def create_coco_pipeline(file_root,
annotations_file,
batch_size=1,
device_id=0,
num_threads=4,
local_rank=0,
world_size=1):
pipeline = Pipeline(batch_size, num_threads,
local_rank, seed=42 + device_id)
with pipeline:
images, bboxes, labels = fn.coco_reader(file_root=file_root,
annotations_file=annotations_file,
skip_empty=True,
shard_id=local_rank,
num_shards=world_size,
ratio=True,
ltrb=True,
random_shuffle=False,
shuffle_after_epoch=True,
name="Reader")
crop_begin, crop_size, bboxes, labels = fn.random_bbox_crop(bboxes, labels,
device="cpu",
aspect_ratio=[0.5, 2.0],
thresholds=[0, 0.1, 0.3, 0.5, 0.7, 0.9],
scaling=[0.3, 1.0],
bbox_layout="xyXY",
allow_no_crop=True,
num_attempts=50)
images = fn.image_decoder_slice(images, crop_begin, crop_size, device="mixed", output_type=types.RGB)
flip_coin = fn.coin_flip(probability=0.5)
images = fn.resize(images,
resize_x=300,
resize_y=300,
min_filter=types.DALIInterpType.INTERP_TRIANGULAR)
# use float to avoid clipping and quantizing the intermediate result
images = fn.hsv(images, dtype=types.FLOAT, hue=fn.uniform(range=[-0.5, 0.5]),
saturation=fn.uniform(range=[0.5, 1.5]))
images = fn.brightness_contrast(images,
contrast_center = 128, # input is in float, but in 0..255 range
dtype = types.UINT8,
brightness = fn.uniform(range=[0.875, 1.125]),
contrast = fn.uniform(range=[0.5, 1.5]))
bboxes = fn.bb_flip(bboxes, ltrb=True, horizontal=flip_coin)
images = fn.crop_mirror_normalize(images,
mean=[104., 117., 123.],
std=[1., 1., 1.],
mirror=flip_coin,
dtype=types.FLOAT,
output_layout="CHW",
pad_output=False)
pipeline.set_outputs(images, bboxes, labels)
return pipeline
def image_decoder_slice_pipe(max_batch_size, input_data, device):
pipe = Pipeline(batch_size=max_batch_size, num_threads=4, device_id=0)
encoded = fn.external_source(source=input_data,
cycle=False,
device='cpu')
anch = fn.constant(fdata=.1)
sh = fn.constant(fdata=.4)
decoded = fn.image_decoder_slice(encoded,
anch,
sh,
axes=0,
device=device)
pipe.set_outputs(decoded)
return pipe
def test_image_decoder_slice_cpu():
anch_shape = [1, 2]
def get_anchors():
out = [(np.random.randint(1, 128, size=anch_shape, dtype=np.uint8) /
255).astype(dtype=np.float32) for _ in range(batch_size)]
return out
def get_shape():
out = [(np.random.randint(1, 128, size=anch_shape, dtype=np.uint8) /
255).astype(dtype=np.float32) for _ in range(batch_size)]
return out
pipe = Pipeline(batch_size=batch_size, num_threads=4, device_id=None)
input, _ = fn.file_reader(file_root=images_dir, shard_id=0, num_shards=1)
anchors = fn.external_source(source=get_anchors)
shape = fn.external_source(source=get_shape)
processed = fn.image_decoder_slice(input, anchors, shape)
pipe.set_outputs(processed)
pipe.build()
for _ in range(3):
pipe.run()
def create_coco_pipeline(default_boxes, args):
try:
shard_id = torch.distributed.get_rank()
num_shards = torch.distributed.get_world_size()
except RuntimeError:
shard_id = 0
num_shards = 1
images, bboxes, labels = fn.readers.coco(
file_root=args.train_coco_root,
annotations_file=args.train_annotate,
skip_empty=True,
shard_id=shard_id,
num_shards=num_shards,
ratio=True,
ltrb=True,
random_shuffle=False,
shuffle_after_epoch=True,
name="Reader")
crop_begin, crop_size, bboxes, labels = fn.random_bbox_crop(
bboxes,
labels,
device="cpu",
aspect_ratio=[0.5, 2.0],
thresholds=[0, 0.1, 0.3, 0.5, 0.7, 0.9],
scaling=[0.3, 1.0],
bbox_layout="xyXY",
allow_no_crop=True,
num_attempts=50)
images = fn.image_decoder_slice(images,
crop_begin,
crop_size,
device="mixed",
output_type=types.RGB)
flip_coin = fn.random.coin_flip(probability=0.5)
images = fn.resize(images,
resize_x=300,
resize_y=300,
min_filter=types.DALIInterpType.INTERP_TRIANGULAR)
saturation = fn.uniform(range=[0.5, 1.5])
contrast = fn.uniform(range=[0.5, 1.5])
brightness = fn.uniform(range=[0.875, 1.125])
hue = fn.uniform(range=[-0.5, 0.5])
images = fn.hsv(images, dtype=types.FLOAT, hue=hue,
saturation=saturation) # use float to avoid clipping and
# quantizing the intermediate result
images = fn.brightness_contrast(
images,
contrast_center=128, # input is in float, but in 0..255 range
dtype=types.UINT8,
brightness=brightness,
contrast=contrast)
dtype = types.FLOAT16 if args.fp16 else types.FLOAT
bboxes = fn.bb_flip(bboxes, ltrb=True, horizontal=flip_coin)
images = fn.crop_mirror_normalize(
images,
crop=(300, 300),
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
std=[0.229 * 255, 0.224 * 255, 0.225 * 255],
mirror=flip_coin,
dtype=dtype,
output_layout="CHW",
pad_output=False)
bboxes, labels = fn.box_encoder(bboxes,
labels,
criteria=0.5,
anchors=default_boxes.as_ltrb_list())
labels = labels.gpu()
bboxes = bboxes.gpu()
return images, bboxes, labels
def check_bbox_random_crop_adjust_polygons(file_root,
annotations_file,
batch_size=3,
num_iters=4,
num_threads=4,
device_id=0,
seed=1234):
pipe = Pipeline(batch_size=batch_size,
num_threads=num_threads,
device_id=device_id,
seed=seed)
with pipe:
# Read data from COCO
# ratio=True means both bboxes and masks coordinates will be relative to the image dimensions (range [0.0, 1.0])
inputs, in_bboxes, labels, in_polygons, in_vertices = \
fn.readers.coco(
file_root=file_root, annotations_file=annotations_file, shard_id=0, num_shards=1,
ratio=True, ltrb=True, polygon_masks=True
)
# Generate a random crop. out_bboxes are adjusted to the crop window
slice_anchor, slice_shape, out_bboxes, labels, bbox_indices = \
fn.random_bbox_crop(
in_bboxes, labels,
aspect_ratio=[0.5, 2.0], thresholds=[0, 0.1, 0.3, 0.5, 0.7, 0.9],
scaling=[0.3, 1.0], bbox_layout='xyXY', output_bbox_indices=True
)
# Crop the image
images = fn.image_decoder_slice(inputs,
slice_anchor,
slice_shape,
device='mixed',
axis_names='WH')
sel_polygons, sel_vertices = fn.segmentation.select_masks(
bbox_indices, in_polygons, in_vertices)
# Adjust masks coordinates to the coordinate space of the cropped image
MT = fn.transforms.crop(from_start=slice_anchor,
from_end=(slice_anchor + slice_shape))
out_vertices = fn.coord_transform(sel_vertices, MT=MT)
# Converting to absolute coordinates (demo purposes)
image_shape = fn.peek_image_shape(inputs, type=types.FLOAT)
h = fn.slice(image_shape, 0, 1, axes=[0])
w = fn.slice(image_shape, 1, 1, axes=[0])
# Original bboxes
bbox_x = fn.slice(in_bboxes, 0, 1, axes=[1])
bbox_y = fn.slice(in_bboxes, 1, 1, axes=[1])
bbox_X = fn.slice(in_bboxes, 2, 1, axes=[1])
bbox_Y = fn.slice(in_bboxes, 3, 1, axes=[1])
in_bboxes_abs = fn.cat(bbox_x * w,
bbox_y * h,
bbox_X * w,
bbox_Y * h,
axis=1)
# Transform to convert relative coordinates to absolute
scale_rel_to_abs = fn.transforms.scale(scale=fn.cat(w, h))
# Selected vertices (relative coordinates)
sel_vertices_abs = fn.coord_transform(out_vertices,
MT=scale_rel_to_abs)
# Output bboxes
bbox2_x = fn.slice(out_bboxes, 0, 1, axes=[1])
bbox2_y = fn.slice(out_bboxes, 1, 1, axes=[1])
bbox2_X = fn.slice(out_bboxes, 2, 1, axes=[1])
bbox2_Y = fn.slice(out_bboxes, 3, 1, axes=[1])
out_bboxes_abs = fn.cat(bbox2_x * w,
bbox2_y * h,
bbox2_X * w,
bbox2_Y * h,
axis=1)
# Output vertices (absolute coordinates)
out_vertices_abs = fn.coord_transform(out_vertices,
MT=scale_rel_to_abs)
# Clamped coordinates
out_vertices_clamped = math.clamp(out_vertices, 0.0, 1.0)
out_vertices_clamped_abs = fn.coord_transform(out_vertices_clamped,
MT=scale_rel_to_abs)
pipe.set_outputs(in_vertices, sel_vertices, sel_vertices_abs, out_vertices,
out_vertices_clamped, out_vertices_abs,
out_vertices_clamped_abs, in_bboxes, in_bboxes_abs,
out_bboxes, out_bboxes_abs, in_polygons, sel_polygons,
image_shape, slice_anchor, slice_shape, bbox_indices)
pipe.build()
for i in range(
num_iters
): # Enough iterations to see an example with more than one bounding box
outs = pipe.run()
for j in range(batch_size):
in_vertices, sel_vertices, sel_vertices_abs, \
out_vertices, out_vertices_clamped, out_vertices_abs, out_vertices_clamped_abs, \
in_bboxes, in_bboxes_abs, out_bboxes, out_bboxes_abs, \
in_polygons, sel_polygons, \
image_shape, slice_anchor, slice_shape, bbox_indices = \
(outs[k].at(j) for k in range(len(outs)))
# Checking that the output polygon descriptors are the ones associated with the
# selected bounding boxes
expected_polygons_list = []
expected_vertices_list = []
ver_count = 0
for k in range(in_polygons.shape[0]):
mask_id = in_polygons[k][0]
in_ver_start_idx = in_polygons[k][1]
in_ver_end_idx = in_polygons[k][2]
pol_nver = in_ver_end_idx - in_ver_start_idx
if mask_id in bbox_indices:
expected_polygons_list.append(
[mask_id, ver_count, ver_count + pol_nver])
for j in range(in_ver_start_idx, in_ver_end_idx):
expected_vertices_list.append(in_vertices[j])
ver_count = ver_count + pol_nver
expected_sel_polygons = np.array(expected_polygons_list)
np.testing.assert_equal(expected_sel_polygons, sel_polygons)
# Checking the selected vertices correspond to the selected masks
expected_sel_vertices = np.array(expected_vertices_list)
np.testing.assert_equal(expected_sel_vertices, sel_vertices)
# Chekc that the vertices are correctly mapped to the cropping window
expected_out_vertices = np.copy(expected_sel_vertices)
crop_x, crop_y = slice_anchor
crop_w, crop_h = slice_shape
for v in range(expected_out_vertices.shape[0]):
expected_out_vertices[v, 0] = (expected_out_vertices[v, 0] -
crop_x) / crop_w
expected_out_vertices[v, 1] = (expected_out_vertices[v, 1] -
crop_y) / crop_h
np.testing.assert_allclose(expected_out_vertices,
out_vertices,
rtol=1e-4)
# Checking the conversion to absolute coordinates
h, w, c = image_shape
expected_out_vertices_abs = np.copy(expected_out_vertices)
for v in range(expected_out_vertices_abs.shape[0]):
expected_out_vertices_abs[v,
0] = expected_out_vertices[v, 0] * w
expected_out_vertices_abs[v,
1] = expected_out_vertices[v, 1] * h
np.testing.assert_allclose(expected_out_vertices_abs,
out_vertices_abs,
rtol=1e-4)
# Checking clamping of the relative coordinates
expected_out_vertices_clamped = np.copy(expected_out_vertices)
np.clip(expected_out_vertices_clamped, a_min=0.0, a_max=1.0)
np.testing.assert_allclose(expected_out_vertices_clamped,
out_vertices,
rtol=1e-4)
# Checking clamping of the absolute coordinates
expected_out_vertices_clamped_abs = np.copy(
expected_out_vertices_abs)
for v in range(expected_out_vertices_clamped_abs.shape[0]):
expected_out_vertices_clamped_abs[v, 0] = np.clip(
expected_out_vertices_abs[v, 0], a_min=0, a_max=w)
expected_out_vertices_clamped_abs[v, 1] = np.clip(
expected_out_vertices_abs[v, 1], a_min=0, a_max=h)
np.testing.assert_allclose(expected_out_vertices_clamped_abs,
out_vertices_clamped_abs,
rtol=1e-4)
# Checking scaling of the bounding boxes
expected_in_bboxes_abs = np.copy(in_bboxes)
for v in range(expected_in_bboxes_abs.shape[0]):
expected_in_bboxes_abs[v, 0] = expected_in_bboxes_abs[v, 0] * w
expected_in_bboxes_abs[v, 1] = expected_in_bboxes_abs[v, 1] * h
expected_in_bboxes_abs[v, 2] = expected_in_bboxes_abs[v, 2] * w
expected_in_bboxes_abs[v, 3] = expected_in_bboxes_abs[v, 3] * h
np.testing.assert_allclose(expected_in_bboxes_abs,
in_bboxes_abs,
rtol=1e-4)
# Check box selection and mapping to the cropping window
expected_out_bboxes = np.copy(in_bboxes[bbox_indices, :])
for k in range(expected_out_bboxes.shape[0]):
expected_out_bboxes[k, 0] = (expected_out_bboxes[k, 0] -
crop_x) / crop_w
expected_out_bboxes[k, 1] = (expected_out_bboxes[k, 1] -
crop_y) / crop_h
expected_out_bboxes[k, 2] = (expected_out_bboxes[k, 2] -
crop_x) / crop_w
expected_out_bboxes[k, 3] = (expected_out_bboxes[k, 3] -
crop_y) / crop_h
expected_out_bboxes = np.clip(expected_out_bboxes,
a_min=0.0,
a_max=1.0)
np.testing.assert_allclose(expected_out_bboxes,
out_bboxes,
rtol=1e-4)
expected_out_bboxes_abs = np.copy(expected_out_bboxes)
expected_out_bboxes_abs[:, 0] = expected_out_bboxes_abs[:, 0] * w
expected_out_bboxes_abs[:, 1] = expected_out_bboxes_abs[:, 1] * h
expected_out_bboxes_abs[:, 2] = expected_out_bboxes_abs[:, 2] * w
expected_out_bboxes_abs[:, 3] = expected_out_bboxes_abs[:, 3] * h
np.testing.assert_allclose(expected_out_bboxes_abs,
out_bboxes_abs,
rtol=1e-4)