Skip to content

Commit

Permalink
Add DALI proxy option to EfficientNet example (#5791)
Browse files Browse the repository at this point in the history 
Add DALI proxy option to EfficientNet example
Add performance check of DALI proxy variant to efficient net performance test

Signed-off-by: Joaquin Anton Guirao <[email protected]>
  • Loading branch information
@jantonguirao
jantonguirao authored Jan 27, 2025
1 parent 6f0e6b5 commit f625871
Show file tree
Hide file tree
Showing 5 changed files with 433 additions and 81 deletions.
2 changes: 1 addition & 1 deletion docs/examples/use_cases/pytorch/efficientnet/image_classification/autoaugment.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -96,7 +96,7 @@ def solarize_add(image, addition=0, threshold=128):
for i in range(256):
if i < threshold:
res = i + addition if i + addition <= 255 else 255
res = res if res >= 0 else 0
res = int(res if res >= 0 else 0)
lut.append(res)
else:
lut.append(i)
Expand Down
191 changes: 151 additions & 40 deletions docs/examples/use_cases/pytorch/efficientnet/image_classification/dali.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,31 +20,37 @@
from nvidia.dali.auto_aug import auto_augment, trivial_augment


@pipeline_def(enable_conditionals=True)
def training_pipe(data_dir, interpolation, image_size, output_layout, automatic_augmentation,
dali_device="gpu", rank=0, world_size=1):
rng = fn.random.coin_flip(probability=0.5)

jpegs, labels = fn.readers.file(name="Reader", file_root=data_dir, shard_id=rank,
num_shards=world_size, random_shuffle=True, pad_last_batch=True)

if dali_device == "gpu":
decoder_device = "mixed"
resize_device = "gpu"
else:
decoder_device = "cpu"
resize_device = "cpu"

images = fn.decoders.image_random_crop(jpegs, device=decoder_device, output_type=types.RGB,
random_aspect_ratio=[0.75, 4.0 / 3.0],
random_area=[0.08, 1.0])

images = fn.resize(images, device=resize_device, size=[image_size, image_size],
interp_type=interpolation, antialias=False)
def efficientnet_processing_training(
jpegs_input,
interpolation,
image_size,
output_layout,
automatic_augmentation,
dali_device="gpu",
):
"""
Image processing part of the ResNet training pipeline (excluding data loading)
"""
decoder_device = "mixed" if dali_device == "gpu" else "cpu"
images = fn.decoders.image_random_crop(
jpegs_input,
device=decoder_device,
output_type=types.RGB,
random_aspect_ratio=[0.75, 4.0 / 3.0],
random_area=[0.08, 1.0],
)

images = fn.resize(
images,
size=[image_size, image_size],
interp_type=interpolation,
antialias=False,
)

# Make sure that from this point we are processing on GPU regardless of dali_device parameter
images = images.gpu()

rng = fn.random.coin_flip(probability=0.5)
images = fn.flip(images, horizontal=rng)

# Based on the specification, apply the automatic augmentation policy. Note, that from the point
Expand All @@ -53,33 +59,138 @@ def training_pipe(data_dir, interpolation, image_size, output_layout, automatic_
# We pass the shape of the image after the resize so the translate operations are done
# relative to the image size.
if automatic_augmentation == "autoaugment":
output = auto_augment.auto_augment_image_net(images, shape=[image_size, image_size])
output = auto_augment.auto_augment_image_net(
images, shape=[image_size, image_size]
)
elif automatic_augmentation == "trivialaugment":
output = trivial_augment.trivial_augment_wide(images, shape=[image_size, image_size])
output = trivial_augment.trivial_augment_wide(
images, shape=[image_size, image_size]
)
else:
output = images

output = fn.crop_mirror_normalize(output, dtype=types.FLOAT, output_layout=output_layout,
crop=(image_size, image_size),
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
std=[0.229 * 255, 0.224 * 255, 0.225 * 255])
output = fn.crop_mirror_normalize(
output,
dtype=types.FLOAT,
output_layout=output_layout,
crop=(image_size, image_size),
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
std=[0.229 * 255, 0.224 * 255, 0.225 * 255],
)

return output, labels
return output


@pipeline_def
def validation_pipe(data_dir, interpolation, image_size, image_crop, output_layout, rank=0,
world_size=1):
jpegs, label = fn.readers.file(name="Reader", file_root=data_dir, shard_id=rank,
num_shards=world_size, random_shuffle=False, pad_last_batch=True)
@pipeline_def(enable_conditionals=True)
def training_pipe(
data_dir,
interpolation,
image_size,
output_layout,
automatic_augmentation,
dali_device="gpu",
rank=0,
world_size=1,
):
jpegs, labels = fn.readers.file(
name="Reader",
file_root=data_dir,
shard_id=rank,
num_shards=world_size,
random_shuffle=True,
pad_last_batch=True,
)
outputs = efficientnet_processing_training(
jpegs,
interpolation,
image_size,
output_layout,
automatic_augmentation,
dali_device,
)
return outputs, labels


@pipeline_def(enable_conditionals=True)
def training_pipe_external_source(
interpolation,
image_size,
output_layout,
automatic_augmentation,
dali_device="gpu",
rank=0,
world_size=1,
):
filepaths = fn.external_source(name="images", no_copy=True)
jpegs = fn.io.file.read(filepaths)
outputs = efficientnet_processing_training(
jpegs,
interpolation,
image_size,
output_layout,
automatic_augmentation,
dali_device,
)
return outputs


def efficientnet_processing_validation(
jpegs, interpolation, image_size, image_crop, output_layout
):
"""
Image processing part of the ResNet validation pipeline (excluding data loading)
"""
images = fn.decoders.image(jpegs, device="mixed", output_type=types.RGB)

images = fn.resize(images, resize_shorter=image_size, interp_type=interpolation,
antialias=False)
images = fn.resize(
images,
resize_shorter=image_size,
interp_type=interpolation,
antialias=False,
)

output = fn.crop_mirror_normalize(
images,
dtype=types.FLOAT,
output_layout=output_layout,
crop=(image_crop, image_crop),
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
std=[0.229 * 255, 0.224 * 255, 0.225 * 255],
)
return output


output = fn.crop_mirror_normalize(images, dtype=types.FLOAT, output_layout=output_layout,
crop=(image_crop, image_crop),
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
std=[0.229 * 255, 0.224 * 255, 0.225 * 255])
return output, label
@pipeline_def
def validation_pipe(
data_dir,
interpolation,
image_size,
image_crop,
output_layout,
rank=0,
world_size=1,
):
jpegs, label = fn.readers.file(
name="Reader",
file_root=data_dir,
shard_id=rank,
num_shards=world_size,
random_shuffle=False,
pad_last_batch=True,
)
outputs = efficientnet_processing_validation(
jpegs, interpolation, image_size, image_crop, output_layout
)
return outputs, label


@pipeline_def
def validation_pipe_external_source(
interpolation, image_size, image_crop, output_layout
):
filepaths = fn.external_source(name="images", no_copy=True)
jpegs = fn.io.file.read(filepaths)
outputs = efficientnet_processing_validation(
jpegs, interpolation, image_size, image_crop, output_layout
)
return outputs
Loading

0 comments on commit f625871

Please sign in to comment.