-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathutils.py
174 lines (146 loc) · 7.57 KB
/
utils.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
import logging
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable
import pandas as pd
import seaborn as sn
import matplotlib
import matplotlib.pyplot as plt
from matplotlib.pyplot import figure
matplotlib.use('Agg')
from prdc import compute_prdc
import datasets
def get_scores(x, score_model, noise_conditioned=False, std_value=7.5):
with torch.no_grad():
std = torch.ones((x.shape[0],)).to(x.device) * std_value
score = score_model(x, std) if noise_conditioned else score_model(x)
return score
def get_conditional_scores(x, score_model, classifier_model, class_id=None, noise_conditioned=False, std_value=7.5):
with torch.no_grad():
# Get prior scores
std = torch.ones((x.shape[0],)).to(x.device) * std_value
score = score_model(x, std) if noise_conditioned else score_model(x)
grad = torch.zeros(score.shape).to(x.device)
# Get likelihood scores
torch.set_grad_enabled(True)
x_grad = Variable(x.detach().clone(), requires_grad=True)
sm = nn.Softmax(dim=1)
pred = sm(classifier_model(x_grad, std)) if noise_conditioned else sm(classifier_model(x_grad))
pred = torch.log(pred)
pred[torch.arange(pred.shape[0]), class_id].sum().backward()
torch.set_grad_enabled(False)
grad = x_grad.grad
return score, grad
def oracle_score_denominator(point, batch, sigma):
return torch.exp( - ( ( (batch[:, 0]-point[0])**2 + (batch[:, 1]-point[1])**2 ) / (2*(sigma**2)) ) ) / ( 2*np.pi*(sigma**2) )
def oracle_score_numerator(point, batch, sigma):
p = torch.exp( - ( ( (batch[:, 0]-point[0])**2 + (batch[:, 1]-point[1])**2 ) / (2*(sigma**2)) ) ) / ( 2*np.pi*(sigma**2) )
diff = ( batch-point ) / (sigma**2)
diff[:, 0] *= p
diff[:, 1] *= p
return diff
def get_oracle_score(points, batch, sigma, eps=1e-8):
points = torch.tensor(points)
mean_numerator = torch.mean( oracle_score_numerator(points, batch, sigma = sigma), dim=0)
mean_denominator = torch.mean( oracle_score_denominator(points, batch, sigma = sigma), dim=0)
return mean_numerator / (mean_denominator + eps)
def get_oracle_scores(points, config):
config.data.dataset = 'inter_twinning_moon_upper'
upper_ds = datasets.get_dataset(config)
config.data.dataset = 'inter_twinning_moon_lower'
lower_ds = datasets.get_dataset(config)
iter_upper_ds = iter(upper_ds)
iter_lower_ds = iter(lower_ds)
batch_upper = torch.tensor(next(iter_upper_ds)['position']._numpy()).to(config.device)
batch_lower = torch.tensor(next(iter_lower_ds)['position']._numpy()).to(config.device)
config.data.dataset = 'inter_twinning_moon'
config.training.batch_size = config.training.batch_size * 2
ds = datasets.get_dataset(config)
iter_ds = iter(ds)
batch_full = torch.tensor(next(iter_ds)['position']._numpy()).to(config.device)
score_oracle_upper = np.zeros(points.shape)
score_oracle_lower = np.zeros(points.shape)
for i in range(points.shape[0]):
score_oracle_upper[i, :] = get_oracle_score(points[i], batch_upper, config.model.std).cpu().numpy()
score_oracle_lower[i, :] = get_oracle_score(points[i], batch_lower, config.model.std).cpu().numpy()
score_oracle_full = np.zeros(points.shape)
for i in range(points.shape[0]):
score_oracle_full[i, :] = get_oracle_score(points[i], batch_full, config.model.std).cpu().numpy()
score_oracle_likelihood_upper = score_oracle_upper - score_oracle_full
score_oracle_likelihood_lower = score_oracle_lower - score_oracle_full
return score_oracle_upper, score_oracle_lower, score_oracle_likelihood_upper, score_oracle_likelihood_lower, score_oracle_full
def calculate_prdc(config, samples_upper, samples_lower):
with torch.no_grad():
config.data.dataset = 'inter_twinning_moon_upper'
upper_ds = datasets.get_dataset(config)
config.data.dataset = 'inter_twinning_moon_lower'
lower_ds = datasets.get_dataset(config)
iter_upper_ds = iter(upper_ds)
iter_lower_ds = iter(lower_ds)
batch_upper = torch.tensor(next(iter_upper_ds)['position']._numpy())
batch_lower = torch.tensor(next(iter_lower_ds)['position']._numpy())
z = torch.randn_like(batch_upper)
batch_upper = batch_upper + z * config.model.std
batch_lower = batch_lower + z * config.model.std
nearest_k = 5
metrics_upper = compute_prdc(real_features=batch_upper.cpu().detach().numpy(),
fake_features=samples_upper.cpu().detach().numpy(),
nearest_k=nearest_k)
metrics_lower = compute_prdc(real_features=batch_lower.cpu().detach().numpy(),
fake_features=samples_lower.cpu().detach().numpy(),
nearest_k=nearest_k)
return metrics_upper, metrics_lower
def draw_and_save_data_points(w, h, samples_upper, samples_lower, filename):
cond_upper = torch.zeros(samples_upper.shape[0], dtype=torch.long)
cond_lower = torch.ones(samples_lower.shape[0], dtype=torch.long)
figure(figsize=(w/4, h/4), dpi=100)
plt.xlim((-w, w))
plt.ylim((-h, h))
data = np.concatenate((samples_upper.cpu().numpy(), samples_lower.cpu().numpy()), axis=0)
label = np.concatenate((cond_upper.cpu().numpy(), cond_lower.cpu().numpy()), axis=0)
plot_data = np.vstack((data.T, label)).T
df = pd.DataFrame(data=plot_data, columns=("x", "y", "label"))
sn.scatterplot(data=df, x="x", y="y", hue="label", alpha=0.8)
plt.xticks([])
plt.yticks([])
plt.legend('')
plt.savefig(filename)
def draw_and_save_vector_field(x, y, w, h, scores, filename):
figure(figsize=(w/4, h/4), dpi=100)
plt.quiver(x, y, scores[:,0], scores[:,1])
plt.xticks([])
plt.yticks([])
plt.savefig(filename)
def plot_vector_field(config, score_fn, dir_file, w=40, h=25, density=35, noise_conditioned=False, std_value=7.5):
logging.info("Plotting Vector Field")
with torch.no_grad():
x, y = np.meshgrid(np.linspace(-w, w, density, dtype=np.float32), np.linspace(-h, h, density, dtype=np.float32))
points = np.concatenate((np.expand_dims(x.flatten(), axis=1), np.expand_dims(y.flatten(), axis=1)), axis=1)
points = torch.from_numpy(points).to(config.device)
if noise_conditioned:
std = torch.ones((points.shape[0],)).to(config.device) * std_value
points_vf = score_fn(points, std).cpu().numpy()
else:
points_vf = score_fn(points).cpu().numpy()
draw_and_save_vector_field(x, y, w, h, points_vf, dir_file)
def plot_vector_field_likelihood(config, classifier_model, dir_file, w=40, h=25, density=35, class_id=0, noise_conditioned=False, std_value=7.5):
def calculate_vector_field_likelihood(classifier_model, points, class_id, noise_conditioned, std_value):
with torch.no_grad():
torch.set_grad_enabled(True)
points_grad = Variable(points.detach().clone(), requires_grad=True)
sm = nn.Softmax(dim=1)
std = torch.ones((points.shape[0],)).to(points.device) * std_value
pred = sm(classifier_model(points_grad, std)) if noise_conditioned else sm(classifier_model(points_grad))
pred = torch.log(pred + 1e-8)
pred[torch.arange(pred.shape[0]), class_id].sum().backward()
torch.set_grad_enabled(False)
grad = points_grad.grad
points_vf = grad.cpu().numpy()
return points_vf
logging.info("Plotting Vector Field")
x, y = np.meshgrid(np.linspace(-w, w, density, dtype=np.float32), np.linspace(-h, h, density, dtype=np.float32))
points = np.concatenate((np.expand_dims(x.flatten(), axis=1), np.expand_dims(y.flatten(), axis=1)), axis=1)
points = torch.from_numpy(points).to(config.device)
points_vf = calculate_vector_field_likelihood(classifier_model, points, class_id, noise_conditioned, std_value)
draw_and_save_vector_field(x, y, w, h, points_vf, dir_file)