tsukar / evolo

Evolutionary YOLO

evolo

Evolutionary YOLO

Requirements

• Python 3.6.7
• Jupyter Notebook

Usage

Clone this repository and AlexeyAB/darknet, respectively.

$ git clone https://github.com/tsukar/evolo.git
$ git clone https://github.com/AlexeyAB/darknet.git
$ cd darknet/

Edit the Makefile so that you can use your GPU.

@@ -1,7 +1,7 @@
-GPU=0
-CUDNN=0
+GPU=1
+CUDNN=1
 CUDNN_HALF=0
-OPENCV=0
+OPENCV=1
 AVX=0
 OPENMP=0
 LIBSO=0
@@ -30,7 +30,7 @@ OS := $(shell uname)
 # ARCH= -gencode arch=compute_72,code=[sm_72,compute_72]
 
 # GTX 1080, GTX 1070, GTX 1060, GTX 1050, GTX 1030, Titan Xp, Tesla P40, Tesla P4
-# ARCH= -gencode arch=compute_61,code=sm_61 -gencode arch=compute_61,code=compute_61
+ARCH= -gencode arch=compute_61,code=sm_61 -gencode arch=compute_61,code=compute_61
 
 # GP100/Tesla P100 <96> DGX-1
 # ARCH= -gencode arch=compute_60,code=sm_60

Just make and copy the compiled binary into the root directory of this repository.

$ make
$ cp darknet ../evolo/
$ cd ../evolo/

Download the pre-trained weights from YOLOv2 official site.

$ wget https://pjreddie.com/media/files/darknet19_448.conv.23

Download train.zip and test.zip (currently not available - please contact me at tsukada [at] iba.t.u-tokyo.ac.jp if you need them). Put the decompressed *.jpg files in data/x-ray/train/, data/x-ray/test/, respectively.

Run Jupyter Notebook and open evolution.ipynb -> Run All to start evolution.

$ jupyter notebook

After evolution, open evaluation.ipynb -> Run All for evaluation.

evolo's Issues

Some layers are not implemented

The following layers are not compatible with the ones on the darknet framework:

• DropoutLayer
• ConnectedLayer
• SkipLayer

Add test code for input/output dimensions

The logic for calculating input/output dimensions has not yet been tested.

A possible way to confirm it works is:

• To let the darknet binary parse the .cfg files mutated by this program, and check if no error occurs.