search

doonny / PipeCNN

An OpenCL-based FPGA Accelerator for Convolutional Neural Networks
Apache License 2.0
1.22k stars 370 forks source link

Fixed point quantization for input/output data #64

Closed META-DREAMER closed 5 years ago

META-DREAMER commented 6 years ago

I'm working on implementing YOLO with PipeCNN, and was wondering how to determine the correct precision_config for these layers. For the weights, it is easy to determine the range of values from the caffemodel file, but for frac_din and frac_dout I am not sure how to determine these values. I read the quantization flow in the paper, but don't know how I should actually perform those steps. Any help would be much appreciated.

doonny commented 6 years ago

The featuremap should be truncated at runtime. You need modification on Darknet itself. I don't think either Keras nor TensorFlow support fixed-point convolutions, you really need to write your own code for fixed-point computations. We have done this in Matlab, writing everything from scratch using fixed-point toolbox.

doonny commented 5 years ago

I am assuming this problem has been solved.