[Operator] Bilinear upsamping

[antinucleon]

@piiswrong @winstywang Init is not implemented and not test yet, seems don't need concat

[piiswrong]

what do you mean dont need concat? does this handle multiple input without duplicating memory?

[antinucleon]

After init the bilinear weight with

weight = np.zeros(np.prod(shape))
f = shape[3] / 2.
c = (2 * f - 1 - f % 2) / (2. * f)
for i in range(np.prod(shape)):
    x = i % shape[3]
    y = (i / shape[3]) % shape[2]
    weight[i] = (1 - abs(x / f - c)) * (1 - abs(y / f - c));

the result is:

original image:

nearest result:

bilinear result:

[antinucleon]

@piiswrong I have finished bilinear upsampling. For hypercolumn or concat, I am not very clear about these and feel free to refactor for your purpose.

BTW. Travis lint system is not working correctly.

[piiswrong]

I think it's easy to do for nearest neighbor. I'll do it after you merge this.

[piiswrong]

What happened to travis lint? Did the rule set change?

[antinucleon]

https://travis-ci.org/dmlc/mxnet/jobs/93842415 ,seems are errors with numpy and some more strict rules.

[antinucleon]

Right