# I spent two weeks tuning a paper reproduction and their checkpoint scored the same as mine — retained record

> Retained reddit record from reddit:computervision.

Canonical: https://highsignal.app/data/records/45d58f9295609167

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:computervision)  published 2026-10-14 12:48 ingested 2026-10-15 02:37

[original source ↗](https://www.reddit.com/r/computervision/comments/1x5qcbv/i_spent_two_weeks_tuning_a_paper_reproduction_and/)

## retained text



```
Three weeks ago I started reproducing a fine-grained classification paper from 2024, we are going to build on it at work if it holds up. The results table in the paper says 87.3. I could not get my runs past about 85.8 and most of the next two weeks went into the learning rate and after that into augmentation strength. Then I ran their released checkpoint through their eval script on our training box, which I should have done on day one, and it got 85.9, basically what I was getting. That machine is shared with other teams, it is still on PyTorch 2.1 and torchvision 0.16 and nobody wants to touch it.

So I tried the same checkpoint in a fresh PyTorch notebook on HyperAI, and installing their repo and uploading the checkpoint and the test split took longer than the eval did. It gave 87.2 there. I diffed pip freeze between the two machines and spent an evening going down it line by line before I got to torchvision.

Their repo loads images with read\_image, which gives you tensors, and Resize is called without antialias. On torchvision 0.16 that means tensors get resized with no antialiasing at all, it only became the default in 0.17. The warning about this is the first line of every eval log I have. Their requirements.txt says torchvision>=0.15, which our box meets, and I do not get why anyone puts a >= on a paper repo they expect people to reproduce.

With antialias=True their checkpoint matches on our box, but my runs were all trained on the aliased images, so they need redoing and with our queue it will be next week before I know how much of my gap this was.
```



document fetched 2026-10-15 00:24 · canonical https://reddit.com/r/computervision/comments/1x5qcbv/i_spent_two_weeks_tuning_a_paper_reproduction_and
