We introduce the very first NPU compilation workflow, called NPiler, which automatically converts annotated regions of imperative code to a neural network representation. First, the programmer annotates the regions of imperative code which he/she wants to transform to a neural representation. NPiler accepts inputs from the programmer to train the network. During this step, NPiler automatically observes the input and output pairs to the annotated regions to collect training and testing data. Then, NPiler trains each possible NPU topology given constraints provide by the programmer. The outcome of this exploration provides the best possible NPU topology in terms of minimum root mean square error (RMSE) on test data. Finally, our compiler replaces the annotated regions with the final neural network representation. We use FANN library to execute the neural network representation. We released NPiler with seven representative benchmarks from diverse domains to evaluate our NPU compilation workflow.
Citation
If you use this work, please cite our paper published in MICRO'2012 in which we developed the NPU compilation workflow (NPiler) along with the seven representative benchmarks from diverse domains.
H. Esmaeilzadeh, A. Sampson, L. Ceze, D. Burger, "Neural acceleration for general-purpose approximate programs", MICRO 2012.
Sponsers
ACT Lab
