SJTU CS433 Parallel and Distributed Programming

Professor

Xiaoyao Liang

Basic Info

An optimized MobileNet-V2 Model implemented by Cuda C

My Work

Set up MobileNet-V2 Model Structure and Test Binary Parameters in Each Layer, Write all Convolution(Conv2d, Depth-wise Conv and Point-wise Conv) Layers and Matrix Gemm Methods.

Structure

\Codes
	\models
		-- mobilenet_v2.onnx	onnx model files
	\parameters
		-- bias_data.bin	model bias	(binary)
		-- bias_data.txt	model bias	(text)
		-- weight_data.bin	model weights	(binary)
		-- weight_data.txt	model weights	(text)
	\tmpfiles	Store Layer Outputs, Just for Test
	\preprocess
		-- CheckData.py	check layer output
		-- CodeGenerateFunctions.py	generate codes
		-- Img2Col.py	image to column methods
		-- InferModel.py infer standard layer data for check
		-- ReadONNX.py	read onnx model and store weights and bias
	-- mobilenet_main.cc	program entrance
	-- layers.cuh	model layers headfile
	-- layers.cu	model layers implements
	-- init_model.cuh	model init headfile
	-- init_model.cu	model init implements
	-- mobilenetInput.txt	Input Image Data
	-- mobilenetOutput.txt	Standard Outputs

Running

Please put mobilenetInput.txt, mobilenetOutput.txt in /Code/ folder and put weight_data.bin and bias_data.bin in /Code/parameters/ folder.

• mobilenetInput.txt contains a batch of images data, each line represents an images (c * h * w).

• mobilenetOutput.txt contains inference result from model/mobilenet_v2.onnx model, which can be seen as ground truth. The dimension of each line is 1000.

• weight_data.bin is W of WX + b for each layer; bias_data.bin is b of WX + b for each layer. These two binary files can be gotten from preprocess/ReadONNX.py, from where you can get two text files and then convert them to binary files.

​ Then execute in terminal:

$ make
$ ./mobilenet_main