Pytorch frameworks, a few comparissons

Catalyst, Fastai, Ignite and Pytorch-Lightning are all amazing frameworks but which one should I use for project x? I have been asking myself the same question and it is not an easy answer.

There are several factors at play and framework selection also depends on your background. I will outline some basic statistics and library codestyle examples. I believe it is important to be able to delve into the source code of these libraries when you inevitably get stuck on problem/can’t work out how to implement something or want to debug your code.


I started to use Catalyst last year and really liked the pre-built notebooks showing examples for image classification and segmentation. I have placed in top 1% in a couple of image segmentation and an image classification competition using it. The catalyst github repo documents code from mutiple high competition placings.

An example of the starting code for the SupervisedRunner in catalyst is shown above.

What I like: The dataloading, transforms, model are all pure pytorch, only the training loop is different, with lots of boilerplate pytorch code simplified. The team implements latest best training practices.

What is hard: Catalyst relies heavily on callbacks and it requires a bit of work to understand how to use and modify these. Also some of the built-in features may not have clear examples on how to use them.

When to use: If you are at an intermediate level+ researcher or kaggle (or other platform) competitor.


I have been using Fastai since 2017 which then was using Tensorflow, and first started using Pytorch though the training corses (parts 1&2 in 2017,2018,2019 and 2020). Out of the libraries here, Fastai to me feels the higest level. Not only this but Fastai manages to do this with a fairly small codebase which is impressive. Note however the code-style is more densely packed than the other libraries, and if formatted in a black the lines of code would exand out significantly.

Example code from the Learner class in fastai is shown above.

What I like: The community is amazing. You can find a solution to pretty much any problem there. The courses are really good too, and at the moment I am working through the Fastai book. I am really grateful to Jeremy and the Fastai team for introducing me to a lot of new concepts. The team keeps up-to date with research and implements latest best training practices.

What is hard: The Fastai dataloder is different to the other 3 frameworks where (which all use the pytorch dataloader), and is a core piece of Fastai. Maybe not so much hard as different, it takes time to get your head around the datablocks API.

When to use: If you are a beginner through expert and like the way you can very quickly implement working code.


I have only breifly used Ignite, the library does have some interesting features. Out of the 4 lbraries reviewed, Ingite seems to allow you the closest coupling to pure pytorch, and I am looking forward to experimenting more with it.

The Ignite codebase is the smallest of the 4 frameworks.

Code from the run method of the Engine class is shown above.


I have been using Pytorch-Lightning on and off for a few months. I have been doing some experimenting with pre-trained transformers from the huggingface library with it.

An example of code from the fit method in pytorch Trainer is shown above

What I like: The examples for porting pytorch code to pl. For small codebases it is fairly easily to port over pytorch code.

What is hard: I have found it tricky to debug for example my implementation of loading a pre-trained checkpoint into a new model for inference. Also there are only a few example implementations in the codebase.

When to use: Intermediate+ when you have mastered the basics of training models in pytorch.

The winner

I don’t think there is a clear winner, each of the libraries has it’s strengths and weaknesses (to me). I’d encourage you to explore each and see which library gels with you.

Geophysicist and Deep Learning Practitioner