def make_pipe():
image = fn.external_source(source=image_gen)
if device == "gpu":
image = image.gpu()
if get_dynamic_axes:
axes, rel_start, rel_shape = fn.external_source(
source=get_dynamic_axes, num_outputs=3)
else:
axes = types.Constant(np.array([0, 1], dtype=np.int32),
device="cpu")
rel_start = fn.random.uniform(range=(0.1, 0.2),
shape=(2, ),
dtype=types.FLOAT,
device=args_device)
rel_shape = fn.random.uniform(range=(0.4, 0.6),
shape=(2, ),
dtype=types.FLOAT,
device=args_device)
if args_device == "gpu":
sliced = fn.slice(image, rel_start, rel_shape, axes=axes)
return image, axes, rel_start, rel_shape, sliced
else:
sliced1 = fn.slice(image,
rel_start=rel_start,
rel_shape=rel_shape,
axes=axes)
sliced2 = fn.slice(image, rel_start, rel_shape, axes=axes)
return image, axes, rel_start, rel_shape, sliced1, sliced2
def check_slice_named_args_default_start_or_end(device, batch_size):
test_data_shape = [5, 4, 3]
def get_data():
out = [
np.random.randint(0, 255, size=test_data_shape, dtype=np.uint8)
for _ in range(batch_size)
]
return out
pipe = Pipeline(batch_size=batch_size, num_threads=4, device_id=0)
with pipe:
data = fn.external_source(source=get_data, layout="HWC")
in_shape = np.array([5, 4])
start = np.array([1, 2])
shape = np.array([3, 1])
end = start + shape
rel_start = start / in_shape
rel_shape = shape / in_shape
rel_end = end / in_shape
outs = [
fn.slice(data, start=start, end=in_shape, axes=(0, 1)),
fn.slice(data, start=[0, 0], end=end, axes=(0, 1)),
fn.slice(data, start=start, axes=(0, 1)),
fn.slice(data, end=end, axes=(0, 1)),
]
pipe.set_outputs(*outs)
pipe.build()
for _ in range(3):
outs = pipe.run()
for sample in range(batch_size):
np.testing.assert_equal(np.array(outs[0][sample]),
np.array(outs[2][sample]))
np.testing.assert_equal(np.array(outs[1][sample]),
np.array(outs[3][sample]))
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 make_pipe():
image = fn.external_source(source=image_gen)
if device == 'gpu':
image = image.gpu()
axes, rel_start, rel_shape = fn.external_source(source=get_dynamic_axes, num_outputs=3)
sliced1 = fn.slice(image, rel_start=rel_start, rel_shape=rel_shape, axes=axes)
sliced2 = fn.slice(image, rel_start, rel_shape, axes=axes)
return image, axes, rel_start, rel_shape, sliced1, sliced2
def crop_fn(self, img, lbl):
center = fn.segmentation.random_mask_pixel(lbl, foreground=fn.coin_flip(probability=self.oversampling))
crop_anchor = self.slice_fn(center, 1, self.dim) - self.crop_shape // 2
adjusted_anchor = math.max(0, crop_anchor)
max_anchor = self.slice_fn(fn.shapes(lbl), 1, self.dim) - self.crop_shape
crop_anchor = math.min(adjusted_anchor, max_anchor)
img = fn.slice(img.gpu(), crop_anchor, self.crop_shape, axis_names=self.axis_name, out_of_bounds_policy="pad")
lbl = fn.slice(lbl.gpu(), crop_anchor, self.crop_shape, axis_names=self.axis_name, out_of_bounds_policy="pad")
return img, lbl
def make_pipe():
encoded, _ = fn.readers.caffe(path=caffe_db_folder, random_shuffle=False)
image = fn.decoders.image(encoded, device="cpu", output_type=types.RGB)
if device == 'gpu':
image = image.gpu()
image = fn.cast(image, dtype=dtype)
sliced1 = fn.slice(image, 0, 3, axes=(2,))
sliced2 = fn.slice(image, rel_start=(0, 0, 0), rel_end=(1, 1, 1), axis_names="HWC")
return image, sliced1, sliced2
def make_pipe():
fake_data = fn.constant(idata=0, shape=[10, 10, 3], dtype=types.FLOAT, device=device)
axes = fn.random.uniform(range=wrong_axes_range, shape=(2,), dtype=types.INT32)
rel_start = fn.random.uniform(range=[0.0, 0.3], shape=(2,), dtype=types.FLOAT)
rel_shape = fn.random.uniform(range=[0.4, 0.6], shape=(2,), dtype=types.FLOAT)
if named_args:
sliced = fn.slice(fake_data, rel_start=rel_start, rel_shape=rel_shape, axes=axes)
else:
sliced = fn.slice(fake_data, rel_start, rel_shape, axes=axes)
return sliced
def test_slice_fn():
pipe = dali.pipeline.Pipeline(1,1,0)
src = fn.external_source([[np.array([[10,11,12],[13,14,15],[16,17,18]], dtype=np.float32)]])
out_cpu = fn.slice(src, np.array([1,1]), np.array([2,1]), axes=[0,1])
out_gpu = fn.slice(src.gpu(), np.array([1,1]), np.array([2,1]), axes=[0,1])
pipe.set_outputs(out_cpu, out_gpu)
pipe.build()
o = pipe.run()
assert np.array_equal(o[0].at(0), np.array([[14],[17]]))
assert np.array_equal(o[1].as_cpu().at(0), np.array([[14],[17]]))
def crop_fn(self, img, lbl):
center = fn.segmentation.random_mask_pixel(lbl, foreground=fn.coin_flip(probability=self.oversampling, **self.aug_seed_kwargs),
**self.aug_seed_kwargs)
crop_anchor = self.slice_fn(center) - self.crop_shape // 2
adjusted_anchor = math.max(0, crop_anchor)
max_anchor = self.slice_fn(fn.shapes(lbl)) - self.crop_shape
crop_anchor = math.min(adjusted_anchor, max_anchor)
img = fn.slice(img, crop_anchor, self.crop_shape, axis_names="DHW", out_of_bounds_policy="pad")
lbl = fn.slice(lbl, crop_anchor, self.crop_shape, axis_names="DHW", out_of_bounds_policy="pad")
return img, lbl
def dali_reflect_pad_graph(x, x_len, pad_amount):
def flip_1d(x):
# TODO(janton): remove the layout trick when Flip supports arbitrary data layouts
x = fn.reshape(x, shape=(-1, 1, 1), layout="HWC")
x = fn.flip(x, vertical=1)
x = fn.reshape(x, shape=(-1, ), layout="t")
return x
pad_start = fn.slice(x, 1, pad_amount, axes=(0, ))
pad_start = flip_1d(pad_start)
pad_end = fn.slice(x, x_len - pad_amount - 1, pad_amount, axes=(0, ))
pad_end = flip_1d(pad_end)
x = fn.cat(pad_start, x, pad_end, axis=0)
return x
def slice_pipeline(get_anchor, get_shape):
data = fn.external_source(source=get_data, layout='HWC')
anchors = fn.external_source(source=get_anchor)
shape = fn.external_source(source=get_shape)
processed = fn.slice(data, anchors, shape, out_of_bounds_policy='pad')
return processed
def check_slice_named_args_errors(device, batch_size):
test_data_shape = [5, 4, 3]
def get_data():
out = [
np.random.randint(0, 255, size=test_data_shape, dtype=np.uint8)
for _ in range(batch_size)
]
return out
pipe = Pipeline(batch_size=batch_size, num_threads=4, device_id=0)
with pipe:
data = fn.external_source(source=get_data, layout="HWC")
in_shape = np.array([5, 4])
start = np.array([1, 2])
shape = np.array([3, 1])
outs = [
fn.slice(data,
start,
shape,
start=start,
end=start + shape,
shape=shape,
axes=(0, 1)),
]
pipe.set_outputs(*outs)
with assert_raises(RuntimeError):
pipe.build()
for _ in range(1):
outs = pipe.run()
def create_decoder_slice_pipeline(data_path, device):
jpegs, _ = fn.readers.file(file_root = data_path,
shard_id = 0,
num_shards = 1,
name = "Reader")
anchor = fn.random.uniform(range=[0.05, 0.15], shape=(2,))
shape = fn.random.uniform(range=[0.5, 0.7], shape=(2,))
images_sliced_1 = fn.decoders.image_slice(jpegs,
anchor,
shape,
device = device,
hw_decoder_load = 0.7,
output_type = types.RGB,
axes = (0, 1))
images = fn.decoders.image(jpegs,
device = device,
hw_decoder_load = 0.7,
output_type = types.RGB)
images_sliced_2 = fn.slice(images,
anchor,
shape,
axes = (0, 1))
return images_sliced_1, images_sliced_2
def define_graph(self):
jpegs, dummy_labels = self.input()
self.labels = self.label()
self.crop_dim = self.crops()
anchor = fn.reshape(fn.slice(self.crop_dim, 0, 2, axes=[1]), shape=[-1])
shape = fn.reshape(fn.slice(self.crop_dim, 2, 2, axes = [1]), shape= [-1])
anchor = self.cast(anchor)
shape = self.cast(shape)
images = self.decode(jpegs)
images = self.res(images)
# decode and slicing
jpegs = fn.slice(jpegs, anchor, shape, axes= [0,1], device= 'gpu')
jpegs = self.res(jpegs)
return (images, self.labels, self.crop_dim)
def jpeg_distortion_pipe(device='cpu', quality=None):
iii = InputImagesIter(seq_len)
inputs = fn.external_source(source=iii, layout='FHWC', batch=False)
if device == 'gpu':
inputs = inputs.gpu()
if quality is None:
quality = fn.random.uniform(range=[1, 99], dtype=types.INT32)
tmp = fn.jpeg_compression_distortion(inputs, quality=quality)
outs = []
for i in range(seq_len):
# First, slice of the distorted sequence
outs.append(fn.slice(tmp, axes=(0, ), start=(i, ), end=(i + 1, )))
# Second, distorted slice of the input
slice_in = fn.slice(inputs, axes=(0, ), start=(i, ), end=(i + 1, ))
outs.append(
fn.jpeg_compression_distortion(slice_in, quality=quality))
return tuple(outs)
def pipe(max_batch_size, input_data, device):
pipe = Pipeline(batch_size=max_batch_size, num_threads=4, device_id=0)
anch = fn.constant(fdata=.1, device='cpu')
sh = fn.constant(fdata=.5, device='cpu')
data = fn.external_source(source=input_data, cycle=False, device=device)
processed = fn.slice(data, anch, sh, axes=0, device=device)
pipe.set_outputs(processed)
return pipe
def test_pipe():
data = fn.external_source(source=get_data)
shape = types.ScalarConstant(10)
anchor = types.ScalarConstant(5)
if device != 'cpu':
data = data.gpu()
sliced = fn.slice(data, start = anchor, shape = shape, axes=[0], device=device)
return data, sliced, shape, anchor
def biased_crop_fn(self, img, label):
roi_start, roi_end = fn.segmentation.random_object_bbox(
label,
format="start_end",
foreground_prob=self.oversampling,
background=0,
seed=self.internal_seed,
device="cpu",
cache_objects=True,
)
anchor = fn.roi_random_crop(label, roi_start=roi_start, roi_end=roi_end, crop_shape=[1, *self.patch_size])
anchor = fn.slice(anchor, 1, 3, axes=[0]) # drop channels from anchor
img, label = fn.slice(
[img, label], anchor, self.crop_shape, axis_names="DHW", out_of_bounds_policy="pad", device="cpu"
)
return img.gpu(), label.gpu()
def check_random_mask_pixel(ndim=2, batch_size=3,
min_extent=20, max_extent=50):
pipe = dali.pipeline.Pipeline(batch_size=batch_size, num_threads=4, device_id=0, seed=1234)
with pipe:
# Input mask
in_shape_dims = [fn.cast(fn.random.uniform(range=(min_extent, max_extent + 1)),
dtype=types.INT32) for _ in range(ndim)]
in_shape = fn.stack(*in_shape_dims)
in_mask = fn.cast(fn.random.uniform(range=(0, 2), shape=in_shape), dtype=types.INT32)
# > 0
fg_pixel1 = fn.segmentation.random_mask_pixel(in_mask, foreground=1)
# >= 0.99
fg_pixel2 = fn.segmentation.random_mask_pixel(in_mask, foreground=1, threshold=0.99)
# == 2
fg_pixel3 = fn.segmentation.random_mask_pixel(in_mask, foreground=1, value=2)
rnd_pixel = fn.segmentation.random_mask_pixel(in_mask, foreground=0)
coin_flip = fn.random.coin_flip(probability=0.7)
fg_biased = fn.segmentation.random_mask_pixel(in_mask, foreground=coin_flip)
# Demo purposes: Taking a random pixel and produce a valid anchor to feed slice
# We want to force the center adjustment, thus the large crop shape
crop_shape = in_shape - 2
anchor = fn.cast(fg_pixel1, dtype=types.INT32) - crop_shape // 2
anchor = math.min(math.max(0, anchor), in_shape - crop_shape)
out_mask = fn.slice(in_mask, anchor, crop_shape, axes=tuple(range(ndim)))
pipe.set_outputs(in_mask, fg_pixel1, fg_pixel2, fg_pixel3, rnd_pixel, coin_flip, fg_biased,
anchor, crop_shape, out_mask)
pipe.build()
for iter in range(3):
outputs = pipe.run()
for idx in range(batch_size):
in_mask = outputs[0].at(idx)
fg_pixel1 = outputs[1].at(idx).tolist()
fg_pixel2 = outputs[2].at(idx).tolist()
fg_pixel3 = outputs[3].at(idx).tolist()
rnd_pixel = outputs[4].at(idx).tolist()
coin_flip = outputs[5].at(idx).tolist()
fg_biased = outputs[6].at(idx).tolist()
anchor = outputs[7].at(idx).tolist()
crop_shape = outputs[8].at(idx).tolist()
out_mask = outputs[9].at(idx)
assert in_mask[tuple(fg_pixel1)] > 0
assert in_mask[tuple(fg_pixel2)] > 0.99
assert in_mask[tuple(fg_pixel3)] == 2
assert in_mask[tuple(fg_biased)] > 0 or not coin_flip
for d in range(ndim):
assert 0 <= anchor[d] and anchor[d] + crop_shape[d] <= in_mask.shape[d]
assert out_mask.shape == tuple(crop_shape)
def check_pad_to_square(device='cpu', batch_size=3, ndim=2, num_iter=3):
pipe = Pipeline(batch_size=batch_size,
num_threads=3,
device_id=0,
seed=1234)
axes = (0, 1)
with pipe:
in_shape = fn.cast(fn.random.uniform(range=(10, 20), shape=(ndim, )),
dtype=types.INT32)
in_data = fn.reshape(fn.random.uniform(range=(0., 1.), shape=in_shape),
layout="HW")
shape = fn.shapes(in_data, dtype=types.INT32)
h = fn.slice(shape, 0, 1, axes=[0])
w = fn.slice(shape, 1, 1, axes=[0])
side = math.max(h, w)
if device == 'gpu':
in_data = in_data.gpu()
out_data = fn.pad(in_data,
axis_names="HW",
shape=fn.cat(side, side, axis=0))
pipe.set_outputs(in_data, out_data)
pipe.build()
for _ in range(num_iter):
outs = [
out.as_cpu() if isinstance(out, TensorListGPU) else out
for out in pipe.run()
]
for i in range(batch_size):
in_data, out_data = \
[outs[out_idx].at(i) for out_idx in range(len(outs))]
in_shape = in_data.shape
max_side = max(in_shape)
for s in out_data.shape:
assert s == max_side
np.testing.assert_equal(out_data[:in_shape[0], :in_shape[1]],
in_data)
np.testing.assert_equal(out_data[in_shape[0]:, :], 0)
np.testing.assert_equal(out_data[:, in_shape[1]:], 0)
def make_pipe():
fake_data = fn.constant(idata=0,
shape=[10, 10, 3],
dtype=types.FLOAT,
device="cpu")
rel_start = fn.random.uniform(range=[0.0, 0.3],
shape=(2, ),
dtype=types.FLOAT,
device="gpu")
rel_shape = fn.random.uniform(range=[0.4, 0.6],
shape=(2, ),
dtype=types.FLOAT,
device="gpu")
sliced = fn.slice(fake_data, rel_start, rel_shape, device="cpu")
return sliced
def create_dali_pipe(channel_first, seq_len, interp, dtype, w, h, batch_size=2):
pipe = dali.pipeline.Pipeline(batch_size, 1, 0, 0)
with pipe:
layout = "FCHW" if channel_first else "FHWC"
ext = fn.external_source(GetSequences(channel_first, seq_len, batch_size), layout=layout)
resize_cpu_out = fn.resize(ext, resize_x=w, resize_y=h, interp_type=interp,
dtype=dtype, save_attrs=True)
resize_gpu_out = fn.resize(ext.gpu(), resize_x=w, resize_y=h, interp_type=interp,
minibatch_size=4, dtype=dtype, save_attrs=True)
dali_resized_cpu, size_cpu = resize_cpu_out
dali_resized_gpu, size_gpu = resize_gpu_out
# extract just HW part from the input shape
ext_size = fn.slice(fn.cast(fn.shapes(ext), dtype=types.INT32),
2 if channel_first else 1, 2, axes=[0])
pipe.set_outputs(dali_resized_cpu, dali_resized_gpu, ext_size, size_cpu, size_gpu)
return pipe
def numpy_reader_roi_pipe(file_root,
device="cpu",
file_filter='*.npy',
roi_start=None,
rel_roi_start=None,
roi_end=None,
rel_roi_end=None,
roi_shape=None,
rel_roi_shape=None,
roi_axes=None,
default_axes=[],
out_of_bounds_policy=None,
fill_value=None):
data = fn.readers.numpy(device=device,
file_root=file_root,
file_filter=file_filter,
shard_id=0,
num_shards=1,
cache_header_information=False)
roi_data = fn.readers.numpy(device=device,
file_root=file_root,
file_filter=file_filter,
roi_start=roi_start,
rel_roi_start=rel_roi_start,
roi_end=roi_end,
rel_roi_end=rel_roi_end,
roi_shape=roi_shape,
rel_roi_shape=rel_roi_shape,
roi_axes=roi_axes,
out_of_bounds_policy=out_of_bounds_policy,
fill_value=fill_value,
shard_id=0,
num_shards=1,
cache_header_information=False)
sliced_data = fn.slice(
data,
start=roi_start,
rel_start=rel_roi_start,
end=roi_end,
rel_end=rel_roi_end,
shape=roi_shape,
rel_shape=rel_roi_shape,
axes=roi_axes
or default_axes, # Slice has different default (axis_names="WH")
out_of_bounds_policy=out_of_bounds_policy,
fill_values=fill_value)
return roi_data, sliced_data
def _test_empty_input(dim, device):
batch_size = 8
pipe = Pipeline(batch_size=batch_size,
num_threads=8,
device_id=0,
seed=1234)
if dim == 2:
files, labels = dali.fn.readers.caffe(path=db_2d_folder,
random_shuffle=True)
images_cpu = dali.fn.decoders.image(files, device="cpu")
else:
images_cpu = dali.fn.external_source(
source=random_3d_loader(batch_size), layout="DHWC")
images = images_cpu if device == "cpu" else images_cpu.gpu()
in_rel_shapes = np.ones([batch_size, dim], dtype=np.float32)
in_rel_shapes[::2, :] *= 0 # all zeros in every second sample
degenerate_images = fn.slice(images,
np.zeros([dim]),
fn.external_source(lambda: in_rel_shapes),
axes=list(range(dim)))
sizes = np.random.randint(20, 50, [batch_size, dim], dtype=np.int32)
size_inp = fn.external_source(
lambda: [x.astype(np.float32) for x in sizes])
resize_no_empty = fn.resize(images, size=size_inp, mode="not_larger")
resize_with_empty = fn.resize(degenerate_images,
size=size_inp,
mode="not_larger")
pipe.set_outputs(resize_no_empty, resize_with_empty)
pipe.build()
for it in range(3):
out_no_empty, out_with_empty = pipe.run()
if device == "gpu":
out_no_empty = out_no_empty.as_cpu()
out_with_empty = out_with_empty.as_cpu()
for i in range(batch_size):
if i % 2 != 0:
assert np.array_equal(out_no_empty.at(i), out_with_empty.at(i))
else:
assert np.prod(out_with_empty.at(i).shape) == 0
def check_slice_named_args(device, batch_size):
test_data_shape = [5, 4, 3]
def get_data():
out = [np.random.randint(0, 255, size = test_data_shape, dtype = np.uint8) for _ in range(batch_size)]
return out
pipe = Pipeline(batch_size=batch_size, num_threads=4, device_id=0)
with pipe:
data = fn.external_source(source = get_data, layout = "HWC")
in_shape_list = [5, 4]
start_list = [1, 2]
shape_list = [3, 1]
in_shape = np.array(in_shape_list)
start = np.array(start_list)
shape = np.array(shape_list)
end_list = [start_list[i] + shape_list[i] for i in range(2)]
end = start + shape
rel_start_list = [start_list[i] / in_shape_list[i] for i in range(2)]
rel_start = start / in_shape
rel_shape_list = [shape_list[i] / in_shape_list[i] for i in range(2)]
rel_shape = shape / in_shape
rel_end_list = [end_list[i] / in_shape_list[i] for i in range(2)]
rel_end = end / in_shape
outs = [
fn.slice(data, start, shape, axes = (0, 1)),
fn.slice(data, rel_start, rel_shape, axes = (0, 1)),
]
for start_arg in [start, start_list]:
for shape_arg in [shape, shape_list]:
outs += [fn.slice(data, start=start_arg, shape=shape_arg, axes = (0, 1))]
for end_arg in [end, end_list]:
outs += [fn.slice(data, start=start_arg, end=end_arg, axes = (0, 1))]
for rel_start_arg in [rel_start, rel_start_list]:
for rel_shape_arg in [rel_shape, rel_shape_list]:
outs += [fn.slice(data, rel_start=rel_start_arg, rel_shape=rel_shape_arg, axes = (0, 1))]
for rel_end_arg in [rel_end, rel_end_list]:
outs += [fn.slice(data, rel_start=rel_start_arg, rel_end=rel_end_arg, axes = (0, 1))]
for shape_arg in [shape, shape_list]:
outs += [fn.slice(data, rel_start=rel_start_arg, shape=shape_arg, axes = (0, 1))]
pipe.set_outputs(*outs)
pipe.build()
for _ in range(3):
outs = pipe.run()
for out_idx in range(1, len(outs)):
for sample in range(batch_size):
np.testing.assert_equal(np.array(outs[0][sample]), np.array(outs[out_idx][sample]))
def create_dali_pipe(channel_first,
seq_len,
interp,
dtype,
w,
h,
batch_size=2):
pipe = dali.pipeline.Pipeline(batch_size, 1, 0, 0)
with pipe:
layout = "FCHW" if channel_first else "FHWC"
ext = fn.external_source(GetSequences(channel_first, seq_len,
batch_size),
layout=layout)
resize_cpu_out = fn.resize(ext,
resize_x=w,
resize_y=h,
interp_type=interp,
dtype=dtype,
save_attrs=True)
resize_gpu_out = fn.resize(ext.gpu(),
resize_x=w,
resize_y=h,
interp_type=interp,
minibatch_size=4,
dtype=dtype,
save_attrs=True)
dali_resized_cpu, size_cpu = resize_cpu_out
dali_resized_gpu, size_gpu = resize_gpu_out
# extract just HW part from the input shape
shape_anchor = np.array([2 if channel_first else 1], dtype=np.float32)
shape_shape = np.array([2], dtype=np.float32)
ext_size = fn.slice(fn.cast(fn.shapes(ext), dtype=types.INT32),
types.Constant(shape_anchor, device="cpu"),
types.Constant(shape_shape, device="cpu"),
normalized_anchor=False,
normalized_shape=False,
axes=[0])
pipe.set_outputs(dali_resized_cpu, dali_resized_gpu, ext_size,
size_cpu, size_gpu)
return pipe
def dali_frame_splicing_graph(x, nfeatures, x_len, stacking=1, subsampling=1):
if stacking > 1:
seq = [x]
for n in range(1, stacking):
f = fn.slice(x,
n,
x_len,
axes=(1, ),
out_of_bounds_policy='pad',
fill_values=0)
seq.append(f)
x = fn.cat(*seq, axis=0)
nfeatures = nfeatures * stacking
if subsampling > 1:
out_len = (x_len + subsampling - 1) // subsampling
m = fn.transforms.scale(scale=[subsampling, 1], center=[0.5, 0])
x = fn.reshape(x, rel_shape=[1, 1, -1],
layout="HWC") # Layout required by WarpAffine
size = fn.cat(nfeatures, out_len)
x = fn.warp_affine(x, matrix=m, size=size, interp_type=types.INTERP_NN)
x = fn.reshape(x, rel_shape=[1, 1], layout="ft")
return x
def test_slice_cpu():
anch_shape = [2]
def get_anchors():
out = [(np.random.randint(1, 256, 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, 256, 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)
data = fn.external_source(source=get_data, layout="HWC")
anchors = fn.external_source(source=get_anchors)
shape = fn.external_source(source=get_shape)
processed = fn.slice(data, anchors, shape, out_of_bounds_policy="pad")
pipe.set_outputs(processed)
pipe.build()
for _ in range(3):
pipe.run()
def rnnt_train_pipe(files,
sample_rate,
pad_amount=0,
preemph_coeff=.97,
window_size=.02,
window_stride=.01,
window="hann",
nfeatures=64,
nfft=512,
frame_splicing_stack=1,
frame_splicing_subsample=1,
lowfreq=0.0,
highfreq=None,
normalize_type='per_feature',
speed_perturb=False,
silence_trim=False,
device='cpu'):
assert normalize_type == 'per_feature' or normalize_type == 'all_features'
norm_axes = [1] if normalize_type == 'per_feature' else [0, 1]
win_len, win_hop = win_args(sample_rate, window_size, window_stride)
window_fn = torch_windows.get(window, None)
window_fn_arg = window_fn(
win_len, periodic=False).numpy().tolist() if window_fn else None
data, _ = fn.readers.file(files=files,
device="cpu",
random_shuffle=False,
shard_id=0,
num_shards=1)
audio, _ = fn.decoders.audio(data, dtype=types.FLOAT, downmix=True)
# splicing with subsampling doesn't work if audio_len is a GPU data node
if device == 'gpu' and frame_splicing_subsample == 1:
audio = audio.gpu()
# Speed perturbation 0.85x - 1.15x
if speed_perturb:
target_sr_factor = fn.random.uniform(device="cpu",
range=(1 / 1.15, 1 / 0.85))
audio = fn.experimental.audio_resample(audio, scale=target_sr_factor)
# Silence trimming
if silence_trim:
begin, length = fn.nonsilent_region(audio, cutoff_db=-80)
audio = fn.slice(audio, begin, length, axes=[0])
audio_shape = fn.shapes(audio, dtype=types.INT32)
orig_audio_len = fn.slice(audio_shape, 0, 1, axes=(0, ))
# If we couldn't move to GPU earlier, do it now
if device == 'gpu' and frame_splicing_subsample > 1:
audio = audio.gpu()
if pad_amount > 0:
audio_len = orig_audio_len + 2 * pad_amount
padded_audio = dali_reflect_pad_graph(audio, orig_audio_len,
pad_amount)
else:
audio_len = orig_audio_len
padded_audio = audio
# Preemphasis filter
preemph_audio = fn.preemphasis_filter(padded_audio,
preemph_coeff=preemph_coeff,
border='zero')
# Spectrogram
spec_len = audio_len // win_hop + 1
spec = fn.spectrogram(preemph_audio,
nfft=nfft,
window_fn=window_fn_arg,
window_length=win_len,
window_step=win_hop,
center_windows=True,
reflect_padding=True)
# Mel spectrogram
mel_spec = fn.mel_filter_bank(spec,
sample_rate=sample_rate,
nfilter=nfeatures,
freq_low=lowfreq,
freq_high=highfreq)
# Log
log_features = fn.to_decibels(mel_spec + 1e-20,
multiplier=np.log(10),
reference=1.0,
cutoff_db=-80)
# Frame splicing
if frame_splicing_stack > 1 or frame_splicing_subsample > 1:
log_features_spliced = dali_frame_splicing_graph(
log_features,
nfeatures,
spec_len,
stacking=frame_splicing_stack,
subsampling=frame_splicing_subsample)
else:
log_features_spliced = log_features
# Normalization
if normalize_type:
norm_log_features = fn.normalize(log_features_spliced,
axes=norm_axes,
device=device,
epsilon=4e-5,
ddof=1)
else:
norm_log_features = log_features_spliced
return norm_log_features, log_features_spliced, log_features, mel_spec, spec, preemph_audio, padded_audio, audio
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)