def encode_image(self, inputs):
conv1 = model_utils.conv2d(inputs,
16,
3,
2,
scope='conv1',
max_pool=False)
conv2 = model_utils.conv2d(conv1,
32,
3,
2,
scope='conv2',
max_pool=False)
conv3 = model_utils.conv2d(conv2,
64,
3,
2,
scope='conv3',
max_pool=False)
conv4 = model_utils.conv2d(conv3,
128,
3,
2,
scope='conv4',
max_pool=False)
conv5 = model_utils.conv2d(conv4,
256,
3,
2,
scope='conv5',
max_pool=False)
shape = conv5.get_shape().as_list()
outputs = tf.reshape(conv5, shape=[-1, shape[1] * shape[2] * shape[3]
]) # b * l, -1
return outputs
def Model(self, inputs):
input_depth, input_cmd, input_prev_a, input_goal, input_action = inputs
# encode depth image
conv1 = model_utils.conv2d(input_depth,
4,
5,
4,
scope='conv1',
max_pool=False)
conv2 = model_utils.conv2d(conv1,
16,
5,
4,
scope='conv2',
max_pool=False)
conv3 = model_utils.conv2d(conv2,
32,
3,
2,
scope='conv3',
max_pool=False)
shape = conv3.get_shape().as_list()
depth_vect = tf.reshape(conv3,
shape=[-1,
shape[1] * shape[2] * shape[3]]) # b,d
# encode cmd
embedding_cmd = tf.get_variable('cmd_embedding',
[self.n_cmd_type, self.dim_emb])
cmd_vect = tf.reshape(tf.nn.embedding_lookup(embedding_cmd, input_cmd),
[-1, self.dim_emb])
# encode action
embedding_w_action = tf.get_variable('embedding_w_action',
[self.dim_action, self.dim_emb])
embedding_b_action = tf.get_variable('embedding_b_action',
[self.dim_emb])
prev_a_vect = tf.matmul(input_prev_a,
embedding_w_action) + embedding_b_action
action_vect = tf.matmul(
input_action, embedding_w_action) + embedding_b_action # b, d
# encode goal
embedding_w_goal = tf.get_variable('embedding_w_goal',
[self.dim_action, self.dim_emb])
embedding_b_goal = tf.get_variable('embedding_b_goal', [self.dim_emb])
goal_vect = tf.matmul(input_goal, embedding_w_goal) + embedding_b_goal
input_vect = tf.concat(
[depth_vect, cmd_vect, prev_a_vect, goal_vect, action_vect],
axis=1)
hidden_1 = model_utils.dense_layer(input_vect, self.n_hidden,
'hidden_1')
hidden_2 = model_utils.dense_layer(hidden_1, self.n_hidden / 2,
'hidden_2')
q = model_utils.dense_layer(
hidden_2,
1,
'q',
activation=None,
w_init=tf.initializers.random_uniform(-0.003, 0.003),
b_init=tf.initializers.random_uniform(-0.003, 0.003))
return q
def Model(self, inputs):
input_depth, input_cmd, input_prev_a, rnn_h_in = inputs
# encode depth image
conv1 = model_utils.conv2d(input_depth,
4,
5,
4,
scope='conv1',
max_pool=False)
conv2 = model_utils.conv2d(conv1,
16,
5,
4,
scope='conv2',
max_pool=False)
conv3 = model_utils.conv2d(conv2,
32,
3,
2,
scope='conv3',
max_pool=False)
shape = conv3.get_shape().as_list()
depth_vect = tf.reshape(conv3,
shape=[-1,
shape[1] * shape[2] * shape[3]]) # b,d
# encode cmd
embedding_cmd = tf.get_variable('cmd_embedding',
[self.n_cmd_type, self.dim_emb])
cmd_vect = tf.reshape(tf.nn.embedding_lookup(embedding_cmd, input_cmd),
[-1, self.dim_emb])
# encode prev action
embedding_w_action = tf.get_variable('embedding_w_action',
[self.dim_action, self.dim_emb])
embedding_b_action = tf.get_variable('embedding_b_action',
[self.dim_emb])
prev_a_vect = tf.matmul(input_prev_a,
embedding_w_action) + embedding_b_action
input_vect = tf.concat([depth_vect, cmd_vect, prev_a_vect], axis=1)
# rnn
if self.rnn_type == 'lstm':
rnn_cell = model_utils._lstm_cell(self.n_hidden,
1,
name='rnn_cell')
else:
rnn_cell = model_utils._gru_cell(self.n_hidden, 1, name='rnn_cell')
rnn_output, rnn_h_out = rnn_cell(input_vect, self.rnn_h_in)
# action
a_linear = model_utils.dense_layer(
rnn_output, 1, 'a_linear',
activation=tf.nn.sigmoid) * self.action_range[0]
a_angular = model_utils.dense_layer(
rnn_output, 1, 'a_angular',
activation=tf.nn.tanh) * self.action_range[1]
pred_action = tf.concat([a_linear, a_angular], axis=1)
return pred_action, rnn_h_out
def encode_image(self, inputs, activation=tf.nn.leaky_relu, scope='rgb'):
trainable = True
conv1 = model_utils.conv2d(inputs,
16,
3,
2,
scope=scope + '/conv1',
max_pool=False,
trainable=trainable,
activation=activation)
conv2 = model_utils.conv2d(conv1,
32,
3,
2,
scope=scope + '/conv2',
max_pool=False,
trainable=trainable,
activation=activation)
conv3 = model_utils.conv2d(conv2,
64,
3,
2,
scope=scope + '/conv3',
max_pool=False,
trainable=trainable,
activation=activation)
conv4 = model_utils.conv2d(conv3,
128,
3,
2,
scope=scope + '/conv4',
max_pool=False,
trainable=trainable,
activation=activation)
conv5 = model_utils.conv2d(conv4,
256,
3,
2,
scope=scope + '/conv5',
max_pool=False,
trainable=trainable,
activation=None)
shape = conv5.get_shape().as_list()
outputs = tf.reshape(conv5, shape=[-1, shape[1] * shape[2] * shape[3]
]) # b*l, dim_img_feat
return outputs
def Model(self, inputs):
input_depth, input_cmd, input_prev_a, input_action, gru_h_in, length = inputs
# encode depth image
conv1 = model_utils.conv2d(input_depth,
4,
5,
4,
scope='conv1',
max_pool=False)
conv2 = model_utils.conv2d(conv1,
16,
5,
4,
scope='conv2',
max_pool=False)
conv3 = model_utils.conv2d(conv2,
32,
3,
2,
scope='conv3',
max_pool=False)
shape = conv3.get_shape().as_list()
depth_vect = tf.reshape(conv3,
shape=[-1, shape[1] * shape[2] * shape[3]
]) # b*l,d
# encode cmd
embedding_cmd = tf.get_variable('cmd_embedding',
[self.n_cmd_type, self.dim_emb])
cmd_vect = tf.reshape(tf.nn.embedding_lookup(embedding_cmd, input_cmd),
[-1, self.dim_emb])
# encode prev action and action
embedding_w_action = tf.get_variable('embedding_w_action',
[self.dim_action, self.dim_emb])
embedding_b_action = tf.get_variable('embedding_b_action',
[self.dim_emb])
prev_a_vect = tf.matmul(input_prev_a,
embedding_w_action) + embedding_b_action
action_vect = tf.matmul(input_action,
embedding_w_action) + embedding_b_action
input_vect = tf.concat(
[depth_vect, cmd_vect, prev_a_vect, action_vect], axis=1)
rnn_cell = model_utils._gru_cell(self.n_hidden, 1, name='gru_cell')
shape = input_vect.get_shape().as_list()
input_vect_reshape = tf.reshape(input_vect,
[-1, self.max_step, shape[-1]])
rnn_output, _ = tf.nn.dynamic_rnn(rnn_cell,
input_vect_reshape,
sequence_length=length,
dtype=tf.float32) # b, l, h
rnn_output_reshape = tf.reshape(rnn_output,
[-1, self.n_hidden]) # b*l, h
# q
q = model_utils.dense_layer(
rnn_output_reshape,
1,
'q',
activation=None,
w_init=tf.initializers.random_uniform(-0.003, 0.003),
b_init=tf.initializers.random_uniform(-0.003, 0.003))
return [q]
def Model(self, inputs):
input_depth, input_cmd, input_prev_a, gru_h_in, length = inputs
# encode depth image
conv1 = model_utils.conv2d(input_depth,
4,
5,
4,
scope='conv1',
max_pool=False)
conv2 = model_utils.conv2d(conv1,
16,
5,
4,
scope='conv2',
max_pool=False)
conv3 = model_utils.conv2d(conv2,
32,
3,
2,
scope='conv3',
max_pool=False)
shape = conv3.get_shape().as_list()
depth_vect = tf.reshape(conv3,
shape=[-1,
shape[1] * shape[2] * shape[3]]) # b,d
# encode cmd
embedding_cmd = tf.get_variable('cmd_embedding',
[self.n_cmd_type, self.dim_emb])
cmd_vect = tf.reshape(tf.nn.embedding_lookup(embedding_cmd, input_cmd),
[-1, self.dim_emb])
# encode prev action
embedding_w_action = tf.get_variable('embedding_w_action',
[self.dim_action, self.dim_emb])
embedding_b_action = tf.get_variable('embedding_b_action',
[self.dim_emb])
prev_a_vect = tf.matmul(input_prev_a,
embedding_w_action) + embedding_b_action
input_vect = tf.concat([depth_vect, cmd_vect, prev_a_vect], axis=1)
gru_cell = model_utils._gru_cell(self.n_hidden, 1, name='gru_cell')
# training
shape = input_vect.get_shape().as_list()
input_vect_reshape = tf.reshape(input_vect,
[-1, self.max_step, shape[-1]])
gru_output, _ = tf.nn.dynamic_rnn(gru_cell,
input_vect_reshape,
sequence_length=length,
dtype=tf.float32) # b, l, h
gru_output_reshape = tf.reshape(gru_output,
[-1, self.n_hidden]) # b*l, h
# action
a_linear = model_utils.dense_layer(
gru_output_reshape, 1, 'a_linear',
activation=tf.nn.sigmoid) * self.action_range[0]
a_angular = model_utils.dense_layer(
gru_output_reshape, 1, 'a_angular',
activation=tf.nn.tanh) * self.action_range[1]
action = tf.concat([a_linear, a_angular], axis=1)
# testing
gru_output, gru_h_out = gru_cell(input_vect, gru_h_in)
# action
a_linear = model_utils.dense_layer(
gru_output, 1, 'a_linear',
activation=tf.nn.sigmoid) * self.action_range[0]
a_angular = model_utils.dense_layer(
gru_output, 1, 'a_angular',
activation=tf.nn.tanh) * self.action_range[1]
action_test = tf.concat([a_linear, a_angular], axis=1)
return [action, action_test, gru_h_out]
def Model(self, inputs):
input_depth, input_cmd, input_prev_a = inputs
# encode depth image
conv1 = model_utils.conv2d(input_depth,
4,
5,
4,
scope='conv1',
max_pool=False)
conv2 = model_utils.conv2d(conv1,
16,
5,
4,
scope='conv2',
max_pool=False)
conv3 = model_utils.conv2d(conv2,
32,
3,
2,
scope='conv3',
max_pool=False)
shape = conv3.get_shape().as_list()
depth_vect = tf.reshape(conv3,
shape=[-1,
shape[1] * shape[2] * shape[3]]) # b,d
# encode cmd
embedding_cmd = tf.get_variable('cmd_embedding',
[self.n_cmd_type, self.dim_emb])
cmd_vect = tf.reshape(tf.nn.embedding_lookup(embedding_cmd, input_cmd),
[-1, self.dim_emb])
# encode prev action
embedding_w_action = tf.get_variable('embedding_w_action',
[self.dim_action, self.dim_emb])
embedding_b_action = tf.get_variable('embedding_b_action',
[self.dim_emb])
prev_a_vect = tf.matmul(input_prev_a,
embedding_w_action) + embedding_b_action
input_vect = tf.concat([depth_vect, cmd_vect, prev_a_vect], axis=1)
hidden_1 = model_utils.dense_layer(input_vect, self.n_hidden,
'hiddent_1')
hidden_2 = model_utils.dense_layer(input_vect, self.n_hidden / 2,
'hiddent_2')
a_linear = model_utils.dense_layer(
hidden_2,
1,
'a_linear',
activation=tf.nn.sigmoid,
w_init=tf.initializers.random_uniform(-0.003, 0.003),
b_init=tf.initializers.random_uniform(
-0.003, 0.003)) * self.action_range[0]
a_angular = model_utils.dense_layer(
hidden_2,
1,
'a_angular',
activation=tf.nn.tanh,
w_init=tf.initializers.random_uniform(-0.003, 0.003),
b_init=tf.initializers.random_uniform(
-0.003, 0.003)) * self.action_range[1]
pred_action = tf.concat([a_linear, a_angular], axis=1)
return pred_action