def _inputs(problem, lstm_states, lstm_state_dims):
# Set up inputs.
with tf.name_scope('inputs'):
n_views = problem.n_views
inputs = []
inputs.append(('orig_maps', tf.float32,
(problem.batch_size, 1, None, None, 1)))
inputs.append(('goal_loc', tf.float32,
(problem.batch_size, problem.num_goals, 2)))
# For initing LSTM.
inputs.append(('rel_goal_loc_at_start', tf.float32,
(problem.batch_size, problem.num_goals,
problem.rel_goal_loc_dim)))
common_input_data, _ = tf_utils.setup_inputs(inputs)
inputs = []
inputs.append(('imgs', tf.float32, (problem.batch_size, None, n_views,
problem.img_height, problem.img_width,
problem.img_channels)))
# Goal location as a tuple of delta location and delta theta.
inputs.append(('rel_goal_loc', tf.float32, (problem.batch_size, None,
problem.rel_goal_loc_dim)))
if problem.outputs.visit_count:
inputs.append(('visit_count', tf.int32, (problem.batch_size, None, 1)))
inputs.append(('last_visit', tf.int32, (problem.batch_size, None, 1)))
for i, (state, dim) in enumerate(zip(lstm_states, lstm_state_dims)):
inputs.append((state, tf.float32, (problem.batch_size, 1, dim)))
if problem.outputs.egomotion:
inputs.append(('incremental_locs', tf.float32,
(problem.batch_size, None, 2)))
inputs.append(('incremental_thetas', tf.float32,
(problem.batch_size, None, 1)))
inputs.append(('step_number', tf.int32, (1, None, 1)))
inputs.append(('node_ids', tf.int32, (problem.batch_size, None,
problem.node_ids_dim)))
inputs.append(('perturbs', tf.float32, (problem.batch_size, None,
problem.perturbs_dim)))
# For plotting result plots
inputs.append(('loc_on_map', tf.float32, (problem.batch_size, None, 2)))
inputs.append(('gt_dist_to_goal', tf.float32, (problem.batch_size, None, 1)))
step_input_data, _ = tf_utils.setup_inputs(inputs)
inputs = []
inputs.append(('executed_actions', tf.int32, (problem.batch_size, None)))
inputs.append(('rewards', tf.float32, (problem.batch_size, None)))
inputs.append(('action_sample_wts', tf.float32, (problem.batch_size, None)))
inputs.append(('action', tf.int32, (problem.batch_size, None,
problem.num_actions)))
train_data, _ = tf_utils.setup_inputs(inputs)
train_data.update(step_input_data)
train_data.update(common_input_data)
return common_input_data, step_input_data, train_data
def _inputs(problem, lstm_states, lstm_state_dims):
# Set up inputs.
with tf.name_scope('inputs'):
n_views = problem.n_views
inputs = []
inputs.append(('orig_maps', tf.float32,
(problem.batch_size, 1, None, None, 1)))
inputs.append(('goal_loc', tf.float32,
(problem.batch_size, problem.num_goals, 2)))
# For initing LSTM.
inputs.append(('rel_goal_loc_at_start', tf.float32,
(problem.batch_size, problem.num_goals,
problem.rel_goal_loc_dim)))
common_input_data, _ = tf_utils.setup_inputs(inputs)
inputs = []
inputs.append(('imgs', tf.float32, (problem.batch_size, None, n_views,
problem.img_height, problem.img_width,
problem.img_channels)))
# Goal location as a tuple of delta location and delta theta.
inputs.append(('rel_goal_loc', tf.float32, (problem.batch_size, None,
problem.rel_goal_loc_dim)))
if problem.outputs.visit_count:
inputs.append(('visit_count', tf.int32, (problem.batch_size, None, 1)))
inputs.append(('last_visit', tf.int32, (problem.batch_size, None, 1)))
for i, (state, dim) in enumerate(zip(lstm_states, lstm_state_dims)):
inputs.append((state, tf.float32, (problem.batch_size, 1, dim)))
if problem.outputs.egomotion:
inputs.append(('incremental_locs', tf.float32,
(problem.batch_size, None, 2)))
inputs.append(('incremental_thetas', tf.float32,
(problem.batch_size, None, 1)))
inputs.append(('step_number', tf.int32, (1, None, 1)))
inputs.append(('node_ids', tf.int32, (problem.batch_size, None,
problem.node_ids_dim)))
inputs.append(('perturbs', tf.float32, (problem.batch_size, None,
problem.perturbs_dim)))
# For plotting result plots
inputs.append(('loc_on_map', tf.float32, (problem.batch_size, None, 2)))
inputs.append(('gt_dist_to_goal', tf.float32, (problem.batch_size, None, 1)))
step_input_data, _ = tf_utils.setup_inputs(inputs)
inputs = []
inputs.append(('executed_actions', tf.int32, (problem.batch_size, None)))
inputs.append(('rewards', tf.float32, (problem.batch_size, None)))
inputs.append(('action_sample_wts', tf.float32, (problem.batch_size, None)))
inputs.append(('action', tf.int32, (problem.batch_size, None,
problem.num_actions)))
train_data, _ = tf_utils.setup_inputs(inputs)
train_data.update(step_input_data)
train_data.update(common_input_data)
return common_input_data, step_input_data, train_data
def _inputs(problem):
# Set up inputs.
with tf.name_scope('inputs'):
inputs = []
inputs.append(('orig_maps', tf.float32,
(problem.batch_size, 1, None, None, 1)))
inputs.append(('goal_loc', tf.float32,
(problem.batch_size, problem.num_goals, 2)))
common_input_data, _ = tf_utils.setup_inputs(inputs)
inputs = []
if problem.input_type == 'vision':
# Multiple images from an array of cameras.
inputs.append(('imgs', tf.float32,
(problem.batch_size, None, len(problem.aux_delta_thetas)+1,
problem.img_height, problem.img_width,
problem.img_channels)))
elif problem.input_type == 'analytical_counts':
for i in range(len(problem.map_crop_sizes)):
inputs.append(('analytical_counts_{:d}'.format(i), tf.float32,
(problem.batch_size, None, problem.map_crop_sizes[i],
problem.map_crop_sizes[i], problem.map_channels)))
if problem.outputs.readout_maps:
for i in range(len(problem.readout_maps_crop_sizes)):
inputs.append(('readout_maps_{:d}'.format(i), tf.float32,
(problem.batch_size, None,
problem.readout_maps_crop_sizes[i],
problem.readout_maps_crop_sizes[i],
problem.readout_maps_channels)))
for i in range(len(problem.map_crop_sizes)):
inputs.append(('ego_goal_imgs_{:d}'.format(i), tf.float32,
(problem.batch_size, None, problem.map_crop_sizes[i],
problem.map_crop_sizes[i], problem.goal_channels)))
for s in ['sum_num', 'sum_denom', 'max_denom']:
inputs.append(('running_'+s+'_{:d}'.format(i), tf.float32,
(problem.batch_size, 1, problem.map_crop_sizes[i],
problem.map_crop_sizes[i], problem.map_channels)))
inputs.append(('incremental_locs', tf.float32,
(problem.batch_size, None, 2)))
inputs.append(('incremental_thetas', tf.float32,
(problem.batch_size, None, 1)))
inputs.append(('step_number', tf.int32, (1, None, 1)))
inputs.append(('node_ids', tf.int32, (problem.batch_size, None,
problem.node_ids_dim)))
inputs.append(('perturbs', tf.float32, (problem.batch_size, None,
problem.perturbs_dim)))
# For plotting result plots
inputs.append(('loc_on_map', tf.float32, (problem.batch_size, None, 2)))
inputs.append(('gt_dist_to_goal', tf.float32, (problem.batch_size, None, 1)))
step_input_data, _ = tf_utils.setup_inputs(inputs)
inputs = []
inputs.append(('action', tf.int32, (problem.batch_size, None, problem.num_actions)))
train_data, _ = tf_utils.setup_inputs(inputs)
train_data.update(step_input_data)
train_data.update(common_input_data)
return common_input_data, step_input_data, train_data
def _inputs(problem):
# Set up inputs.
with tf.name_scope('inputs'):
inputs = []
inputs.append(('orig_maps', tf.float32,
(problem.batch_size, 1, None, None, 1)))
inputs.append(('goal_loc', tf.float32,
(problem.batch_size, problem.num_goals, 2)))
common_input_data, _ = tf_utils.setup_inputs(inputs)
inputs = []
if problem.input_type == 'vision':
# Multiple images from an array of cameras.
inputs.append(('imgs', tf.float32,
(problem.batch_size, None, len(problem.aux_delta_thetas)+1,
problem.img_height, problem.img_width,
problem.img_channels)))
elif problem.input_type == 'analytical_counts':
for i in range(len(problem.map_crop_sizes)):
inputs.append(('analytical_counts_{:d}'.format(i), tf.float32,
(problem.batch_size, None, problem.map_crop_sizes[i],
problem.map_crop_sizes[i], problem.map_channels)))
if problem.outputs.readout_maps:
for i in range(len(problem.readout_maps_crop_sizes)):
inputs.append(('readout_maps_{:d}'.format(i), tf.float32,
(problem.batch_size, None,
problem.readout_maps_crop_sizes[i],
problem.readout_maps_crop_sizes[i],
problem.readout_maps_channels)))
for i in range(len(problem.map_crop_sizes)):
inputs.append(('ego_goal_imgs_{:d}'.format(i), tf.float32,
(problem.batch_size, None, problem.map_crop_sizes[i],
problem.map_crop_sizes[i], problem.goal_channels)))
for s in ['sum_num', 'sum_denom', 'max_denom']:
inputs.append(('running_'+s+'_{:d}'.format(i), tf.float32,
(problem.batch_size, 1, problem.map_crop_sizes[i],
problem.map_crop_sizes[i], problem.map_channels)))
inputs.append(('incremental_locs', tf.float32,
(problem.batch_size, None, 2)))
inputs.append(('incremental_thetas', tf.float32,
(problem.batch_size, None, 1)))
inputs.append(('step_number', tf.int32, (1, None, 1)))
inputs.append(('node_ids', tf.int32, (problem.batch_size, None,
problem.node_ids_dim)))
inputs.append(('perturbs', tf.float32, (problem.batch_size, None,
problem.perturbs_dim)))
# For plotting result plots
inputs.append(('loc_on_map', tf.float32, (problem.batch_size, None, 2)))
inputs.append(('gt_dist_to_goal', tf.float32, (problem.batch_size, None, 1)))
step_input_data, _ = tf_utils.setup_inputs(inputs)
inputs = []
inputs.append(('action', tf.int32, (problem.batch_size, None, problem.num_actions)))
train_data, _ = tf_utils.setup_inputs(inputs)
train_data.update(step_input_data)
train_data.update(common_input_data)
return common_input_data, step_input_data, train_data