Neural network training for accelerator with computing errors

With the advancements of neural networks, customized accelerators are increasingly adopted in massive AI applications. To gain higher energy efficiency or performance, many hardware design optimizations such as near-threshold logic or overclocking can be utilized. In these cases, computing errors may happen and the computing errors are difficult to be captured by conventional training on general purposed processors (GPPs). Applying the offline trained neural network models to the accelerators with errors directly may lead to considerable prediction accuracy loss.

To address the above problem, we opt to integrate the CNN accelerator into Caffe such that the application data and the computing errors can be trained and exposed in the rsulting model. In this project, we take overcloked CNN accelerator as an example. Basically, the CNN accelerator (PipeCNN) is overclocked regardless the timing errors. While we have forward computing deployed on the overclocked CNN accelerator and backward computing on host CPU, the retrained models can tolerate the computing errors according to our experiments on ImageNet.

User guide:

1. Install Caffe and Xilinx SDAccel 2017.1 on ubuntu16.04.

2. Compile PepeCNN models to bitstream. PipeCNN can be found in https://github.com/doonny/PipeCNN. We have compiled bitstream stored in ./bitstreams. You can find the implementations with different frequency. Relace "binary_container_1.xclbin" file with the one you want to test. Do keep the "binary_container_1.xclbin" when you change the bitstream.

3. Update SDAcccel path to Makefile, run "make all" completes the compilation

4. Download AlexNet model from http://dl.caffe.berkeleyvision.org/bvlc_reference_caffenet.caffemodel to ./alexNet You can compile the model and generate the bistream by yourself.

5. Setup the ImageNet dataset path in train_val_caffe.prototxt

6. Run “alexNet/alexnet.sh" to performance the inference.

Reference:

[1] Xing, Kouzi, Dawen Xu, Cheng Liu, Ying Wang, Huawei Li and Xiaowei Li, Squeezing the Last MHz for CNN Acceleration on FPGAs, The 3rd International Test Conference in Asia, 2019

[2] Xu, Dawen, Kouzi Xing, Cheng Liu, Ying Wang, Yulin Dai, Long Cheng, Huawei Li, Lei Zhang, Resilient Neural Network Training for Accelerators with Computing Errors, The 30th IEEE International Conference on Application-specific Systems, Architectures and Processors (ASAP), July 15-17, 2019