jungin500/snpe repository overview

Unofficial Dockerized Qualcomm Neural Processing SDK (SNPE - Snapdragon Neural Processing Engine)

⁠Description

• I've packaged Qualcomm's SNPE SDK into dockerized environment, as installing Caffe over different environment took ALMOST FULL DAY for me! (BVLC/caffe didn't supplied python wheel file, and caffe V1 repository was deprecated as there was no updates in recent 2 years!)

⁠License

• Qualcomm's PKLA(Product Kit License Aggrement) indicates that SDK should not be included in open source in any forms (refer to the license text:

  3.5. Open Source Prohibition.
  LICENSEE shall not ... incorporate, link, distribute or use any third party software or code in conjunction with any part of a PKLA Product Kit ...
  

  Copy
• If you agree the Full License⁠, you could download⁠ SNPE SDK. Recommending to use it for educational/study purpose.

⁠Requirements

• Windows(Docker Desktop with WSL2), Linux
• Docker
• Qualcomm Neural Processing SDK for AI v1.53.2 (or newer): Download snpe-1.53.2.zip⁠
  • WARNING: Version under (or same as) SNPE 1.48.0 would not work. Qualcomm changed host platform from Ubuntu 16.04 to 18.04, and a lot of libraries are changed. Try at your own adventure.
• Boost 1.77.0 source: Download boost_1_77_0.tar.gz⁠
  (Be aware that boost's distribution platform jfrog didn't let me automatically download their sourcecode using wget!)

⁠Usage

1. Download snpe-1.53.2.zip or newer.

2. Create new empty directory and unzip file snpe-1.53.2.zip into that directory.

3. Run following command (${SNPE_ROOT} refers to current directory, you could replace it with $(pwd) on linux, or full path on windows WSL2 environment)

   docker run \
       -it --rm \
       -v ${SNPE_ROOT}:/snpe \
       jungin500/snpe:1.5x
   

   Copy

4. (Optional) Run examples inside SNPE SDK's /snpe/models directory:

   cd /snpe
   
   # It will take some time (5~10mins) to download AlexNet from Caffe repository
   python models/alexnet/scripts/setup_alexnet.py -a models/alexnet/assets -d
   

   Copy

   Output should show up like following:

   ... (Download sequence)
   
   Copying Caffe model
   Modiying prototxt to use a batch size of 1
   Creating DLC
   2021-09-09 08:45:54,886 - 188 - INFO - across channels=True
   2021-09-09 08:45:54,887 - 188 - INFO - across channels=True
   2021-09-09 08:45:55,901 - 188 - INFO - INFO_DLC_SAVE_LOCATION: Saving model at /snpe/models/alexnet/dlc/bvlc_alexnet.dlc
   2021-09-09 08:46:18,776 - 188 - INFO - INFO_CONVERSION_SUCCESS: Conversion completed successfully
   Getting imagenet aux data
   Creating ilsvrc_2012_mean.npy
   Creating %s ilsvrc_2012_mean_cropped.bin
   Creating ilsvrc_2012_labels.txt
   Create SNPE alexnet input
   processing /snpe/models/alexnet/data/chairs.jpg
   processing /snpe/models/alexnet/data/notice_sign.jpg
   processing /snpe/models/alexnet/data/plastic_cup.jpg
   processing /snpe/models/alexnet/data/trash_bin.jpg
   Create file lists
   /snpe/models/alexnet/data/cropped/raw_list.txt created listing 4 files.
   /snpe/models/alexnet/data/target_raw_list.txt created listing 4 files.
   Setup alexnet completed.
   

   Copy

   and result model .dlc file would be created on models/alexnet/dlc/bvlc_alexnet.dlc:

   (snpe) root@fc6441ac1308:/snpe# ls -alh models/alexnet/dlc/
   total 233M
   drwxrwxrwx 1 1000 1000 4.0K Sep  9 07:18 .
   drwxrwxrwx 1 1000 1000 4.0K Sep  8 08:40 ..
   -rwxrwxrwx 1 1000 1000 233M Sep  9 08:46 bvlc_alexnet.dlc
   

   Copy

5. You can use SNPE's python tools listed below:

   snpe-caffe-to-dlc
   snpe-caffe2-to-dlc
   snpe-diagview
   snpe-dlc-diff
   snpe-dlc-info
   snpe-dlc-quantize
   snpe-dlc-reorder
   snpe-dlc-viewer
   snpe-net-run
   snpe-onnx-to-dlc
   snpe-parallel-run
   snpe-platform-validator-py
   snpe-tensorflow-to-dlc
   snpe-tflite-to-dlc
   snpe-throughput-net-run
   snpe-udo-package-generator
   

   Copy