tlpss / keypoint-detection

2D keypoint detection with Pytorch Lightning and wandb
MIT License
70 stars 9 forks source link

Resizeable data #23

Closed lucasvandijck closed 1 year ago

lucasvandijck commented 1 year ago

By adding the --image_size argument, you can dynamically resize the input data before training. I use this option to retrain large images at lower sizes without having to store the resized data.

I did not add a test function.

tlpss commented 1 year ago

@lucasvandijck, It should indeed by easier to add a custom augmentation pipeline... I've been thinking about how to make this possibly from the CLI, but I think at that point it becomes reasonable to create a custom train script.

Anyways, I'm not sure it is a good idea to do the resizing online, this tends to be an expensive operation and therefore it is usually better to to this before training. To make things even worse, you would also do it in each epoch for the same image instead of doing it once, which would require over more compute.