Autoencoder with fastai V2

Part 1 (2020) Non-beginner
#1

Hello!
I put a small example on how to construct an autoencoder using V2.
This is a Image2Image model using the DataBlockAPI
It runs on colab, you can view it here
I struggled a little bit, so it may be useful for someone else.

9 Likes
#2

Great work! Could you put it over here in the projects thread? :slight_smile:

https://forums.fast.ai/t/fastai2-blog-posts-projects-and-tutorials/65827

#3

Thanks very much for your input. I am using fastai to built my own AE and this has been very helpful!

#4

I updated the notebooks to V2 release (was on beta V2).

1 Like
#5

Thanks tcapelle for sharing your autoencoder solution. Interesting solution.

Maybe there has been version updates. When running it now and installing newest versions I had to make following two modifications:

‘arch = create_body(xresnet18, n_in=1).cuda()’
to
‘arch = create_body(xresnet18(pretrained=True), n_in=1).cuda()’

and

‘get_y = lambda o: o,’
to
‘get_y = named_function_to_make_learn_export_work,’

Without named function I ended up to error when saving the learner. This seems to be generic fastai issue and not just with my environment:

1 Like
#6

Yeah, this is very old. It does not take into account that fastai now integrates timm, and other cool stuff!