-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathrnd.py
173 lines (123 loc) · 7.67 KB
/
rnd.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
import numpy as np
import tensorflow as tf
from constants import *
from noisy_dense import noisy_dense
class RND(object):
def __init__(self, id, config, session, lr):
self.id = id + '_RND'
self.session = session
# Extract relevant configuration:
self.config = {}
self.config['env_name'] = config['env_name']
self.config['env_n_actions'] = config['env_n_actions']
self.config['env_obs_dims'] = config['env_obs_dims']
self.config['env_obs_form'] = config['env_obs_form']
self.config['experiment_setup'] = config['experiment_setup']
self.config['n_training_frames'] = config['n_training_frames']
# Hyperparameters
self.config['rnd_learning_rate'] = lr
# Hardcoded hyperparameters
self.config['rnd_adam_epsilon'] = 0.00015
self.config['rnd_output_size'] = 6
self.config['rnd_hidden_size'] = 128
normalization_coefficients = self.get_normalization_coefficients(self.config['env_name'])
self.obs_mean = np.clip(np.array(normalization_coefficients[0]), a_min=0.00001, a_max=None)
self.obs_std = np.clip(np.array(normalization_coefficients[1]), a_min=0.00001, a_max=None)
self.name_fixed = self.id + '/RND_FIXED'
self.name_online = self.id + '/RND_ONLINE'
self.input_fixed, self.output_fixed, self.evaluation_fixed = \
self.build_model(self.name_fixed, self.config['rnd_hidden_size'], self.config['rnd_output_size'])
self.input_online, self.output_online, self.evaluation_online = \
self.build_model(self.name_online, self.config['rnd_hidden_size'], self.config['rnd_output_size'])
self.labels, \
self.losses, \
self.minimises, \
self.error = self.build_training_op()
self.training_steps = 0
# ------------------------------------------------------------------------------------------------------------------
def dense_net(self, scope, inputs, hidden_size, output_size):
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
layer_1 = tf.layers.dense(inputs, hidden_size, use_bias=True,
kernel_initializer=tf.contrib.layers.variance_scaling_initializer(),
activation=tf.nn.relu, name='DENSE_LAYER_1')
layer_2 = tf.layers.dense(layer_1, output_size, use_bias=True,
kernel_initializer=tf.contrib.layers.variance_scaling_initializer(),
activation=None, name='DENSE_LAYER_2')
return layer_2
def noisy_dense_net(self, scope, inputs, hidden_size, output_size, evaluation):
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
layer_1, _, _ = noisy_dense(inputs, size=hidden_size, bias=True, evaluation=evaluation,
activation_fn=tf.nn.relu, name='N_DENSE_LAYER_1')
layer_2, _, _ = noisy_dense(layer_1, size=output_size, bias=True, evaluation=evaluation,
name='N_DENSE_LAYER_2')
return layer_2
def convolutional_net(self, scope, inputs):
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
layer_1 = tf.layers.conv2d(inputs=inputs,
filters=16,
kernel_size=(3, 3),
strides=(1, 1),
padding='VALID',
kernel_initializer=tf.contrib.layers.variance_scaling_initializer(),
activation=tf.nn.relu,
name='CONV_LAYER_1')
layer_out = tf.contrib.layers.flatten(layer_1)
print(layer_out.get_shape())
return layer_out
# ------------------------------------------------------------------------------------------------------------------
def build_model(self, name, dense_hidden_size, output_size):
input = None
if self.config['env_obs_form'] == NONSPATIAL:
input = tf.placeholder(tf.float32, [None, self.config['env_obs_dims'][0]], name=name + '_OBS')
elif self.config['env_obs_form'] == SPATIAL:
input = tf.placeholder(tf.float32, [None, self.config['env_obs_dims'][0],
self.config['env_obs_dims'][1],
self.config['env_obs_dims'][2]], name=name + '_OBS')
evaluation = tf.placeholder(tf.bool, name=name + '_EVALUATION')
latent_features = None
if self.config['env_obs_form'] == NONSPATIAL:
latent_features = input
elif self.config['env_obs_form'] == SPATIAL:
latent_features = self.convolutional_net(name, input)
output = self.dense_net(name, latent_features, dense_hidden_size, output_size)
return input, output, evaluation
# ------------------------------------------------------------------------------------------------------------------
def build_training_op(self):
label = tf.placeholder(tf.float32, [None, self.config['rnd_output_size']], name='LABELS_' + str(self.id))
error = tf.abs(label - self.output_online)
optimizer = tf.train.AdamOptimizer(self.config['rnd_learning_rate'], epsilon=self.config['rnd_adam_epsilon'])
losses = []
losses.append(tf.losses.mean_squared_error(labels=label, predictions=self.output_online))
minimises = []
for loss in losses:
minimises.append(optimizer.minimize(loss))
return label, losses, minimises, error
# ------------------------------------------------------------------------------------------------------------------
def train_model(self, obs_batch_in, loss_id, is_batch=True, normalize=True):
self.training_steps += 1
if normalize:
obs_batch_in = self.normalize_obs(obs_batch_in)
obs_batch = obs_batch_in if isinstance(obs_batch_in, list) else obs_batch_in
obs_batch = obs_batch if is_batch else [obs_batch]
feed_dict = {self.input_fixed: obs_batch, self.evaluation_fixed: False}
output_fixed = self.session.run([self.output_fixed], feed_dict=feed_dict)[0]
feed_dict = {self.input_online: obs_batch, self.evaluation_online: False,
self.labels: output_fixed}
loss, _, error = self.session.run([self.losses[loss_id], self.minimises[loss_id], self.error],
feed_dict=feed_dict)
return loss, np.abs(error)
# ------------------------------------------------------------------------------------------------------------------
def get_error(self, obs_in, evaluation=False, normalize=True):
obs = self.normalize_obs(obs_in) if normalize else obs_in
feed_dict = {self.input_fixed: [obs],
self.input_online: [obs],
self.evaluation_fixed: evaluation,
self.evaluation_online: evaluation}
output_fixed, output_online = self.session.run([self.output_fixed, self.output_online], feed_dict=feed_dict)
return ((output_fixed - output_online) ** 2).mean(axis=1)[0]
# ------------------------------------------------------------------------------------------------------------------
def normalize_obs(self, obs):
return np.clip(((obs - self.obs_mean) / self.obs_std), -5, 5)
# ------------------------------------------------------------------------------------------------------------------
def get_normalization_coefficients(self, env_name):
return RND_MEAN_COEFFS[env_name], RND_STD_COEFFS[env_name]