-
Notifications
You must be signed in to change notification settings - Fork 0
/
discover_subspace.py
352 lines (307 loc) · 16.2 KB
/
discover_subspace.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
import os
import argparse
import torch
import torch.nn as nn
import cv2
import copy
import pickle as pkl
import numpy as np
from utils.file_utils import image_files, load_as_tensor, Tensor2PIL, split_to_batches
from utils.image_precossing import _sigmoid_to_tanh, _tanh_to_sigmoid, _add_batch_one
from derivable_models.derivable_generator import get_derivable_generator
from models.model_settings import MODEL_POOL
from utils.file_utils import create_experiments_directory
import torch.optim as optim
import torchvision
from scipy.sparse.linalg import svds
from sklearn import cluster
from sklearn.preprocessing import normalize
import matplotlib.pyplot as plt
# threshold the matrix.
def thrC(C,ro):
if ro < 1:
N = C.shape[1]
Cp = np.zeros((N,N))
S = np.abs(np.sort(-np.abs(C), axis=0))
Ind = np.argsort(-np.abs(C), axis=0)
for i in range(N):
cL1 = np.sum(S[:, i]).astype(float)
stop = False
csum = 0
t = 0
while(stop == False):
csum = csum + S[t, i]
if csum > ro*cL1:
stop = True
Cp[Ind[0:t+1,i],i] = C[Ind[0:t+1,i],i]
t = t + 1
else:
Cp = C
return Cp
def post_proC(C, K, d, alpha):
# C: coefficient matrix, K: number of clusters, d: dimension of each subspace
C = 0.5*(C + C.T)
r = min(d*K + 1, C.shape[0]-1)
U, S, _ = svds(C,r,v0 = np.ones(C.shape[0]))
U = U[:,::-1]
S = np.sqrt(S[::-1])
S = np.diag(S)
U = U.dot(S)
U = normalize(U, norm='l2', axis = 1)
Z = U.dot(U.T)
Z = Z * (Z>0)
L = np.abs(Z ** alpha)
L = L/L.max()
L = 0.5 * (L + L.T)
spectral = cluster.SpectralClustering(n_clusters=K, eigen_solver='arpack', affinity='precomputed',assign_labels='discretize')
spectral.fit(L)
grp = spectral.fit_predict(L) + 1
return grp, L
def switch_matrix(C, predict):
n = C.shape[0]
n_classes = max(predict)
class_counter = [sum(predict == (1 + i)) for i in range(n_classes)]
print(class_counter)
class_counter = list(np.cumsum(class_counter))
print(class_counter)
class_counter = [0] + class_counter
print(class_counter)
tmp_counter = [0 for i in range(n_classes)]
def exchange_element(slice, i, j):
kk = slice[i].copy()
slice[i] = slice[j].copy()
slice[j] = kk
return slice.copy()
for i in range(n):
class_id = predict[i] - 1
exchange_idx = class_counter[class_id] + tmp_counter[class_id]
# exchange i and exchange_idx
row_i = exchange_element(C[i, :].copy(), i, exchange_idx)
col_i = exchange_element(C[:, i].copy(), i, exchange_idx)
row_j = exchange_element(C[exchange_idx, :].copy(), i, exchange_idx)
col_j = exchange_element(C[:, exchange_idx].copy(), i, exchange_idx)
C[i, :] = row_j
C[:, i] = col_j
C[exchange_idx, :] = row_i
C[:, exchange_idx] = col_i
tmp_counter[class_id] += 1
return C.copy()
def main(args):
os.makedirs(args.outputs, exist_ok=True)
out_dir, exp_name = create_experiments_directory(args, args.exp_id)
print(out_dir)
print(exp_name)
generator = get_derivable_generator(args.gan_model, args.inversion_type, args)
generator.cuda()
print('There are %d blocks in this generator.' % generator.n_blocks) # 19 for pggan
latent_space = generator.PGGAN_LATENT[args.layer]
print('The latent space is ', latent_space)
Matrix = torch.ones([latent_space[0], latent_space[0]], dtype=torch.float32).cuda() * 0.0001
Matrix.requires_grad = True
gan_type, image_type = args.gan_model.split("_")
print('The gan type is %s, and the image type is %s' % (gan_type, image_type))
test_image_dir = os.path.join('./bin/', gan_type, image_type)
print(test_image_dir)
if args.optimization == 'GD':
optimizer = torch.optim.SGD([Matrix], lr=args.lr)
elif args.optimization == 'Adam':
optimizer = torch.optim.Adam([Matrix], lr=args.lr)
else:
raise NotImplemented('Not Implemented.')
files = os.listdir(test_image_dir)
test_zs = []
for i in range(len(files)):
if files[i].endswith('.pkl'):
with open(os.path.join(test_image_dir, files[i]), 'rb') as file_in:
test_zs.append(pkl.load(file_in))
test_zs = torch.from_numpy(np.concatenate(test_zs, axis=0).astype(np.float32)).cuda()
print(test_zs.size())
losses_dict = {'total_loss': [], 'feature_loss': [], 'data_loss': [], 'sparse_loss': []}
for iter in range(args.iterations):
if args.beta0 > 0:
Z = torch.randn([args.batch_size, 512, 1, 1], dtype=torch.float32).cuda()
F = generator([Z], which_block=args.layer, pre_model=True)
F_reshape = F.transpose(0, 1).reshape((latent_space[0], -1))
F_rec = torch.matmul(Matrix * (1. - torch.eye(n=latent_space[0], m=latent_space[0]).cuda()), F_reshape).\
reshape((latent_space[0], args.batch_size,) + tuple(latent_space[1:])).transpose(0, 1)
X_rec = generator([F_rec], which_block=args.layer, post_model=True)
X = generator([F], which_block=args.layer, post_model=True)
optimizer.zero_grad()
if args.sparse_type == 'L2':
sparse_loss = torch.mean(torch.pow(Matrix * (1. - torch.eye(n=latent_space[0], m=latent_space[0]).cuda()), 2.0))
elif args.sparse_type == 'L1':
sparse_loss = torch.mean(torch.abs(Matrix * (1. - torch.eye(n=latent_space[0], m=latent_space[0]).cuda())))
else:
raise NotImplemented('Type not implemented.')
feature_loss = torch.mean(torch.pow(F - F_rec, 2.0))
data_loss = torch.mean(torch.pow(X - X_rec, 2.0))
loss = feature_loss + args.beta0 * data_loss + args.beta1 * sparse_loss
loss.backward()
optimizer.step()
elif args.beta0 == 0:
Z = torch.randn([args.batch_size, 512, 1, 1], dtype=torch.float32).cuda()
F = generator([Z], which_block=args.layer, pre_model=True)
F_reshape = F.transpose(0, 1).reshape((latent_space[0], -1))
F_rec = torch.matmul(Matrix * (1. - torch.eye(n=latent_space[0], m=latent_space[0]).cuda()), F_reshape). \
reshape((latent_space[0], args.batch_size,) + tuple(latent_space[1:])).transpose(0, 1)
optimizer.zero_grad()
if args.sparse_type == 'L2':
sparse_loss = torch.mean(
torch.pow(Matrix * (1. - torch.eye(n=latent_space[0], m=latent_space[0]).cuda()), 2.0))
elif args.sparse_type == 'L1':
sparse_loss = torch.mean(
torch.abs(Matrix * (1. - torch.eye(n=latent_space[0], m=latent_space[0]).cuda())))
else:
raise NotImplemented('Type not implemented.')
feature_loss = torch.mean(torch.pow(F - F_rec, 2.0))
loss = feature_loss + args.beta1 * sparse_loss
loss.backward()
optimizer.step()
else:
raise NotImplemented()
if iter % args.report_value == 0:
if args.beta0 == 0:
print('Iter %d, Layer %d, loss=%.6f, f_loss=%.6f, sparse_loss=%.6f' %
(iter, args.layer, float(loss.item()), float(feature_loss.item()), float(sparse_loss.item())))
else:
print('Iter %d, Layer %d, loss=%.6f, f_loss=%.6f, x_loss=%.6f, sparse_loss=%.6f' %
(iter, args.layer, float(loss.item()), float(feature_loss.item()),
float(data_loss.item()), float(sparse_loss.item())))
losses_dict['total_loss'].append(float(loss.item()))
losses_dict['feature_loss'].append(float(feature_loss.item()))
losses_dict['data_loss'].append(float(data_loss.item()))
losses_dict['sparse_loss'].append(float(sparse_loss.item()))
if iter % args.report_image == 0:
# save reconstruction images.
if args.beta0 == 0:
X_rec = generator([F_rec], which_block=args.layer, post_model=True).detach()
X = generator([F], which_block=args.layer, post_model=True).detach()
test_feature = generator([test_zs[:4]], which_block=args.layer, pre_model=True).detach()
test_feature_reshape = test_feature.transpose(0, 1).reshape((latent_space[0], -1))
test_feature_rec = torch.matmul(Matrix * (1. - torch.eye(n=latent_space[0], m=latent_space[0]).cuda()),
test_feature_reshape). \
reshape((latent_space[0], args.batch_size,) + tuple(latent_space[1:])).transpose(0, 1)
test_rec = generator([test_feature_rec], which_block=args.layer, post_model=True).detach()
test_images = generator([test_feature], which_block=args.layer, post_model=True).detach()
image_number = min(64, args.batch_size)
Xs = torch.clamp(_tanh_to_sigmoid(torch.cat([test_rec,
test_images], dim=0).detach().cpu()), min=0.0, max=1.0)
torchvision.utils.save_image(Xs, os.path.join(out_dir, 'ReconstructionImages_%d.png' % iter), nrow=4)
# visualize the affinity matrix.
Matrix_abs = torch.relu(Matrix)
S_val = Matrix_abs.detach().cpu().numpy()
S_val = thrC(S_val.copy().T, args.alpha).T
#S_val_norm = S_val / np.sum(S_val, axis=1, keepdims=True)
#S_val_norm = S_val_norm / np.sum(S_val_norm, axis=0, keepdims=True)
#S_val_visual = np.clip(S_val_norm, 0, args.times * 1/512)
#S_val_visual = S_val_visual / np.max(S_val_visual, axis=1, keepdims=True)
#plt.figure()
#plt.imshow(S_val_visual, cmap='Oranges')
#plt.savefig(os.path.join(out_dir, 'thrd_matrix_iter_%d.png' % iter))
predict, L_val = post_proC(S_val, args.n_subs, args.d_subs, args.power)
p_sum = [sum(predict == k) for k in range(1, args.n_subs+1, 1)]
p_sum = np.array(p_sum)
print(p_sum)
p_sort = np.argsort(p_sum)[::-1]
predict_new = predict.copy()
for i in range(1, args.n_subs+1, 1):
predict_new[predict == (p_sort[i - 1] + 1)] = i
predict = predict_new.copy()
p_sum = [sum(predict == k) for k in range(1, args.n_subs+1, 1)]
print(predict)
print(p_sum)
S_val_blockized = switch_matrix(S_val.copy(), predict.copy())
# S_val_blockized = np.clip(S_val_blockized, 0, S_val_blockized.mean() * args.times)
torchvision.utils.save_image(torch.from_numpy(S_val_blockized + S_val_blockized.T),
os.path.join(out_dir, 'SwitchedMatrix%d.png' % iter), nrow=1,
normalize=True)
sel_idx = np.random.choice(test_zs.shape[0], size=[image_number], replace=False)
F = generator([test_zs[sel_idx]], which_block=args.layer, pre_model=True).detach()
features = F.detach().cpu().numpy()
for class_i in range(1, args.n_subs+1, 1):
ex_images = []
for ii in range(image_number):
ex_rows = []
for jj in range(image_number):
f_a = features[ii].copy()
f_b = features[jj].copy()
f_a[predict == class_i] = f_b[predict == class_i]
f_a = f_a.reshape((1, ) + latent_space)
ex_rows.append(f_a)
ex_rows = np.concatenate(ex_rows, axis=0).astype(np.float32)
ex_rows = torch.from_numpy(ex_rows).cuda()
ex_ys = generator([ex_rows], which_block=args.layer, post_model=True).detach()
ex_ys = torch.clamp(_tanh_to_sigmoid(ex_ys), min=0.0, max=1.0)
ex_images.append(ex_ys)
ex_images = torch.cat(ex_images, dim=0)
torchvision.utils.save_image(ex_images.detach().cpu(), os.path.join(out_dir,
'layer_%d_class_%d_iter_%d.png' %
(args.layer, class_i, iter)),
nrow=image_number)
# the subspace mask of class_i, save it.
subspace_i = predict == class_i
with open(os.path.join(out_dir, 'subspace_mask_layer%d_iter%d_class%d.pkl' %
(args.layer, iter, class_i)), 'wb') as file_out:
pkl.dump(subspace_i, file_out)
print('Save layer %d iter %d class %d, out_dir=%s.' % (args.layer, iter, class_i, out_dir))
if iter % args.report_model == 0:
with open(os.path.join(out_dir, 'value%d_layer%d.pkl'%(iter, args.layer)), 'wb') as file_out:
pkl.dump(Matrix.detach().cpu().numpy(), file_out)
print('Save S.')
with open(os.path.join(out_dir, 'loss_dict%d_layer%d.pkl'%(iter, args.layer)), 'wb') as file_out:
pkl.dump(losses_dict, file_out)
print('Save dict.')
def str2bool(s):
if s.lower() in ['yes', 'true', 'y', 't']:
return True
elif s.lower() in ['no', 'false', 'n', 'f']:
return False
else:
raise NotImplemented()
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Layer-Wise Inversion')
# Image Path and Saving Path
parser.add_argument('-o', '--outputs',
default='./TRAIN2',
help='Path to save results.')
# Parameters for Multi-Code GAN Inversion
parser.add_argument('--inversion_type', default='PGGAN-Layerwise',
help='Inversion type, "PGGAN-Multi-Z" for Multi-Code-GAN prior.')
# Generator Setting, Check models/model_settings for available GAN models
parser.add_argument('--gan_model', default='pggan_celebahq',
help='The name of model used.', type=str)
# Loss Parameters
parser.add_argument('--image_size', type=int, default=1024,
help='Size of images for perceptual model')
parser.add_argument('--optimization', default='Adam',
help="['GD', 'Adam']. Optimization method used.")
parser.add_argument('--lr', default=1e-4,
help='Learning rate.', type=float)
parser.add_argument('--sparse_type', default='L2', type=str, help='the type of sparse term.')
parser.add_argument('--iterations', default=10000,
help='Number of optimization steps.', type=int)
parser.add_argument('--batch_size', default=4,
type=int, help='The number of batchsize')
parser.add_argument('--layer', default=3,
type=int, help='The number of batchsize')
parser.add_argument('--exp_id', default="SubspaceGAN2",
type=str, help='The prefix of the experiment. ')
parser.add_argument('--beta', default=1.0, type=float, help='...')
parser.add_argument('--n_subs', type=int, default=6, help='The number of cluster')
parser.add_argument('--d_subs', type=int, default=6, help='The number of subspace dimension.')
parser.add_argument('--power', type=float, default=3.0, help='The power of the alpha.')
parser.add_argument('--alpha', type=float, default=0.2, help='The power of the alpha.')
parser.add_argument('--times', type=float, default=3, help='the times')
# Video Settings
parser.add_argument('--video', type=str2bool, default=False, help='Save video. False for no video.')
parser.add_argument('--fps', type=int, default=24, help='Frame rate of the created video.')
parser.add_argument('--report_image', type=int, default=50)
parser.add_argument('--report_value', type=int, default=10)
parser.add_argument('--report_model', type=int, default=100)
parser.add_argument('--beta0', default=1.0,
help='Learning rate.', type=float)
parser.add_argument('--beta1', default=1.0,
help='Learning rate.', type=float)
args = parser.parse_args()
main(args)
#