Compress caffemodel using multiple methods written by Dezhi Jiang.
K-means: Alexnet down to 14M. quantize(float32 to int8): Alexnet down to 41M.
K-Means: We use 6-bit for Convolution layer weights which means we will use 64 clusters to represent all weights per convolution layer. We use 2-bit for fc layer. So, every convolution layer weight fp number is stored as 6-bit and 32-bit can store 5 ones. To fc layer weight, the number is 16.
Round down: We also tried another quantization method(round quantization). Every float point number will be converted to a unint8 number.
This is a code written by python + cpp. To Compile cpp code: ./build.sh To just run caffemodel_compress.py: python caffemodel_compress.py [FileOption] [ConvBit] [FcBit] [FileOption]: alex vgg16 vgg19 resnet50 [ConvBit]: better be 6 [FcBit]: better be 2
Example: python caffemodel_compress.py resnet50 6 2 2>&1 | tee resulttmp.md This command compress resnet50.caffemodel using parameter 6 and 2.The caffemodel path is configed in compress_resnet50(convbit=6, fcbit=2) in caffemodel_compress.py. Some Explanation: This code configs origin caffemodel and compressed *zip.npz path in hard code, it compress *.caffemodel to *zip.npz, than decompress *zip.npz to *_xx.caffemodel. Those 3 file paths can be found in code.