Using value from batch within pipeline transforms

[tavisshore]

My dict contains an image and a heading angle float. Is it possible to use the heading tensors from the same dict to crop my images with an FFCV custom fast transform?

Thanks

[andrewilyas]

If it really is just a single float per example, I'd recommend keeping the array of floats in memory, and then using a custom transform with inds (https://docs.ffcv.io/ffcv_examples/transform_with_inds.html) to access the right one. Let me know if this doesn't work!