def pipe(device, out_type, files):
encoded, _ = fn.readers.file(files=files)
decoded = fn.decoders.image(encoded,
device=device,
output_type=out_type)
peeked_shape = fn.peek_image_shape(encoded)
return decoded, peeked_shape
def run_decode(data_path, out_type):
batch_size = 4
pipe = Pipeline(batch_size=batch_size, num_threads=4, device_id=0)
input, _ = fn.file_reader(file_root=data_path,
shard_id=0,
num_shards=1,
name="reader")
decoded = fn.image_decoder(input, output_type=types.RGB)
decoded_shape = fn.shapes(decoded)
raw_shape = fn.peek_image_shape(input, type=out_type)
pipe.set_outputs(decoded, decoded_shape, raw_shape)
pipe.build()
samples = 0
length = pipe.reader_meta(name="reader")['epoch_size']
while samples < length:
samples += batch_size
(images, decoded_shape, raw_shape) = pipe.run()
for i in range(batch_size):
# as we are asking for a particular color space it may differ from the source image, so don't compare it
image = images.at(i)
shape_type = dali_types_to_np(out_type)
for d in range(len(image.shape) - 1):
assert image.shape[d] == decoded_shape.at(
i)[d], "{} vs {}".format(image.shape[d],
decoded_shape.at(i)[d])
assert image.shape[d] == raw_shape.at(i)[d], "{} vs {}".format(
image.shape[d],
raw_shape.at(i)[d])
assert raw_shape.at(
i)[d].dtype == shape_type, "{} vs {}".format(
raw_shape.at(i)[d].dtyp, shape_type)
def define_graph(self):
inputs, bboxes, labels, polygons, vertices = fn.readers.coco(
file_root=self.file_root,
annotations_file=self.annotation_file,
skip_empty=True,
shard_id=self.share_id,
num_shards=self.num_gpus,
ratio=True,
ltrb=True,
polygon_masks = True,
random_shuffle=self.random_shuffle,
shuffle_after_epoch=self.shuffle_after_epoch,
name="Reader")
input_shape = fn.slice(fn.cast(fn.peek_image_shape(inputs), dtype=types.INT32), 0, 2, axes=[0])
h = fn.slice(input_shape, 0, 1, axes = [0], dtype=types.FLOAT)
w = fn.slice(input_shape, 1, 1, axes = [0], dtype=types.FLOAT)
short_side = math.min(w, h)
scale = fn.random.uniform(range=[0.3, 1.])
crop_side = fn.cast(math.ceil(scale * short_side), dtype=types.INT32)
crop_shape = fn.cat(crop_side, crop_side)
anchor_rel, shape_rel, bboxes, labels, bbox_indices = fn.random_bbox_crop(
bboxes,
labels,
input_shape=input_shape,
crop_shape=crop_shape,
shape_layout="HW",
thresholds=[0.], # No minimum intersection-over-union, for demo purposes
allow_no_crop=False, # No-crop is disallowed, for demo purposes
seed=-1, # Fixed random seed for deterministic results
bbox_layout="xyXY", # left, top, right, back
output_bbox_indices=True, # Output indices of the filtered bounding boxes
total_num_attempts=1024,
)
polygons, vertices = fn.segmentation.select_masks(
bbox_indices, polygons, vertices
)
images = fn.decoders.image_slice(
inputs, anchor_rel, shape_rel, normalized_anchor=False, normalized_shape=False, device='mixed'
)
images = fn.color_space_conversion(images, image_type=types.RGB, output_type=types.BGR)
MT_1_vertices = fn.transforms.crop(
to_start=(0.0, 0.0), to_end=fn.cat(w, h)
)
MT_2_vertices = fn.transforms.crop(
from_start=anchor_rel, from_end=(anchor_rel + shape_rel),
to_start=(0.0, 0.0), to_end=(1., 1.)
)
vertices = fn.coord_transform(fn.coord_transform(vertices, MT=MT_1_vertices), MT=MT_2_vertices)
targets = fn.cat( bboxes, fn.reshape(vertices, shape=[-1, 10]), axis=1)
interp_methods = [types.INTERP_LINEAR, types.INTERP_CUBIC, types.INTERP_LANCZOS3, types.INTERP_GAUSSIAN, types.INTERP_NN, types.INTERP_TRIANGULAR]
interp_method = fn.random.uniform(values=[int(x) for x in interp_methods], dtype=types.INT32)
interp_method = fn.reinterpret(interp_method, dtype=types.INTERP_TYPE)
images = fn.resize(images, dtype=types.FLOAT, size=self.input_dim, interp_type=interp_method)
labels = labels.gpu()
targets = targets.gpu()
return (images, targets, labels)
def peek_image_shape_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')
shape = fn.peek_image_shape(encoded, device=device)
pipe.set_outputs(shape)
return pipe
def test_peek_image_shape_cpu():
pipe = Pipeline(batch_size=batch_size, num_threads=4, device_id=None)
input, _ = fn.readers.file(file_root=images_dir, shard_id=0, num_shards=1)
shapes = fn.peek_image_shape(input)
pipe.set_outputs(shapes)
pipe.build()
for _ in range(3):
pipe.run()
def setup_dali(
image_file='/mnt/data/DATASETS/samples/images/image_110.jpg',
image_dim=[800, 1600],
batch_size=1,
num_threads=4,
device='mixed',
device_id=0,
output_dir='./out/',
):
os.makedirs(os.path.dirname(output_dir), exist_ok=True)
pipeline = dali.pipeline.Pipeline(batch_size=batch_size,
num_threads=num_threads,
device_id=device_id)
with pipeline:
data, _ = fn.file_reader(files=[image_file])
# image preprocess
images = fn.image_decoder(data, device=device)
images = fn.resize(images,
size=image_dim,
mode="not_larger",
max_size=image_dim)
images = fn.pad(images,
fill_value=0,
shape=[image_dim[0], image_dim[1], 1])
images = fn.transpose(images, perm=[2, 0, 1])
images = fn.cast(images, dtype=dali.types.FLOAT)
images = images / 255.
# input shape
input_shape = np.float32((image_dim[0], image_dim[1], 1))
# original shape
shapes = fn.peek_image_shape(data)
shapes = fn.cast(shapes, dtype=dali.types.FLOAT)
# gather outputs
out = [images, input_shape, shapes]
pipeline.set_outputs(*out)
pipeline.build()
output = pipeline.run()
img = output[0].at(0) if device == 'cpu' else output[0].as_cpu().at(0)
img = img.transpose(1, 2, 0) # HWC
img = img[:, :, ::-1] # BGR
print(img)
quit()
cv2.imwrite(os.path.join(output_dir, 'dali_image.jpg'), img)
def setup_dali(
input_name='DALI_INPUT_0',
image_dim=[896, 1536],
batch_size=1,
num_threads=4,
device='cpu',
device_id=0,
output_dir='./out/',
):
pipeline = dali.pipeline.Pipeline(batch_size=batch_size,
num_threads=num_threads,
device_id=device_id)
with pipeline:
data = fn.external_source(name=input_name, device="cpu")
# image preprocess
images = fn.image_decoder(data, device=device)
images = fn.resize(images,
size=image_dim,
mode="not_larger",
max_size=image_dim)
images = fn.pad(images,
fill_value=0,
shape=[image_dim[0], image_dim[1], 1])
images = fn.transpose(images, perm=[2, 0, 1])
images = fn.cast(images, dtype=dali.types.FLOAT)
images = images / 255.
# input shape
input_shape = np.float32((image_dim[0], image_dim[1], 1))
# original shape
shapes = fn.peek_image_shape(data)
shapes = fn.cast(shapes, dtype=dali.types.FLOAT)
# gather outputs
out = [images, input_shape, shapes]
pipeline.set_outputs(*out)
os.makedirs(os.path.dirname(output_dir), exist_ok=True)
pipeline.serialize(filename=os.path.join(output_dir, 'model.dali'))
def pipe():
encoded, _ = fn.readers.file(files=[normal, palette])
peeked_shapes = fn.peek_image_shape(encoded)
decoded = fn.decoders.image(encoded, device='cpu')
return decoded, peeked_shapes
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
_ = fn.decoders.image_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, dtype=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()
# Enough iterations to see an example with more than one bounding box
for i in range(num_iters):
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, _ = image_shape
wh = np.array([w, h])
whwh = np.array([w, h, w, h])
expected_out_vertices_abs = expected_out_vertices * wh
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.clip(expected_out_vertices,
a_min=0.0,
a_max=1.0)
np.testing.assert_allclose(expected_out_vertices_clamped,
out_vertices_clamped,
rtol=1e-4)
# Checking clamping of the absolute coordinates
expected_out_vertices_clamped_abs = np.clip(
expected_out_vertices_abs, 0, wh)
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 = in_bboxes * whwh
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 = expected_out_bboxes * whwh
np.testing.assert_allclose(expected_out_bboxes_abs,
out_bboxes_abs,
rtol=1e-4)
