def automatic_dpatch(hidden):
"""
"""
if self.parameters['obs_type'] == 'image':
shape = hidden.get_shape().as_list()
num_regions = shape[1]*shape[2]
hidden = tf.reshape(hidden, [-1, num_regions, shape[3]])
inf_hidden = net.fcn(hidden, 1, self.parameters["inf_num_predictors"],
None, 'fcn_auto_dset')
# for predictor in range(self.parameters['inf_num_predictors']):
# name = "weights"+str(predictor)
# weights = tf.get_variable(name, shape=(num_regions,1), dtype=tf.dtypes.float32,
# initializer=tf.ones_initializer, trainable=True)
# softmax_weights = tf.contrib.distributions.RelaxedOneHotCategorical(0.1, weights)
# softmax_weights = tf.reshape(softmax_weights,[num_regions,1])
# softmax_weights = tf.nn.softmax(weights, axis=0)
# inf_hidden.append(tf.reduce_sum(softmax_weights*hidden, axis=1))
# inf_hidden = tf.stack(inf_hidden, axis=1)
hidden_size = inf_hidden.get_shape().as_list()[2]*self.parameters['inf_num_predictors']
inf_hidden = tf.reshape(inf_hidden, shape=[-1, hidden_size])
else:
inf_hidden = net.fcn(hidden, 1, self.parameters["inf_num_predictors"],
None, 'fcn_auto_dset')
# shape = hidden.get_shape().as_list()
# num_variables = shape[1]
# for predictor in range(self.parameters['inf_num_predictors']):
# name = "weights"+str(predictor)
# weights = tf.get_variable(name, shape=(1, num_variables), dtype=tf.dtypes.float32,
# initializer=tf.ones_initializer, trainable=True)
# softmax_weights = tf.nn.softmax(weights, axis=0)
# inf_hidden.append(tf.reduce_sum(softmax_weights*hidden, axis=1))
# inf_hidden = tf.stack(inf_hidden, axis=1)
return inf_hidden#, softmax_weights
def attention(hidden_conv, inf_hidden):
"""
"""
shape = hidden_conv.get_shape().as_list()
num_regions = shape[1] * shape[2]
hidden_conv = tf.reshape(hidden_conv, [-1, num_regions, shape[3]])
inf_hidden_vec = []
for head in range(self.parameters['num_heads']):
linear_conv = net.fcn(hidden_conv, 1,
self.parameters['num_att_units'], None,
'att', 'att1_' + str(head))
linear_hidden = net.fcn(inf_hidden, 1,
self.parameters['num_att_units'], None,
'att', 'att2_' + str(head))
context = tf.nn.tanh(linear_conv +
tf.expand_dims(linear_hidden, 1))
attention_weights = net.fcn(context, 1, [1], None, 'att',
'att3_' + str(head))
attention_weights = tf.nn.softmax(attention_weights, axis=1)
d_patch = tf.reduce_sum(attention_weights * hidden_conv,
axis=1)
inf_hidden_vec.append(
tf.concat([
d_patch,
tf.reshape(attention_weights, shape=[-1, num_regions])
],
axis=1))
inf_hidden = tf.concat(inf_hidden_vec, axis=1)
return inf_hidden
def build_main_model(self):
"""
Builds neural network model to approximate policy and value functions
"""
if self.parameters['obs_type'] == 'image':
self.observation = tf.placeholder(shape=[
None, self.parameters["frame_height"],
self.parameters["frame_width"], self.parameters["num_frames"]
],
dtype=tf.float32,
name='observation')
else:
self.observation = tf.placeholder(
shape=[None, self.parameters['obs_size']],
dtype=tf.float32,
name='observation')
# normalize input
if self.parameters['env_type'] == 'atari':
self.observation = tf.cast(self.observation, tf.float32) / 255.
if self.convolutional:
self.feature_vector = net.cnn(self.observation,
self.parameters["num_conv_layers"],
self.parameters["num_filters"],
self.parameters["kernel_sizes"],
self.parameters["strides"],
tf.nn.relu, False, 'cnn')
network_input = c_layers.flatten(self.feature_vector)
else:
self.feature_vector = self.observation
network_input = self.feature_vector
if self.fully_connected:
hidden = net.fcn(network_input, self.parameters["num_fc_layers"],
self.parameters["num_fc_units"], tf.nn.relu,
'fcn')
if self.recurrent:
self.prev_action = tf.placeholder(shape=[None],
dtype=tf.int32,
name='prev_action')
self.prev_action_onehot = c_layers.one_hot_encoding(
self.prev_action, self.act_size)
# network_input = tf.concat([network_input, self.prev_action_onehot], axis=1)
c_in = tf.placeholder(tf.float32,
[None, self.parameters['num_rec_units']],
name='c_state')
h_in = tf.placeholder(tf.float32,
[None, self.parameters['num_rec_units']],
name='h_state')
self.seq_len = tf.placeholder(shape=None,
dtype=tf.int32,
name='sequence_length')
self.state_in = tf.contrib.rnn.LSTMStateTuple(c_in, h_in)
hidden, self.state_out = net.rnn(network_input, self.state_in,
self.parameters['num_rec_units'],
self.seq_len, 'rnn')
self.hidden = hidden