-
Notifications
You must be signed in to change notification settings - Fork 11
/
hyperopt_celebA.py
140 lines (113 loc) · 5.69 KB
/
hyperopt_celebA.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
from __future__ import print_function
import argparse
import os
import anndata
import numpy as np
import scanpy as sc
from hyperas import optim
from hyperas.distributions import choice
from hyperopt import Trials, STATUS_OK, tpe
from matplotlib import pyplot as plt
from scipy import sparse
import trvae
def data():
DATASETS = {
"CelebA": {"name": 'celeba', "gender": "Male", 'attribute': "Smiling", 'source_key': -1, "target_key": 1,
"width": 64, 'height': 64, "n_channels": 3},
}
data_key = "CelebA"
data_dict = DATASETS[data_key]
data_name = data_dict['name']
source_key = data_dict.get('source_key', None)
target_key = data_dict.get('target_key', None)
img_width = data_dict.get("width", None)
img_height = data_dict.get("height", None)
n_channels = data_dict.get("n_channels", None)
attribute = data_dict.get('attribute', None)
gender = data_dict.get('gender', None)
data = trvae.prepare_and_load_celeba(file_path="./data/celeba/img_align_celeba.zip",
attr_path="./data/celeba/list_attr_celeba.txt",
landmark_path="./data/celeba/list_landmarks_align_celeba.txt",
gender=gender,
attribute=attribute,
max_n_images=50000,
img_width=img_width,
img_height=img_height,
restore=True,
save=True)
if sparse.issparse(data.X):
data.X = data.X.A
source_images = data.copy()[data.obs['condition'] == source_key].X
target_images = data.copy()[data.obs['condition'] == target_key].X
source_images = np.reshape(source_images, (-1, img_width, img_height, n_channels))
target_images = np.reshape(target_images, (-1, img_width, img_height, n_channels))
source_images /= 255.0
target_images /= 255.0
source_labels = np.zeros(shape=source_images.shape[0])
target_labels = np.ones(shape=target_images.shape[0])
train_labels = np.concatenate([source_labels, target_labels], axis=0)
train_images = np.concatenate([source_images, target_images], axis=0)
train_images = np.reshape(train_images, (-1, np.prod(source_images.shape[1:])))
preprocessed_data = anndata.AnnData(X=train_images)
preprocessed_data.obs['condition'] = train_labels
if data.obs.columns.__contains__('labels'):
preprocessed_data.obs['labels'] = data.obs['condition'].values
data = preprocessed_data.copy()
train_size = int(data.shape[0] * 0.85)
indices = np.arange(data.shape[0])
np.random.shuffle(indices)
train_idx = indices[:train_size]
test_idx = indices[train_size:]
data_train = data[train_idx, :]
data_valid = data[test_idx, :]
print(data_train.shape, data_valid.shape)
train_data = data_train.copy()[
~((data_train.obs['labels'] == -1) & (data_train.obs['condition'] == target_key))]
valid_data = data_valid.copy()[
~((data_valid.obs['labels'] == -1) & (data_valid.obs['condition'] == target_key))]
return train_data, valid_data, data_name
def create_model(train_data, valid_data, data_name):
z_dim_choices = {{choice([20, 40, 50, 60, 80, 100])}}
mmd_dim_choices = {{choice([64, 128, 256])}}
alpha_choices = {{choice([1.0, 0.1, 0.01, 0.001, 0.0001, 0.00001, 0.000001])}}
beta_choices = {{choice([1, 5, 10, 50, 100, 500, 1000])}}
# gamma_choices = {{choice([0.001, 0.01, 0.1, 1, 10.0])}}
batch_size_choices = {{choice([256, 512, 1024])}}
dropout_rate_choices = {{choice([0.1, 0.2, 0.5, 0.75])}}
network = trvae.DCtrVAE(x_dimension=(64, 64, 3),
z_dimension=z_dim_choices,
mmd_dimension=mmd_dim_choices,
alpha=alpha_choices,
beta=beta_choices,
gamma=0,
kernel='rbf',
arch_style=3,
train_with_fake_labels=False,
learning_rate=0.001,
model_path=f"./models/RCCVAE/hyperopt/{data_name}-{64}x{64}-{True}/{3}-{z_dim_choices}/",
gpus=4,
dropout_rate=dropout_rate_choices)
history = network.train(train_data,
use_validation=True,
valid_adata=valid_data,
n_epochs=10000,
batch_size=batch_size_choices,
verbose=2,
early_stop_limit=200,
shuffle=True,
save=True)
print(f'Best Reconstruction Loss of model: ({history.history["val_kl_reconstruction_loss"][0]})')
return {'loss': history.history["val_kl_reconstruction_loss"][0], 'status': STATUS_OK}
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Sample a trained autoencoder.')
arguments_group = parser.add_argument_group("Parameters")
arguments_group.add_argument('-n', '--max_evals', type=int, required=True,
help='name of dataset you want to train')
args = vars(parser.parse_args())
best_run, best_network = optim.minimize(model=create_model,
data=data,
algo=tpe.suggest,
max_evals=args['max_evals'],
trials=Trials())
print("All Done!")
print(best_run)