def Planner(self, training_input, testing_input, label_status, length,
mask):
with tf.variable_scope('planner'):
batch_size = self.batch_size / self.gpu_num
rnn_cell = model_utils._lstm_cell(self.n_hidden, self.n_layers)
w_status = tf.get_variable(
'w_status', [self.n_hidden, 2],
initializer=tf.contrib.layers.xavier_initializer())
b_status = tf.get_variable(
'b_status', [2],
initializer=tf.contrib.layers.xavier_initializer())
# training
training_input_dropout = tf.nn.dropout(training_input,
self.keep_prob) # b*l, h
shape = training_input_dropout.get_shape().as_list()
training_input_reshape = tf.reshape(
training_input_dropout,
[batch_size, self.max_step, shape[1]]) # b, l, h
rnn_output, _ = tf.nn.dynamic_rnn(rnn_cell,
training_input_reshape,
sequence_length=length,
dtype=tf.float32) # b, l, h
rnn_output_dropout = tf.nn.dropout(rnn_output, self.keep_prob)
rnn_output_reshape = tf.reshape(rnn_output_dropout,
[-1, self.n_hidden]) # b*l, h
logits = tf.reshape(tf.matmul(rnn_output_reshape, w_status),
[-1, 2]) + b_status # b*l, n
label_status_reshape = tf.reshape(label_status, [-1])
loss_status = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_status_reshape, logits=logits)
loss_status_scalar = tf.reduce_sum(loss_status * mask)
# testing
prev_state = []
for l in xrange(self.n_layers):
prev_state.append(
LSTMStateTuple(
tf.placeholder(tf.float32,
shape=[None, self.n_hidden],
name='initial_state{0}.c'.format(l)),
tf.placeholder(tf.float32,
shape=[None, self.n_hidden],
name='initial_state{0}.h'.format(l))))
if self.n_layers == 1:
prev_state = prev_state[0]
rnn_output_test, state = rnn_cell(testing_input,
prev_state) # b*l, h
prob = tf.reshape(
tf.nn.softmax(tf.matmul(rnn_output_test, w_status) + b_status),
[-1, 2])
# pred_status_test = tf.argmax(prob, axis=1)
return loss_status_scalar, prob, state, prev_state
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 Model(self, inputs):
laser, cmd, cmd_next, cmd_skip, prev_action, obj_goal, prev_state_2 = inputs
with tf.variable_scope('encoder'):
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])
embedding_status = tf.get_variable(
'embedding_status', [self.n_cmd_type**2, self.dim_emb])
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])
embedding_w_status = tf.get_variable('embedding_w_status',
[self.dim_cmd, self.dim_emb])
embedding_b_status = tf.get_variable('embedding_b_status',
[self.dim_emb])
# training input
conv1 = model_utils.Conv1D(laser, 2, 5, 4, scope='conv1')
conv2 = model_utils.Conv1D(conv1, 4, 5, 4, scope='conv2')
conv3 = model_utils.Conv1D(conv2, 8, 5, 4, scope='conv3')
shape = conv3.get_shape().as_list()
vector_laser = tf.reshape(conv3, (-1, shape[1] * shape[2]))
curr_status = cmd * self.n_cmd_type + cmd_next
next_status = cmd_next * self.n_cmd_type + cmd_skip
vector_curr_status = tf.reshape(
tf.nn.embedding_lookup(embedding_status, curr_status),
(-1, self.dim_emb))
vector_prev_action = tf.matmul(
prev_action, embedding_w_action) + embedding_b_action
vector_obj_goal = tf.matmul(obj_goal,
embedding_w_goal) + embedding_b_goal
input_vector = tf.concat([
vector_laser, vector_curr_status, vector_prev_action,
vector_obj_goal
],
axis=1)
with tf.variable_scope('controller'):
shape = input_vector.get_shape().as_list()
w_hidden = tf.get_variable(
'w_hidden', [shape[1], self.n_hidden],
initializer=tf.contrib.layers.xavier_initializer())
b_hidden = tf.get_variable(
'b_hidden', [self.n_hidden],
initializer=tf.contrib.layers.xavier_initializer())
w_action_linear = tf.get_variable(
'w_action_linear', [self.n_hidden, self.dim_action / 2],
initializer=tf.contrib.layers.xavier_initializer())
b_action_linear = tf.get_variable(
'b_action_linear', [self.dim_action / 2],
initializer=tf.contrib.layers.xavier_initializer())
w_action_angular = tf.get_variable(
'w_action_angular', [self.n_hidden, self.dim_action / 2],
initializer=tf.contrib.layers.xavier_initializer())
b_action_angular = tf.get_variable(
'b_action_angular', [self.dim_action / 2],
initializer=tf.contrib.layers.xavier_initializer())
hidden = tf.nn.leaky_relu(
tf.matmul(input_vector, w_hidden) + b_hidden)
a_linear = tf.nn.sigmoid(
tf.matmul(hidden, w_action_linear) +
b_action_linear) * self.action_range[0]
a_angular = tf.nn.tanh(
tf.matmul(hidden, w_action_angular) +
b_action_angular) * self.action_range[1]
pred_action = tf.concat([a_linear, a_angular], axis=1)
with tf.variable_scope('planner'):
rnn_cell_2 = model_utils._lstm_cell(self.n_hidden,
self.n_layers,
name='rnn/basic_lstm_cell')
w_status_matrix = tf.get_variable(
'w_status_matrix', [self.n_cmd_type**2, self.n_hidden],
initializer=tf.contrib.layers.xavier_initializer())
b_status_matrix = tf.get_variable(
'b_status_matrix', [self.n_cmd_type**2],
initializer=tf.contrib.layers.xavier_initializer())
status_curr = tf.reshape(cmd * self.n_cmd_type + cmd_next,
[-1]) # b*l, 1 -> (1)
status_next = tf.reshape(cmd_next * self.n_cmd_type + cmd_skip,
[-1])
w_status_curr = tf.reshape(tf.gather(w_status_matrix, status_curr),
[-1, self.n_hidden, 1]) # b, h, 1
w_status_next = tf.reshape(tf.gather(w_status_matrix, status_next),
[-1, self.n_hidden, 1])
b_status_curr = tf.reshape(tf.gather(b_status_matrix, status_curr),
[-1, 1]) # b, 1
b_status_next = tf.reshape(tf.gather(b_status_matrix, status_next),
[-1, 1])
w_status = tf.concat([w_status_curr, w_status_next],
axis=2) # b, h, 2
b_status = tf.concat([b_status_curr, b_status_next],
axis=1) # b, 2
rnn_output_2, state_2 = rnn_cell_2(input_vector, prev_state_2)
rnn_output_expand = tf.expand_dims(rnn_output_2, 1) # b, h, 1
logits = tf.reshape(tf.matmul(rnn_output_expand, w_status),
[-1, 2]) + b_status
return pred_action, logits, state_2
def ControllerLSTM(self, training_input, testing_input, label_action,
length, mask):
with tf.variable_scope('controller'):
batch_size = self.batch_size / self.gpu_num
rnn_cell = model_utils._lstm_cell(self.n_hidden, self.n_layers)
w_action_linear = tf.get_variable(
'w_action_linear', [self.n_hidden, self.dim_action / 2],
initializer=tf.contrib.layers.xavier_initializer())
b_action_linear = tf.get_variable(
'b_action_linear', [self.dim_action / 2],
initializer=tf.contrib.layers.xavier_initializer())
w_action_angular = tf.get_variable(
'w_action_angular', [self.n_hidden, self.dim_action / 2],
initializer=tf.contrib.layers.xavier_initializer())
b_action_angular = tf.get_variable(
'b_action_angular', [self.dim_action / 2],
initializer=tf.contrib.layers.xavier_initializer())
training_input_dropout = tf.nn.dropout(training_input,
self.keep_prob) # b*l, h
shape = training_input_dropout.get_shape().as_list()
training_input_reshape = tf.reshape(
training_input_dropout,
[batch_size, self.max_step, shape[1]]) # b, l, h
rnn_output, _ = tf.nn.dynamic_rnn(rnn_cell,
training_input_reshape,
sequence_length=length,
dtype=tf.float32) # b, l, h
rnn_output_dropout = tf.nn.dropout(rnn_output, self.keep_prob)
rnn_output_reshape = tf.reshape(rnn_output_dropout,
[-1, self.n_hidden]) # b*l, h
a_linear = tf.nn.sigmoid(
tf.matmul(rnn_output_reshape, w_action_linear) +
b_action_linear) * self.action_range[0]
a_angular = tf.nn.tanh(
tf.matmul(rnn_output_reshape, w_action_angular) +
b_action_angular) * self.action_range[1]
pred_action = tf.concat([a_linear, a_angular], axis=1)
# calculate the mean and variance of the masked error
mask_reshape = tf.reshape(
mask, [batch_size * self.max_step, 1]) # b*l, 1
mask_tile = tf.tile(mask_reshape, [1, 2]) # b*l, 2
masked_error = (pred_action - label_action) * mask_tile # b*l, 2
mean = tf.reduce_sum(masked_error, axis=0) / tf.cast(
tf.reduce_sum(length), tf.float32) # 2
mean_expand = tf.expand_dims(mean, axis=0) # 1, 2
mean_tile = tf.tile(mean_expand,
[batch_size * self.max_step, 1]) # b*l, 2
variance = tf.square(
tf.reduce_sum(
(masked_error - mean_tile) * mask_tile, axis=0)) / tf.cast(
tf.reduce_sum(length), tf.float32)
loss_action = tf.losses.mean_squared_error(
labels=label_action,
predictions=pred_action,
reduction=tf.losses.Reduction.NONE)
loss_action = tf.reduce_sum(loss_action, axis=1)
loss_action_scalar = tf.reduce_sum(loss_action * mask)
# testing
prev_state = []
for l in xrange(self.n_layers):
prev_state.append(
LSTMStateTuple(
tf.placeholder(tf.float32,
shape=[None, self.n_hidden],
name='initial_state{0}.c'.format(l)),
tf.placeholder(tf.float32,
shape=[None, self.n_hidden],
name='initial_state{0}.h'.format(l))))
if self.n_layers == 1:
prev_state = prev_state[0]
rnn_output_test, state = rnn_cell(testing_input,
prev_state) # b*l, h
a_linear_test = tf.nn.sigmoid(
tf.matmul(rnn_output_test, w_action_linear) +
b_action_linear) * self.action_range[0]
a_angular_test = tf.nn.tanh(
tf.matmul(rnn_output_test, w_action_angular) +
b_action_angular) * self.action_range[1]
pred_action_test = tf.concat([a_linear_test, a_angular_test],
axis=1)
return loss_action_scalar, pred_action_test, state, prev_state, mean, variance
def Model(self, inputs):
laser, cmd, cmd_next, prev_action, obj_goal, action = inputs
with tf.variable_scope('encoder'):
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])
embedding_status = tf.get_variable(
'embedding_status', [self.n_cmd_type**2, self.dim_emb])
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])
embedding_w_status = tf.get_variable('embedding_w_status',
[self.dim_cmd, self.dim_emb])
embedding_b_status = tf.get_variable('embedding_b_status',
[self.dim_emb])
conv1 = model_utils.Conv1D(self.input_laser,
2,
5,
4,
scope='conv1')
conv2 = model_utils.Conv1D(conv1, 4, 5, 4, scope='conv2')
conv3 = model_utils.Conv1D(conv2, 8, 5, 4, scope='conv3')
shape = conv3.get_shape().as_list()
vector_laser = tf.reshape(conv3, (-1, shape[1] * shape[2]))
curr_status = cmd * self.n_cmd_type + cmd_next
vector_curr_status = tf.reshape(
tf.nn.embedding_lookup(embedding_status, curr_status),
(-1, self.dim_emb))
vector_prev_action = tf.matmul(
prev_action, embedding_w_action) + embedding_b_action
vector_obj_goal = tf.matmul(obj_goal,
embedding_w_goal) + embedding_b_goal
vector_action = tf.matmul(action,
embedding_w_action) + embedding_b_action
input_vector = tf.concat([
vector_laser, vector_curr_status, vector_prev_action,
vector_obj_goal, vector_action
],
axis=1)
with tf.variable_scope('q'):
rnn_cell = model_utils._lstm_cell(self.n_hidden,
self.n_layers,
name='rnn/basic_lstm_cell')
w_q = tf.get_variable('w_q', [self.n_hidden, 1],
initializer=tf.initializers.random_uniform(
-0.003, 0.003))
b_q = tf.get_variable('b_q', [1],
initializer=tf.initializers.random_uniform(
-0.003, 0.003))
shape = input_vector.get_shape().as_list()
input_vector_reshape = tf.reshape(
input_vector, [self.batch_size, self.max_step, shape[1]])
rnn_output, _ = tf.nn.dynamic_rnn(rnn_cell,
input_vector_reshape,
sequence_length=self.length,
dtype=tf.float32) # b, l, h
rnn_output_reshape = tf.reshape(rnn_output,
[-1, self.n_hidden]) # b*l, h
q = tf.matmul(rnn_output_reshape, w_q) + b_q
return q
def Planner(self, training_input, testing_input, input_cmd, input_cmd_next,
input_cmd_skip, label_status, length, mask):
with tf.variable_scope('planner'):
batch_size = self.batch_size / self.gpu_num
rnn_cell = model_utils._lstm_cell(self.n_hidden, self.n_layers)
w_status_matrix = tf.get_variable(
'w_status_matrix', [self.n_cmd_type**2, self.n_hidden],
initializer=tf.contrib.layers.xavier_initializer())
b_status_matrix = tf.get_variable(
'b_status_matrix', [self.n_cmd_type**2],
initializer=tf.contrib.layers.xavier_initializer())
status_curr = tf.reshape(input_cmd * self.n_cmd_type +
input_cmd_next, [-1]) # b*l, 1 -> (1)
status_next = tf.reshape(
input_cmd_next * self.n_cmd_type + input_cmd_skip, [-1])
w_status_curr = tf.reshape(tf.gather(w_status_matrix, status_curr),
[-1, self.n_hidden, 1]) # b*l, h, 1
w_status_next = tf.reshape(tf.gather(w_status_matrix, status_next),
[-1, self.n_hidden, 1])
b_status_curr = tf.reshape(tf.gather(b_status_matrix, status_curr),
[-1, 1]) # b*l, 1
b_status_next = tf.reshape(tf.gather(b_status_matrix, status_next),
[-1, 1])
w_status = tf.concat([w_status_curr, w_status_next],
axis=2) # b*l, h, 2
b_status = tf.concat([b_status_curr, b_status_next],
axis=1) # b*l, 2
# training
training_input_dropout = tf.nn.dropout(training_input,
self.keep_prob) # b*l, h
shape = training_input_dropout.get_shape().as_list()
training_input_reshape = tf.reshape(
training_input_dropout,
[batch_size, self.max_step, shape[1]]) # b, l, h
rnn_output, _ = tf.nn.dynamic_rnn(rnn_cell,
training_input_reshape,
sequence_length=length,
dtype=tf.float32) # b, l, h
rnn_output_dropout = tf.nn.dropout(rnn_output, self.keep_prob)
rnn_output_reshape = tf.reshape(rnn_output_dropout,
[-1, self.n_hidden]) # b*l, h
rnn_output_expand = tf.expand_dims(rnn_output_reshape,
1) # b*l, 1, h
# 1. dot product distance
logits = tf.reshape(tf.matmul(rnn_output_expand, w_status),
[-1, 2]) + b_status # b*l, 2
self.training_logits = logits
self.training_pred = tf.argmax(logits, axis=1)
# # 2. eucl distance
# w_status_curr_reshape = tf.reshape(w_status_curr, [-1, self.n_hidden]) # b*l, h
# w_status_next_reshape = tf.reshape(w_status_next, [-1, self.n_hidden]) # b*l, h
# squared_dist_curr = tf.reduce_sum(tf.square(rnn_output_reshape - w_status_curr_reshape), axis=1, keepdims=True) # b*l
# squared_dist_next = tf.reduce_sum(tf.square(rnn_output_reshape - w_status_next_reshape), axis=1, keepdims=True) # b*l
# logits = tf.concat([squared_dist_curr, squared_dist_next], axis=1)
# # 3. binary prediction
# w_binary = tf.get_variable('w_status', [self.n_hidden, 2], initializer=tf.contrib.layers.xavier_initializer())
# b_binary = tf.get_variable('b_status', [2], initializer=tf.contrib.layers.xavier_initializer())
# logits = tf.matmul(rnn_output_reshape, w_binary) + b_binary
# # 4. n^2 precition
# w_square = tf.get_variable('w_square', [self.n_hidden, self.n_cmd_type**2 * 2], initializer=tf.contrib.layers.xavier_initializer())
# b_square = tf.get_variable('b_square', [self.n_cmd_type**2 * 2], initializer=tf.contrib.layers.xavier_initializer())
# logits = tf.matmul(rnn_output_reshape, w_square) + b_square
# label_status_reshape = tf.reshape(label_status, [-1])
# label_status = (1 - label_status_reshape) * status_curr + label_status_reshape * status_next
label_status_reshape = tf.reshape(label_status, [-1])
loss_status = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_status_reshape, logits=logits)
loss_status_scalar = tf.reduce_sum(loss_status * mask)
# testing
prev_state = []
for l in xrange(self.n_layers):
prev_state.append(
LSTMStateTuple(
tf.placeholder(tf.float32,
shape=[None, self.n_hidden],
name='initial_state{0}.c'.format(l)),
tf.placeholder(tf.float32,
shape=[None, self.n_hidden],
name='initial_state{0}.h'.format(l))))
if self.n_layers == 1:
prev_state = prev_state[0]
rnn_output_test, state = rnn_cell(testing_input,
prev_state) # b, h
rnn_output_test_expand = tf.expand_dims(rnn_output_test,
1) # b, 1, h
w_status_curr = tf.reshape(
tf.gather(w_status_matrix, self.test_status),
[-1, self.n_hidden, 1]) # b, h, 1
w_status_next = tf.reshape(
tf.gather(w_status_matrix, self.test_status_next),
[-1, self.n_hidden, 1])
b_status_curr = tf.reshape(
tf.gather(b_status_matrix, self.test_status), [-1, 1]) # b, 1
b_status_next = tf.reshape(
tf.gather(b_status_matrix, self.test_status_next), [-1, 1])
w_status = tf.concat([w_status_curr, w_status_next],
axis=2) # b, h, 2
b_status = tf.concat([b_status_curr, b_status_next], axis=1)
logits = tf.reshape(tf.matmul(rnn_output_test_expand, w_status),
[-1, 2]) + b_status
pred_done = tf.argmax(logits, axis=1)
return loss_status_scalar, pred_done, logits, state, prev_state