/Quantitative-and-ternary-caffe-for-network-compression

Quantitative and ternary caffe for network compression.

Primary LanguageC++OtherNOASSERTION

Quantitative-and-ternary-caffe-for-network-compression

Quantitative and ternary caffe for network compression.

Compiler

You need to change the configuration file (Makeconfig) according to your platform.

Usage

Once the compilation is complete, you can add the specified compression parameters to each layer to indicate the compression operations the layer needs to perform:

weights_compress:

Optional parameter: ""(default),"Ternary","Ternary_Quantize","Quantize"

weights_caompress_param:

Optional parameter:

delta: Threshold of the step function (for ternary)

alpha: Scaling factor (for ternary)

fixedpos: Fixed position (for quantize)

maxbits: Store the maximum number of quantized integers (for quantize)

activations_compress:

Optional parameter: Same as weights_compress

activations_compress_param:

Optional parameter: Same as weights_compress_param