Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add AutoAugment and ImageNet policy #4702

Merged
merged 7 commits into from
Mar 13, 2023
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 9 additions & 4 deletions dali/python/nvidia/dali/auto_aug/augmentations.py
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,12 @@
# See the License for the specific language governing permissions and
# limitations under the License.

import numpy as np
try:
import numpy as np
except ImportError:
raise RuntimeError(
"Could not import numpy. DALI's automatic augmentation examples depend on numpy. "
"Please install numpy to use the examples.")
Comment on lines +15 to +20
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This appears in multiple places throughout our codebase. Wouldn't it be better to just add a dependency to the DALI wheel?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I kind of followed a pattern here used elsewhere in DALI, but that makes a lot of sense. If we decide to change it, we can remove all such import wrapper at once.


from nvidia.dali import fn
from nvidia.dali import types
Expand All @@ -23,7 +28,7 @@

The `@augmentation` decorator handles computation of the decorated transformations's parameter.
When called, the decorated augmentation expects:
* a single positional argument: batch o samples
* a single positional argument: batch of samples
* `magnitude_bin` and `num_magnitude_bins` instead of the parameter.
The parameter is computed as if by calling
`as_param(magnitudes[magnitude_bin] * ((-1) ** random_sign))`, where
Expand Down Expand Up @@ -60,7 +65,7 @@ def shear_y(sample, shear, fill_value=128, interp_type=None):

@augmentation(mag_range=(0., 1.), randomly_negate=True, as_param=warp_x_param)
def translate_x(sample, rel_offset, shape, fill_value=128, interp_type=None):
offset = rel_offset * shape[-2]
offset = rel_offset * shape[1]
mt = fn.transforms.translation(offset=offset)
return fn.warp_affine(sample, matrix=mt, fill_value=fill_value, interp_type=interp_type,
inverse_map=False)
Expand All @@ -75,7 +80,7 @@ def translate_x_no_shape(sample, offset, fill_value=128, interp_type=None):

@augmentation(mag_range=(0., 1.), randomly_negate=True, as_param=warp_y_param)
def translate_y(sample, rel_offset, shape, fill_value=128, interp_type=None):
offset = rel_offset * shape[-3]
offset = rel_offset * shape[0]
mt = fn.transforms.translation(offset=offset)
return fn.warp_affine(sample, matrix=mt, fill_value=fill_value, interp_type=interp_type,
inverse_map=False)
Expand Down
237 changes: 237 additions & 0 deletions dali/python/nvidia/dali/auto_aug/auto_augment.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,237 @@
# Copyright (c) 2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from typing import List, Optional, Tuple

from nvidia.dali import fn
from nvidia.dali import types
from nvidia.dali.auto_aug import augmentations as a
from nvidia.dali.auto_aug.core import _Augmentation, Policy, signed_bin
from nvidia.dali.auto_aug.core._args import forbid_unused_kwargs as _forbid_unused_kwargs
from nvidia.dali.auto_aug.core._utils import \
parse_validate_offset as _parse_validate_offset, \
pretty_select as _pretty_select
from nvidia.dali.data_node import DataNode as _DataNode

try:
import numpy as np
except ImportError:
raise RuntimeError(
"Could not import numpy. DALI's automatic augmentation examples depend on numpy. "
"Please install numpy to use the examples.")


def auto_augment_image_net(sample: _DataNode, shape: Optional[_DataNode] = None,
fill_value: Optional[int] = 128,
interp_type: Optional[types.DALIInterpType] = None,
max_translate_abs: Optional[int] = None,
max_translate_rel: Optional[float] = None, seed: Optional[int] = None):
"""
Applies `auto_augment_image_net_policy` in AutoAugment (https://arxiv.org/abs/1805.09501)
fashion to the provided batch of samples.

Parameter
---------
sample : DataNode
A batch of samples to be processed. The samples should be images of `HWC` layout,
`uint8` type and reside on GPU.
shapes: DataNode, optional
A batch of shapes of the `sample`. If specified, the magnitude of `translation`
operations depends on the image shape and spans from 0 to `max_translate_rel * shape`.
Otherwise, the magnitude range is `[0, max_translate_abs]` for any sample.
fill_value: int, optional
A value to be used as a padding for images transformed with warp_affine ops
(translation, shear and rotate). If `None` is specified, the images are padded
with the border value repeated (clamped).
interp_type: types.DALIInterpType, optional
Interpolation method used by the warp_affine ops (translation, shear and rotate).
Supported values are `types.INTERP_LINEAR` (default) and `types.INTERP_NN`.
seed: int, optional
Seed to be used to randomly sample operations (and to negate magnitudes).

Returns
-------
DataNode
A batch of transformed samples.
"""
aug_kwargs = {"fill_value": fill_value, "interp_type": interp_type}
use_shape = shape is not None
if use_shape:
aug_kwargs["shape"] = shape
image_net_policy = get_image_net_policy(use_shape=use_shape,
max_translate_abs=max_translate_abs,
max_translate_rel=max_translate_rel)
return apply_auto_augment(image_net_policy, sample, seed, **aug_kwargs)


def apply_auto_augment(policy: Policy, sample: _DataNode, seed: Optional[int] = None,
**kwargs) -> _DataNode:
"""
Applies AutoAugment (https://arxiv.org/abs/1805.09501) augmentation scheme to the
provided batch of samples.

Parameter
---------
policy: Policy
Set of sequences of augmentations to be applied in AutoAugment fashion.
sample : DataNode
A batch of samples to be processed.
seed: int, optional
Seed to be used to randomly sample operations (and to negate magnitudes).
kwargs:
A dictionary of extra parameters to be passed when calling augmentations.
The signature of each augmentation is checked for any extra arguments and if
the name of the argument matches one from the `kwargs`, the value is
passed as an argument. For example, some augmentations from the default
random augment suite accept `shapes`, `fill_value` and `interp_type`.

Returns
-------
DataNode
A batch of transformed samples.
"""
if len(policy.sub_policies) == 0:
raise Exception(f"Cannot run empty policy. Got {policy} in `apply_auto_augment` call.")
max_policy_len = max(len(sub_policy) for sub_policy in policy.sub_policies)
should_run = fn.random.uniform(range=[0, 1], shape=(max_policy_len, ), dtype=types.FLOAT)
sub_policy_id = fn.random.uniform(values=list(range(len(policy.sub_policies))), seed=seed,
dtype=types.INT32)
run_probabilities = _sub_policy_to_probability_map(policy)[sub_policy_id]
magnitude_bins = _sub_policy_to_magnitude_bin_map(policy)[sub_policy_id]
aug_ids, augmentations = _sub_policy_to_augmentation_map(policy)
aug_ids = aug_ids[sub_policy_id]
use_signed_magnitudes = any(aug.randomly_negate for aug in policy.augmentations.values())
_forbid_unused_kwargs(augmentations, kwargs, 'apply_auto_augment')
for stage_id in range(max_policy_len):
magnitude_bin = magnitude_bins[stage_id]
if use_signed_magnitudes:
magnitude_bin = signed_bin(magnitude_bin)
if should_run[stage_id] < run_probabilities[stage_id]:
op_kwargs = dict(sample=sample, magnitude_bin=magnitude_bin,
num_magnitude_bins=policy.num_magnitude_bins, **kwargs)
sample = _pretty_select(augmentations, aug_ids[stage_id], op_kwargs,
auto_aug_name='apply_auto_augment',
ref_suite_name='get_image_net_policy')
return sample


def get_image_net_policy(use_shape: bool = False, max_translate_abs: int = None,
max_translate_rel: float = None) -> Policy:
"""
Creates augmentation policy tuned for the ImageNet as described in AutoAugment
(https://arxiv.org/abs/1805.09501).
The returned policy can be run with `apply_auto_augment`.

Parameter
---------
use_shape : bool
If true, the translation offset is computed as a percentage of the image. Useful if the
images processed with the auto augment have different shapes. If false, the offsets range
is bounded by a constant (`max_translate_abs`).
max_translate_abs: int or (int, int), optional
Only valid with use_shape=False, specifies the maximal shift (in pixels) in the translation
augmentations. If tuple is specified, the first component limits height, the second the
width.
max_translate_rel: float or (float, float), optional
Only valid with use_shape=True, specifies the maximal shift as a fraction of image shape
in the translation augmentations. If tuple is specified, the first component limits
height, the second the width.
"""
translate_y = _get_translate_y(use_shape, max_translate_abs, max_translate_rel)
shear_x = a.shear_x.augmentation((0, 0.3), True)
shear_y = a.shear_y.augmentation((0, 0.3), True)
rotate = a.rotate.augmentation((0, 30), True)
color = a.color.augmentation((0.1, 1.9), False, None)
posterize = a.posterize.augmentation((0, 4), False, a.poster_mask_uint8)
solarize = a.solarize.augmentation((0, 256), False)
solarize_add = a.solarize_add.augmentation((0, 110), False)
invert = a.invert
equalize = a.equalize
auto_contrast = a.auto_contrast
return Policy(
name="ImageNetPolicy", num_magnitude_bins=11, sub_policies=[
[(equalize, 0.8, 1), (shear_y, 0.8, 4)],
[(color, 0.4, 9), (equalize, 0.6, 3)],
[(color, 0.4, 1), (rotate, 0.6, 8)],
[(solarize, 0.8, 3), (equalize, 0.4, 7)],
[(solarize, 0.4, 2), (solarize, 0.6, 2)],
[(color, 0.2, 0), (equalize, 0.8, 8)],
[(equalize, 0.4, 8), (solarize_add, 0.8, 3)],
[(shear_x, 0.2, 9), (rotate, 0.6, 8)],
[(color, 0.6, 1), (equalize, 1.0, 2)],
[(invert, 0.4, 9), (rotate, 0.6, 0)],
[(equalize, 1.0, 9), (shear_y, 0.6, 3)],
[(color, 0.4, 7), (equalize, 0.6, 0)],
[(posterize, 0.4, 6), (auto_contrast, 0.4, 7)],
[(solarize, 0.6, 8), (color, 0.6, 9)],
[(solarize, 0.2, 4), (rotate, 0.8, 9)],
[(rotate, 1.0, 7), (translate_y, 0.8, 9)],
[(shear_x, 0.0, 0), (solarize, 0.8, 4)],
[(shear_y, 0.8, 0), (color, 0.6, 4)],
[(color, 1.0, 0), (rotate, 0.6, 2)],
[(equalize, 0.8, 4)],
[(equalize, 1.0, 4), (auto_contrast, 0.6, 2)],
[(shear_y, 0.4, 7), (solarize_add, 0.6, 7)],
[(posterize, 0.8, 2), (solarize, 0.6, 10)],
[(solarize, 0.6, 8), (equalize, 0.6, 1)],
[(color, 0.8, 6), (rotate, 0.4, 5)],
])


def _get_translate_y(use_shape: bool = False, max_translate_abs: int = None,
max_translate_rel: float = None):
max_translate_height, _ = _parse_validate_offset(use_shape, max_translate_abs=max_translate_abs,
max_translate_rel=max_translate_rel,
default_translate_abs=250,
default_translate_rel=1.)
if use_shape:
return a.translate_y.augmentation((0, max_translate_height), True)
else:
return a.translate_y_no_shape.augmentation((0, max_translate_height), True)


def _sub_policy_to_probability_map(policy: Policy) -> _DataNode:
sub_policies = policy.sub_policies
max_policy_len = max(len(sub_policy) for sub_policy in sub_policies)
prob = np.array([[0. for _ in range(max_policy_len)] for _ in range(len(sub_policies))],
dtype=np.float32)
for sub_policy_id, sub_policy in enumerate(sub_policies):
for stage_idx, (aug_name, p, mag) in enumerate(sub_policy):
prob[sub_policy_id, stage_idx] = p
return types.Constant(prob)


def _sub_policy_to_magnitude_bin_map(policy: Policy) -> _DataNode:
sub_policies = policy.sub_policies
max_policy_len = max(len(sub_policy) for sub_policy in sub_policies)
magnitude_bin = np.array([[0 for _ in range(max_policy_len)] for _ in range(len(sub_policies))],
dtype=np.int32)
for sub_policy_id, sub_policy in enumerate(sub_policies):
for stage_idx, (aug_name, p, mag) in enumerate(sub_policy):
magnitude_bin[sub_policy_id, stage_idx] = mag
return types.Constant(magnitude_bin)


def _sub_policy_to_augmentation_map(policy: Policy) -> Tuple[_DataNode, List[_Augmentation]]:
sub_policies = policy.sub_policies
max_policy_len = max(len(sub_policy) for sub_policy in sub_policies)
augmentations = list(policy.augmentations.values()) + [a.identity]
identity_id = len(augmentations) - 1
augment_to_id = {augmentation: i for i, augmentation in enumerate(augmentations)}
augments_by_id = np.array([[identity_id for _ in range(max_policy_len)]
for _ in range(len(sub_policies))], dtype=np.int32)
for sub_policy_id, sub_policy in enumerate(sub_policies):
for stage_idx, (augment, p, mag) in enumerate(sub_policy):
augments_by_id[sub_policy_id, stage_idx] = augment_to_id[augment]
return types.Constant(augments_by_id), augmentations
3 changes: 2 additions & 1 deletion dali/python/nvidia/dali/auto_aug/core/__init__.py
Original file line number Diff line number Diff line change
Expand Up @@ -15,5 +15,6 @@
from nvidia.dali.auto_aug.core._augmentation import signed_bin, Augmentation as _Augmentation
from nvidia.dali.auto_aug.core.decorator import augmentation
from nvidia.dali.auto_aug.core._select import select
from nvidia.dali.auto_aug.core.policy import Policy

__all__ = ("signed_bin", "augmentation", "select", "_Augmentation")
__all__ = ("signed_bin", "augmentation", "select", "Policy", "_Augmentation")
14 changes: 7 additions & 7 deletions dali/python/nvidia/dali/auto_aug/core/_augmentation.py
Original file line number Diff line number Diff line change
Expand Up @@ -76,8 +76,8 @@ def signed_magnitude_idx(self):
return self._signed_magnitude_idx


def signed_bin(magnitude_bin: Optional[Union[int, _DataNode]],
random_sign: Optional[_DataNode] = None, seed=None) -> _SignedMagnitudeBin:
def signed_bin(magnitude_bin: Union[int, _DataNode], random_sign: Optional[_DataNode] = None,
seed=None) -> _SignedMagnitudeBin:
"""
Combines the `magnitude_bin` with information about the sign of the magnitude.
The Augmentation wrapper can generate and handle the random sign on its own. Yet,
Expand All @@ -90,8 +90,8 @@ def signed_bin(magnitude_bin: Optional[Union[int, _DataNode]],
magnitude_bin: int or DataNode
The magnitude bin from range `[0, num_magnitude_bins - 1]`. Can be plain int or
a batch (_DataNode) of ints.
random_sign : DataNode
A batch of {0, 1} integers. For augmentations declared with `random_negate=True`,
random_sign : DataNode, optional
A batch of {0, 1} integers. For augmentations declared with `randomly_negate=True`,
it determines if the magnitude is negated (for 1) or not (for 0).
"""
return _SignedMagnitudeBin(magnitude_bin, random_sign, seed)
Expand Down Expand Up @@ -295,9 +295,9 @@ def _get_param(self, magnitude_bin, num_magnitude_bins):
f"but unsigned `magnitude_bin` was passed to the augmentation call. "
f"The augmentation will randomly negate the magnitudes manually. "
f"However, for better performance, if you conditionally split batch "
f"between multiple augmentations, please call "
f"`signed_magnitude_bin = signed_bin(magnitude_bin)` and pass the "
f"signed bins instead.", Warning)
f"between multiple augmentations, it is better to call "
f"`signed_magnitude_bin = signed_bin(magnitude_bin)` before the split "
f"and pass the signed bins instead.", Warning)
if self.randomly_negate:
assert isinstance(magnitude_bin, _SignedMagnitudeBin) # by the two checks above
if isinstance(magnitude_bin.bin, int):
Expand Down
71 changes: 71 additions & 0 deletions dali/python/nvidia/dali/auto_aug/core/_utils.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,71 @@
# Copyright (c) 2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from typing import List

from nvidia.dali.data_node import DataNode as _DataNode

from nvidia.dali.auto_aug.core._select import select
from nvidia.dali.auto_aug.core._args import MissingArgException
from nvidia.dali.auto_aug.core._augmentation import Augmentation
import nvidia.dali.auto_aug.augmentations as a


def max_translate_hw(max_translate):
if isinstance(max_translate, (tuple, list)):
height, width = max_translate
return height, width
return max_translate, max_translate


def parse_validate_offset(use_shape, max_translate_abs=None, max_translate_rel=None,
default_translate_abs=250, default_translate_rel=1.):
# if one passes DataNode (with shapes for instance), the error message would be very vague
if not isinstance(use_shape, bool):
raise Exception(f"The `use_shape` is a flag that should be set to either True or False, "
f"got {use_shape}.")
if use_shape:
if max_translate_abs is not None:
raise Exception("The argument `max_translate_abs` cannot be used with image shapes. "
"You may use `max_translate_rel` instead.")
if max_translate_rel is None:
max_translate_rel = default_translate_rel
return max_translate_hw(max_translate_rel)
else:
if max_translate_rel is not None:
raise Exception("The argument `max_translate_rel` cannot be used without image shapes. "
"You may use `max_translate_abs` instead.")
if max_translate_abs is None:
max_translate_abs = default_translate_abs
return max_translate_hw(max_translate_abs)


def pretty_select(augmentations: List[Augmentation], aug_ids: _DataNode, op_kwargs,
auto_aug_name: str, ref_suite_name: str):
try:
return select(augmentations, aug_ids, **op_kwargs)
except MissingArgException as e:
if e.missing_args != ['shape'] or e.augmentation.op not in [
a.translate_x.op, a.translate_y.op
]:
raise
else:
raise Exception(
f"The augmentation `{e.augmentation.name}` requires `shape` argument that "
f"describes image shape (in HWC layout). Please provide it as `shape` argument "
f"to `{auto_aug_name}` call. You can get the image shape from encoded "
f"images with `fn.peek_image_shape`. Alternatively, you can use "
f"`translate_x_no_shape`/`translate_y_no_shape` that does not rely on image "
f"shape, but uses offset from fixed range: for reference see `{ref_suite_name}` "
f"and its `use_shape` argument. ")
Loading