def process_empty_map(self, map_feature):
# In the returned empty map, 1 represents empty space
empty_map = np.atleast_3d(decode_image(map_feature))
ones, zeros = np.ones_like(empty_map), np.zeros_like(empty_map)
empty_map = np.where(np.greater(empty_map, 1), ones, zeros)
empty_map = np.transpose(empty_map, axes=[1, 0, 2])
return empty_map
def raw_images_to_array(images):
"""
Decode and normalize multiple images from tfrecord data
:param images: list of images encoded as a png in a string
:return: a numpy array of size (N, 56, 56, channels), normalized for training
"""
image_list = []
for image_str in images:
image = decode_image(image_str, (56, 56))
image = scale_observation(np.atleast_3d(image.astype(np.float32)))
image = remove_depth_noise(image)
image_list.append(image)
return np.stack(image_list, axis=0)
def process_roomtype_map(self, map_feature):
output = np.atleast_3d(decode_image(map_feature))
# transpose and invert
output = np.transpose(output, axes=[1, 0, 2])
return output
def process_roomid_map(self, roomidmap_feature):
# this is not transposed, unlike other maps
roomidmap = np.atleast_3d(decode_image(roomidmap_feature))
return roomidmap
def process_door_map(self, map_feature):
wall_map = np.atleast_3d(decode_image(map_feature))
wall_map = np.transpose(wall_map, axes=[1, 0, 2])
wall_map = wall_map.astype(np.float32) * (1.0 / 255.0)
return wall_map