Using your own dataloaders (fastai's data utility is too much evolved around images saved on disk)
Using your own architecture instead of the predefined ones (e.g.
FASTAI IMAGE CLEANER HOW TO
It's tiresome to figure out how to do things differently, for example: HOWEVER, if you want to do anything with fastai that isn't exactly documented, it's a pain in the ass. I could produce very good models in a very short amount of time. and putting on a new multi-layer head that fits your number of classes. For example 1-cycle learning rate scheduler, learning rate finder, differential learning rates, automatically cutting off the head of a pre-trained resnet/inception etc. I guess most importantly, fastai implemented "state of the art" training techniques that are done under the hood and either happen automatically or are just a 1-liner for the user. PyTorch has a nice api to write your own data-loaders and parallelise them.įastai is basically another layer on top of PyTorch that gives you a lot new functionality around your neural network, such as visualization methods for your data, more ways of loading and splitting data, infering the number of classes from the dataset your provide and it extends the training utilities by a concept called "callbacks" (which keras also has but pytorch doesnt). Both frameworks come with utility functions like loading data in a certain format from disk. In Keras, training/testing/predicting is as easy as calling 1 function, in PyTorch you have to write a few more lines to train your network.
In both frameworks it is easy to define neural networks and use implemented versions of different optimizers and loss functions. The better comparison would be PyToch = Keras. I have used PyTorch, Keras and fastai, here is my point of view:įastai for PyTorch is NOT what Keras is for TF. I don't agree with some of the posts here to be honest.
0 kommentar(er)
