sabipeople / tegra

cnn

cnntest

how to install digits 1. git clone https://github.com/NVIDIA/DIGITS.git 2. digits server "https://localhost" or https://127.0.0.1:5000

digits functions: 1. create database to train and test 1.1. create database 1.1.1 connect to https://127.0.0.1:5000 1.1.2 move to [datasets]->[new dataset]->[images]->[classification]->[input username-whatever you want]->[set image type and image size]->[move to use image folder]->[set training image->put /path/to/train/data/*.jpg]->[set min sample per class and % for validation]->[set db backend-the way how to treat database]->[set image format]->[set dataset name]->[create] 1.2. create model -> lenet for mnist db

	2. finetuning
		2.1 train model or get trained model
		2.2 digits->[new model]->[classification]->[select dataset]->[move to previous network and choose one]->[click customize(it will be shown in the right side of the selected network]->[add prototxt]
	3. python layer
		3.1 digits->[new model]->[classification]->[select dataset]->[user side python add *.py file]->[move to previous network and choose one]->[click customize(it will be shown in the right side of the selected network]->[add prototxt]
		3.1.1
				layer {
				  name: "blank_square"
				  type: "Python"
				  bottom: "scale"
				  top: "scale"
				  python_param {
				    module: "digits_python_layers"   #this module name must be same with python file name which defines the player
				    layer: "BlankSquareLayer"
				  }
				  include {
				    phase: TRAIN
				  }
				}

tensorRT type: optimizer properties: -use network parameter from caffe and change the parameter to int8 , float16 so that user can reduce the inference time

inference : 1. download pretrained model 2. unzip tar.gz(model zip file) in the proper folder 3. set env_var=/path/to/unzip/file 4. deploy.prototxt -> forward network architecture, original.prototxt -> original training network architecture deploy.prototxt->forward network architecture
snapshot_iter_22620.caffemodel -> weight matrix info.json
solver.prototxt -> mean.binaryproto -> mean image used in training phase
train_val.prototxt ->

---

ps

caffe install

1. Set up the Caffe environment.

   a) Install packages from APT with the following commands:

    $ sudo add-apt-repository universe
    $ sudo add-apt-repository multiverse
    $ sudo apt-get update
    $ sudo apt-get install libboost-all-dev libprotobuf-dev libleveldb-dev libsnappy-dev
    $ sudo apt-get install libhdf5-serial-dev protobuf-compiler libgflags-dev libgoogle-glog-dev
    $ sudo apt-get install liblmdb-dev libblas-dev libatlas-base-dev
   

   b) Download Caffe source package from the following website:

    https://github.com/BVLC/caffe
   
   And copy the package to $HOME directory on the target board
   with the following command:
   
    $ mkdir -pv $HOME/Work/caffe
    $ cp caffe-master.zip $HOME/Work/caffe/
    $ cd $HOME/Work/caffe/ && unzip caffe-master.zip
   

   c) Build Caffe source with the following commands:

    $ cd $HOME/Work/caffe/caffe-master
    $ vi Makefile.config.example
   
    Uncomment the following line to enable cuDNN acceleration:
   
        USE_CUDNN := 1
   
    And modify the following two lines, save, and exit.
   
        INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial
        LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/aarch64-linux-gnu/hdf5/serial
   
    $ cp Makefile.config.example Makefile.config
    $ make -j4
   
    The library libcaffe.so is generated in the build/lib directory.