-
Notifications
You must be signed in to change notification settings - Fork 1
/
adversarial_domain_adaptation_utils.py
226 lines (189 loc) · 9.21 KB
/
adversarial_domain_adaptation_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
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
import numpy as np
import pandas as pd
import torch
from torch import nn
import matplotlib.pyplot as plt
from openTSNE import TSNE
from PIL import Image
from torch import autograd
def save_backbone_features(model, data_loader):
print('--- extract_backbone_features...')
model.eval()
features = []
for index, (raindrop, runoff_history, runoff) in enumerate(data_loader):
feature = model.extract_feature(raindrop.cuda().float())
features.append(feature.detach().cpu().numpy())
features = np.array(features)
features = np.reshape(features, [-1, 36])
print(features.shape)
np.save('TunXi_features.npy', features)
print('--- feature saved to TunXi_features.npy')
class GANDataset(torch.utils.data.Dataset):
def __init__(self, forecast_range, mode, train_test_split_ratio, sample_length, dataset='ChangHua', training_set_scale=1):
assert mode in ['train', 'test']
assert forecast_range >= 1
self.mode = mode
self.sample_length = sample_length
self.forecast_range = forecast_range
self.training_set_scale = training_set_scale
self.overlapping_split = 20
# 加载屯溪编码器输出的特征
self.TunXi_feature = np.load(r'TunXi_features.npy')
self.TunXi_feature_cursor = 0
csv_file = f'dataset/{dataset}/data.csv'
print('--- reading', csv_file)
df = pd.read_csv(csv_file)
df = df.dropna()
data = np.array(df)
raindrop = data[:-self.forecast_range, 1:]
raindrop = (raindrop - np.min(raindrop, axis=0)) / (np.max(raindrop, axis=0) - np.min(raindrop, axis=0))
runoff_history = data[:-self.forecast_range, 0]
runoff = data[self.forecast_range:, 0]
train_test_spilt = int(train_test_split_ratio * len(raindrop))
training_subset = int(train_test_spilt * self.training_set_scale)
if self.mode == 'train':
self.raindrop = raindrop[:training_subset, :]
self.runoff_history = runoff_history[:training_subset]
self.runoff = runoff[:training_subset]
elif self.mode == 'test':
self.raindrop = raindrop[train_test_spilt:, :]
self.runoff_history = runoff_history[train_test_spilt:]
self.runoff = runoff[train_test_spilt:]
print(f'--- loaded [{dataset}] [{self.mode}] set: '
f'input shape: {self.raindrop.shape}, '
f'target shape: {self.runoff.shape}, len(dataset)={len(self)}')
def __getitem__(self, index):
start = int(index / self.overlapping_split * self.sample_length) # 20-overlapping samples
if self.TunXi_feature_cursor >= self.TunXi_feature.shape[0]:
self.TunXi_feature_cursor = 0
feature_start = self.TunXi_feature_cursor
feature = self.TunXi_feature[feature_start:feature_start + self.sample_length]
self.TunXi_feature_cursor = self.TunXi_feature_cursor + self.sample_length
return torch.from_numpy(self.raindrop[start:start + self.sample_length]), \
torch.from_numpy(self.runoff_history[start:start + self.sample_length]), \
torch.from_numpy(self.runoff[start:start + self.sample_length]), \
torch.from_numpy(feature)
def __len__(self):
return int(len(self.raindrop) / self.sample_length * self.overlapping_split) - self.overlapping_split
def get_input_size(self):
return self.raindrop.shape[1]
def get_TunXi_feature(self):
return self.TunXi_feature
class Discriminator_1DCNN(nn.Module):
def __init__(self):
super(Discriminator_1DCNN, self).__init__()
self.encoder = nn.Sequential(
nn.Conv1d(36, 64, kernel_size=5, stride=1, padding=2),
nn.ReLU(),
nn.MaxPool1d(kernel_size=5, stride=3),
nn.Conv1d(64, 32, kernel_size=5, stride=1, padding=2),
nn.ReLU(),
nn.MaxPool1d(kernel_size=5, stride=2),
nn.Conv1d(32, 32, kernel_size=5, stride=1, padding=2),
nn.ReLU(),
nn.MaxPool1d(kernel_size=5, stride=2),
)
self.fc = nn.Sequential(
nn.Linear(32, 32),
nn.ReLU(),
nn.Linear(32, 32),
nn.ReLU(),
nn.Linear(32, 1)
)
def forward(self, x):
x = x.flatten(start_dim=2).transpose(1, 2)
x = self.encoder(x)
x = self.fc(x.transpose(1, 2))
x = torch.mean(x, dim=1)
return x
def calc_gradient_penalty_ST(D, real_data, fake_data, term=None):
if term is None:
term = ['real', 'fake', 'real_fake', 'real_motion', 'fake_motion']
loss = 0
center = 0
if 'real' in term:
output = D(real_data.requires_grad_(True))
gradients = autograd.grad(outputs=output, inputs=real_data, grad_outputs=torch.ones(output.size()).cuda(),
create_graph=True, retain_graph=True, only_inputs=True)[0]
norm_real = gradients.norm(2, dim=1)
GP_real = ((norm_real - center) ** 2).mean()
loss += GP_real
if 'fake' in term:
output = D(fake_data.requires_grad_(True))
gradients = autograd.grad(outputs=output, inputs=fake_data, grad_outputs=torch.ones(output.size()).cuda(),
create_graph=True, retain_graph=True, only_inputs=True)[0]
norm_fake = gradients.norm(2, dim=1)
GP_fake = ((norm_fake - center) ** 2).mean()
loss += GP_fake
if 'real_motion' in term:
real_motion = real_data - real_data.mean(dim=1).unsqueeze(1)
real_structure = real_data.mean(dim=1).unsqueeze(1)
fake_structure = fake_data.mean(dim=1).unsqueeze(1)
alpha = torch.rand(1).cuda()
input = (alpha * real_motion + alpha * fake_structure + (1 - alpha) * real_structure).requires_grad_(True)
output = D(input)
gradients = autograd.grad(outputs=output, inputs=input, grad_outputs=torch.ones(output.size()).cuda(),
create_graph=True, retain_graph=True, only_inputs=True)[0]
norm_real_motion = gradients.norm(2, dim=1)
GP_real_motion = ((norm_real_motion - center) ** 2).mean()
loss += GP_real_motion
if 'fake_motion' in term:
fake_motion = fake_data - fake_data.mean(dim=1).unsqueeze(1)
real_structure = real_data.mean(dim=1).unsqueeze(1)
fake_structure = fake_data.mean(dim=1).unsqueeze(1)
alpha = torch.rand(1).cuda()
input = (alpha * fake_motion + alpha * fake_structure + (1 - alpha) * real_structure).requires_grad_(True)
output = D(input)
gradients = autograd.grad(outputs=output, inputs=input, grad_outputs=torch.ones(output.size()).cuda(),
create_graph=True, retain_graph=True, only_inputs=True)[0]
norm_fake_motion = gradients.norm(2, dim=1)
GP_fake_motion = ((norm_fake_motion - center) ** 2).mean()
loss += GP_fake_motion
if 'real_fake' in term:
alpha = torch.Tensor(np.random.random((real_data.size(0), 1, 1))).cuda()
interpolates = (alpha * real_data + ((1 - alpha) * fake_data)).requires_grad_(True)
disc_interpolates = D(interpolates)
gradients = autograd.grad(outputs=disc_interpolates, inputs=interpolates,
grad_outputs=torch.ones(disc_interpolates.size()).cuda(),
create_graph=True, retain_graph=True, only_inputs=True)[0]
norm_real_fake = gradients.norm()
gradient_penalty = ((norm_real_fake - center) ** 2).mean()
loss += gradient_penalty
return loss
def plot_feature_tsne(prediction, feature):
print('fitting tsne')
tsne = TSNE() # TSNE降维，降到2
# 降维后的数据
data = np.concatenate([prediction, feature], axis=0)
data = tsne.fit(data)
x_min, x_max = np.min(data), np.max(data)
data = (data - x_min) / (x_max - x_min) # 对数据进行归一化处理
plt.scatter(data[:prediction.shape[0], 0], data[:prediction.shape[0], 1], color='r', s=1, label='ChangHua',
alpha=0.5)
plt.scatter(data[prediction.shape[0]:, 0], data[prediction.shape[0]:, 1], color='b', s=1, label='TunXi', alpha=0.5)
plt.legend(loc="best", fontsize=6)
image_path = 'temp_tsne.png'
plt.savefig(image_path)
plt.close()
image_PIL = Image.open(image_path)
img = np.array(image_PIL)
plt.close()
print('tsne ok')
return img
def plot_backbone_features(prediction, feature):
fig, axes = plt.subplots(nrows=1, ncols=6, figsize=(20, 5))
for i, plt_image in enumerate(
[prediction[0], feature[0], prediction[32], feature[32], prediction[63], feature[63]]):
axes[i].matshow(plt_image, aspect='auto')
if i % 2 == 0:
axes[i].set_title("ChangHua feature")
else:
axes[i].set_title("TunXi feature")
axes[i].grid(b=False)
image_path = 'temp_matshow.png'
plt.savefig(image_path, bbox_inches='tight')
plt.close()
image_PIL = Image.open(image_path)
img = np.array(image_PIL)
plt.close()
return img