Use FixRes with fastai

Part 1 (2019)
1

I am trying to use FixRes method from https://github.com/facebookresearch/FixRes within fastai learner.
There is an example code for loading pretrained models provided in the repo:
https://github.com/facebookresearch/FixRes#pre-trained-networks
How could I use this and integrate it in cnn_learner like: arch = resnext101_32x48d_wsl and learner = cnn_learner(data, arch, [accuracy])? Obviously I tried and it didn’t work.
I searched the forum for a way to load custom weights for custom model but couldn’t make it work

2

I usually pass my custom models in the Learner like so learn = Learner(data, model) and it works fine. Have you tried it ? I think that cnn_learner only works for some architectures predefined in the library.

Hope that helps :slight_smile:

1 Like
3

Hi Nathan,
Is the model an architecture like resnet50 or an instance like resnet50()?

4

I tried Learner(data, model) as you suggested and encounterer this error: `/usr/local/lib/python3.6/dist-packages/fastai/basic_train.py in post_init(self)
163 “Setup path,metrics, callbacks and ensure model directory exists.”
164 self.path = Path(ifnone(self.path, self.data.path))
–> 165 self.model = self.model.to(self.data.device)
166 self.loss_func = self.loss_func or self.data.loss_func
167 self.metrics=listify(self.metrics)

AttributeError: module ‘FixRes.imnet_finetune.resnext_wsl’ has no attribute ‘to’`
Can you show me an example of your custom model or how to properly construct one @NathanHub?
Thank you

5

I was able to use the method with arch = resnext101_32x48d_wsl and pass it to learner = cnn_learner(data, arch), then load the model provided in FixRes repo by calling learner.model blah blah

Actually it didn’t work out, cause when the arch is passed in cnn_learner, the keys in its state_dict don’t hold the original name, so nothing matched between the 2 state_dict

I’m still looking for an workaround

6

Here is a simple gist where I create a Learner from a custom model.

Also, when you use the cnn_learner, you should specify pretrained=False, as it is True by default.

1 Like
7

I revisited the docs and could clarify that cnn_learner take base_arch as a Callable, while Learner take model as a module, your suggestion worked for me. Thank you, have a good day @NathanHub