FixCaffe: Training CNN with Low Precision Arithmetic Operations by Fixed Point Caffe

摘要

The convolutional neural networks are widely used in deep learning model because of its advantages in image classification, speech recognition and natural language processing. However, training large-scale networks is very time and resource consuming, because it is both compute-intensive and memory-intensive. In this paper, we proposed to use the fixed point arithmetic to train CNN with popular deep learning framework Caffe. We propose our framework FixCaffe (Fixed Point Caffe), where fixed point matrix multiply function is substitute for part of the original floating point matrix multiply function in Caffe. We analyze the range of the operands during the training process, and choose the proper scaling factor for transform floating point operands to fixed point operands. Training LeNet-S model, obtained by modifying LeNet-5, on the MNIST benchmark, the result shows that after training 1000 iterations, FixCaffe with 8-bit fixed point multiplications only leads to about 0.5% loss in the classification accuracy compared to the single-precision floating point Caffe baseline. Using Xilinx V7 690T to implement the multiplier, the cost of computing resource can save up to 83.3%, and the on-chip storage overhead for the LeNet-S model's parameters can save 75%.