atlas-wrapper

a simple, efficient, easy-to-use HUAWEI Ascend310 Atlas wrapper for cnn with c++ and python api. You will be able use atlas wrapper to fast demo or deploy your FP16 model with few lines of code!

//c++
// create Net
atlas::Net *net = new atlas::Net(modelPath, enginePath, deviceId);
// get Net information
net->getNetInfo()
// inferece
// accept input images as vector<cv::Mat> or vector<uint8_t> with vector<int> for image shape.
// recieve the output and outputDims from atlas::NetResult
net->inference(opencv_img_vec, results);
// net->inference(uint8_t_img_vec, image_shape_vec, results);

// destroy net
delete net;

# python
# import
import pyatlas
# create net
net = pyatlas.net(modelPath, enginePath, deviceId)
# get net information
netInfo = net.getNetInfo()
# inference
result = net.DoInference(img, imgSize)

Features

• Support Atlas300. Make little change on CMakeLists.txt to support Atlas500, Atlas200SoC.
• Resize in DVPP, so you can use images of any shape in a batch inference.
• Auto alignUp for DVPP input.
• Do mean and std opteration in the first layer (AIPP)
• Dynamic graph.
• High performance data transfer between host and device.
• get net information such as batchSize, net input shapes, output dims and size.
• Python api support.
• Set device.

System Requirements

Atlas DDK B893+ and driver

opencv_world 3.4+

for python api, python 2.x/3.x and numpy in needed

Installation

Make sure you had install dependencies list above.

# clone project and submodule
git clone --recurse-submodules ...

cd atlaswrapper

bash build.sh A300 #(options: A500, A200)

then you can intergrate it into your own project with libatlasWrapper.so and Net.h, for python module, you get pyatlas.so

Model Conersion

Please follow the details from the documents of Atlas300

Demo

Please check the c++ & python demo on test folder.

Liscense

AtlasWrapper is released under the [Apache 2.0 license].

TODO

1. Add demo for model conversion.
2. Modify CMakeLists.txt to support A500, A200SoC.