shellhue / Cifar10

This is an interview project for AI engineer of a SECRET COMPANY.

Targets

• reproduce the results of resnet on Cifar10 with pytorch.
• convert pytorch model to caffe2 model
• predict image with c++ project organized by cmake

Reproduce

Reproducing is easy, just implement the resnet and train it with suggested super parameters.

To evaluate the accuracy on test dataset for resnet20, just run:

git clone https://github.com/shellhue/Cifar10.git && cd Cifar10
python evaluate.py --layers=20 --weights='./weights/resnet20_164.pth'

The corresponding error is:

8.33%  (target is 8.75%)

To evaluate the accuracy on test dataset for resnet56, just run:

git clone https://github.com/shellhue/Cifar10.git && cd Cifar10
python evaluate.py --layers=56 --weights='./weights/resnet56_164.pth'

The corresponding error is:

6.83%  (target is 6.97%)

Convert to caffe2

To convert pytorch model to caffe2 model, two steps are needed.

First, convert pytorch model to onnx model:

python convert2onnx.py --layers=20 --pretrained_weights='./weights/resnet20_164.pth' // resnet20.onnx will be created

Second, convert onnx model to caffe2 model:

python onnx2pb.py --layers=20 --onnx_proto_file='pathToOnnxProtoFile' // onnx-init-20.pb and onnx-predict-20.pb will be created

Predict with c++

I haven’t finished this. But i know how to do it. Just follow caffe2_cpp_tutorial, and then change the pretrained.cc in dir /src/caffe2/binaries/ to use the onnx-init-20.pb onnx-predict-20.pb files. Unfortunately, it is hard to make caffe2_cpp_tutorialrun. I find that installing caffe by building the source is needed, but the installation is very slow. And i have no time!!