/Gomah.jl

model converter from Chainer to Flux

Primary LanguageJuliaMIT LicenseMIT

Gomah.jl

Stable Dev CircleCI Build Status Codecov Coveralls

About this repo

  • This repo provides train/inference procedure of MNIST model known as Hello World of Deep Learning on Julia runtime. These techniques are based on Chainer and PyCall.jl.
  • NEW: Add feature to convert ResNet50/Chainer -> ResNet50/Flux.jl

Usage

How to install

  • This package is not registered as official julia package, so called 野良(nora), which means we should specify repository url:
  • Note that Package Gomah.jl depends on PyCall.jl. So before installing, We recommend set environment variable in Julia.
$ julia
julia> ENV["PYTHON"] = Sys.which("python3")
pkg> add https://github.com/terasakisatoshi/Gomah.jl.git
julia> using Gomah

Call Chainer script from Julia via Gomah.jl

  • PyCall.jl provides interface between Python and Julia.
  • This means you can construct training script of Chainer on Julia environment.
  • If you are familiar with some Deep learnig framework, checkout our src/mnist.jl.
  • We provide an example of training MNIST classifier.
$ julia
julia> using Gomah
julia> train()
epoch       main/loss   main/accuracy  validation/main/loss  validation/main/accuracy  elapsed_time
1           0.56446     0.840621       0.28946               0.914727                  3.49408       
2           0.248238    0.927429       0.212789              0.937381                  5.90154       
3           0.189395    0.945183       0.175694              0.948423                  8.43833       
4           0.152917    0.95592        0.145494              0.957898                  10.832        
5           0.128008    0.963307       0.128785              0.962386                  13.2714       
6           0.108765    0.968306       0.121768              0.961222                  15.6906       
7           0.0948146   0.97286        0.103434              0.969201                  18.361        
8           0.0854993   0.974762       0.10229               0.96818                   20.7444       
9           0.0746463   0.977655       0.0916977             0.972739                  23.1146       
10          0.0663983   0.980598       0.0889726             0.972573                  25.5239    
julia> predict()
accuracy for test set = 97.31 [%]

Convert ResNet/Chainer -> ResNet/Flux.jl

  • We found the structure (shape) of parameter i.e. weight of Chainer is similar that of Flux.
  • The structure of weight of Convolution of Chainer is NCHW. On the other hand, Conv of Flux.jl has weight its shape is WHCN, where N is batchsize, H (resp. W) is height (resp. width) of kernel and C is num of channel.
  • We provided script to convert ResNet50 of Chainer to that of Flux.jl
  • Here is a example of How to use converted model. What you have to do is ...
    • Install chainer and chainercv
    • Install Flux.jl, PyCall, Gomah.jl
    • Prepare sample RGB image. e.g. pineapple.png
    • Run the following the script.
using Gomah
using Gomah: L, np, reversedims
using Flux
using PyCall

using Test
using BenchmarkTools

py"""
import chainer
import chainercv
import numpy as np
num = 50
PyResNet = chainercv.links.model.resnet.ResNet
resnet = PyResNet(num, pretrained_model = "imagenet")
img=chainercv.utils.read_image("pineapple.png",dtype=np.float32,alpha="ignore")
img=chainercv.transforms.resize(img,(224,224))
_imagenet_mean = np.array(
            [123.15163084, 115.90288257, 103.0626238],
            dtype=np.float32
        )[:, np.newaxis, np.newaxis]
img=img-_imagenet_mean
img=np.expand_dims(img,axis=0)
resnet.pick=resnet.layer_names
with chainer.using_config('train', False):
    pyret=resnet(img)
    result=np.squeeze(pyret[-1].array)
    chidx=int(np.argmax(result))
    chprob=100*float(result[chidx])
    print(chidx)
    print(chprob)
"""

@testset "regression" begin
    num = 50
    myres = ResNet(num)
    Flux.testmode!.(myres.layers)
    img = reversedims(py"img")
    @show size(img), typeof(img)
    ret, name2data = myres(img)
    for (i,name) in enumerate(myres.layer_names)
        pyr = reversedims(py"pyret[$i-1].array")
        flr = name2data[name]
        @show name, size(flr)
        @test size(pyr) == size(flr)
        @show maximum(abs.(pyr .- flr))
    end
    flidx = argmax(ret)
    flprob = 100ret[argmax(ret)]
    @show flidx,flprob
    @test Int(py"chidx") == flidx[1]-1
    @show Float32(py"chprob") - flprob
end

@testset "benchmark" begin
    num=50
    img = reversedims(py"img")
    myres = ResNet(num)
    chainmodel = Chain(myres.layers...)
    Flux.testmode!(chainmodel)
    @time chainmodel(img)
    @time chainmodel(img)
    @time chainmodel(img)
    @time chainmodel(img)
    @time chainmodel(img)
    @time chainmodel(img)
end

Why Gomah(ごまぁ）?

  • My favorite thing ごまちゃん
  • Inspired by Menoh
  • Gomah.jl will be promoted as DNN inference library (in the future)

Acknowledgement