implement Conv2DTranspose
reduce memory usage during model loading https://www.reddit.com/r/cpp/comments/7c91n0/frugallydeep_a_headeronly_library_for_using_keras/dq529a6/
empty node-result cache when possible to save RAM?
add tests for Conv2DTranspose with dilation when it actually supports it: keras-team/keras#8159
test also with tf.keras