fasterrcnn-pytorch-training-pipeline

fasterrcnn-pytorch-training-pipeline copied to clipboard

Hi,

I get the following error when I train with --use-train-aug:

It stems from albumentaions' gamma correction method - Numpy can't handle a fractional power of a negative number (see this - https://github.com/numpy/numpy/issues/20186)

Note that I don't have negative pixel values. My input is in pascal_voc format and I think some zero pixel values are changing to very small negative values during normalization due to floating point error.

I was able to suppress this error by changing normalized pixel values to zero during gamma correction/transformation but would like to hear your suggestion on this. Added the following code block before np.power statement in albumentaions' gamma correction method:

img[img < 0] = 0

Thanks!

It may be dataset specific. Kind of hard to say much here actually. In fact, this is the first time I am seeing this error. Still, will try to research more.