Best way to load a pretrained model and start fine tuning it?

Part 1 (2020)
1

Right now, when I need to load a pre-trained image model, I just do it this way-

learn = cnn_learner(
    dls,
    arch,
    loss_func=loss_func,
    metrics=some_metrics    
)

What this triggers is, the ImageNet pre-trained model is downloaded and that’s how the learner is created.

But it requires internet access. Without internet access, this doesn’t work.

What I do is, I notice the path into which the .pth file is attempted to be downloaded. I manually download the .pth file from another computer and then put it to the directory the learner would look for in. It is usually- /root/.cache/torch/hub/checkpoints/.

If I do this, it works without the internet. But it is manual and quite messy. Isn’t there a better way to do it?

I want to work like an architecture, not a learner in itself. If I use the .load_learner() method, then it is the learner in itself. I don’t want that.

Please suggest what I should do.

2

Many pretrained weights are already available as Kaggle datasets, you can click on ‘upload data’ and search by the name of the architecture.

If you are trying to use a custom model below are what I feel some good examples.
https://www.kaggle.com/muellerzr/recreating-abhishek-s-tez-with-fastai

3

This is not about Kaggle. Well, at least not entirely. I also have writing code for edge devices in my mind.

4

oops! I was some mistaken. But those links still have info on creating custom models.