-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_rnn.py
132 lines (98 loc) · 4.38 KB
/
train_rnn.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
import sys
import math
import numpy as np
from keras.layers import Input, LSTM, Dense
from keras.models import Model
from keras import backend as K
from keras.callbacks import EarlyStopping
import config
import tensorflow as tf
HIDDEN_UNITS = 256
def get_mixture_coef(y_pred):
d = config.GAUSSIAN_MIXTURES * config.Z_DIM
rollout_length = K.shape(y_pred)[1]
pi = y_pred[:, :, :d]
mu = y_pred[:, :, d:(2 * d)]
log_sigma = y_pred[:, :, (2 * d):(3 * d)]
pi = K.reshape(pi, [-1, rollout_length, config.GAUSSIAN_MIXTURES, config.Z_DIM])
mu = K.reshape(mu, [-1, rollout_length, config.GAUSSIAN_MIXTURES, config.Z_DIM])
log_sigma = K.reshape(log_sigma, [-1, rollout_length, config.GAUSSIAN_MIXTURES, config.Z_DIM])
pi = K.exp(pi) / K.sum(K.exp(pi), axis=2, keepdims=True)
sigma = K.exp(log_sigma)
return pi, mu, sigma # , discrete
def tf_normal(y_true, mu, sigma, pi):
rollout_length = K.shape(y_true)[1]
y_true = K.tile(y_true, (1, 1, config.GAUSSIAN_MIXTURES))
y_true = K.reshape(y_true, [-1, rollout_length, config.GAUSSIAN_MIXTURES, config.Z_DIM])
oneDivSqrtTwoPI = 1 / math.sqrt(2 * math.pi)
result = y_true - mu
result = result * (1 / (sigma + 1e-8))
result = -K.square(result) / 2
result = K.exp(result) * (1 / (sigma + 1e-8)) * oneDivSqrtTwoPI
result = result * pi
result = K.sum(result, axis=2)
return result
class RNN():
def __init__(self):
self.models = self._build()
self.model = self.models[0]
self.forward = self.models[1]
self.hidden_units = HIDDEN_UNITS
def _build(self):
# Training model
input_x = Input(shape=(None, config.Z_DIM + config.ACTION_DIM))
lstm = LSTM(HIDDEN_UNITS, return_sequences=True, return_state=True)
lstm_output, _, _ = lstm(input_x)
mdn = Dense(config.GAUSSIAN_MIXTURES * (3 * config.Z_DIM))(lstm_output) # + discrete_dim
rnn = Model(input_x, mdn)
# Predictive model
input_h = Input(shape=(HIDDEN_UNITS,))
input_c = Input(shape=(HIDDEN_UNITS,))
inputs = [input_h, input_c]
_, state_h, state_c = lstm(input_x, initial_state=[input_h, input_c])
forward = Model([input_x] + inputs, [mdn, state_h, state_c])
def r_loss(y_true, y_pred):
pi, mu, sigma = get_mixture_coef(y_pred)
res = tf_normal(y_true, mu, sigma, pi)
res = -K.log(res + 1e-8)
res = K.mean(res, axis=(1, 2))
return res
def kl_loss(y_true, y_pred):
pi, mu, sigma = get_mixture_coef(y_pred)
kl_loss = - 0.5 * K.mean(1 + K.log(K.square(sigma)) - K.square(mu) - K.square(sigma), axis=[1, 2, 3])
return kl_loss
def loss_func(y_true, y_pred):
return r_loss(y_true, y_pred)
rnn.compile(loss=loss_func, optimizer='rmsprop', metrics=[r_loss, kl_loss])
return (rnn, forward)
def set_weights(self, filepath):
self.model.load_weights(filepath)
def train(self, rnn_input, rnn_output, validation_split=0.2):
earlystop = EarlyStopping(monitor='val_loss', min_delta=0.0001, patience=5, verbose=2, mode='auto')
callbacks_list = [earlystop]
print('--------------')
self.model.fit(rnn_input, rnn_output,
shuffle=True,
epochs=config.RNN_EPOCHS,
batch_size=config.RNN_BATCH_SIZE,
validation_split=validation_split,
callbacks=callbacks_list)
self.model.save_weights('./weights/rnn/weights_' + sys.argv[1] + '.h5')
def save_weights(self, filepath):
self.model.save_weights(filepath)
def main():
rnn = RNN()
if not config.NEW_MODEL:
rnn.set_weights('./weights/rnn/weights_' + sys.argv[1] + '.h5')
for batch_num in range(config.START_BATCH, config.MAX_BATCH + 1):
new_rnn_input = np.load('./data/' + sys.argv[1] + '/rnn_input_' + str(batch_num) + '.npy')
new_rnn_output = np.load('./data/' + sys.argv[1] + '/rnn_output_' + str(batch_num) + '.npy')
if batch_num > config.START_BATCH:
rnn_input = np.concatenate([rnn_input, new_rnn_input])
rnn_output = np.concatenate([rnn_output, new_rnn_output])
else:
rnn_input = new_rnn_input
rnn_output = new_rnn_output
rnn.train(rnn_input, rnn_output)
if __name__ == "__main__":
main()