Adapt to different image sizes
Breeze-Zero opened this issue · 5 comments
My task requires that the images do not scale and crop uniformly, so I can only batch size 1 to import different sizes of data. Can FFCV writer achieve this?
Have you used the RandomResizedCrop() function, or ResizedCropRGBImageDecoder() class?
Have you used the
RandomResizedCrop()function, orResizedCropRGBImageDecoder()class?
No, Just reading the data, the data itself is not consistent and I don't want to resize it.
What are you actually trying to do
What are you actually trying to do
Here's the size range of my data: {(448, 132, 10, 9, 12), (448, 132, 10, 10, 12), (512, 204, 10, 11, 12), (512, 204, 10, 12, 12), (512, 204, 10, 13, 12). (512, 246, 10, 11, 12), (512, 204, 10, 9, 12), (512, 204, 10, 10, 12), (512, 246, 10, 9, 12), (512, 162, 10, 9, 12), (512, 246, 10, 10, 12), (512, 162, 10, 10, 12), (512, 246, 10, 12, 12)}, some dimensions are not consistent. I can't use cropping or scaling operations due to the peculiarity of the data, so I have to read it into the dataloader originally. The model is trained in a batch size. Whether the Writer of FFCV can support this kind of data without changing the size of the data?
As far as I'm aware the writer can create the beton file fine with varying sizes, but you won't be able to read it effectively without having to crop / resize to make all tensors within a batch the same shape.