decode_ind_ab

[whartley94]

in utils decode_ind_ab(), the calculations is

    data_a = data_q/opt.A
    data_b = data_q - data_a*opt.A
    data_ab = torch.cat((data_a, data_b), dim=1)

however I believe according to how the encoding was done we should have instead something like
data_a = (data_q - data_b)/opt.A

I'm imagining this would have to be solved using linear programming or something.
I was just wondering if this is something you're aware of and whether I am missing something?

My issue is that when I use decode_ind_ab currently all my b values come through as -1 as
with

    data_a = data_q/opt.A (eq1)
    data_b = data_q - data_a*opt.A (eq2)

we can sub eq1 into eq2 to show that

data_b = data_q - data_q = 0

which then gets scaled and shifted to -1 before being returned.

[whartley94]

This seems to work

def decode_ind_ab(data_q, opt):
    # Decode index into ab value
    # INPUTS
    #   data_q      Nx1xHxW \in [0,Q)
    # OUTPUTS
    #   data_ab     Nx2xHxW \in [-1,1]

    assert isinstance(opt.A, (int, float))
    data_b = torch.fmod(data_q, opt.A)
    data_a = (data_q-data_b)/opt.A

    data_ab = torch.cat((data_a, data_b), dim=1)

    if data_q.is_cuda:
        type_out = torch.cuda.FloatTensor
    else:
        type_out = torch.FloatTensor
    data_ab = ((data_ab.type(type_out)*opt.ab_quant) - opt.ab_max)/opt.ab_norm

    return data_ab