Refactor DARTS code to support non-square images

[JiaweiZhuang]

The original DARTS implementation (https://github.com/quark0/darts) requires the input image to be a perfect n*n square. Otherwise it crashes as:

RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 7 and 8 in dimension 2 at /pytorch/torch/lib/THC/generic/THCTensorMath.cu:111

This is fine for CIFAR (32x32) and MNIST (28x28), but most scientific images won't be a perfect square. The graphene data is 3x5 / 30x80 for example.

If the code is too hard to refactor, we can also resample all input images to a perfect square, which is less ideal.

[JiaweiZhuang]

Ideally there should be a parameterized test for the DARTS code to check all kinds of input shapes. This can be done more easily if the code runs on CPU (#8).

Here's how the test code would look like, in an abstract level:

import pytest
from darts import train_search  # some sort of DARTS wrapper; TBD

# (channel, x, y)
shape_list = [
    (3, 32, 32), # CIFAR
    (3, 28, 28), # MNIST
    (1, 3, 5),   # coarse-grid graphene
    (1, 30, 80)  # fine-grid graphene
    ]  # add more shapes here

@pytest.mark.parametrize('input_shape', shape_list)
def test_input_shape(input_shape):
    train_search(input_shape)  # should train on synthetic numpy data with this shape

Such test can be particularly useful when applying DARTS to other scientific data. New datasets are basically just numpy arrays with a different shape. The actual data content doesn't matter much, in terms of making DARTS running.

[JiaweiZhuang]

@dylanrandle You should be able to run the test suite on GPU via #10.

[dylanrandle]

Solved via https://github.com/capstone2019-neuralsearch/darts/commit/c54e3b4b2b70eb572f2ca159fd9935f4716deb0c