################################################################################
# Copyright (c) 2021 ContinualAI. #
# Copyrights licensed under the MIT License. #
# See the accompanying LICENSE file for terms. #
# #
# Date: 13-02-2021 #
# Author(s): Jary Pomponi #
################################################################################
from pathlib import Path
from typing import Optional, Sequence, Any, Union
import torch
from torch import Tensor
from torchvision.transforms import (
ToTensor,
Compose,
Normalize,
ToPILImage,
RandomRotation,
)
from PIL.Image import Image
from avalanche.benchmarks import nc_benchmark, NCScenario
from avalanche.benchmarks.classic.classic_benchmarks_utils import (
check_vision_benchmark,
)
from avalanche.benchmarks.datasets import default_dataset_location
from avalanche.benchmarks.datasets.omniglot import Omniglot
from avalanche.benchmarks.utils import AvalancheDataset
import numpy as np
_default_omniglot_train_transform = Compose(
[ToTensor(), Normalize((0.9221,), (0.2681,))]
)
_default_omniglot_eval_transform = Compose(
[ToTensor(), Normalize((0.9221,), (0.2681,))]
)
class PixelsPermutation(object):
"""
Apply a fixed permutation to the pixels of the given image.
Works with both Tensors and PIL images. Returns an object of the same type
of the input element.
"""
def __init__(self, index_permutation: Sequence[int]):
self.permutation = index_permutation
self._to_tensor = ToTensor()
self._to_image = ToPILImage()
def __call__(self, img: Union[Image, Tensor]):
is_image = isinstance(img, Image)
if (not is_image) and (not isinstance(img, Tensor)):
raise ValueError("Invalid input: must be a PIL image or a Tensor")
if is_image:
img = self._to_tensor(img)
img = img.view(-1)[self.permutation].view(*img.shape)
if is_image:
img = self._to_image(img)
return img
[docs]def SplitOmniglot(
n_experiences: int,
*,
return_task_id=False,
seed: Optional[int] = None,
fixed_class_order: Optional[Sequence[int]] = None,
shuffle: bool = True,
train_transform: Optional[Any] = _default_omniglot_train_transform,
eval_transform: Optional[Any] = _default_omniglot_eval_transform,
dataset_root: Union[str, Path] = None
):
"""
Creates a CL benchmark using the OMNIGLOT dataset.
If the dataset is not present in the computer, this method will
automatically download and store it.
The returned benchmark will return experiences containing all patterns of a
subset of classes, which means that each class is only seen "once".
This is one of the most common scenarios in the Continual Learning
literature. Common names used in literature to describe this kind of
scenario are "Class Incremental", "New Classes", etc.
By default, an equal amount of classes will be assigned to each experience.
OMNIGLOT consists of 964 classes, which means that the number of
experiences can be 1, 2, 4, 241, 482, 964.
This generator doesn't force a choice on the availability of task labels,
a choice that is left to the user (see the `return_task_id` parameter for
more info on task labels).
The benchmark instance returned by this method will have two fields,
`train_stream` and `test_stream`, which can be iterated to obtain
training and test :class:`Experience`. Each Experience contains the
`dataset` and the associated task label.
The benchmark API is quite simple and is uniform across all benchmark
generators. It is recommended to check the tutorial of the "benchmark" API,
which contains usage examples ranging from "basic" to "advanced".
:param n_experiences: The number of incremental experiences in the current
benchmark. The value of this parameter should be a divisor of 10.
:param return_task_id: if True, a progressive task id is returned for every
experience. If False, all experiences will have a task ID of 0.
:param seed: A valid int used to initialize the random number generator.
Can be None.
:param fixed_class_order: A list of class IDs used to define the class
order. If None, value of ``seed`` will be used to define the class
order. If non-None, ``seed`` parameter will be ignored.
Defaults to None.
:param shuffle: If true, the class order in the incremental experiences is
randomly shuffled. Default to True.
:param train_transform: The transformation to apply to the training data,
e.g. a random crop, a normalization or a concatenation of different
transformations (see torchvision.transform documentation for a
comprehensive list of possible transformations).
If no transformation is passed, the default train transformation
will be used.
:param eval_transform: The transformation to apply to the test data,
e.g. a random crop, a normalization or a concatenation of different
transformations (see torchvision.transform documentation for a
comprehensive list of possible transformations).
If no transformation is passed, the default test transformation
will be used.
:param dataset_root: The root path of the dataset. Defaults to None, which
means that the default location for 'omniglot' will be used.
:returns: A properly initialized :class:`NCScenario` instance.
"""
omniglot_train, omniglot_test = _get_omniglot_dataset(dataset_root)
if return_task_id:
return nc_benchmark(
train_dataset=omniglot_train,
test_dataset=omniglot_test,
n_experiences=n_experiences,
task_labels=True,
seed=seed,
fixed_class_order=fixed_class_order,
shuffle=shuffle,
class_ids_from_zero_in_each_exp=True,
train_transform=train_transform,
eval_transform=eval_transform,
)
else:
return nc_benchmark(
train_dataset=omniglot_train,
test_dataset=omniglot_test,
n_experiences=n_experiences,
task_labels=False,
seed=seed,
fixed_class_order=fixed_class_order,
shuffle=shuffle,
train_transform=train_transform,
eval_transform=eval_transform,
)
[docs]def PermutedOmniglot(
n_experiences: int,
*,
seed: Optional[int] = None,
train_transform: Optional[Any] = _default_omniglot_train_transform,
eval_transform: Optional[Any] = _default_omniglot_eval_transform,
dataset_root: Union[str, Path] = None
) -> NCScenario:
"""
Creates a Permuted Omniglot benchmark.
If the dataset is not present in the computer, this method will
automatically download and store it.
Random pixel permutations are used to permute the Omniglot images in
``n_experiences`` different manners. This means that each experience is
composed of all the original 964 Omniglot classes, but the pixel in the
images are permuted in a different way.
The benchmark instance returned by this method will have two fields,
`train_stream` and `test_stream`, which can be iterated to obtain
training and test :class:`Experience`. Each Experience contains the
`dataset` and the associated task label.
A progressive task label, starting from "0", is applied to each experience.
The benchmark API is quite simple and is uniform across all benchmark
generators. It is recommended to check the tutorial of the "benchmark" API,
which contains usage examples ranging from "basic" to "advanced".
:param n_experiences: The number of experiences (tasks) in the current
benchmark. It indicates how many different permutations of the Omniglot
dataset have to be created.
:param seed: A valid int used to initialize the random number generator.
Can be None.
:param train_transform: The transformation to apply to the training data
before the random permutation, e.g. a random crop, a normalization or a
concatenation of different transformations (see torchvision.transform
documentation for a comprehensive list of possible transformations).
If no transformation is passed, the default train transformation
will be used.
:param eval_transform: The transformation to apply to the test data
before the random permutation, e.g. a random crop, a normalization or a
concatenation of different transformations (see torchvision.transform
documentation for a comprehensive list of possible transformations).
If no transformation is passed, the default test transformation
will be used.
:param dataset_root: The root path of the dataset. Defaults to None, which
means that the default location for 'omniglot' will be used.
:returns: A properly initialized :class:`NCScenario` instance.
"""
list_train_dataset = []
list_test_dataset = []
rng_permute = np.random.RandomState(seed)
omniglot_train, omniglot_test = _get_omniglot_dataset(dataset_root)
# for every incremental experience
for _ in range(n_experiences):
# choose a random permutation of the pixels in the image
idx_permute = torch.from_numpy(rng_permute.permutation(11025)).type(
torch.int64
)
permutation = PixelsPermutation(idx_permute)
permutation_transforms = dict(
train=(permutation, None), eval=(permutation, None)
)
# Freeze the permutation
permuted_train = AvalancheDataset(
omniglot_train,
transform_groups=permutation_transforms,
initial_transform_group="train",
).freeze_transforms()
permuted_test = AvalancheDataset(
omniglot_test,
transform_groups=permutation_transforms,
initial_transform_group="eval",
).freeze_transforms()
list_train_dataset.append(permuted_train)
list_test_dataset.append(permuted_test)
return nc_benchmark(
list_train_dataset,
list_test_dataset,
n_experiences=len(list_train_dataset),
task_labels=True,
shuffle=False,
class_ids_from_zero_in_each_exp=True,
one_dataset_per_exp=True,
train_transform=train_transform,
eval_transform=eval_transform,
)
[docs]def RotatedOmniglot(
n_experiences: int,
*,
seed: Optional[int] = None,
rotations_list: Optional[Sequence[int]] = None,
train_transform: Optional[Any] = _default_omniglot_train_transform,
eval_transform: Optional[Any] = _default_omniglot_eval_transform,
dataset_root: Union[str, Path] = None
) -> NCScenario:
"""
Creates a Rotated Omniglot benchmark.
If the dataset is not present in the computer, this method will
automatically download and store it.
Random angles are used to rotate the Omniglot images in ``n_experiences``
different manners. This means that each experience is
composed of all the original 964 Omniglot classes, but each image is
rotated in a different way.
The benchmark instance returned by this method will have two fields,
`train_stream` and `test_stream`, which can be iterated to obtain
training and test :class:`Experience`. Each Experience contains the
`dataset` and the associated task label.
A progressive task label, starting from "0", is applied to each experience.
The benchmark API is quite simple and is uniform across all benchmark
generators. It is recommended to check the tutorial of the "benchmark" API,
which contains usage examples ranging from "basic" to "advanced".
:param n_experiences: The number of experiences (tasks) in the current
benchmark. It indicates how many different rotations of the Omniglot
dataset have to be created.
:param seed: A valid int used to initialize the random number generator.
Can be None.
:param rotations_list: A list of rotations values in degrees (from -180 to
180) used to define the rotations. The rotation specified in position
0 of the list will be applied to the task 0, the rotation specified in
position 1 will be applied to task 1 and so on.
If None, value of ``seed`` will be used to define the rotations.
If non-None, ``seed`` parameter will be ignored.
Defaults to None.
:param train_transform: The transformation to apply to the training data
after the random rotation, e.g. a random crop, a normalization or a
concatenation of different transformations (see torchvision.transform
documentation for a comprehensive list of possible transformations).
If no transformation is passed, the default train transformation
will be used.
:param eval_transform: The transformation to apply to the test data
after the random rotation, e.g. a random crop, a normalization or a
concatenation of different transformations (see torchvision.transform
documentation for a comprehensive list of possible transformations).
If no transformation is passed, the default test transformation
will be used.
:param dataset_root: The root path of the dataset. Defaults to None, which
means that the default location for 'omniglot' will be used.
:returns: A properly initialized :class:`NCScenario` instance.
"""
if rotations_list is None:
rng_rotate = np.random.RandomState(seed)
rotations_list = [
rng_rotate.randint(-180, 181) for _ in range(n_experiences)
]
else:
assert len(rotations_list) == n_experiences, (
"The number of rotations"
" should match the number"
" of incremental experiences."
)
assert all(
-180 <= rotations_list[i] <= 180 for i in range(len(rotations_list))
), ("The value of a rotation" " should be between -180" " and 180 degrees.")
list_train_dataset = []
list_test_dataset = []
omniglot_train, omniglot_test = _get_omniglot_dataset(dataset_root)
# for every incremental experience
for experience in range(n_experiences):
rotation_angle = rotations_list[experience]
rotation = RandomRotation(degrees=(rotation_angle, rotation_angle))
rotation_transforms = dict(
train=(rotation, None), eval=(rotation, None)
)
# Freeze the rotation
rotated_train = AvalancheDataset(
omniglot_train,
transform_groups=rotation_transforms,
initial_transform_group="train",
).freeze_transforms()
rotated_test = AvalancheDataset(
omniglot_test,
transform_groups=rotation_transforms,
initial_transform_group="eval",
).freeze_transforms()
list_train_dataset.append(rotated_train)
list_test_dataset.append(rotated_test)
return nc_benchmark(
list_train_dataset,
list_test_dataset,
n_experiences=len(list_train_dataset),
task_labels=True,
shuffle=False,
class_ids_from_zero_in_each_exp=True,
one_dataset_per_exp=True,
train_transform=train_transform,
eval_transform=eval_transform,
)
def _get_omniglot_dataset(dataset_root):
if dataset_root is None:
dataset_root = default_dataset_location("omniglot")
train = Omniglot(root=dataset_root, train=True, download=True)
test = Omniglot(root=dataset_root, train=False, download=True)
return train, test
__all__ = ["SplitOmniglot", "PermutedOmniglot", "RotatedOmniglot"]
if __name__ == "__main__":
import sys
print("Split Omniglot")
benchmark_instance = SplitOmniglot(
4, train_transform=None, eval_transform=None
)
check_vision_benchmark(benchmark_instance)
print("Permuted Omniglot")
benchmark_instance = PermutedOmniglot(
5, train_transform=None, eval_transform=None
)
check_vision_benchmark(benchmark_instance)
print("Rotated Omniglot")
benchmark_instance = RotatedOmniglot(
5, train_transform=None, eval_transform=None
)
check_vision_benchmark(benchmark_instance)
sys.exit(0)