Resnet 152 not included in Fast.ai libraries?


(James Dietle) #1

I was trying to run through the different Resnet models (32,50,101,152) on the Multi-label classification set, and it appears that 152 is not supported. Just curious, is this true or am I reading the code wrong?

I went through the following troubleshooting included updating my libraries. I see PyTorch supports Resnet 152 in torchvision/resnet.py.

Turning back to the Fast.ai folder I used:

grep -nr resnet

to find that both torch_imports.py and conv_learner.py are calling the other Resnet models but not 152.

I tried altering the code in both files to add resnet152 but could not trace down all the dependencies.


(yinterian) #2

It is not supported. You can try to add it. Jeremy recently added VGG16. Here is his CL.


(Ramesh Sampath) #3

I created a pull request to add these two networks (Resnet152 and VGG19). Tested via Lesson 1 Notebook.


Wiki: Fastai Library Feature Requests
(Kevin Bird) #4

Great work Ramesh, that will be very nice to have. :slight_smile:


(Jeremy Howard (Admin)) #5

I guess I never used it since resnext101_64 seems like the best ‘big model’ choice. But thanks for the PR - no harm having the option! :slight_smile:


(James Requa) #6

Agreed the more model options the better! (imo)