Bodo Kaiser

Requires us to use [`tf.layers.Conv3D(..., data_format='channels_first')`][1] and `mrtoct.data.transform` to change between `CDHW` and `DHWC`. [1]: https://www.tensorflow.org/versions/master/api_docs/python/tf/layers/Conv3D

Use `with tf.device('/cpu:0')` for pre processing and `with tf.device('/gpu:0')` for training and inference.

Should improve performance and be simple to enable just by passing `fuse=True` to [`tf.layers.BatchNormalization`][1]. [1]: https://www.tensorflow.org/versions/master/api_docs/python/tf/layers/BatchNormalization

If we use `padding='same'` for the first convolution layer with `kernel_size=9` we get output shape `[batch_size, 16, 16, 16, 1]` as intended from the original paper.

Requires consent on how to approach #4 as well as #6 to be solved.

Some Inspiration from [`tf.image.resize_image_with_crop_or_pad`][1]. [1]: https://github.com/tensorflow/tensorflow/blob/r1.3/tensorflow/python/ops/image_ops_impl.py#L572

Basically relates to * [ ] `mrtoct.data.transform.CenterPad` * [ ] `mrtoct.data.transform.CenterCrop` * [ ] `mrtoct.data.transform.CenterExctractSlice` * [ ] `mrtoct.data.transform.CenterExctractPatch` to check if * [ ] dimensions of input and target...

In the end we replicated some functionality of `tf.contrib.gan` with our `mrtoct.model.GeneratorSpec` approach. We should still keep this up for supporting different models but adapt `tf.contrib.gan` where possible in it.