Keras has a low-level compatibility library[1] (`from keras import backend as K`) that people have reported as useful independently.
The interface offered by K[2] seems quite similar to that of TensorFuse[3]
As the K backend is used for Keras, it also offers proof that K can be used for sizable and complex projects, plus can take advantage of pre-existing testing for Keras.
Whilst I love the idea of CGT, it has not yet taken off.
I'd be far more interested in seeing a Neon[4] backend considering it has the fastest performance across the board on existing hardware and they're planning to release their own hardware soon.
Edit: Incorrectly thought TensorFuse didn't support RNNs, thanks dementrock! Also excited that Lasagne has someone working on being backend independent!
Actually TensorFuse does support RNN by porting a subset of scan to TensorFlow. I've been using it to port Lasagne to support TensorFlow: https://github.com/dementrock/Lasagne-tf. The examples/recurrent.py there actually works.
Whilst I love the idea of CGT, it has not yet taken off. I'd be far more interested in seeing a Neon[4] backend considering it has the fastest performance across the board on existing hardware and they're planning to release their own hardware soon.
[1]: http://keras.io/backend/
[2]: https://github.com/fchollet/keras/blob/master/keras/backend/...
[3]: https://github.com/dementrock/tensorfuse/tree/master/tensorf...
[4]: https://github.com/NervanaSystems/neon
Edit: Incorrectly thought TensorFuse didn't support RNNs, thanks dementrock! Also excited that Lasagne has someone working on being backend independent!