.module
Generator network that synthesizes images.
layers : list of int
Subset of layers to apply the semantics' effect. Used only in the
case of StyleGAN/StyleGAN2.
gan_type : {'pggan', 'stylegan', 'stylegan2'}
GAN model type.
basis : numpy.ndarray
Basis matrix of shape (|z|, |z|) in the general case. Each column
of this matrix corresponds to a discovered direction.
proj_code : torch.Tensor
The result of the 1st generation step G_1(z)=y. Has shape (|z|).
magnitudes : numpy.ndarray
Array containing the values of the magnitudes to use when editing.
n_directions : int
Number of directions to use for visualization.
title : str
Figure title (the default is None).
begin, end : int
Direction indices to place on the left of each interpolation row
(the default is None).
show_original : bool
Whether to show the original synthesized image in the top row or not
(the default is True).
**kwargs : dict
Extra arguments passed to plt.subplots().
Returns
-------
matplotlib.figure.Figure
A figure showcasing how the original synthesized
image is modified by editing it with the
discovered directions.
"""
# prepare `n_directions` interpolation rows
rows_per_page = []
for i in range(n_directions):
direction = basis[:, i]
rows_per_page.append(interpolation(G, layers, gan_type, proj_code, direction, magnitudes))
# create a figure to showcase the interpolation rows
rows_num = n_directions + int(show_original)
fig, ax = plt.subplots(nrows=rows_num, **kwargs) # **kwargs are passed to pyplot.figure()
if title is not None:
fig.suptitle(title)
# show original image in 1st row
if show_original:
original_image = semantic_edit(G, layers, gan_type, proj_code, direction, 0).squeeze(0)
ax[0].axis('off')
ax[0].imshow(postprocess(original_image))
# plot each linear interpolation sequence on the corresponing row
if begin is not None and end is not None:
desc = range(begin, end)
else:
desc = range(n_directions)
for axis, direction_interp, text in zip(ax[int(show_original):], rows_per_page, desc):
axis.axis('off')
#plt.subplots_adjust(left=0.25) # setting left=0.2 or lower eliminates whitespace between charts
axis.imshow(postprocess(make_grid(direction_interp, nrow=magnitudes.shape[0])))
axis.text(0, 0.5, str(text), horizontalalignment='right',
verticalalignment='center', fontsize='xx-small',
transform=axis.transAxes)
return fig
def lerp_matrix(G,
gan_type,
layers,
basis_list,
proj_codes,
n_samples,
magnitudes,
results_dir,
max_columns=45,
directions_per_page=15):
"""
Create an interpolation chart demonstrating the effects of the first
`max_columns` discovered directions. It can also compare MddGAN to SeFa.
Parameters
----------
G : torch.nn.module
Generator network that synthesizes images.
gan_type : {'pggan', 'stylegan', 'stylegan2'}
GAN model type.
layers : list of int
Subset of layers to apply the semantic's effect. Used only in the
case of StyleGAN/StyleGAN2.
basis_list : list of numpy.ndarray
Contains the basis matrix of shape (|z|, |z|) in the general case.
Each column of this matrix corresponds to a discovered direction.
When comparing MddGAN to SeFa, this list contains 2 such matrices.
proj_codes : torch.Tensor
The result of the 1st generation step G_1(z)=y. Has shape (B, |z|).
n_samples : int
Number of generated samples to use for visualization.
magnitudes : numpy.ndarray
Array containing the values of the magnitudes to use when editing.
results_dir : str
Path to directory where the results will be saved.
max_columns : int
Number of directions to visualize (the default is 45).
directions_per_page : int
Number of directions on each visualization page (the default is 15).
Returns
-------
None
"""
assert all(basis.ndim == 2 for basis in basis_list)
max_columns = min(max_columns, basis_list[0].shape[1])
if len(basis_list) == 2:
method_names = ['MddGAN', 'SeFa']
n_samples = 1
else:
method_names = [None]
# visualize the effect of `directions_per_page` directions on each page
pbar = tqdm(range(0, max_columns, directions_per_page))
for begin in pbar:
end = min(max_columns, begin + directions_per_page)
pbar.set_description(desc=f'Creating chart for directions {begin}-{end}... ')
charts = []
# create an interpolation chart for each sample
for sample_id in range(n_samples):
code = proj_codes[sample_id:sample_id + 1]
for idx, name in enumerate(method_names):
basis_matrix = basis_list[idx]
submatrix = basis_matrix[:, begin:end]
# create figure
fig = interpolation_chart(G, layers, gan_type, submatrix, code,
magnitudes, end - begin, title=name,
begin=begin, end=end, dpi=600,
constrained_layout=True)
# convert figure to numpy and append it to `charts` list
charts.append(draw_chart(fig))
# conserve memory
fig.clf()
plt.close(fig)
# concatenate charts (horizontally) into a single grid, save the grid
out_file = os.path.join(results_dir, f'directions_{begin}_{end}.jpg')
print('Saving chart to ', out_file)
Image.fromarray(np.hstack(charts)).save(out_file) # concatenate figures column-wise
def lerp_tensor(G,
gan_type,
layers,
basis,
basis_dims,
proj_codes,
n_samples,
magnitudes,
results_dir,
directions_per_page=15,
n_secondary_bases=3):
"""
Investigate how the discovered directions are separated into the dimensions
of the multilinear basis \mathcal{B}.
Parameters
----------
G : torch.nn.module
Generator network that synthesizes images.
gan_type : {'pggan', 'stylegan', 'stylegan2'}
GAN model type.
layers : list of int
Subset of layers to apply the semantics' effect. Used only in the
case of StyleGAN/StyleGAN2.
basis : numpy.ndarray
Basis matrix of shape (|z|, |z|) in the general case. Each column
of the matrix is a discovered direction.
basis_dims : list of int
The dimensions K_2, K_3, ..., K_M of the multilinear basis.
proj_codes : torch.Tensor
The result of the 1st generation step G_1(z)=y. Has shape (B, |z|).
n_samples : int
Number of generated samples to use for visualization.
magnitudes : numpy.ndarray
Array containing the values of the magnitudes to use when editing.
results_dir : str
Path to directory where the results will be saved.
directions_per_page : int
Number of directions on each visualization page (the default is 15).
n_secondary_bases : int
How many bases of the secondary tensor mode to explore.
Returns
-------
None
"""
# tensorize `basis`, in case it is in matrix form
if basis.ndim == 2:
basis = basis.reshape(basis.shape[0], *basis_dims)
assert (basis.ndim > 2 and basis.ndim < 6), ('More than 4 modes of '
'variation are not supported yet.')
# investigate each mode of the `basis` tensor
# tensor has shape (d, K_2, K_3, ..., K_M)
pbar = tqdm(enumerate(basis_dims))
for primary_mode_idx, primary_mode_dim in pbar:
pbar.set_description(desc='Creating chart for tensor mode'
f' {primary_mode_idx + 2}... ')
mode_dir = os.path.join(results_dir, f'Mode_{primary_mode_idx + 1}')
os.makedirs(mode_dir, exist_ok=True)
for secondary_mode_idx, secondary_mode_dim in enumerate(basis_dims):
if primary_mode_idx == secondary_mode_idx:
continue
for base_idx in range(min(n_secondary_bases, secondary_mode_dim)):
#for secondary_base_idx in range(n_secondary_bases):
charts = []
for sample_id in range(n_samples):
code = proj_codes[sample_id:sample_id + 1]
#bases, subscript = select_bases(basis, primary_mode_idx, secondary_base_idx, len(basis_dims))
bases, subscript = select_bases(basis, primary_mode_idx, secondary_mode_idx, base_idx, len(basis_dims))
directions_num = min(directions_per_page, bases.shape[1])
# create figure
fig = interpolation_chart(G, layers, gan_type, bases, code,
magnitudes, directions_num, dpi=600,
constrained_layout=True)
# draw chart and append it to `charts` list
charts.append(draw_chart(fig))
# conserve memory
fig.clf()
plt.close(fig)
# concatenate charts (horizontally) into a single grid, save the grid
out_file = os.path.join(mode_dir, f'B[{subscript}].jpg')
print('Saving chart to ', out_file)
Image.fromarray(np.hstack(charts)).save(out_file) # concatenate figures column-wise
def create_comparison_chart(G,
gan_type,
trunc_psi,
trunc_layers,
layers,
semantics,
magnitudes,
text,
reverse=False,
n_samples=2):
"""
Compare 2 semantics discovered by different methods (e.g MddGAN and SeFa),
but achieving the same effect.
Create and visualize a comparison chart demonstrating both produced effects
on the same generated sample, via linear interpolation. Top row corresponds
to the effect produced by the competing method (InterFaceGAN or SeFa), while
the bottom row corresponds to the effect produced by MddGAN.
Parameters
----------
G : torch.nn.module
Generator network that synthesizes images.
gan_type : {'pggan', 'stylegan', 'stylegan2'}
GAN model type.
trunc_psi : float
A StyleGAN/StyleGAN2 specific value used to "cutoff" some regions
of the generator distribution p_g, aka the "truncation trick" (the
default is 0.7).
trunc_layers : int
A StyleGAN/StyleGAN2 specific value indicating the number of layers
to apply the "truncation trick" (the default is 8).
layers : list of int
Subset of layers to apply the semantics' effect. Used only in the
case of StyleGAN/StyleGAN2.
semantics : list of torch.Tensor
Semantics used to edit. Index [0] corresponds to the semantic
discovered by the competing method, index [1] to the semantic
discovered by MddGAN.
magnitudes : numpy.ndarray
Array containing the values of the magnitudes to use when editing.
text : list of str
Names of the methods to compare. They are placed on the left of
each interpolation row. Index [1] always corresponds to MddGAN.
reverse : bool
Used for certain semantics, in order to evenly compare the results
(the default is False).
n_samples : int
Number of generated samples to use for the comparison (the default
is 2).
Returns
-------
None
"""
fig, ax = plt.subplots(nrows=2*n_samples, dpi=600, constrained_layout=True)
latent_vectors = torch.randn(n_samples, G.z_space_dim, device='cuda')
if gan_type == 'pggan':
codes = G.layer0.pixel_norm(latent_vectors)
elif gan_type == 'stylegan' or gan_type == 'stylegan2':
codes = G.mapping(latent_vectors)['w']
codes = G.truncation(codes, trunc_psi=trunc_psi, trunc_layers=trunc_layers)
for i in range(n_samples):
for method_idx in range(2):
ax_idx = i * n_samples + method_idx
ax[ax_idx].axis('off')
ax[ax_idx].imshow(
postprocess(
make_grid(
interpolation(G,
layers[method_idx],
gan_type,
codes[i:i+1],
semantics[method_idx],
magnitudes[::-1] if method_idx == 1 and reverse else magnitudes),
nrow=magnitudes.shape[0])
)
)
ax[ax_idx].text(0, 0.5,
text[method_idx],
horizontalalignment='right',
verticalalignment='center',
rotation='vertical',
fontsize='medium',
fontweight='bold' if text[method_idx] == 'Ours' else 'regular',
transform=ax[ax_idx].transAxes)
# draw horizontal line
if n_samples > 1:
line = plt.Line2D([0.1, 0.9], [0.5, 0.5], color="k", linewidth=1)
fig.add_artist(line)
plt.show()
def fid_plot(title, x_axis, competing_fid, mddgan_fid, competing_name, fname='fid.png'):
"""
Plot FID scores of the 2 methods for the same discovered direction.
Parameters
----------
title : str
Title of the plot.
x_axis : list of float
Common magnitude values used during FID calculation.
competing_fid : list of float
Resulting FID scores for the direction discovered by the
competing method.
mddgan_fid : list of float
Resulting FID scores for the direction discovered by MddGAN.
competing_name : str
Name of the competing method (InterFaceGAN or SeFa).
Returns
-------
None
"""
fig, ax = plt.subplots()
fig.suptitle(title)
ax.set_xlabel('Magnitude')
ax.set_ylabel('FID')
ax.plot(x_axis, competing_fid, 'o-', label=competing_name)
ax.plot(x_axis, mddgan_fid, 'o-', label='mddgan')
ax.legend()
fig.savefig('fid.png')
print('Saved plot to {}.'.format(fname))
def create_attribute_chart(proj_codes,
layers,
generator,
magnitude,
gan_type,
semantic,
attr_name):
interpolations = []
for code_idx in range(proj_codes.shape[0]):
interpolations.append(interpolation(generator, layers, gan_type,
proj_codes[code_idx:code_idx+1], semantic, [magnitude]))
assert len(interpolations) == (proj_codes.shape[0])
fig, axs = plt.subplots(nrows=proj_codes.shape[0], dpi=600, constrained_layout=True)
fig.suptitle(key_to_title(attr_name))
for ax, interp in zip(axs, interpolations):
ax.axis('off')
ax.imshow(postprocess(interp[0]))
return fig
def create_semantic_chart(G,
gan_type,
proj_codes,
attr_dict,
args,
n_samples_per_page=4):
total_samples = proj_codes.size()[0]
n_pages = int(total_samples / n_samples_per_page)
for i in range(n_pages):
start = i * n_samples_per_page
end = min(start + n_samples_per_page, total_samples)
codes = proj_codes[start:end]
charts = []
for idx, (key, item) in enumerate(attr_dict.items()):
print(f'Creating {key} chart...')
if idx == 0:
fig = create_attribute_chart(codes,
list(range(G.num_layers)),
G,
0.0,
gan_type,
torch.zeros(G.z_space_dim),
key)
else:
# load the corresponding semantic
semantic = np.load(f'{args.semantic_dir}/{args.method_name}/{args.model_name}_{key}.npy')
semantic = torch.from_numpy(semantic)
#
if args.method_name == 'interfacegan':
layers = list(range(G.num_layers))
else:
layer_idx = item[0]
layers = parse_indices(layer_idx, min_val=0, max_val=G.num_layers - 1)
magnitude = item[1]
# create attribute chart
fig = create_attribute_chart(codes,
layers,
G,
magnitude,
gan_type,
semantic,
key)
# draw vertical dotted line
if idx != len(attr_dict) - 1:
line = plt.Line2D([1.0, 1.0], [0, 1], color="k", linewidth=5, transform=fig.transFigure)
fig.add_artist(line)
# draw chart figure
charts.append(draw_chart(fig))
# conserve memory
fig.clf()
plt.close(fig)
# save chart
out_file = os.path.join(args.save_dir, f'{args.method_name}_{args.model_name}_{start}_{end}.jpg')
print(f'Saving chart to {out_file}\n')
Image.fromarray(np.hstack(charts)).save(out_file) # concatenate figures column-wis