xvdp

""" I was writing a dataloader from a video stream. I ran some numbers.
# in a nutshell. 
-> np.transpose() or torch.permute() is faster as uint8, no difference between torch and numpy
-> np.uint8/number results in np.float64, never do it, if anything cast as np.float32
-> convert to pytorch before converting uint8 to float32
-> contiguous() is is faster in torch than numpy
-> contiguous() is faster for torch.float32 than for torch.uint8
-> convert to CUDA in the numpy to pytorch conversion, if you can.
-> in CPU tensor/my_float is > 130% more costly than tensor.div_(myfloat), however tensor.div_()
does not keep track of gradients, so be careful using it.

xvdp

"""testing vram in pytorch cuda
every time a variable is put inside a container in python, to remove it completely
one needs to delete variable and container,
this can be problematic when using pytorch cuda if one doesnt clear all containers
Three tests:

>>> python memory_tests list
    # creates 2 tensors puts them in a list, modifies them in place, deletes them
    # in place mod changes original tensors
    # list and both tensors need to be deleted