torch_moving_mnist

Install

pip install -e .

How to use

from types import SimpleNamespace
from torch_moving_mnist.data import MovingMNIST
from torch_moving_mnist.utils import show_images

affine_params = SimpleNamespace(
    angle=(-5, 5), # rotation in degrees (min and max values)
    translate=((-5, 5), (-5, 5)), # translation in pixels x and y
    scale=(.9, 1.1), # scaling in percentage (1.0 = no scaling)
    shear=(-2, 2), # deformation on the z-plane
)

Create a MovingMNIST dataset with affine_params, with 10 frames and may include up to 3 digitis. Image size is 64.

from nbdev.showdoc import show_doc

---

source

MovingMNIST

 MovingMNIST (path='.', affine_params:dict=namespace(angle=(-4, 4),
              translate=((-5, 5), (-5, 5)), scale=(0.8, 1.2), shear=(-3,
              3)), num_digits:list[int]=[1, 2], num_frames:int=4,
              img_size=64, concat=True, normalize=False)

Initialize self. See help(type(self)) for accurate signature.

	Type	Default	Details
path	str	.	path to store the MNIST dataset
affine_params	dict	namespace(angle=(-4, 4), translate=((-5, 5), (-5, 5)), scale=(0.8, 1.2), shear=(-3, 3))	affine transform parameters, refer to torchvision.transforms.functional.affine
num_digits	list	[1, 2]	how many digits to move, random choice between the value provided
num_frames	int	4	how many frames to create
img_size	int	64	the canvas size, the actual digits are always 28x28
concat	bool	True	if we concat the final results (frames, 1, 28, 28) or a list of frames.
normalize	bool	False	scale images in [0,1] and normalize them with MNIST stats. Applied at batch level. Have to take care of the canvas size that messes up the stats!

ds = MovingMNIST(affine_params=affine_params, num_frames=10, num_digits=[1,2,3], img_size=64)

when you index the dataset, it generates a random set of MNIST digits and trajectories. You could basically only call ds[0]

sequence = ds[0]

show_images(sequence, figsize=(20,10))

t = sequence
type(t), t.shape

(torch.Tensor, torch.Size([10, 1, 64, 64]))

Dataloader

This dataset is randomly creating sequences on the fly, so the dataloader is just going to generate a batch…

batch = ds.get_batch(bs=128)

batch.shape

torch.Size([128, 10, 1, 64, 64])

show_images(batch[0])

the dataloader also is normalizing the inputs for you, after constructing the batch.

ds.batch_tfms

Compose(
    ConvertImageDtype()
)

This site is open source. Improve this page.