def image_decode_job(images_def: oft.ListListNumpy.Placeholder(
shape=static_shape, dtype=flow.int8)):
images_buffer = flow.tensor_list_to_tensor_buffer(images_def)
decoded_images_buffer = flow.image_decode(images_buffer)
return flow.tensor_buffer_to_tensor_list(decoded_images_buffer,
shape=(640, 640, 3),
dtype=flow.uint8)
def image_normalize_job(images_def: oft.ListListNumpy.Placeholder(
shape=image_shape, dtype=flow.float)):
images_buffer = flow.tensor_list_to_tensor_buffer(images_def)
norm_images = flow.image_normalize(images_buffer, std, mean)
return flow.tensor_buffer_to_tensor_list(norm_images,
shape=image_shape[1:],
dtype=flow.float)
def image_flip_job(images_def: oft.ListListNumpy.Placeholder(
shape=image_shape, dtype=flow.float)):
images_buffer = flow.tensor_list_to_tensor_buffer(images_def)
flip_images = flow.image_flip(images_buffer, flip_code)
return flow.tensor_buffer_to_tensor_list(flip_images,
shape=image_shape[1:],
dtype=flow.float)
def image_resize_keep_aspect_ratio(
image_list: otp.ListListNumpy.Placeholder(shape=image_static_shape, dtype=dtype)
) -> tp.Tuple[otp.ListListNumpy, otp.ListNumpy, otp.ListNumpy]:
image_buffer = flow.tensor_list_to_tensor_buffer(image_list)
(res_image, scale, new_size) = flow.image.resize(
image_buffer,
target_size=target_size,
min_size=min_size,
max_size=max_size,
keep_aspect_ratio=True,
resize_side=resize_side,
interpolation_type=interpolation_type,
)
out_shape = image_test_util.infer_keep_aspect_ratio_resized_images_static_shape(
target_size=target_size,
min_size=min_size,
max_size=max_size,
aspect_ratio_list=aspect_ratio_list,
resize_side=resize_side,
channels=channels,
)
if print_debug_info:
print("resized image_static_shape: {}".format(out_shape))
res_image = flow.tensor_buffer_to_tensor_list(
res_image, shape=out_shape, dtype=dtype
)
return (res_image, scale, new_size)
def coco_load_fn():
with flow.scope.placement("cpu", "0:0-{}".format(nthread - 1)):
(
image,
image_id,
image_size,
gt_bbox,
gt_label,
gt_segm,
gt_segm_index,
) = flow.data.coco_reader(
annotation_file=anno_file,
image_dir=image_dir,
batch_size=batch_size,
shuffle=shuffle_after_epoch,
stride_partition=stride_partition,
name="COCOReader",
)
if ret_image_id_only:
return image_id
decoded_image = flow.image_decode(image, dtype=flow.float)
image_list = flow.tensor_buffer_to_tensor_list(decoded_image,
shape=(800, 1333,
3),
dtype=flow.float)
bbox_list = flow.tensor_buffer_to_tensor_list(gt_bbox,
shape=(128, 4),
dtype=flow.float)
label_list = flow.tensor_buffer_to_tensor_list(gt_label,
shape=(128, ),
dtype=flow.int32)
segm_list = flow.tensor_buffer_to_tensor_list(gt_segm,
shape=(1024, 2),
dtype=flow.float)
segm_index_list = flow.tensor_buffer_to_tensor_list(
gt_segm_index, shape=(1024, 3), dtype=flow.int32)
return (
image_id,
image_size,
image_list,
bbox_list,
label_list,
segm_list,
segm_index_list,
)
def object_bbox_flip_job(
bbox_def: oft.ListListNumpy.Placeholder(shape=tuple(bbox_shape),
dtype=flow.float),
image_size_def: oft.ListNumpy.Placeholder(shape=image_size.shape,
dtype=flow.int32),
):
bbox_buffer = flow.tensor_list_to_tensor_buffer(bbox_def)
flip_bbox = flow.object_bbox_flip(bbox_buffer, image_size_def,
flip_code)
return flow.tensor_buffer_to_tensor_list(flip_bbox,
shape=bbox_shape[1:],
dtype=flow.float)
def object_segm_poly_flip_job(
poly_def: oft.ListListNumpy.Placeholder(shape=tuple(poly_shape),
dtype=flow.float),
image_size_def: oft.ListNumpy.Placeholder(shape=image_size.shape,
dtype=flow.int32),
):
poly_buffer = flow.tensor_list_to_tensor_buffer(poly_def)
flip_poly = flow.object_segmentation_polygon_flip(
poly_buffer, image_size_def, flip_code)
return flow.tensor_buffer_to_tensor_list(flip_poly,
shape=poly_shape[1:],
dtype=flow.float)
def target_resize_bbox_scale_job(
image_def: oft.ListListNumpy.Placeholder(shape=tuple(image_shape),
dtype=flow.float),
bbox_def: oft.ListListNumpy.Placeholder(shape=tuple(bbox_shape),
dtype=flow.float),
):
images_buffer = flow.tensor_list_to_tensor_buffer(image_def)
(resized_images_buffer, new_size,
scale) = flow.image_target_resize(images_buffer,
target_size=target_size,
max_size=max_size)
bbox_buffer = flow.tensor_list_to_tensor_buffer(bbox_def)
scaled_bbox = flow.object_bbox_scale(bbox_buffer, scale)
scaled_bbox_list = flow.tensor_buffer_to_tensor_list(
scaled_bbox, shape=bbox_shape[1:], dtype=flow.float)
return (scaled_bbox_list, new_size)
def image_target_resize_job(image: otp.ListListNumpy.Placeholder(
shape=image_static_shape, dtype=flow.float)) -> tp.Tuple[
otp.ListListNumpy, otp.ListNumpy, otp.ListNumpy]:
image_buffer = flow.tensor_list_to_tensor_buffer(image)
(res_image_buffer, new_size, scale) = flow.image_target_resize(
image_buffer,
target_size=target_size,
max_size=max_size,
resize_side="shorter",
)
out_shape = image_test_util.infer_keep_aspect_ratio_resized_images_static_shape(
target_size=target_size,
min_size=None,
max_size=max_size,
aspect_ratio_list=aspect_ratio_list,
resize_side="shorter",
channels=3,
)
res_image = flow.tensor_buffer_to_tensor_list(res_image_buffer,
shape=out_shape,
dtype=flow.float)
return (res_image, new_size, scale)