huangyaling / caffe_boost

Caffe Boost includes scripts running demo programs of Cambricon Caffe.

Caffe Boost

Caffe Boost includes scripts running demo programs, downloading library set, and generating image list. Demo programs for both classification and detection are provided.

Prerequisites

Please clone Cambricon Caffe and Caffe Boost repository, download ImageNet, COCO and VOC2012 datasets.

Cambricon Caffe

Please download Cambricon Caffe from repository:

• For serials of MLU200:

You need to switch to the master branch.

git clone [e-mail]:Cambricon/caffe.git
git checkout master

• For serials of MLU100:

You need to switch to the release_v1.0.0 branch.

git clone [e-mail]:Cambricon/caffe.git
git checkout release_v1.0.0

Then build Cambricon Caffe. Please refer to instructions listed in Cambricon Caffe repository.

Caffe Boost

Please download Caffe Boost from repository:

• For serials of MLU200:

You need to switch to the master branch first, and see below for how to use it.

git clone [e-mail]:Cambricon/caffe_boost.git
git checkout master

• For serials of MLU100:

You need to switch to the release_v1.0.0 branch first, and see below for how to use it.

git clone [e-mail]:Cambricon/caffe_boost.git
git checkout release_v1.0.0

Datasets

ImageNet, COCO and VOC2012 validation datasets are needed to run the demo programs.so you may need to download them before running demo programs.

ImageNet

Please download ILSVRC 2012 for validation data , caffe_ilsvrc12 and ILSVRC_2015.

VOC2012

Please download VOC 2012 for validation data.

COCO

Please download COCO val2017 and annotations 2017 for validation data.

Directory Layout

The structure of directory is shown as below:

-caffe  
-caffe_boost   
  -examples  
  -scripts  
-datasets  
  -IMAGENET  
    -ILSVRC2012_img_val
    -caffe_ilsvrc12
        -ILSVRC_2015
  -VOC2012  
    -Annotations  
    -ImageSets  
    -JPEGImages  
    -SegmentationClass  
    -SegmentationObject  
  -COCO  
    -val2017  
    -annotations  
-caffe_mp_c20
  -alexnet
  -googlenet
  -inception-v3
  -mobilenet
  -resnet101
  -resnet152
  -resnet18
  -resnet34
  -resnet50
  -squeezenet
  -ssd
  -vgg16
  -vgg19
  -yolov2
  -yolov3

Demo program scripts assume above structure and respective prototxt and model file names.

Cambricon Caffe support int8 and int16 data formats.If you want to use official model or your owner model, you need to refer to userguideto quantify the model first.

Generating File List

Please generate the file list before running the example. It can be done by running script gen_val.sh. It has 3 parameters, which are dataset_type, val_path and image_path.

Usage: ./gen_val.sh [dataset_type] [val_path] [image_path]

Parameter description:
dataset_type: the type of datasets, eg:imagenet, voc, coco
val_path: the path of val.txt, this parameter is only for imagenet and voc
image_path: the path of image.

example:

   <imagenet>      ./gen_val.sh imagenet ../../datasets/IMAGENET/caffe_ilsvrc12/val.txt ../../datasets/IMAGENET/ILSVRC2012_img_val

   <imagenet_2015> ./gen_val.sh imagenet_2015 ../../datasets/IMAGENET/ILSVRC_2015/val.txt ../../datasets/IMAGENET/ILSVRC2012_img_val

   <voc>           ./gen_val.sh voc ../../datasets/VOC2012/ImageSets/Main/val.txt ../../datasets/VOC2012/JPEGImages 

   <coco>          ./gen_val.sh coco ../../datasets/COCO/val2017

Running

To run the example on platform x86, please go to examples folder. Cambricon Caffe supports running a network with multiple cores in either offline or online mode.

• run_all_offline_mc.sh: running networks with multiple cores in offline mode.
• run_all_offline_sc.sh: running networks with single core in offline mode.
• run_all_online_sc.sh: running networks with single core in online mode.
• run_all_online_mc.sh: running networks with multiple core in online mode.

Environment Setup

You need to setup some environment variables so that scripts could find Cambricon Caffe binaries and datasets. There are 2 environment variables, please setup them according to below commands:

export CAFFE_EXTERNAL=Y
export CAFFE_DIR=your_cambriconcaffe_path  // please replace your_cambriconcaffe_path with your actual path

x86

e.g. Please go to clas_offline_singlecore directory if you'd like to run networks with single core in offline mode. Argument 1 means that the network is run in int8 mode, argument MLU270 means work is run in MLU270 device.

cd caffe_boost/examples/clas_offline_singlecore
./run_all_offline_sc.sh 1 MLU270