lpa_cnn

Low Precision Arithmetic For Convolutional Neural Network Inference

lpa_cnn is a benchmarking tool for comparing accuracies and speeds of convolutional neural networks run with different arithmetic precision modes for the convolutions. The first mode is the baseline Caffe implentation, the second is floating point arithmetic with eigen, and the third is quantized mode, which uses integer airthmetic through gemmlowp.

Dependencies

gcc 5.4 w/ Eigen 3 & Armadillo

Python 2.7 w/ NumPy & PIL

R w/ gtools & stringr

Caffe (see Setup below for installation)

Setup

Install Caffe as caffe/ (in root directory), following the guide @ https://chunml.github.io/ChunML.github.io/project/Installing-Caffe-CPU-Only/.

Have the following files in place for each desired model:

  models/<model_name>/<model_name.caffemodel>
  models/<model_name>/<model_name.prototxt>

adjusting the .prototxt input layer to receive one image as follows:

  1 x <depth> x <width> x <height>

Have the following input file in place for each installed model:

  inputs/<model_name>/production/<input_file_name.csv>

having the form:

  <img_0_label><img_0_channel_1>...<img_0_channel_2><img_0_channel_3>
  <img_1_label><img_1_channel_1>...<img_1_channel_2><img_1_channel_3>
  ...

Reproduction

To run experiments with the installed models, call $ bash run_routine.sh.

Results are written to results/.

Installing new models

A great resource for finding new Caffe models is Model Zoo @ https://github.com/BVLC/caffe/wiki/Model-Zoo

To install a new model, follow the Setup directions above, providing an appropriate and consistent model name as <model_name>.

NOTE that when preparing .prototxt files, lpa_cnn supports the following parameters:

  layer_types = ['convolution', 'pooling', 'relu', 'eltwise', 'innerproduct', 'scale', 'batchnorm']
  param_types = ['num_output', 'pad', 'kernel_size', 'stride', 'bias_term', 'pool']
  special_types = ['shape', 'input_dim']
  shape_dims = ['n','d','w','h']

NOTE that batch processing is not supported.