/dnn-acceleration

Primary LanguageC++

Development of C programs for Convolutional Neural Network Accelerators

Purpose

Creating a new toolkit for people who are using fpga.

Prerequisites

  • python 2.7
  • tensorflow 1.10.0
  • keras(2.2)
  • numpy

System structure

Usage

1. Download the repository

git clone "git@github.com:Hamidu68/ML-acceleration.git"

2. Set condition

in ML-acceleration/config.json file, change the value of various keys.

{
  "model": "vgg19",
  "extract_configs": "False",
  "model_info": "./model_info/vgg19_test.csv",
  "sw_test": "False",
  "vivado_test": "True",
  "use_trained_weight": "False",
  "data_type": "int",
  "random_range": "5",
  "batch": "False",
  "weight_file_path": "./variables/init_weight.bin",
  "input_file_path": "./variables/init_input.bin",
  "trained_weight_file_path": "./variables/vgg19_weights.bin",
  "image_file_path": "./variables/image.bin",
  "output_path": "./output/",
  "template_path": "src/model/template/"
}
  • model : name of the model(network) (ex.vgg19, resnet50)
  • extract_configs : if you don't have layer information of the model(True, False)
  • model_info : location of layer information (ex."./model_info/vgg19.csv")
  • sw_test : keras and c code test(True/False)
  • vivado_test : generate code for Vivado HLS (True/False)
  • use_trained_weight : if you don't need variable generator (True/False)                                    
  • data_type : data type (int, unsinged int, float, ap_uint<16>)
  • random_range : number of range that will be used to variable generator (ex. 5 means input.bin, weight.bin files consist of value between 1 to 5)
  • batch : batch layer or not (True/False)
  • weight_file_path : location of weight file ("./variables/init_weight.bin")
  • input_file_path : location of input file ("./variables/init_input.bin")
  • trained_weight_file_path : location of trained weight file ("./variables/vgg19_weights.bin")
  • image_file_path : location of image input file ("./variables/image.bin.bin")
  • output_path : output path ("./output/")
  • template_path : template path ("src/model/template/")

Example

The following options can be combined at the same time.

  1. Extract layer information "extract_configs": "True"
  2. Generate random variables "use_trained_weight": "False"
  3. Run sw test (keras and C verifier) "sw_test": "True"
  4. Generate hw optimized code for Vivado HLS "vivado_test": "True"
  5. Skip Batch layer "batch": "False"

3. Run run.sh

in ML-acceleration folder, use the command below to run run.sh script file.

./run.sh

in run.sh file,

python main.py

script file 'run.sh' will run main.py file.

Ongoing work

  1. add new models(architecture)
  2. apply quantization/weight prunning