def split_train_val(dataset: TaskSet,
val_split: float = 0.1) -> Tuple[TaskSet, TaskSet]:
"""Split train dataset into two datasets, one for training and one for validation.
:param dataset: A torch dataset, with .x and .y attributes.
:param val_split: Percentage to allocate for validation, between [0, 1[.
:return: A tuple a dataset, respectively for train and validation.
"""
random_state = np.random.RandomState(seed=1)
indexes = np.arange(len(dataset))
random_state.shuffle(indexes)
train_indexes = indexes[int(val_split * len(indexes)):]
val_indexes = indexes[:int(val_split * len(indexes))]
x_train, y_train, t_train = dataset.get_raw_samples(train_indexes)
train_dataset = TaskSet(x_train,
y_train,
t_train,
trsf=dataset.trsf,
data_type=dataset.data_type)
x_val, y_val, t_val = dataset.get_raw_samples(val_indexes)
val_dataset = TaskSet(x_val,
y_val,
t_val,
trsf=dataset.trsf,
data_type=dataset.data_type)
return train_dataset, val_dataset
def test_get_random_samples(nb_samples):
x = np.ones((10, 2, 2, 3))
y = np.ones((10, ))
t = np.ones((10, ))
base_set = TaskSet(x, y, t, None)
base_set.get_random_samples(nb_samples=nb_samples)
def test_get_raw_samples(nb_samples):
x = np.ones((10, 2, 2, 3))
y = np.ones((10, ))
t = np.ones((10, ))
base_set = TaskSet(x, y, t, None)
data, y_, t_ = base_set.get_raw_samples(indexes=range(nb_samples))
assert (x[:nb_samples] == data).all()
assert (y[:nb_samples] == y_).all()
assert (t[:nb_samples] == t_).all()
def __getitem__(self, task_index: Union[int, slice]):
"""Returns a task by its unique index.
:param task_index: The unique index of a task. As for List, you can use
indexing between [0, len], negative indexing, or
even slices.
:return: A train PyTorch's Datasets.
"""
if isinstance(task_index, slice):
raise NotImplementedError(
f"You cannot select multiple task ({task_index}) on OnlineFellowship scenario yet"
)
self.cl_dataset = self.cl_datasets[task_index]
if isinstance(self.cl_dataset, _ContinuumDataset):
x, y, _ = self.cl_dataset.get_data()
t = np.ones(len(y)) * task_index
taskset = TaskSet(x,
y,
t,
trsf=self._get_trsf(task_index,
self.transformations),
target_trsf=self._get_label_trsf(task_index),
data_type=self.cl_dataset.data_type,
bounding_boxes=self.cl_dataset.bounding_boxes)
else:
if not isinstance(self.cl_dataset, BaseTaskSet):
raise ValueError(
"self.cl_datasets can only contain _ContinuumDataset or TaskSet"
)
taskset = self.cl_dataset
return taskset
def concat(task_sets: List[TaskSet]) -> TaskSet:
"""Concatenate a dataset A with one or many *other* datasets.
The transformations will be those of the first dataset.
:param Tasksets: A list of task sets.
:return: A concatenated task set.
"""
x, y, t = [], [], []
data_type = task_sets[0].data_type
for task_set in task_sets:
if task_set.data_type != data_type:
raise Exception(
f"Invalid data type {task_set.data_type} != {data_type}")
x.append(task_set._x)
y.append(task_set._y)
t.append(task_set._t)
return TaskSet(np.concatenate(x),
np.concatenate(y),
np.concatenate(t),
trsf=task_sets[0].trsf,
data_type=data_type)
def test_concat_method(nb_others):
x = np.random.rand(10, 2, 2, 3)
y = np.ones((10, ))
t = np.ones((10, ))
base_set = TaskSet(x, y, t, None)
initial_len = len(base_set)
others = [
TaskSet(np.copy(x), np.copy(y), np.copy(t), None)
for _ in range(nb_others)
]
base_set.concat(*others)
assert len(base_set) == initial_len + nb_others * initial_len
loader = DataLoader(base_set)
for x, y, t in loader:
pass
def test_split_train_val(val_split, nb_val):
x = np.random.rand(10, 2, 2, 3)
y = np.ones((10, ))
t = np.ones((10, ))
base_set = TaskSet(x, y, t, None)
train_set, val_set = split_train_val(base_set, val_split)
assert len(val_set) == nb_val
assert len(train_set) + len(val_set) == len(base_set)
def __getitem__(self, task_index: Union[int, slice]):
"""Returns a task by its unique index.
:param task_index: The unique index of a task. As for List, you can use
indexing between [0, len], negative indexing, or
even slices.
:return: A train PyTorch's Datasets.
"""
x, y, t = self._select_data_by_task(task_index)
return TaskSet(x, y, t, self.trsf, data_type=self.cl_dataset.data_type)
def taskset_subset(taskset: TaskSet, indices: np.ndarray) -> TaskSet:
# x, y, t = taskset.get_raw_samples(indices)
x, y, t = taskset.get_raw_samples(indices)
# TODO: Not sure if/how to handle the `bounding_boxes` attribute here.
bounding_boxes = taskset.bounding_boxes
if bounding_boxes is not None:
bounding_boxes = bounding_boxes[indices]
return replace_taskset_attributes(
taskset, x=x, y=y, t=t, bounding_boxes=bounding_boxes
)
def test_split_train_val_loading():
x = np.random.rand(10, 2, 2, 3)
y = np.ones((10, ))
t = np.ones((10, ))
base_set = TaskSet(x, y, t, None)
train_set, val_set = split_train_val(base_set, 0.2)
for task_set in (train_set, val_set):
loader = DataLoader(task_set, batch_size=32)
for x, y, t in loader:
pass
def __getitem__(self, task_index):
"""Returns a task by its unique index.
:param task_index: The unique index of a task, between 0 and len(loader) - 1. Or it could
be a list or a numpy array or even a slice.
:return: A train PyTorch's Datasets.
"""
x, y, _ = self.dataset
if isinstance(task_index, slice):
# Convert a slice to a list and respect the Python's advanced indexing conventions
start = task_index.start if task_index.start is not None else 0
stop = task_index.stop if task_index.stop is not None else len(self) + 1
step = task_index.step if task_index.step is not None else 1
task_index = list(range(start, stop, step))
if len(task_index) == 0:
raise ValueError(f"Invalid slicing resulting in no data (start={start}, end={stop}, step={step}).")
elif isinstance(task_index, np.ndarray):
task_index = list(task_index)
elif isinstance(task_index, int):
task_index = [task_index]
else:
raise TypeError(f"Invalid type of task index {type(task_index).__name__}.")
task_index = set([_handle_negative_indexes(ti, len(self)) for ti in task_index])
t = np.concatenate([
(np.ones(len(x)) * ti).astype(np.int32) for ti in task_index
])
x = np.concatenate([
x for _ in range(len(task_index))
])
if self.shared_label_space:
y = np.concatenate([
y for _ in range(len(task_index))
])
else:
# Different transformations have different labels even though
# the original images were the same
y = np.concatenate([
y + ti * self.num_classes_per_task for ti in task_index
])
trsf = [ # Non-used tasks have a None trsf
self.get_task_transformation(ti)
if ti in task_index else None
for ti in range(len(self))
]
return TaskSet(x, y, t, trsf, data_type=self.cl_dataset.data_type)
def test_target_trsf(nb_classes):
x = np.random.rand(10, 2, 2, 3)
y = np.arange(10)
t = np.ones((10, ))
target_trsf = transforms.Lambda(lambda x: x % nb_classes)
tasket = TaskSet(x, y, t, None, target_trsf=target_trsf)
assert tasket.nb_classes == nb_classes, print(
"target transform not applied in get_classes")
loader = DataLoader(tasket)
for x, y, t in loader:
pass
def test_concat_function(nb_others):
x = np.random.rand(10, 2, 2, 3)
y = np.ones((10, ))
t = np.ones((10, ))
task_sets = [
TaskSet(np.copy(x), np.copy(y), np.copy(t), None)
for _ in range(nb_others)
]
concatenation = concat(task_sets)
assert len(concatenation) == nb_others * 10
loader = DataLoader(concatenation)
for x, y, t in loader:
pass
def test_sampler_function(log):
np.random.seed(1)
torch.manual_seed(1)
x = np.random.rand(100, 2, 2, 3)
y = np.ones((100, ), dtype=np.int64)
y[0] = 0
t = np.ones((100, ))
taskset = TaskSet(x, y, t, None)
sampler = get_balanced_sampler(taskset, log=log)
loader = DataLoader(taskset, sampler=sampler, batch_size=1)
nb_0 = 0
for x, y, t in loader:
if 0 in y:
nb_0 += 1
assert nb_0 > 1
def __getitem__(self, task_index: Union[int, slice]) -> TaskSet:
"""Returns a task by its unique index.
:param task_index: The unique index of a task. As for List, you can use
indexing between [0, len], negative indexing, or
even slices.
:return: A train PyTorch's Datasets.
"""
if isinstance(task_index, slice) and task_index.step is not None:
raise ValueError("Step in slice for segmentation is not supported.")
x, y, t, task_index = self._select_data_by_task(task_index)
t = self._get_task_ids(t, task_index)
return TaskSet(
x, y, t,
self.trsf,
target_trsf=self._get_label_transformation(task_index),
data_type=self.cl_dataset.data_type
)
def __getitem__(self, task_index):
"""Returns a task by its unique index.
:param task_index: The unique index of a task, between 0 and len(loader) - 1.
:return: A train PyTorch's Datasets.
"""
if isinstance(task_index, slice):
raise ValueError(
"Incremental training based on transformations "
"does not support slice, please provide only integer.")
elif task_index < 0: # Support for negative index, e.g. -1 == last
while task_index < 0:
task_index += len(self)
self.update_task_indexes(task_index)
if not self.shared_label_space:
self.update_labels(task_index)
train = self._select_data_by_task(task_index)
trsf = self.get_task_transformation(task_index)
return TaskSet(*train, trsf, data_type=self.cl_dataset.data_type)
def __getitem__(self, task_index: Union[int, slice]):
"""Returns a task by its unique index.
:param task_index: The unique index of a task. As for List, you can use
indexing between [0, len], negative indexing, or
even slices.
:return: A train PyTorch's Datasets.
"""
if isinstance(task_index, slice) and isinstance(self.trsf, list):
raise ValueError(
f"You cannot select multiple task ({task_index}) when you have a "
"different set of transformations per task")
x, y, t, _, data_indexes = self._select_data_by_task(task_index)
return TaskSet(x,
y,
t,
trsf=self.trsf[task_index] if isinstance(
self.trsf, list) else self.trsf,
data_type=self.cl_dataset.data_type,
bounding_boxes=self.cl_dataset.bounding_boxes,
data_indexes=data_indexes)
