Revert "Fix training crash with custom batch size"

[andy-yun]

Reverts andy-yun/pytorch-0.4-yolov3#2

Because the getitem is called in every iteration, the image augmented for every 64 images. Some cases, batch_size is 1, then all images are resized. This is not the case we expected.

[mickolka]

That is exactly what is expected: each new batch (in your proposed case, of size 1, but there is nothing special about this value) will contain images of different size. First 4000*64 seen images won't be resized. After that, all images are resized anyway, but the target size is changed each batch_size images.

[andy-yun]

In most cases, you are right because of the width and height are same and assigned to 13*32. But, I trained and tested at the different image size. for example, I use 480x800 images. In that cases, the images are not correctly trained by many resized images.

[mickolka]

Would be useful to see the actual code you used to discuss it further, can't really understand what you mean.