-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_autoencoder.py
195 lines (163 loc) · 5.34 KB
/
train_autoencoder.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
import itertools
import os
import re
import numpy as np
import seaborn as sns
import sklearn.model_selection
import sklearn.preprocessing
import tensorflow as tf
import lib.math
import lib.models
from lib.tfcustom import (
AnnealingVariableCallback,
VariableTimeseriesBatchGenerator,
)
sns.set_style("darkgrid")
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
def _get_data(path):
"""
Loads all traces
"""
X = np.load(path, allow_pickle=True)["data"]
if X.shape[0] < 100:
raise ValueError("File is suspiciously small. Recheck!")
return X
def _preprocess(X, n_features, max_batch_size, train_size):
"""
Preprocess data into tensors and appropriate train/test sets
"""
idx = np.arange(0, len(X), 1)
(
X_train,
X_test,
idx_train,
idx_test,
) = sklearn.model_selection.train_test_split(X, idx, train_size=train_size)
mu, sg, *_ = lib.math.array_stats(X)
X_train, X_test = [
lib.math.standardize(X, mu, sg) for X in (X_train, X_test)
]
data, steps_per_epoch, sizes = [], [], []
for X in X_train, X_test:
# Batch into variable batches to speed up (but see caveats)
Xi = VariableTimeseriesBatchGenerator(
X=X.tolist(),
y=None,
max_batch_size=max_batch_size,
shuffle_samples=True,
shuffle_batches=True,
)
steps = Xi.steps_per_epoch
batch_sizes = Xi.batch_sizes
X = tf.data.Dataset.from_generator(
generator=Xi,
output_types=(tf.float64, tf.float64),
output_shapes=((None, None, n_features), (None, None, n_features)),
)
sizes.append(batch_sizes)
steps_per_epoch.append(steps)
data.append(X)
info = idx_train, idx_test, mu, sg
return data, steps_per_epoch, info
if __name__ == "__main__":
MODELF = (lib.models.lstm_vae_bidir,)
INPUT_NPZ = ("data/preprocessed/combined_filt5_var.npz",)
N_TIMESTEPS = None
EARLY_STOPPING = 3
EPOCHS = 100
TRAIN_TEST_SIZE = 0.8
BATCH_SIZE = (4,)
CONTINUE_DIR = None
# Remember to end everything with a comma to make single values iterable
LSTM_UNITS = (128,) # LSTM memory capacity. Set as high as possible to avoid bottleneck
LATENT_DIM = (64,) # default at least 16, but higher may work better
EPS = (1,) # default 1
ANNEAL_TIME = (1,) # default 1
KEEP_ONLY = (None,) # select channel to keep, 'None' if keep all
ACTIVATION = (None,) # experimental, keep to 'None'
# Add iterables here
for (
_input_npz,
_batch_size,
_latent_dim,
_activation,
_eps,
_zdim,
_anneal_time,
_keep_only,
_modelf,
) in itertools.product(
INPUT_NPZ,
BATCH_SIZE,
LSTM_UNITS,
ACTIVATION,
EPS,
LATENT_DIM,
ANNEAL_TIME,
KEEP_ONLY,
MODELF,
):
X_raw = _get_data(_input_npz)
if _keep_only is not None:
X_raw = np.array([x[:, _keep_only].reshape(-1, 1) for x in X_raw])
N_FEATURES = X_raw[0].shape[-1]
# Pre-define loss so it gets compiled in the graph
KL_LOSS = tf.Variable(0.0)
build_args = [
N_TIMESTEPS,
N_FEATURES,
_latent_dim,
KL_LOSS,
_eps,
_zdim,
_activation,
]
TAG = "_{}".format(_modelf.__name__)
TAG += "_data={}".format(_input_npz.split("/")[-1]) # input data
TAG += "_dim={}".format(_latent_dim) # LSTM latent dimension
TAG += "_act={}".format(_activation) # activation function
TAG += "_bat={}".format(_batch_size) # batch size
TAG += "_eps={}_zdim={}_anneal={}".format(
_eps, _zdim, _anneal_time
) # vae parameters
if _keep_only is not None:
TAG += "_single={}".format(
_keep_only
) # Keep only one of the features
data, steps_per_epoch, info = _preprocess(
X=X_raw,
n_features=N_FEATURES,
max_batch_size=_batch_size,
train_size=TRAIN_TEST_SIZE,
)
(X_train, X_test), (X_train_steps, X_test_steps) = data, steps_per_epoch
model, callbacks, initial_epoch, model_dir = lib.models.model_builder(
model_dir=CONTINUE_DIR,
tag =TAG,
weights_only=False,
patience=EARLY_STOPPING,
model_build_f=_modelf,
build_args=build_args,
)
if re.search(string=_modelf.__name__, pattern="vae") is not None:
print("re-initialized KL loss")
KL_LOSS.assign(value=0.0)
callbacks.append(
AnnealingVariableCallback(
var=KL_LOSS,
anneal_over_n_epochs=_anneal_time,
anneals_starts_at=2,
)
)
model.summary()
model.fit(
x=X_train.repeat(),
validation_data=X_test.repeat(),
epochs=EPOCHS,
steps_per_epoch=X_train_steps,
validation_steps=X_test_steps,
initial_epoch=initial_epoch,
callbacks=callbacks,
)
# Save indices and normalization values to the newly created model directory
np.savez(os.path.join(model_dir, "info.npz"), info=info)