I've managed to get training up and running without issue as long as I keep image_shape square.
If I try to change it something like [480,272] I get an error like the following:
RuntimeError: The size of tensor a (128) must match the size of tensor b (64) at non-singleton dimension 2
Is it possible to train on non-square aspect ratios, and if so, what parameters do I need to adjust to do this?
If not, what's the recommended way of using the network on input data that isn't square? Should I crop images to square, or does the network expect squished people because of training and I should just pass it a squished image?
Hello,
I've managed to get training up and running without issue as long as I keep
image_shapesquare. If I try to change it something like[480,272]I get an error like the following:RuntimeError: The size of tensor a (128) must match the size of tensor b (64) at non-singleton dimension 2Is it possible to train on non-square aspect ratios, and if so, what parameters do I need to adjust to do this? If not, what's the recommended way of using the network on input data that isn't square? Should I crop images to square, or does the network expect squished people because of training and I should just pass it a squished image?
Thanks!