MXNET-Scala 实现估算卷积神经网络浮点数运算量

MXNET-Scala Useful Tools

Implementation of the estimation of model size and flop counts for convolutional neural networks with MXNET-Scala.

https://github.com/albanie/convnet-burden

For now, the estimation of flops only consider Layers: Convolution, Deconvolution, FullyConnected, Pooling, relu

Building

Tested on Ubuntu 14.04

Requirements

• sbt 0.13
• Mxnet

steps

1, compile Mxnet with CUDA, then compile the scala-pkg;

2,

cd Mxnet-Scala/UsefulTools
mkdir lib

3, copy your compiled mxnet-full_2.11-linux-x86_64-gpu-1.3.1-SNAPSHOT.jar into lib folder;

4, run sbt, compile the project

Running

run cal_flops.sh under scripts folder

caffenet
flops: 723.0072 MFLOPS
model size: 232.56387 MB

squeezenet1-0
flops: 861.60394 MFLOPS
model size: 4.7623596 MB

resnet-101
flops: 7818.2407 MFLOPS
model size: 170.28586 MB

resnext-101-64x4d
flops: 15491.882 MFLOPS
model size: 319.13058 MB