I like the idea of a NN library with a focus on very fast graph construction. Reading about this makes me want to use the library for model-selection in the be

I like the idea of a NN library with a focus on very fast graph construction. Reading about this makes me want to use the library for model-selection in the beginning; and switch to TensorFlow later once a good model is found. The Chainer library mentioned also looks interesting; I think I've seen a few other people using it. I like that it shares some of TensorFlow's backend, though I wonder if it's only for CPU: "In order to achieve efficient execution of arithmetic operations,DyNet relies on the Eigen [28] library, which is also used to provide the backend for TensorFlow"