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 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_per_sample_shapes_and_alignment(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.random.uniform(range=(0., 1.), shape=in_shape)
if device == 'gpu':
in_data = in_data.gpu()
req_shape = fn.cast(fn.random.uniform(range=(21, 30), shape=(ndim, )),
dtype=types.INT32)
req_align = fn.cast(fn.random.uniform(range=(3, 5), shape=(ndim, )),
dtype=types.INT32)
out_pad_shape = fn.pad(in_data, axes=axes, align=None, shape=req_shape)
out_pad_align = fn.pad(in_data, axes=axes, align=req_align, shape=None)
out_pad_both = fn.pad(in_data,
axes=axes,
align=req_align,
shape=req_shape)
pipe.set_outputs(in_shape, in_data, req_shape, req_align,
out_pad_shape, out_pad_align, out_pad_both)
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_shape, in_data, req_shape, req_align, out_pad_shape, out_pad_align, out_pad_both = \
[outs[out_idx].at(i) for out_idx in range(len(outs))]
assert (in_shape == in_data.shape).all()
# Pad to explicit shape
assert (out_pad_shape.shape >= in_shape).all()
assert (req_shape == out_pad_shape.shape).all()
# Alignment only
assert (out_pad_align.shape >= in_shape).all()
assert is_aligned(out_pad_align.shape, req_align, axes)
# Explicit shape + alignment
assert (out_pad_both.shape >= in_shape).all()
assert (req_shape <= out_pad_both.shape).all()
assert is_aligned(out_pad_both.shape, req_align, axes)
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 many_input_pipeline(def_for_dataset, device, sources, input_names,
batches):
""" Pipeline accepting multiple inputs via external source
Parameters
----------
def_for_dataset : bool
True if this pipeline will be converted to TF Dataset
device : str
device that the Dataset will be placed ("cpu" or "gpu")
sources : list of callables
callbacks for the external sources in baseline pipeline otherwise None
input_names : list of str
Names of inputs placeholder for TF
"""
inputs = []
if def_for_dataset:
for input_name, batch in zip(input_names, batches):
if batch == "dataset":
# Special value used in tests, reroute it to the default
batch = None
input = fn.external_source(name=input_name, batch=batch)
input = input if device == 'cpu' else input.gpu()
inputs.append(input)
else:
for source in sources:
input = fn.external_source(source=source, batch=False)
input = input if device == 'cpu' else input.gpu()
inputs.append(input)
processed = []
for input in inputs:
processed.append(fn.cast(input + 10, dtype=dali.types.INT32))
results = fn.pad(inputs + processed)
return tuple(results)
def check_per_sample_gaussian_blur(batch_size,
sigma_dim,
window_size_dim,
shape,
layout,
axes,
op_type="cpu"):
pipe = Pipeline(batch_size=batch_size, num_threads=4, device_id=0)
data = RandomlyShapedDataIterator(batch_size, max_shape=shape)
with pipe:
if sigma_dim is not None:
sigma = fn.random.uniform(range=[0.5, 3], shape=[sigma_dim])
sigma_arg = sigma
else:
# placeholder, so we can return something
sigma = fn.random.coin_flip(probability=0)
sigma_arg = None
if window_size_dim is not None:
window_radius = fn.random.uniform(range=[5, 10],
shape=[window_size_dim])
window_size = fn.cast(window_radius, dtype=types.INT32) * 2 + 1
window_arg = window_size
else:
window_size = fn.random.coin_flip(probability=0)
window_arg = None
input = fn.external_source(data, layout=layout)
if op_type == "gpu":
input = input.gpu()
blurred = fn.gaussian_blur(input,
device=op_type,
sigma=sigma_arg,
window_size=window_arg)
pipe.set_outputs(blurred, input, sigma, window_size)
pipe.build()
for _ in range(test_iters):
result, input, sigma, window_size = pipe.run()
if op_type == "gpu":
result = result.as_cpu()
input = input.as_cpu()
input = to_batch(input, batch_size)
sigma = to_batch(sigma, batch_size)
window_size = to_batch(window_size, batch_size)
baseline = []
for i in range(batch_size):
sigma_arg = sigma[i] if sigma is not None else None
window_arg = window_size[i] if window_size_dim is not None else None
skip_axes = count_skip_axes(layout)
baseline.append(
gaussian_baseline(input[i], sigma_arg, window_arg, axes,
skip_axes))
check_batch(result,
baseline,
batch_size,
max_allowed_error=1,
expected_layout=layout)
def define_graph(self):
inputs, labels = self.input(name="Reader")
images = self.decode(inputs)
if self.prime_size:
images = fn.resize(images, resize_x=101, resize_y=43)
images = fn.cast(images, dtype=self.dtype)
images = self.water(images)
return images
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 get_random_pipeline(batch_size):
pipe = Pipeline(batch_size, 4, None)
with pipe:
input, _ = fn.readers.file(file_root=img_dir)
decoded = fn.image_decoder(input, device='cpu', output_type=types.RGB)
tile = fn.cast(fn.uniform(range=(50, 200), shape=[1]), dtype=types.INT32)
ratio = fn.uniform(range=(0.3, 0.7), shape=[1])
angle = fn.uniform(range=(-math.pi, math.pi), shape=[1])
grided = fn.grid_mask(decoded, device='cpu', tile=tile, ratio=ratio, angle=angle)
pipe.set_outputs(grided, decoded, tile, ratio, angle)
return pipe
def define_graph(self):
inputs, labels = self.input(name="Reader")
images = self.decode(inputs)
if self.device == 'gpu':
images = images.gpu()
if self.prime_size:
images = fn.resize(images, resize_x=101, resize_y=43)
mask = fn.random.coin_flip(seed=42) if self.do_mask else None
images = fn.cast(images, dtype=self.dtype)
images = self.water(images, mask=mask)
return images
def create_ref_pipe(channel_first, seq_len, interp, dtype, w, h, batch_size=2):
pipe = dali.pipeline.Pipeline(batch_size, 1, 0, 0, exec_async=False, exec_pipelined=False)
with pipe:
layout = "FCHW" if channel_first else "FHWC"
ext = fn.external_source(GetSequences(channel_first, seq_len, batch_size), layout=layout)
pil_resized = fn.python_function(ext, function=resize_PIL(channel_first, interp, w, h),
batch_processing=False)
if dtype is not None: # unfortunately, PIL can't quite handle that
pil_resized = fn.cast(pil_resized, dtype=dtype)
pil_resized = fn.reshape(pil_resized, layout=layout)
pipe.set_outputs(pil_resized)
return pipe
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 create_dali_pipeline(batch_size, num_threads, device_id, data_dir, crop, size,
shard_id, num_shards, dali_cpu=False, is_training=True):
pipeline = Pipeline(batch_size, num_threads, device_id, seed=12 + device_id)
with pipeline:
images, labels = fn.readers.file(file_root=data_dir,
shard_id=shard_id,
num_shards=num_shards,
random_shuffle=is_training,
pad_last_batch=True,
name="Reader")
dali_device = 'cpu' if dali_cpu else 'gpu'
decoder_device = 'cpu' if dali_cpu else 'mixed'
device_memory_padding = 211025920 if decoder_device == 'mixed' else 0
host_memory_padding = 140544512 if decoder_device == 'mixed' else 0
if is_training:
images = fn.image_decoder_random_crop(images,
device=decoder_device, output_type=types.RGB,
device_memory_padding=device_memory_padding,
host_memory_padding=host_memory_padding,
random_aspect_ratio=[0.8, 1.25],
random_area=[0.1, 1.0],
num_attempts=100)
images = fn.resize(images,
device=dali_device,
resize_x=crop,
resize_y=crop,
interp_type=types.INTERP_TRIANGULAR)
mirror = fn.random.coin_flip(probability=0.5)
else:
images = fn.image_decoder(images,
device=decoder_device,
output_type=types.RGB)
images = fn.resize(images,
device=dali_device,
size=size,
mode="not_smaller",
interp_type=types.INTERP_TRIANGULAR)
mirror = False
images = fn.crop_mirror_normalize(images.gpu(),
dtype=types.FLOAT,
output_layout="CHW",
crop=(crop, crop),
mean=[0.485 * 255,0.456 * 255,0.406 * 255],
std=[0.229 * 255,0.224 * 255,0.225 * 255],
mirror=mirror)
labels = labels.gpu()
labels = fn.cast(labels, dtype=types.INT64)
pipeline.set_outputs(images, labels)
return pipeline
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 get_random_pipeline(device, batch_size):
pipe = Pipeline(batch_size, 4, 0)
with pipe:
input, _ = fn.readers.file(file_root=img_dir)
decoded = fn.decoders.image(input, device='cpu', output_type=types.RGB)
decoded = decoded.gpu() if device == 'gpu' else decoded
tile = fn.cast(fn.uniform(range=(50, 200)), dtype=types.INT32)
ratio = fn.uniform(range=(0.3, 0.7))
angle = fn.uniform(range=(-math.pi, math.pi))
grided = fn.grid_mask(decoded,
device=device,
tile=tile,
ratio=ratio,
angle=angle)
pipe.set_outputs(grided, decoded, tile, ratio, angle)
return pipe
def get_image_pipeline(batch_size, num_threads, device, device_id=0, shard_id=0, num_shards=1,
def_for_dataset=False):
test_data_root = get_dali_extra_path()
file_root = os.path.join(test_data_root, 'db', 'coco_dummy', 'images')
annotations_file = os.path.join(
test_data_root, 'db', 'coco_dummy', 'instances.json')
pipe = Pipeline(batch_size, num_threads, device_id)
with pipe:
jpegs, _, _, image_ids = fn.readers.coco(
file_root=file_root,
annotations_file=annotations_file,
shard_id=shard_id,
num_shards=num_shards,
ratio=False,
image_ids=True)
images = fn.decoders.image(
jpegs,
device=('mixed' if device == 'gpu' else 'cpu'),
output_type=types.RGB)
images = fn.resize(
images,
resize_x=224,
resize_y=224,
interp_type=types.INTERP_LINEAR)
images = fn.crop_mirror_normalize(
images,
dtype=types.FLOAT,
mean=[128., 128., 128.],
std=[1., 1., 1.])
if device == 'gpu':
image_ids = image_ids.gpu()
ids_reshaped = fn.reshape(image_ids, shape=[1, 1])
ids_int16 = fn.cast(image_ids, dtype=types.INT16)
pipe.set_outputs(images, ids_reshaped, ids_int16)
shapes = (
(batch_size, 3, 224, 224),
(batch_size, 1, 1),
(batch_size, 1))
dtypes = (
tf.float32,
tf.int32,
tf.int16)
return pipe, shapes, dtypes
def laplacian_pipe(window_size, in_type, out_type, normalize, grayscale):
# use OpenCV convention - window size 1 implies deriv kernel of size 3 and no smoothing
if window_size == 1:
window_size, smoothing_size = 3, 1
else:
smoothing_size = None
imgs, _ = fn.readers.file(file_root=images_dir, shard_id=0, num_shards=1)
output_type = types.GRAY if grayscale else types.RGB
imgs = fn.decoders.image(imgs, device="cpu", output_type=output_type)
if in_type != types.UINT8:
imgs = fn.cast(imgs, dtype=in_type)
edges = fn.laplacian(imgs,
window_size=window_size,
smoothing_size=smoothing_size,
normalized_kernel=normalize,
dtype=out_type)
return edges, imgs
def ExternalSourcePipeline(params, num_threads, device_id, external_date,
seed):
pipe = Pipeline(params.batch_size, num_threads, device_id, seed=seed)
with pipe:
jpegs, labels = fn.external_source(source=external_date, num_outputs=2)
images = fn.image_decoder(jpegs, device="mixed", output_type=types.RGB)
images = fn.resize(images, resize_x=224, resize_y=224)
images = fn.cast(images, dtype=types.UINT8) / 255
images = fn.normalize(images,
axes=[0, 1],
mean=params.mean,
stddev=params.std,
device='gpu',
batch=False)
output = fn.transpose(images, perm=[2, 0, 1], device='gpu')
pipe.set_outputs(output, labels)
return pipe
def pipe_gaussian_noise(mean, stddev, variable_dist_params, device=None):
encoded, _ = fn.readers.file(file_root=images_dir)
in_data = fn.cast(fn.decoders.image(encoded,
device="cpu",
output_type=types.RGB),
dtype=types.FLOAT)
if device == 'gpu':
in_data = in_data.gpu()
mean_arg = mean
stddev_arg = stddev
if variable_dist_params:
mean_arg = fn.random.uniform(range=(-50.0, 50.0))
stddev_arg = fn.random.uniform(range=(1.0, 10.0))
seed = 12345
out_data1 = fn.noise.gaussian(in_data,
mean=mean_arg,
stddev=stddev_arg,
seed=seed)
out_data2 = in_data + fn.random.normal(
in_data, mean=mean_arg, stddev=stddev_arg, seed=seed)
return out_data1, out_data2
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 get_pipeline(batch_size,
num_threads,
device,
device_id=0,
shard_id=0,
num_shards=1):
test_data_root = os.environ['DALI_EXTRA_PATH']
file_root = os.path.join(test_data_root, 'db', 'coco_dummy', 'images')
annotations_file = os.path.join(test_data_root, 'db', 'coco_dummy',
'instances.json')
pipe = Pipeline(batch_size, num_threads, device_id)
with pipe:
jpegs, _, _, image_ids = fn.coco_reader(
file_root=file_root,
annotations_file=annotations_file,
shard_id=shard_id,
num_shards=num_shards,
ratio=False,
image_ids=True)
images = fn.image_decoder(
jpegs,
device=('mixed' if device == 'gpu' else 'cpu'),
output_type=types.RGB)
images = fn.resize(images,
resize_x=224,
resize_y=224,
interp_type=types.INTERP_LINEAR)
images = fn.crop_mirror_normalize(images,
dtype=types.FLOAT,
mean=[128., 128., 128.],
std=[1., 1., 1.])
if device == 'gpu':
image_ids = image_ids.gpu()
ids_reshaped = fn.reshape(image_ids, shape=[1, 1])
ids_int16 = fn.cast(image_ids, dtype=types.INT16)
pipe.set_outputs(images, ids_reshaped, ids_int16)
return pipe
def one_input_pipeline(def_for_dataset, device, source, external_source_device,
no_copy, batch):
"""Pipeline accepting single input via external source
Parameters
----------
def_for_dataset : bool
True if this pipeline will be converted to TF Dataset
device : str
device that the Dataset will be placed ("cpu" or "gpu")
source : callable
callback for the external source in baseline pipeline otherwise None
external_source_device : str
Device that we want the external source in TF dataset to be placed
"""
if def_for_dataset:
if no_copy is None:
# If no_copy is None, we infer it automatically and we use no_copy=True when
# the input memory is matching the external source placement,
# so the Dataset's placement is the same as external source's device,
# otherwise for cross-backend we use False.
no_copy = (device == external_source_device)
if batch == "dataset":
# Special value used in tests, reroute it to the default
batch = None
input = fn.external_source(name="input_placeholder",
no_copy=no_copy,
device=external_source_device,
batch=batch)
else:
input = fn.external_source(name="actual_input",
source=source,
batch=False,
device=external_source_device)
input = input if device == 'cpu' else input.gpu()
processed = fn.cast(input + 10, dtype=dali.types.INT32)
input_padded, processed_padded = fn.pad([input, processed])
return input_padded, processed_padded
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 cast_pipe():
inp = fn.external_source(src)
inp_dev = inp.gpu() if device == 'gpu' else inp
return inp, fn.cast(inp_dev, dtype=np_type_to_dali(out_dtype))
def build_pipes(device, dim, batch_size, channel_first, mode, interp, dtype,
w_input, h_input, d_input, use_size_arg, use_size_input,
use_roi):
dali_pipe = Pipeline(batch_size=batch_size,
num_threads=8,
device_id=0,
seed=1234)
with dali_pipe:
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_hwc = images_cpu if device == "cpu" else images_cpu.gpu()
if channel_first:
images = dali.fn.transpose(
images_hwc,
perm=[3, 0, 1, 2] if dim == 3 else [2, 0, 1],
transpose_layout=True)
else:
images = images_hwc
roi_start = None
roi_end = None
w = None
h = None
d = None
size = None
minibatch_size = 2 if dim == 3 else 8
if use_roi:
# Calculate absolute RoI
in_size = fn.slice(fn.shapes(images_cpu),
types.Constant(0,
dtype=types.FLOAT,
device="cpu"),
types.Constant(dim,
dtype=types.FLOAT,
device="cpu"),
axes=[0],
normalized_shape=False)
roi_start = fn.random.uniform(range=(0, 0.4), shape=[dim
]) * in_size
roi_end = fn.random.uniform(range=(0.6, 1.0), shape=[dim
]) * in_size
size_range = (10, 200) if dim == 3 else (10, 1000)
if use_size_arg:
if use_size_input:
mask = fn.cast(fn.random.uniform(range=(0.8, 1.9),
shape=[dim]),
dtype=types.INT32)
size = fn.random.uniform(range=size_range, shape=[dim]) * mask
else:
size = [300, 400] if dim == 2 else [80, 100, 120]
resized = resize_dali(images,
channel_first,
dtype,
interp,
mode,
size,
None,
None,
None,
roi_start,
roi_end,
minibatch_size=minibatch_size,
max_size=max_size(dim))
else:
if w_input:
has_w = fn.random.coin_flip(probability=0.8)
w = fn.random.uniform(range=size_range) * has_w
else:
w = 320 # some fixed value
if h_input:
has_h = fn.random.coin_flip(probability=0.8)
h = fn.random.uniform(range=size_range) * has_h
else:
h = 240 # some other fixed value
if dim >= 3:
if d_input:
has_d = fn.random.coin_flip(probability=0.8)
d = fn.random.uniform(range=size_range) * has_d
else:
d = 31 # some other fixed value
resized = resize_dali(images,
channel_first,
dtype,
interp,
mode,
None,
w,
h,
d,
roi_start,
roi_end,
minibatch_size=minibatch_size,
max_size=max_size(dim))
outputs = [images, resized]
if roi_start is not None and roi_end is not None:
outputs += [roi_start, roi_end]
for x in (d, h, w, size):
if x is not None:
if isinstance(x, _DataNode):
outputs.append(x)
else:
outputs.append(
types.Constant(np.array(x, dtype=np.float32)))
dali_pipe.set_outputs(*outputs)
pil_pipe = Pipeline(batch_size=batch_size,
num_threads=8,
device_id=0,
exec_async=False,
exec_pipelined=False)
with pil_pipe:
images = fn.external_source(name="images",
layout=layout_str(dim, channel_first))
sizes = fn.external_source(name="size")
roi_start = fn.external_source(name="roi_start")
roi_end = fn.external_source(name="roi_end")
resized = resize_PIL(dim, channel_first, dtype, interp, images, sizes,
roi_start, roi_end)
resized = fn.reshape(resized, layout=layout_str(dim, channel_first))
pil_pipe.set_outputs(resized)
dali_pipe.build()
pil_pipe.build()
return dali_pipe, pil_pipe
def random_augmentation(self, probability, augmented, original):
condition = fn.cast(fn.coin_flip(probability=probability), dtype=types.DALIDataType.BOOL)
neg_condition = condition ^ True
return condition * augmented + neg_condition * original
def dali_asr_pipeline(
train_pipeline, # True if training, False if validation
file_root,
file_list,
sample_rate,
silence_threshold,
resample_range,
discrete_resample_range,
window_size,
window_stride,
nfeatures,
nfft,
frame_splicing_factor,
dither_coeff,
pad_align,
preemph_coeff,
do_spectrogram_masking=False,
cutouts_generator=None,
shard_id=0,
n_shards=1,
preprocessing_device="gpu",
is_triton_pipeline=False,
):
do_remove_silence = silence_threshold is not None
def _div_ceil(dividend, divisor):
return (dividend + (divisor - 1)) // divisor
if is_triton_pipeline:
assert not train_pipeline, "Pipeline for Triton shall be a validation pipeline"
encoded = fn.external_source(device="cpu",
name="DALI_INPUT_0",
no_copy=True)
else:
encoded, label = fn.readers.file(
device="cpu",
name="file_reader",
file_root=file_root,
file_list=file_list,
shard_id=shard_id,
num_shards=n_shards,
shuffle_after_epoch=train_pipeline,
)
speed_perturbation_coeffs = None
if resample_range is not None:
if discrete_resample_range:
values = [resample_range[0], 1.0, resample_range[1]]
speed_perturbation_coeffs = fn.random.uniform(device="cpu",
values=values)
else:
speed_perturbation_coeffs = fn.random.uniform(device="cpu",
range=resample_range)
if train_pipeline and speed_perturbation_coeffs is not None:
dec_sample_rate_arg = speed_perturbation_coeffs * sample_rate
elif resample_range is None:
dec_sample_rate_arg = sample_rate
else:
dec_sample_rate_arg = None
audio, _ = fn.decoders.audio(encoded,
sample_rate=dec_sample_rate_arg,
dtype=types.FLOAT,
downmix=True)
if do_remove_silence:
begin, length = fn.nonsilent_region(audio, cutoff_db=silence_threshold)
audio = fn.slice(audio, begin, length, axes=[0])
# Max duration drop is performed at DataLayer stage
if preprocessing_device == "gpu":
audio = audio.gpu()
if dither_coeff != 0.0:
audio = audio + fn.random.normal(
device=preprocessing_device) * dither_coeff
audio = fn.preemphasis_filter(audio, preemph_coeff=preemph_coeff)
spec = fn.spectrogram(
audio,
nfft=nfft,
window_length=window_size * sample_rate,
window_step=window_stride * sample_rate,
)
mel_spec = fn.mel_filter_bank(spec,
sample_rate=sample_rate,
nfilter=nfeatures,
normalize=True)
log_features = fn.to_decibels(mel_spec,
multiplier=np.log(10),
reference=1.0,
cutoff_db=math.log(1e-20))
log_features_len = fn.shapes(log_features)
if frame_splicing_factor != 1:
log_features_len = _div_ceil(log_features_len, frame_splicing_factor)
log_features = fn.normalize(log_features, axes=[1])
log_features = fn.pad(log_features,
axes=[1],
fill_value=0,
align=pad_align,
shape=(-1, ))
if train_pipeline and do_spectrogram_masking:
anchors, shapes = fn.external_source(source=cutouts_generator,
num_outputs=2,
cycle=True)
log_features = fn.erase(
log_features,
anchor=anchors,
shape=shapes,
axes=[0, 1],
fill_value=0,
normalized_anchor=True,
)
# When modifying DALI pipeline returns, make sure you update `output_map`
# in DALIGenericIterator invocation
if not is_triton_pipeline:
return log_features.gpu(), label.gpu(), log_features_len.gpu()
else:
return fn.cast(log_features.gpu(), dtype=types.FLOAT16)
device='gpu',
bytes_per_sample_hint=ImageBytes,
size=(H, W))
images = fn.gaussian_blur(images,
device='gpu',
bytes_per_sample_hint=ImageBytes,
sigma=fn.uniform(range=(0.1, 2)),
window_size=11)
images = fn.color_twist(images,
device='gpu',
bytes_per_sample_hint=ImageBytes,
brightness=fn.uniform(range=(0.5, 1.5)),
contrast=fn.uniform(range=(0.5, 2.5)),
saturation=fn.uniform(range=(0.1, 2)))
images = fn.cast(images,
device='gpu',
bytes_per_sample_hint=ImageBytes,
dtype=DALIDataType.FLOAT)
images = fn.normalize(
images,
device='gpu',
bytes_per_sample_hint=ImageBytes,
mean=Constant(numpy.array([[[190.6380, 207.2640, 202.5720]]])),
stddev=Constant(numpy.array([[[85.2720, 68.6970, 81.4215]]])))
images = fn.transpose(images,
device='gpu',
bytes_per_sample_hint=ImageBytes,
perm=[2, 0, 1])
TestingPipe.set_outputs(images, labels)
TestingLoader = DALIClassificationIterator(TestingPipe,
size=1000 * args.bs)
def pipeline():
inp1 = fn.external_source(device='cpu', name='DALI_INPUT_0')
inp2 = fn.external_source(device='gpu', name='DALI_INPUT_1')
return inp1.gpu() / 3, fn.cast(inp2, dtype=dali.types.FLOAT) / 2
def _test_random_object_bbox_with_class(max_batch_size,
ndim,
dtype,
format=None,
fg_prob=None,
classes=None,
weights=None,
background=None,
threshold=None,
k_largest=None,
cache=None):
pipe = dali.Pipeline(max_batch_size, 4, device_id=None, seed=4321)
background_out = 0 if background is None else background
classes_out = np.int32([]) if classes is None else classes
weights_out = np.int32([]) if weights is None else weights
threshold_out = np.int32([]) if threshold is None else threshold
if cache:
source = sampled_dataset(2 * max_batch_size, max_batch_size, ndim,
dtype)
else:
source = batch_generator(max_batch_size, ndim, dtype)
with pipe:
inp = fn.external_source(source)
if isinstance(background,
dali.pipeline.DataNode) or (background is not None
and background >= 0):
inp = fn.cast(inp + (background_out + 1),
dtype=np_type_to_dali(dtype))
# preconfigure
op = ops.segmentation.RandomObjectBBox(format=format,
foreground_prob=fg_prob,
classes=classes,
class_weights=weights,
background=background,
threshold=threshold,
k_largest=k_largest,
seed=1234)
outs1 = op(inp, cache_objects=cache)
outs2 = op(inp, output_class=True)
if not isinstance(outs1, list):
outs1 = [outs1]
# the second instance should have always at least 2 outputs
assert isinstance(outs2, (list, tuple))
outputs = [
inp, classes_out, weights_out, background_out, threshold_out,
*outs1, *outs2
]
pipe.set_outputs(*outputs)
pipe.build()
format = format or "anchor_shape"
for _ in range(50):
inp, classes_out, weights_out, background_out, threshold_out, *outs = pipe.run(
)
nout = (len(outs) - 1) // 2
outs1 = outs[:nout]
outs2 = outs[nout:]
for i in range(len(outs1)):
check_batch(outs1[i], outs2[i])
# Iterate over indices instead of elements, because normal iteration
# causes an exception to be thrown in native code, making debugging near impossible.
outs = tuple([np.array(out[i]) for i in range(len(out))]
for out in outs1)
box_class_labels = [
np.int32(outs2[-1][i]) for i in range(len(outs2[-1]))
]
boxes = convert_boxes(outs, format)
for i in range(len(inp)):
in_tensor = inp.at(i)
class_labels = classes_out.at(i)
if background is not None or classes is None:
background_label = background_out.at(i)
else:
background_label = 0 if 0 not in class_labels else np.min(
class_labels) - 1
label = box_class_labels[i]
if classes is not None:
assert label == background_label or label in list(class_labels)
is_foreground = label != background_label
cls_boxes = class_boxes(in_tensor,
label if is_foreground else None)
if is_foreground:
ref_boxes = cls_boxes
if threshold is not None:
extent = box_extent(boxes[i])
thr = threshold_out.at(i)
assert np.all(extent >= thr)
ref_boxes = list(
filter(lambda box: np.all(box_extent(box) >= thr),
cls_boxes))
if k_largest is not None:
assert box_in_k_largest(ref_boxes, boxes[i], k_largest)
assert contains_box(cls_boxes, boxes[i])