Report

[nayuta-ai]

Background

My research needs to factorize the intermediate layer using the invertible structure such as a flow-based model. In invertible-resnet, it can have an invertible structure and perform high classification accuracy. So, I want to use the invertible-resnet in my research. There is some GitHub repository on invertible-resnet, but there was an error due to a dependency module. Since it was implemented three years ago, I also could not reproduce the implementation. My goal is to conduct a full scratch Implementation of an invertible-resnet.

Goal

This issue will be closed when I finish implementing the invertible-resnet. I will carry out some experiments on an invertible-resnet.

Approach

Overview

I will set up the docker environment (#1 docker), comet-ml (#2 comet-ml), GitHub actions (#3 Github). I will implement the invertible-resnet such as parsing args, getting the dataset, getting the model, etc.

To do

• [x] Set up Docker (#1)
• [x] Set up comet-ml (#2)
• [x] Set up GitHub (#3)
• [x] Implement argparse (#4)
• [ ] Implement get_dataloader (#7)
• [ ] Implement get_model (#8 )
• [ ] Implement the train function(#9 )
• [ ] Install rubber data

Frequency

Every Day

Deadline

2022/8

References

https://github.com/jhjacobsen/invertible-resnet

Notes

None

[nayuta-ai]

Working Time

3.5h

Supplement for the task

I modified #7 and implement #8 and the train function according to this. I didn't clean #8 and the train function because there are many lines, and I focused on implementing the invertible resnet until 8/25.