Skip to content

Latest commit

 

History

History
92 lines (59 loc) · 3.66 KB

darknet.md

File metadata and controls

92 lines (59 loc) · 3.66 KB

Darknet

Joseph Redmon has published some CNN models on his website. Joseph calls his project Darknet, which sounds spooky, but his models are harmless, and really quite fast.

We have provided a Python script that can load those Darknet models. See the darknet_to_ell Python module (found in the tools/importers/darknet directory)

After building ELL and the ELL Python Language Bindings you will find the following folder in your build output. So from your conda environment go into that directory:

cd build/tutorials/vision/gettingStarted

Then download the Darknet ImageNet reference model as follows:

curl -O https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/darknet.cfg
curl -O https://pjreddie.com/media/files/darknet.weights

If your conda environment doesn't have curl just run this:

conda install curl

Class Labels

You will notice we have already provided the following file which contains a list of class labels that correspond to the model prediction output.

darknetImageNetLabels.txt

In this file you will see all sorts of fun classes from dog breeds to birds, fire truck, school bus, and many more. It has 1000 classes it can predict.

Reference Implementation

Ok, so now you can run the following Python script to load up this model and test it out with the ELL Reference Implementation. This will run on your PC (not on the Raspberry Pi yet), just to test if everything is working:

python darknetDemo.py

Note: if you have more than one video camera on your PC and you want to switch to a different one you can change the camera index by passing a command line argument to the script as follows:

python darknetDemo.py 1

You should see some output like this while it is loading the model:

(py35) D:\temp\ELL\tutorials\vision\gettingStarted>python darknet.py 1
convolutional:  224 x 224 x 3 ->  224 x 224 x 16 , pad  1
max_pool:  224 x 224 x 16 ->  112 x 112 x 16 , pad  0
convolutional:  112 x 112 x 16 ->  112 x 112 x 32 , pad  1
max_pool:  112 x 112 x 32 ->  56 x 56 x 32 , pad  0
convolutional:  56 x 56 x 32 ->  56 x 56 x 64 , pad  1
max_pool:  56 x 56 x 64 ->  28 x 28 x 64 , pad  0
convolutional:  28 x 28 x 64 ->  28 x 28 x 128 , pad  1
max_pool:  28 x 28 x 128 ->  14 x 14 x 128 , pad  0
convolutional:  14 x 14 x 128 ->  14 x 14 x 256 , pad  1
max_pool:  14 x 14 x 256 ->  7 x 7 x 256 , pad  0
convolutional:  7 x 7 x 256 ->  7 x 7 x 512 , pad  1
max_pool:  7 x 7 x 512 ->  4 x 4 x 512 , pad  1
convolutional:  4 x 4 x 512 ->  4 x 4 x 1024 , pad  1
convolutional:  4 x 4 x 1024 ->  4 x 4 x 1000 , pad  0
avg_pool:  4 x 4 x 1000 ->  1 x 1 x 1000 , pad  0
softmax:  1 x 1 x 1000 ->  1 x 1 x 1000

Wait for a few seconds (depending on your computer performance) then you will see the video with image classification labels at the top, like this:

screenshot

Compilation

Ok, now on to compiling. A sample script for the the Darknet Reference model was created for you.

On Windows, run

compile_darknetReference.cmd

On Linux / Mac, run

sh compile_darknetReference.sh

This script will:

  • Compile the model from ELL, emitting LLV IR
  • Compiling the LLVM IR, generating model object file
  • Generate SWIG interface (for Python bindings)
  • Running disutils setup to compile and link model + SWIG wrappers into a Python module, and install package locally

The compiled model can now also be accessed from Python. To create a compiled model for the Raspberry Pi, copy the files in this directory over and run the compile script on the Pi.

To test it out, run the compiled model demo app:

python compiledDarknetDemo.py