-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtest_and_plot.py
191 lines (168 loc) · 10.1 KB
/
test_and_plot.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
import torch
from models.masks import SpecificParticleMask, KinematicMask
import utils
from sklearn.metrics import roc_curve, auc, accuracy_score
import scipy
import os
import numpy as np
import matplotlib.pyplot as plt
# Test loop
def test(loader, test_batch_size, X_test_arr, test_labels, names, models, device, mask, scaler, output_vars, information, model_name, lower=[0,-3.2,-1.6,-1], upper=[4,3.2,1.6,1]):
X_test_arr = X_test_arr.copy().reshape(X_test_arr.shape[0], X_test_arr.shape[1], X_test_arr.shape[2])
X_test_arr_tensor = torch.tensor(X_test_arr)
new_b_tags = np.expand_dims(X_test_arr[:,:,4] - X_test_arr[:,:,3], axis=-1)
X_test_arr = np.concatenate((X_test_arr[:,:,:3], new_b_tags), axis=2)
X_test_arr = X_test_arr.reshape(X_test_arr.shape[0], X_test_arr.shape[1] * X_test_arr.shape[2])
X_test_arr_hh = X_test_arr[test_labels==1]
X_test_arr_tt = X_test_arr[test_labels==0]
if information == 'autoencoder':
tae = models[0]
with torch.no_grad():
all_preds = []
for i in range(6):
outputs_arr = torch.zeros(X_test_arr_tensor.size(0), 6, 4+(output_vars % 3))
for batch_idx, batch in enumerate(loader):
# Move the data to the device
inputs, labels = batch
inputs = inputs.to(device)
if mask is not None:
if mask == 0:
mask_layer = SpecificParticleMask(output_vars+(output_vars%3), i)
else:
mask_layer = KinematicMask(mask)
# Mask input data
masked_inputs = mask_layer(inputs)
# Forward pass
outputs = tae(masked_inputs)
# Reset trivial values
mask_999 = (masked_inputs[:, :, 3] == 999).float()
outputs[:,:,3:5] = torch.nn.functional.softmax(outputs[:,:,3:5], dim=2)
outputs[:, :, 3] = (1 - mask_999) * outputs[:, :, 3] + mask_999 * 1
outputs[:, :, 4] = (1 - mask_999) * outputs[:, :, 4]
if output_vars == 3:
outputs_padded = torch.cat((outputs, torch.zeros(outputs.size(0), outputs.size(1), 1).to(device)), axis=2)
outputs_arr[batch_idx*test_batch_size:(batch_idx+1)*test_batch_size] = outputs_padded
else:
outputs_arr[batch_idx*test_batch_size:(batch_idx+1)*test_batch_size] = outputs
outputs_arr = outputs_arr.cpu().numpy()
if output_vars == 4:
new_b_tags = outputs_arr[:,:,3:5]
new_b_tags = np.expand_dims(new_b_tags[:,:,1] - new_b_tags[:,:,0], axis=-1)
outputs_arr = np.concatenate((outputs_arr[:,:,:3], new_b_tags), axis=2)
outputs_arr = outputs_arr.reshape(outputs_arr.shape[0], outputs_arr.shape[1]*outputs_arr.shape[2])
outputs_arr_hh = outputs_arr[test_labels==1]
outputs_arr_tt = outputs_arr[test_labels==0]
# Generate histograms
utils.make_hist2d(i, output_vars, X_test_arr_hh, outputs_arr_hh, scaler, 'di-Higgs', mask=mask_999.int().cpu().numpy(),
file_path='./outputs/' + model_name, lower=lower, upper=upper)
utils.make_hist2d(i, output_vars, X_test_arr_tt, outputs_arr_tt, scaler, 'ttbar',
mask=mask_999.int().cpu().numpy(), file_path='./outputs/' + model_name, lower=lower, upper=upper)
elif information == 'partial':
tae, classifier = models[0], models[1]
tae.eval()
classifier.eval()
with torch.no_grad():
all_preds = []
for i in range(6):
outputs_arr = torch.zeros(X_test_arr_tensor.size(0), 6, 4+(output_vars % 3))
outputs_arr_2 = torch.zeros(X_test_arr_tensor.size(0))
for batch_idx, batch in enumerate(loader):
# Move the data to the device
inputs, labels = batch
inputs = inputs.to(device)
labels = labels.to(device)
if mask is not None:
if mask == 0:
mask_layer = SpecificParticleMask(output_vars+(output_vars%3), i)
else:
mask_layer = KinematicMask(mask)
# Mask input data
masked_inputs = mask_layer(inputs)
# Forward pass
outputs = tae(masked_inputs)
if output_vars == 3:
outputs_padded = torch.cat((outputs, torch.zeros(outputs.size(0), outputs.size(1), 1).to(device)), axis=2)
outputs_arr[batch_idx*test_batch_size:(batch_idx+1)*test_batch_size] = outputs_padded
else:
outputs_arr[batch_idx*test_batch_size:(batch_idx+1)*test_batch_size] = outputs
# Reset trivial values
mask_999 = (masked_inputs[:, :, 3] == 999).float()
outputs[:,:,3:5] = torch.nn.functional.softmax(outputs[:,:,3:5], dim=2)
outputs[:, :, 3] = (1 - mask_999) * outputs[:, :, 3] + mask_999 * 1
outputs[:, :, 4] = (1 - mask_999) * outputs[:, :, 4]
masked_inputs[:,:,3:5] = torch.nn.functional.softmax(masked_inputs[:,:,3:5], dim=2)
masked_inputs[:, :, 3] = (1 - mask_999) * masked_inputs[:, :, 3] + mask_999 * 1
masked_inputs[:, :, 4] = (1 - mask_999) * masked_inputs[:, :, 4]
outputs = torch.reshape(outputs, (outputs.size(0),
outputs.size(1) * outputs.size(2)))
masked_inputs = torch.reshape(masked_inputs, (masked_inputs.size(0),
masked_inputs.size(1) * masked_inputs.size(2)))
outputs_2 = classifier(torch.cat((outputs, masked_inputs), axis=1)).squeeze(1)
outputs_arr_2[batch_idx*test_batch_size:(batch_idx+1)*test_batch_size] = outputs_2
outputs_arr = outputs_arr.cpu().numpy()
outputs_arr_2 = outputs_arr_2.cpu().numpy()
fpr, tpr, _ = roc_curve(test_labels, outputs_arr_2)
roc_auc = auc(fpr, tpr)
plt.plot([0, 1], [0, 1], 'k--')
plt.plot(fpr, tpr, label='(ROC-AUC = {:.3f})'.format(roc_auc))
plt.xlabel('False positive rate')
plt.ylabel('True positive rate')
masked_parts = ['lepton', 'missing energy', 'jet 1', 'jet 2', 'jet 3', 'jet 4']
plt.title('ROC curve masked ' + masked_parts[i])
plt.legend(loc='best')
binary_preds = [1 if p > 0.5 else 0 for p in outputs_arr_2]
acc = accuracy_score(test_labels, binary_preds)
print('Classification Accuracy (masked ', masked_parts[i], '): ', acc)
plt.savefig('./outputs/' + model_name + '/ROC_AUC_partial.png')
plt.show()
plt.close()
elif information == 'full':
tae, classifier = models[0], models[1]
tae.eval()
classifier.eval()
with torch.no_grad():
all_preds = []
outputs_arr_2 = torch.zeros(X_test_arr_tensor.size(0))
for batch_idx, batch in enumerate(loader):
# Move the data to the device
inputs, labels = batch
inputs = inputs.to(device)
labels = labels.to(device)
outputs = torch.zeros(inputs.size(0), 6, output_vars + output_vars % 3).to(device)
for i in range(6):
if mask is not None:
if mask == 0:
mask_layer = SpecificParticleMask(output_vars+ output_vars % 3, i)
else:
mask_layer = KinematicMask(mask)
# Mask input data
masked_inputs = mask_layer(inputs)
# Forward pass for autoencoder
temp_outputs = tae(masked_inputs)
outputs[:,i,:] = temp_outputs[:,i,:]
# Reset trivial values
mask_999 = (masked_inputs[:, :, 3] == 999).float()
outputs[:,:,3:5] = torch.nn.functional.softmax(outputs[:,:,3:5], dim=2)
outputs[:, :, 3] = (1 - mask_999) * outputs[:, :, 3] + mask_999 * 1
outputs[:, :, 4] = (1 - mask_999) * outputs[:, :, 4]
outputs = torch.reshape(outputs, (outputs.size(0),
outputs.size(1) * outputs.size(2)))
inputs = torch.reshape(inputs, (inputs.size(0),
inputs.size(1) * inputs.size(2)))
outputs_2 = classifier(torch.cat((outputs, inputs), axis=1)).squeeze(1)
outputs_arr_2[batch_idx*test_batch_size:(batch_idx+1)*test_batch_size] = outputs_2
outputs_arr_2 = outputs_arr_2.cpu().numpy()
fpr, tpr, _ = roc_curve(test_labels, outputs_arr_2)
roc_auc = auc(fpr, tpr)
plt.plot([0, 1], [0, 1], 'k--')
plt.plot(fpr, tpr, label='(ROC-AUC = {:.3f})'.format(roc_auc))
plt.xlabel('False positive rate')
plt.ylabel('True positive rate')
plt.title('ROC curve (full-information)')
plt.legend(loc='best')
binary_preds = [1 if p > 0.5 else 0 for p in outputs_arr_2]
acc = accuracy_score(test_labels, binary_preds)
print('Classification Accuracy (full-information): ', acc)
plt.savefig('./outputs/' + model_name + '/ROC_AUC_full.png')
plt.show()
plt.close()