fastpapers 0.0.1

Fastai implementation of papers.

Welcome to fastpapers

Play LEGO with papers.

fastpapers is a python library where we reproduce papers on Jupyter Notebooks. We use nbdev to turn these notebooks into modules. The implementations are done using fastai.

Install

pip install your_project_name

How to use

The name of each module is the Bibtexkey of the corresponing paper. For example, if you want to use the FID metric from Heusel, Martin, et al. 2017, you can import it like so:

from fastpapers.heusel2017gans import FIDMetric

If you want to train a pix2pix model from Isola, Phillip, et al you can import a pix2pix_learner

from fastpapers.isola2017image import pix2pix_learner

The core module contains functions and classes that are useful for several papers. For example, you have a ImageNTuple to work with an arbitrary amount of images as input.

path = untar_data(URLs.PETS)
files = get_image_files(path/"images")

it = ImageNTuple.create((files[0], files[1], files[2]))
it = Resize(224)(it)
it = ToTensor()(it)
it.show();

Or useful functions for debuging like explode_shapes or explode_ranges

explode_shapes(it)

[(3, 224, 224), (3, 224, 224), (3, 224, 224)]