Benchmark on Deep Learning Frameworks€¦ · Line # Hits Per 8234. 1ø5ø4. 565. 6512. 5ø17. 368....

Benchmark on Deep Learning Frameworks

김 형 준

https://mli.github.io/cvpr17/gluon_part1.pdf

https://mli.github.io/cvpr17/gluon_part1.pdf

빨강 파랑 노랑

파랑 노랑 빨강

Framework Benchmarkhttps://github.com/tensorflow/benchmarks/tree/master/scripts/keras_benchmarks

https://github.com/awslabs/keras-apache-mxnet/tree/keras2_mxnet_backend/benchmark

Tensorflow CPU (FULL-CORE)

GPU(1-GPU)

DNN(MLP)

CNN

LSTM

RN

MXNET CPU (FULL-CORE)

GPU(1-GPU)

DNN(MLP)

CNN

LSTM

RN

keras-mxnet

File System

[CPU]

ubuntu 16.04

tensorflow==1.8.0 mxnet==1.2.0 keras=2.2.0

kears-mxnet=2.2.0

[CPU]

ubuntu 16.04

tensorflow==1.8.0 mxnet-mkl==1.2.0

keras=2.2.0 kears-mxnet=2.2.0

[GPU]

ubuntu 16.04 (cuda9.0, cudnn7)

tensorflow-gpu==1.8.0 mxnet-cu90==1.2.0

keras=2.2.0 kears-mxnet=2.2.0

DNN(MLP) CNN LSTM

Between Platform - CPU

DNN_TRAIN DNN_INFER(ms) (ms)

Tensorflow - CPU

mxnet-mkl - CPU

mxnet-mkl - CPU

CNN_TRAIN CNN_INFER

LSTM_TRAIN LSTM_INFER

Between Platform - GPU

DNN_TRAIN DNN_INFER(ms) (ms)

Between Platform - GPU

CNN_TRAIN CNN_INFER

LSTM_TRAIN LSTM_INFER

Tensorflow - CPU

Tensorflow - GPU

mxnet - GPU

nvprof

Thank you