-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy patheval.py
173 lines (132 loc) · 5.1 KB
/
eval.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
import numpy as np
import torch
import torch.nn as nn
from torchvision.utils import save_image
from utils.utils import load_model, denorm
import copy
import json
import os
from models.ACGAN import Generator, Discriminator
# 10, 12, 9
epoch = 40180
noise_dim = 100
result_dir = 'results'
G_path = 'runs/acgan/ckpt/G_{}.pth'.format(epoch)
D_path = 'runs/acgan/ckpt/D_{}.pth'.format(epoch)
def load_G(G_path):
prev_state = torch.load(G_path)
G = Generator(noise_dim, sum(class_num))
G.load_state_dict(prev_state['model'])
G.eval()
return G
def load_D(D_path):
prev_state = torch.load(D_path)
D = Discriminator(sum(class_num))
D.load_state_dict(prev_state['model'])
D.eval()
return D
def sample_class_gradient(class_num, change_dim):
labels = []
for i, c in enumerate(class_num):
one_hot = torch.zeros(1, c)
if i != change_dim:
label = np.random.randint(0, c)
one_hot[0, label] = 1
labels.append(one_hot)
batch_labels = [copy.deepcopy(labels) for _ in range(class_num[change_dim])]
for j in range(class_num[change_dim]):
batch_labels[j][change_dim][0, j] = 1
batch_labels = [torch.cat(label, 1) for label in batch_labels]
batch_labels = torch.cat(batch_labels, 0)
return batch_labels
def sample_class_fix(class_num, batch_size, fix):
if fix is not None:
assert(len(fix) == len(class_num))
assert(all([fix[i] < class_num[i] for i in range(len(class_num))]))
labels = []
for i, c in enumerate(class_num):
if fix:
label = torch.LongTensor(batch_size, 1).fill_(fix[i])
else:
label = torch.LongTensor(batch_size, 1).random_() % c
one_hot = torch.zeros(batch_size, c).scatter(1, label, 1)
labels.append(one_hot)
labels = torch.cat(labels, 1)
return labels
def generate_class_gradient(class_num, change_dim):
G = load_G(G_path).cuda()
dim = 'eye' if change_dim == 0 else 'hair'
c = sample_class_gradient(class_num, change_dim).cuda()
z = torch.empty(1, 100).normal_(0, 0.9).repeat(class_num[change_dim], 1).cuda()
img = denorm(G(z, c))
save_image(img, os.path.join(result_dir, 'change_{}.png'.format(dim)), nrow = class_num[change_dim], pad_value = 255)
def generate_class_fix(class_num, batch_size, best_size = 64, fix = None):
eye_label, hair_label = get_label_map()
print(eye_label, hair_label)
if fix:
eye, hair = fix
img_name = '{}_eye_{}_hair.png'.format(eye_label[eye], hair_label[hair])
else:
img_name = 'class_fix.png'
G = load_G(G_path).cuda()
D = load_D(D_path).cuda()
torch.set_printoptions(threshold = 5000)
c = sample_class_fix(class_num, 1, fix).repeat(batch_size, 1).cuda()
z = torch.empty(batch_size, noise_dim).normal_(0, 0.9).cuda()
img = G(z, c)
score, pred = D(img)
print(score)
top_score, idx = torch.topk(score, best_size)
idx = idx.cpu().detach().numpy()
print(idx, top_score)
img = denorm(img[idx])
save_image(img, os.path.join(result_dir, img_name), pad_value = 255)
def generate_class_map(class_num):
G = load_G(G_path).cuda()
eye_class, hair_class = class_num
label_batch = []
for i in range(eye_class):
eye_label = torch.zeros(1, eye_class)
eye_label[0, i] = 1
for j in range(hair_class):
hair_label = torch.zeros(1, hair_class)
hair_label[0, j] = 1
label = torch.cat([eye_label, hair_label], 1)
label_batch.append(label)
c = torch.cat(label_batch, 0).cuda()
z = torch.empty(1, noise_dim).normal_(0, 0.9).repeat(eye_class * hair_class, 1).cuda()
img = G(z, c)
save_image(denorm(img), os.path.join(result_dir, 'class_map.png'), nrow = hair_class, pad_value = 255)
def interpolate(class_num, steps):
"""
Interpolate.
"""
G = load_G(G_path).cuda()
z_batch = []
c_batch = []
c1 = sample_class_fix(class_num, 1, None).cuda()
c2 = sample_class_fix(class_num, 1, None).cuda()
z1 = torch.empty(1, noise_dim).normal_(0, 0.9).cuda()
z2 = torch.empty(1, noise_dim).normal_(0, 0.9).cuda()
c_delta = (c2 - c1) / steps
z_delta = (z2 - z1) / steps
for i in range(steps + 1):
c_batch.append(c1 + i * c_delta)
z_batch.append(z1 + i * z_delta)
c = torch.cat(c_batch, 0)
z = torch.cat(z_batch, 0)
img = G(z, c)
save_image(denorm(img), os.path.join(result_dir, 'interpolate.png'), nrow = (steps + 1), pad_value = 255)
def get_label_map():
eye_label = json.load(open('data/eye_label.json'))
hair_label = json.load(open('data/hair_label.json'))
eye_label = { v: k for k, v in eye_label.items() }
hair_label = { v: k for k, v in hair_label.items() }
return eye_label, hair_label
if __name__ == '__main__':
class_num = (10, 12)
#generate_class_map(class_num)
generate_class_gradient(class_num, 1)
#generate_class_fix(class_num, 1024, best_size = 8, fix = (7, 8))
#interpolate(class_num, 8)
#get_label_map()