def CnnMlpFeatureExtractor(obs, image_height=291, image_width=150):
image_size = image_height * image_width
assert len(obs.shape) == 2 and obs.shape[1] > image_size
feature_num = obs.shape[1] - image_size
grid_map_flat = tf.slice(obs, [0, 0], [1, image_size])
grid_map = tf.reshape(grid_map_flat, [1, image_height, image_width, 1])
extracted_features = nature_cnn(grid_map)
extracted_features = tf.layers.flatten(extracted_features)
# extracted_features = tf.Print(extracted_features,[tf.shape(extracted_features)], 'extracted features shape : ')
features = tf.slice(obs, [0, image_size], [1, feature_num])
# features = tf.Print(features, [tf.shape(features)], "features shape: " )
return tf.concat([extracted_features, features], 1)
def __init__(self,
sess,
ob_space,
ac_space,
n_env,
n_steps,
n_batch,
reuse=False,
**kwargs):
super(CnnMlpPolicy, self).__init__(sess,
ob_space,
ac_space,
n_env,
n_steps,
n_batch,
reuse=reuse,
scale=True)
with tf.variable_scope("model", reuse=reuse):
activ = tf.nn.relu
feature_num = 7
h = 291
w = 150
proc_obs_float = tf.cast(self.processed_obs, dtype=tf.float32)
grid_map_flat = tf.slice(proc_obs_float, [0, 0], [1, 43650])
grid_map = tf.reshape(grid_map_flat, [1, h, w, 1])
# kwargs['data_format']='NCHW'
extracted_features = nature_cnn(grid_map, **kwargs)
extracted_features = tf.layers.flatten(extracted_features)
features = tf.slice(proc_obs_float, [0, 43650], [1, feature_num])
pi_h = tf.concat([extracted_features, features], 1)
for i, layer_size in enumerate([128, 128, 128]):
pi_h = activ(
tf.layers.dense(pi_h, layer_size, name='pi_fc' + str(i)))
pi_latent = pi_h
vf_h = tf.concat([extracted_features, features], 1)
for i, layer_size in enumerate([32, 32]):
vf_h = activ(
tf.layers.dense(vf_h, layer_size, name='vf_fc' + str(i)))
value_fn = tf.layers.dense(vf_h, 1, name='vf')
vf_latent = vf_h
self._proba_distribution, self._policy, self.q_value = \
self.pdtype.proba_distribution_from_latent(pi_latent, vf_latent, init_scale=0.01)
self._value_fn = value_fn
self._setup_init()
def image_and_pose_network(observation, **kwargs):
""" Network to process image and pose data.
Use the stable baselines nature_cnn to process images. The resulting
feature vector is then combined with the pose estimate and given to an
LSTM (LSTM defined in PPO2 below).
Args:
raw_observations (tf.Tensor): 1D tensor containing image and
pose data.
Returns:
tf.Tensor: Feature vector.
"""
imgs, pose = decode_tensor_observations(observation)
image_features = nature_cnn(imgs)
return tf.concat((image_features, pose), axis=-1)
def __init__(self,
sess,
ob_space,
ac_space,
n_env,
n_steps,
n_batch,
reuse=False,
**kwargs):
super(CommunicationPolicy, self).__init__(sess,
ob_space,
ac_space,
n_env,
n_steps,
n_batch,
reuse=reuse,
scale=True)
with tf.variable_scope("model", reuse=reuse):
activ = tf.nn.relu
extracted_features = nature_cnn(self.processed_obs, **kwargs)
extracted_features = tf.layers.flatten(extracted_features)
pi_h = extracted_features
for i, layer_size in enumerate([128, 128, 128]):
pi_h = activ(
tf.layers.dense(pi_h, layer_size, name='pi_fc' + str(i)))
pi_latent = pi_h
vf_h = extracted_features
for i, layer_size in enumerate([32, 32]):
vf_h = activ(
tf.layers.dense(vf_h, layer_size, name='vf_fc' + str(i)))
value_fn = tf.layers.dense(vf_h, 1, name='vf')
vf_latent = vf_h
self._proba_distribution, self._policy, self.q_value = \
self.pdtype.proba_distribution_from_latent(pi_latent, vf_latent, init_scale=0.01)
self._value_fn = value_fn
self._setup_init()