-
Notifications
You must be signed in to change notification settings - Fork 8
/
vae-stats.py
159 lines (124 loc) · 5.02 KB
/
vae-stats.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
import torch
import utils
from utils import tonp
from models import vae as vae_module
import matplotlib.pyplot as plt
import argparse
import os
import numpy as np
from torch.autograd import Variable
import seaborn as sns
from scipy.stats import norm
sns.set()
plt.switch_backend('agg')
def kl_q_vae_ub(n_samples, eps=1e-6):
w = np.concatenate([np.random.randn(*data.shape) * np.sqrt(eps) + data for _ in range(n_samples)])
w_inp = torch.FloatTensor(w).to(device)
z_mean, z_var = vae.encode(w_inp)
z = Variable(torch.randn(*z_mean.shape).cuda()) * torch.sqrt(z_var) + z_mean
w_mu, w_var = map(tonp, vae.decode(z))
z_mean, z_var = map(tonp, [z_mean, z_var])
entq = 0.5 * (np.log(2 * np.pi * eps) + 1) * k2
logp_theta = norm.logpdf(w, loc=w_mu, scale=np.sqrt(w_var)).sum(1)
kl_rp = 0.5 * (z_var + z_mean ** 2 - np.log(z_var) - 1).sum(1)
kl_vae_upper_bound = (kl_rp - entq - logp_theta).sum() / (n_samples * 1.)
return kl_vae_upper_bound
def kl_q_sn(eps=1e-6, lamb=1.):
return (data**2 / lamb + eps/lamb + np.log(lamb/eps) - 1.).sum() * 0.5
def kl_randomq_vae_ub(n_samples, s=1e-6):
w = np.concatenate([np.random.randn(*data.shape) * np.sqrt(s) for _ in range(n_samples)])
w_inp = torch.FloatTensor(w).to(device)
z_mean, z_var = vae.encode(w_inp)
z = Variable(torch.randn(*z_mean.shape).cuda()) * torch.sqrt(z_var) + z_mean
w_mu, w_var = map(tonp, vae.decode(z))
z_mean, z_var = map(tonp, [z_mean, z_var])
entq = 0.5 * (np.log(2 * np.pi * s) + 1) * k2
logp_theta = norm.logpdf(w, loc=w_mu, scale=np.sqrt(w_var)).sum(1)
kl_rp = 0.5 * (z_var + z_mean ** 2 - np.log(z_var) - 1).sum(1)
return (kl_rp - entq - logp_theta).sum() / (n_samples * 1.)
def kl_randq_sn(s=1e-6, lamb=1.):
return (s/lamb + np.log(lamb/s) - 1.) * np.prod(data.shape) * 0.5
parser = argparse.ArgumentParser()
parser.add_argument('--model')
parser.add_argument('--data')
parser.add_argument('--log-dir', default='./vaelog/')
parser.add_argument('--gpu-id', default='0')
parser.add_argument('--z-dim', default=8, type=int)
parser.add_argument('--hidden-dim', default=16, type=int)
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_id
args.cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if not os.path.exists(args.log_dir):
os.makedirs(args.log_dir)
data = np.load(args.data)
D = data.shape[-1]
if D == 3:
decoder = vae_module.Decoder3x3(args.z_dim, args.hidden_dim)
encoder = vae_module.Encoder3x3(args.z_dim, args.hidden_dim)
elif D == 5:
decoder = vae_module.Decoder5x5(args.z_dim, args.hidden_dim)
encoder = vae_module.Encoder5x5(args.z_dim, args.hidden_dim)
vae = vae_module.VAE(encoder, decoder, device=device)
vae.load_state_dict(torch.load(args.model))
# Reconstructions
n, m = 10, 5
np.random.seed(42)
fig, axes = plt.subplots(figsize=(15, 12), nrows=n, ncols=m)
img_idxs = np.random.randint(0, len(data), size=(n * m))
for i, ax in enumerate(axes.flat):
c = data[img_idxs[i]][np.newaxis]
inp = torch.FloatTensor(c).to(device)
_, [rec, _] = vae(inp)
rec = tonp(rec)
c = np.concatenate([c, rec], 3)[0, 0]
sns.heatmap(c, ax=ax)
ax.axis('off')
plt.savefig(os.path.join(args.log_dir, 'reconstruction'), dpi=200)
plt.clf()
# Samples
n, m = 10, 10
z = torch.randn(n * m, args.z_dim, 1, 1).to(device)
mu, var = vae.decode(z)
x = tonp(mu)[:, 0]
fig, axes = plt.subplots(figsize=(5, 5), nrows=n, ncols=m)
for i, ax in enumerate(axes.flat):
im = ax.imshow(x[i], cmap='viridis')
ax.axis('off')
plt.savefig(os.path.join(args.log_dir, 'samples'), dpi=200)
plt.clf()
# KL comparison
k2 = D**2
plt.figure(figsize=(12, 5))
plt.subplot(1, 2, 1)
eps_grid = np.logspace(-9, -3, 20)
kl_standart_normal = []
kl_vae_upper_bound = []
for eps in eps_grid:
kl_vae_upper_bound.append(kl_q_vae_ub(100, eps=eps))
kl_standart_normal.append(kl_q_sn(eps=eps))
plt.semilogx(eps_grid, np.log(np.asarray(kl_standart_normal) + 1), label=r'$KL(N(W, \varepsilon)\, || \,N(0, 1))$')
plt.semilogx(eps_grid, np.log(np.asarray(kl_vae_upper_bound) + 1), label=r'UB $KL(N(W, \varepsilon)\, ||\, p_{VAE})$')
plt.xlabel(r'$\varepsilon$')
plt.ylabel('log(x + 1)')
plt.legend()
plt.subplot(1, 2, 2)
s_grid = np.logspace(-8, -4, 9)
kl_rq_vae = []
kl_rq_sn = []
kl_standart_normal = []
kl_vae_upper_bound = []
for s in s_grid:
kl_rq_vae.append(kl_randomq_vae_ub(100, s=s))
kl_rq_sn.append(kl_randq_sn(s=s))
kl_vae_upper_bound.append(kl_q_vae_ub(100, eps=s))
kl_standart_normal.append(kl_q_sn(eps=s))
plt.semilogx(s_grid, np.log(np.asarray(kl_rq_vae) + 1), label=r'UB KL(N(0, $s$) || $p_{VAE}$ )')
plt.semilogx(s_grid, np.log(np.asarray(kl_vae_upper_bound) + 1), label=r'UB $KL(N(W, s)\, ||\, p_{VAE})$')
plt.semilogx(s_grid, np.log(np.asarray(kl_rq_sn) + 1), label=r'KL(N(0, $s$) || N(0, 1) )')
plt.semilogx(s_grid, np.log(np.asarray(kl_standart_normal) + 1), label=r'$KL(N(W, s)\, || \,N(0, 1))$')
plt.xlabel(r'$s$')
plt.ylabel('log(x + 1)')
plt.legend()
plt.savefig(os.path.join(args.log_dir, 'kl_plots'), dpi=200)
plt.clf()