• Automation framework: CK
  
• Development repository: ck-request-asplos18-caffe-intel
  
• Source: GitHub
  
• Available command lines:
  
  • ck run program:convert-ssd-to-i8 --cmd_key=convert-coco (META)
    
  • ck run program:convert-ssd-to-i8 --cmd_key=convert-voc (META)
    
  • ck run program:convert-ssd-to-i8 --cmd_key=test-f32 (META)
    
  • ck run program:convert-ssd-to-i8 --cmd_key=test-i8 (META)
    
• Support for host OS: any
  
• Support for target OS: any
  
• Tags: caffe,intel,vint8
  
• How to get the stable version via the client:
  
  pip install cbench
  cb download program:convert-ssd-to-i8 --version=1.0.0 --all
  ck run program:convert-ssd-to-i8
• How to get the development version:
  
  pip install ck
  ck pull repo:ck-request-asplos18-caffe-intel
  ck run program:convert-ssd-to-i8
  

convert-ssd-to-i8

Program for conversion of SSD models into Intel Caffe quantized format.

Requirements

Intel Caffe library:

ck install package --tags=lib,caffe,intel

Caffe model to be converted into quantized format:

ck install package --tags=caffemodel,ssd

One of datasets COCO-2014 or VOC-2007 depending on which model you wish to convert:

ck install package:dataset-coco-2014
ck install package:dataset-voc-2007

Run

ck run program:convert-ssd-to-i8 --cmd_key=convert-coco

or

ck run program:convert-ssd-to-i8 --cmd_key=convert-voc

Program takes dataset and prepares LMDB database. Number of images to be included into database is governed by variable CK_IMAGES_PERCENT.

Resulting models are stored into tmp directory as test_quantized.prototxt and deploy_quantized.prototxt.

Test

Program also provide two testing commands:

ck run program:convert-ssd-to-i8 --cmd_key=test-f32
ck run program:convert-ssd-to-i8 --cmd_key=test-i8

Commands take LMDB database that was prepared at previous convert stage and run caffe test operaion using original or converted models.

TODO We should somehow interpret command output. It produce stdout like

I0228 11:02:49.898972 63713 caffe.cpp:472] Running for 171 iterations.
I0228 11:03:08.554591 63713 caffe.cpp:439]     Test net output #0: detection_eval = 0

and original Intel's benchmark.py script treats detection_eval value as mAP.

• 1.0.0 (2019-10-08)

• .cm/desc.json (3 bytes)
• .cm/info.json (313 bytes)
• .cm/meta.json (1884 bytes)
• README.md (1468 bytes)
• convert.py (9496 bytes)
• test.py (1680 bytes)
• utils.py (3613 bytes)