Add random affine transformation

[vfdev-5]

Random affine transformations with tests.
For usage, see sanity_check.ipynb

[vfdev-5]

Any other feedbacks on the PR are highly appreciated :)

[fmassa]

This looks great, thanks!

I have a couple of minor comments, but apart from that this is very good!

[vfdev-5]

@fmassa A precision on matching between transformed points. Take look at the example with scale.
Here is the input image:

which is then transformed with

result = F.affine(img, angle=0, translate=(0, 0), scale=2.0, shear=0.0)

However, manually-verified points gives :

xs = []
ys = []
true_image = np.zeros((100, 100, 3), dtype=np.uint8)
for y, x in pts:
    xs.append(int((x - cnt[0]) * s + cnt[0] + 0.5))
    ys.append(int((y - cnt[1]) * s + cnt[1] + 0.5))
    true_image[ys[-1], xs[-1], :] = (255, 255, 255)

what is correct as input points are transformed to the output and not as an inverse mapping (as in PIL).

So, I think if we keep these tests, we need to match only manually-verified set in the non-null locations of affine transformed image.
What is your opinion ?

[vfdev-5]

I refactored a little the code and tests such that I compare transformation matrices and images with accepted error of max N(=3) different pixels (cause of aliasing). With such approach, tests take about 3 mins to execute because of rather bruteforce tests:

for a in range(-90, 90, 25):
    for t1 in range(-10, 10, 5):
        for s in [0.75, 0.98, 1.0, 1.1, 1.2]:
              for sh in range(-15, 15, 5):
                   _test_transformation(a=a, t=(t1, t1), s=s, sh=sh)

Thoughts @fmassa ?

[vfdev-5]

@fmassa can we merge this ?

[soumith]

thanks @vfdev-5