Giters

Linranran / readme

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Evaccuracy runcnns

Introduction

The evaccuracy is used to run evaccuracy runcnn command and do the evaccuracy in batch manner. you could do evaccuracy runcnn on a example graph and then do the evaccuracy classify/detect/segment by the predefined config file. for the evacccracy runcnn/classify/detect/segment you could reference to readme.
to run evaccuracy runcnns you should do those prepare works

  1. prepare the dataset
  2. prepare the configretion file

Directory structure of dataset

the format of those dataset for classify/detect/segment please refer to imagenet,(coco,voc),(voc,cityscape). we have add another part called calibration which is used to verify the results of evaccuracy.

Classify:

ImageNet

ImageNet/ILSVRC2012(parent_folder)/
├── val
│   ├── *.JPEG images
└── 2012_val.txt/2012_val_cali_2000.txt

Detection:

  • Coco
coco(parent_folder)/
├── val2017
├── test2015
├── val2014
│   ├── *.JPEG images

|─ cocoval2017.txt
  • Voc
VOCdevkit(parent_folder)/
├── VOC2007
│   ├── ImageSets
│       ├── *.JPEG images
│   ├── Annotations
│   ├── SegmentationObject
│   ├── JPEGImages
│   ├── Cali
|─ voc2007_test_imglist.txt/ voc2007_cali_500.txt

Segment:

  • Cityscape
cityscapes(parent_folder)/
├── leftImg8bit
│   ├── *.JPEG images
├── gtCoarse
├── ICNet
|─ icnet_fineval.txt
  • voc
VOCdevkit(arent_folder)/
├── VOC2007
│   ├── ImageSets
│   	├── *.JPEG images
│   ├── Annotations
│   ├── SegmentationObject
│   ├── JPEGImages
│   ├── Cali
|─ seg11valid_test.txt/ seg11valid_test_cali_100.txt
  • CamVid
CamVid(parent_folder)/
├── train
├── trainannot
├── val
├── valannot
│   ├── ImageSets
|─ test.txt/ val.txt

How to run

1 the command

usage: evaccuracy runcnns [-h] --config CONFIG [--output OUTPUT]

MWEV cnn graph examples runner with graph output layer binary dumping function

optional arguments:
-h, --help       show this help message and exit
--config CONFIG  Example runner configuration file in json format
--output OUTPUT  Path of folder to save cnn runner results
  • the config means the configration file that predefined in the input json file. the json file contain the parameters about the evaccuracy runcnn and evaccuracy classify/segment/detect
  • the output is the output folder of evaccuracy runcnns, it contains the results of ecaccuracy runcnn and evaccuracy classify/detect/segment.

2 the format of input json file

The json file contain the configuration information that guide the evaccuracy runcnn and evaccuracy classify/detect/segment，and we will give the format of json file detailly. The json file could divide into two parts, The first part is the global configuration, the second part is local configuration for local examples, when a parameter have been defined in both global and local part, the local part configuration have a high level component. for the means of these pararemeters, please refere to the doc files[].

  1. The parameters of global part
    the gloabal part have the following parameters:
  • run_all_model_names: the option is used to contral whether run all the model exist in example/configs
    • type: string type, choose False or True and the default is False
  • maps_format: this parameters decide how much bit-dumps the evaccuracy runcnns will done, and we only support 8 and 12 bit now.
    • type:: list type, it should be like ["8bit"], ["12bit"] or ["8bit", "12bit"]
  • make_options:
    • type:
  • ev_config: there are four option, they are EVSS_CFG, CNN_ABSTRACTION,EVSS_DBG, run_haps
    • type: list type, you should give a list like ["ev_native", "host_fixed", 0, false]
  • jobs: how many jobs
    • type: int type
  • fixed_dumpdir_root: the root path of fixed_dump
    • type: string type
  • evgencnn_options: we provide two kinds options, tofinno and toronto. for toronto you should add extral parameters
    • type: string type, {"tofino": "", "toronto": "--acc_shift 0 --Zweight_scale true --Zblob_scale true"},
  • target: build option, clean-install or install
    • type: string type, the below is a global config example
 "global_config" : {
        "run_all_model_names": false, 
        "maps_format": ["12bit", "8bit"], 
        "evgencnn_options": {"tofino": "", "toronto": "--acc_shift 0 --Zweight_scale true --Zblob_scale true"},
        "ev_config":
            ["ev_native", "host_fixed", 0, false]
        ], 
        "config": [], 
        "jobs": 10, 
        "cnn_dataset_root": "/slowfs/us01dwt2p832/datasets",
        "target": "clean install"
    },
  1. The local part
    The local part configuration is separate for each graph in cnn_tools/example, and you could add some other parameters combine with gloable parameters to run evaccuracy runcnn.the local part have the following parameters:
  • dumpdir: the dumped file will save in the dumpdir folde
    • type: string type
  • model_name: different model in example/configs. this option has the same function will run_all_model_names in the global part.
    • type:: list type, ["defualt", "compress"] etc.
  • imgfolder: the path for test dataset
    • type: string type
  • imglist: the path for test dataset lables
    • type: string type
  • cali_imgfolder: the path for calibraton dataset
    • type: string type
  • cali_imglist: the path for calibraton dataset lables
    • type: string type
  • maxcnt: sample maxcnt number images for test to accalerate the inference process.
    • type: int type
  • evaccuracy: after run evaccuracy runcnn, we could get the demped files. and the we will use evaccuracy classify/detect/segment to test the accuracy of the dumped file. and the evaccuracy is an extral option to cover differen task, because of the different input parameters for different task (evaccuracy runcnn/classify/detect/segment), we use the option cover the difference of different task. there are three child option in evaccracy option.
    • type: define which kind of task you want to do (classify/detect/segment )
      -- type: string type, "classify", "detect" or "segment"
    • config_file: the config_file is the predefine parameters for differen graphs and you could found the config file in the evgencnn/scripts/evaccuracy/imagenet etc.
      -- type: string type, like "vgg.cfg"
    • extra_options: for evaccuracy classify/detect/segment, because of they are differents task and the commands is little differen ,we use the exa_options to cover the difference. when run evaccuracy classify/segment the extra_options is null and for evaccuracy detect, you should add extral parameters
      -- type: string type, like "--eval_type voc2007 --dataset_type test --min_overlap 0.5"
for 'evaccuracy' there are three child_options, 
* type: classify/detect/segment
* config_file: you could found the config file in the evgencnn/scripts/evaccuracy/imagenet etc. 
* extra_options: for evaccuracy classify/detect/segment, because of they are differents task and the commands is little differen ,we use the exa_optio ns to cover the difference. when run evaccuracy classify/segment the  exa_options is null and for evaccuracy detect, you should add ext ral parameters
the below is a simple example of json file
~~~
{
    "global_config" : {
        "run_all_model_names": false, 
        "maps_format": ["12bit", "8bit"], 
        "evgencnn_options": {"tofino": "", "toronto": "--acc_shift 0 --Zweight_scale true --Zblob_scale true"},
        "ev_config":
            ["ev_native", "host_fixed", 0, false]
        ], 
        "config": [], 
        "jobs": 10, 
        "cnn_dataset_root": "/slowfs/us01dwt2p832/datasets",
        "target": "clean install"
    },
    "examples": {
        "examples/classification/GoogleNet" : { "subset_10": {
            "imgfolder": "ImageNet/ILSVRC2012",
            "imglist": "ImageNet/ILSVRC2012/2012_val.txt",
            "maxcnt": 10,
            "evaccuracy" : {
                "type": "classify",
                "config_file": "googlenet.cfg",
                "extra_options": ""
            }
        }, "subset_10_cali": {
            "imgfolder": "ImageNet/ILSVRC2012",
            "imglist": "ImageNet/ILSVRC2012/2012_val.txt",
            "cali_imgfolder": "ImageNet/ILSVRC2012",
            "cali_imglist": "ImageNet/ILSVRC2012/2012_val_cali_1000.txt",
            "maxcnt": 10,
            "evaccuracy" : {
                "type": "classify",
                "config_file": "googlenet.cfg",
                "extra_options": ""
            }
        } },
"examples/detection/MobileNet-SSD" : { "voc2007_test_fullset": {
            "imgfolder": "VOCdevkit/VOC2007",
            "imglist": "VOCdevkit/voc2007_test_imglist.txt",
			"maxcnt": 10,
            "evaccuracy" : {
                "type": "detect",
                "config_file": "mobilenet_ssd_300x300.cfg",
                "extra_options": "--eval_type voc2007 --dataset_type test --min_overlap 0.5"
            }
        }, "voc2007_test_fullset_cali_500": {
            "imgfolder": "VOCdevkit/VOC2007",
            "imglist": "VOCdevkit/voc2007_test_imglist.txt",
            "cali_imgfolder": "VOCdevkit/VOC2007",
            "cali_imglist": "VOCdevkit/voc2007_cali_500.txt",
			"maxcnt": 10,
            "evaccuracy" : {
                "type": "detect",
                "config_file": "mobilenet_ssd_300x300.cfg",
                "extra_options": "--eval_type voc2007 --dataset_type test --min_overlap 0.5"
            }
        } }
    }
}
~~~
 
# Command Output
Evaccuracy runcnns will create the following file 


1. evaccuracy_analyzed.json

this file contain the parameters of each example

2. evaccuracy_analyzed.xls

this file contain the results of evaccuracy classify/detect/segment

3. evaccuracy_runcnns.json
4. evaccuracy_runcnns.log

this file contain the results show in the terminal,

5. evaccuracy_runcnns_config.json

About