Is it possible to use l2l's meta-train to do standard supervised learning?

[brando90]

e.g. I don't want to do episodic meta-learning but I want to use the data in the meta-train set and sample it randomly. Is that possible? Just normal SL but on the meta-train set.

[seba-1511]

Yes, you can access the underlying dataset of a TaskDataset with taskset.dataset.

[brando90]

Yes, you can access the underlying dataset of a TaskDataset with taskset.dataset.

then pass that to a standard dataloader?

Thanks again for your really nice framework :) Hope to contribute in a more meaningful way soon I hope!

[seba-1511]

Correct, that gives you a standard PyTorch dataset.

[brando90]

Correct, that gives you a standard PyTorch dataset.

amazing! Your library is so nicely designed. Kudos.

[brando90]

decided to do it slightly differently but I it seems right:

def get_sl_l2l_datasets(root,
                        data_augmentation: str = 'rfs2020',
                        device=None
                        ) -> tuple:
    if data_augmentation is None:
        train_data_transforms = transforms.ToTensor()
        test_data_transforms = transforms.ToTensor()
    elif data_augmentation == 'normalize':
        train_data_transforms = Compose([
            lambda x: x / 255.0,
        ])
        test_data_transforms = train_data_transforms
    elif data_augmentation == 'rfs2020':
        train_data_transforms = get_transform(True)
        test_data_transforms = get_transform(False)
    else:
        raise ('Invalid data_augmentation argument.')

    import learn2learn
    train_dataset = learn2learn.vision.datasets.CIFARFS(root=root,
                                                        transform=train_data_transforms,
                                                        mode='train',
                                                        download=True)
    valid_dataset = learn2learn.vision.datasets.CIFARFS(root=root,
                                                        transform=train_data_transforms,
                                                        mode='validation',
                                                        download=True)
    test_dataset = learn2learn.vision.datasets.CIFARFS(root=root,
                                                       transform=test_data_transforms,
                                                       mode='test',
                                                       download=True)
    if device is not None:
        train_dataset = learn2learn.data.OnDeviceDataset(
            dataset=train_dataset,
            device=device,
        )
        valid_dataset = learn2learn.data.OnDeviceDataset(
            dataset=valid_dataset,
            device=device,
        )
        test_dataset = learn2learn.data.OnDeviceDataset(
            dataset=test_dataset,
            device=device,
        )
    return train_dataset, valid_dataset, test_dataset


def get_sl_l2l_cifarfs_dataloader(args: Namespace) -> dict:
    train_dataset, valid_dataset, test_dataset = get_sl_l2l_datasets(root=args.data_path)

    from uutils.torch_uu.dataloaders.common import get_serial_or_distributed_dataloaders
    train_loader, val_loader = get_serial_or_distributed_dataloaders(
        train_dataset=train_dataset,
        val_dataset=valid_dataset,
        batch_size=args.batch_size,
        batch_size_eval=args.batch_size_eval,
        rank=args.rank,
        world_size=args.world_size
    )
    _, test_loader = get_serial_or_distributed_dataloaders(
        train_dataset=test_dataset,
        val_dataset=test_dataset,
        batch_size=args.batch_size,
        batch_size_eval=args.batch_size_eval,
        rank=args.rank,
        world_size=args.world_size
    )
    dataloaders: dict = {'train': train_loader, 'val': val_loader, 'test': test_loader}
    return dataloaders


# - tests

def l2l_sl_dl():
    print('starting...')
    args = Namespace(data_path='~/data/l2l_data/', batch_size=8, batch_size_eval=2, rank=-1, world_size=1)
    args.data_path = Path('~/data/l2l_data/').expanduser()
    dataloaders = get_sl_l2l_cifarfs_dataloader(args)
    max_val = torch.tensor(-1)
    # for i, batch in enumerate(dataloaders['train']):
    #     # print(batch[1])
    #     max_val = max(list(batch[1]) + [max_val])
    #     # print(f'{max_val}')
    #     # if 63 in batch[1]:
    #     #     break
    # print(f'--> TRAIN FINAL: {max_val=}')
    max_val = torch.tensor(-1)
    for i, batch in enumerate(dataloaders['val']):
        # print(batch[1])
        max_val = max(list(batch[1]) + [max_val])
        # print(f'{max_val}')
        # if 15 in batch[1]:
        #     break
    print(f'--> VAL FINAL: {max_val=}')
    max_val = torch.tensor(-1)
    for i, batch in enumerate(dataloaders['test']):
        # print(batch[1])
        max_val = max(list(batch[1]) + [max_val])
        # print(f'{max_val}')
        # if 19 in batch[1]:
        #     break
    print(f'--> TEST FINAL: {max_val=}')

[brando90]

Actually do:

taskset.dataset

[brando90]

@seba-1511 I'm confused, what is the type of taskset in taskset.dataset suppose to be? Is it BenchmarkTasksets

[brando90]

@seba-1511 I'm confused, what is the type of taskset in taskset.dataset suppose to be? Is it BenchmarkTasksets

hmmm... this doesn't seem to be the right type:

args.tasksets.train.dataset
<learn2learn.data.meta_dataset.MetaDataset object at 0x7f89f3a44460>
type(args.tasksets.train.dataset)
<class 'learn2learn.data.meta_dataset.MetaDataset'>

[brando90]

perhaps its this:

args.tasksets.train.dataset
<learn2learn.data.meta_dataset.MetaDataset object at 0x7f89f3a44460>
type(args.tasksets.train.dataset)
<class 'learn2learn.data.meta_dataset.MetaDataset'>
type(args.tasksets.train.dataset.dataset)
<class 'learn2learn.vision.datasets.cifarfs.CIFARFS'>

[brando90]

I think this is better:

def get_standard_pytorch_dataset_from_l2l_taskdatasets(tasksets: BenchmarkTasksets, split: str) -> Dataset:
    """
    Trying to do:
        type(args.tasksets.train.dataset.dataset)
        <class 'learn2learn.vision.datasets.cifarfs.CIFARFS'>

    :param tasksets:
    :param split:
    :return:
    """
    # trying to do something like: args.tasksets.train
    taskset: TaskDataset = getattr(tasksets, split)
    # trying to do: type(args.tasksets.train.dataset.dataset)
    dataset: MetaDataset = taskset.dataset
    dataset: Dataset = dataset.dataset
    # assert isinstance(args.tasksets.train.dataset.dataset, Dataset)
    # asser isinstance(tasksets.train.dataset.dataset, Dataset)
    assert isinstance(dataset, Dataset), f'Expect dataset to be of type Dataset but got {type(dataset)=}.'
    return dataset

[brando90]

if device is not None:
train_dataset = learn2learn.data.OnDeviceDataset(
dataset=train_dataset,
device=device,
)
valid_dataset = learn2learn.data.OnDeviceDataset(
dataset=valid_dataset,
device=device,
)
test_dataset = learn2learn.data.OnDeviceDataset(
dataset=test_dataset,
device=device,
)

@seba-1511 is this really needed?

[brando90]

#385