Skip to content
Branch: master
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.

Edge AI Tutorials

YOLOv3 Tutorial: Darknet to Caffe to Xilinx DNNDK


YOLOv3 is one of the most famous CNN for Object Detection. It was developed in a ML framework different from Caffe which is named Darknet. To run it on the Xilinx® DNNDK release, you need to convert it into a format compliant with Caffe. For this, you will need a special converter to convert Darknet to Caffe and generate the yolov3.prototxt and yolov3.caffemodel files as input to the DNNDK.

This Tutorial describes the process of converting the YOLOv3 CNN (originally trained in Darknet with the COCO dataset (80 classes)) before quantizing it with Xilinx DNNDK 2.0.8 release and run on a ZCU102 target board.

The conversion from Darknet to Caffe supports YOLOv2/tiny, YOLOv2, YOLOv3/tiny, and YOLOv3 basic networks. This conversion relies on a special fork of Caffe (designed by DeePhi), which is placed in the caffe-master folder.


  1. Ubuntu OS 14.04 or 16.04. For more information, see chapter 1 in the DNNDK User Guide UG1327.

  2. DNNDK tools and image for evaluation boards (zcu102 used in this example). For more information, see Xilinx AI Developer Hub.

  3. Python 2.7 and its virtual environments for Ubuntu OS.

  4. The official YOLOv3-608 network model trained with COCO dataset is available here. Download the yolov3.weights file (around 248 MB) and place it in the 0_model_darknet folder.

Project Organization

Assuming the working directory is named yolo_convertor in the folder $ML_DIR, the project is organized with the following directory structure:

├── caffe-master.tar.gz
├── darknet_origin.tar.gz
├── example_yolov3
│   ├──
│   ├── 0_model_darknet
│   │   ├── yolov3.cfg
│   ├──
│   ├── 1_model_caffe
│   │   └── v3.prototxt
│   ├──
│   ├── 2_model_for_quantize
│   │   ├── v3_example.prototxt
│   ├──
│   ├──
│   ├── 3_model_after_quantize
│   │   └── deploy.prototxt
│   ├── 4_model_elf
│   │   ├── yolo_kernel_graph.jpg
│   │   └──
│   ├── 5_file_for_test
│   │   ├── calib_data.tar
│   │   ├── calib.txt
│   │   ├──
│   │   ├── coco.names
│   │   ├── image.txt
│   │   └── test.jpg
│   └── results
├── images
└── yolov3_deploy.tar.gz.partaaa
└── yolov3_deploy.tar.gz.partaab

Preparing the Repository

The script sets the $ML_DIR variable to your working directory (for example /home/root2/ML/YOLOv3/yolov3_convertor is shown in all the examples in this tutorial).

cd <YOUR_WORKING_DIR>/yolov3_convertor
bash -v

The script also performs the following actions:

  • Uncompresses all the *.tar.gz files in the repository, by executing the following commands:
$ tar -xvf caffe-master.tar.gz
$ tar -xvf darknet_origin.tar.gz
$ cat yolov3_deploy.tar.gz.part* > yolov3_deploy.tar.gz
$ tar -xvf yolov3_deploy.tar.gz
$ rm yolov3_deploy.tar.gz.part*
$ cd example_yolov3/5_file_for_test
$ tar -xvf calib_data.tar
$ cd ../../
  • Runs the following commands from the working directory:
$ find . -type f -name "*.txt"   -print0 | xargs -0 dos2unix
$ find . -type f -name "*.data"  -print0 | xargs -0 dos2unix
$ find . -type f -name "*.cfg"   -print0 | xargs -0 dos2unix
$ find . -type f -name "*.names" -print0 | xargs -0 dos2unix

You will need to have the dos2unix utility installed in your Linux PC before executing them.

  • Sets the path in the file in the 5_file_for_test folder, according to the following ( PATH_TO depends on your environment):
valid = /PATH_TO/example_yolov3/5_file_for_test/image.txt
names = /PATH_TO/example_yolov3/5_file_for_test/coco.names
  • Sets the Caffe python interface path in the second line of the script as in the following (PATH_TO depends on your environment):

Processing Flow

Starting from a YOLOv3 CNN trained directly in Darknet with the COCO dataset, in this tutorial you will adopt the following flow:

  1. Convert the Darknet model into a Caffe model using the script.

  2. Test the object detection behavior of either the original Darkenet or the Caffe model with the and scripts respectively.

  3. Quantize the Caffe model generated in the previous step with the DNNDK decent tool by launching the script.

  4. Compile the ELF file for the DPU IP core on the ZCU102 target board with the script.

  5. Build the final application and deploy it on the ZCU102 board. It is archived in the yolov3_deploy.tar.gz file, for your refererence.

Compile Darknet and Caffe

Use the following commands to compile Darknet in the darknet_origin folder:

$ cd darknet_origin
$ make -j
$ cd ..

Use the following commands to compile Caffe in the caffe-master folder:

Note: Do this from your Python virtual environment.

$ cd caffe-master
$ make -j
$ make pycaffe
$ make distribute
$ cd ..

You will use this Caffe fork to convert the model from Darknet to Caffe, but you do not need it for training(as the training was already done in the Darknet framework). Therefore, be sure to set the compilation for CPU only in the Makefile.config file , as illustrated in the following lines:

# cuDNN acceleration switch (uncomment to build with cuDNN).
# CPU-only switch (uncomment to build without GPU support).

Now, use the following commands to set the CAFFE_ROOT environmental variable to the folder of the newly installed Caffe and update the other variables consequently.

export CAFFE_ROOT=~/ML/YOLOv3/yolo_converter/caffe-master
export PYTHONPATH=$CAFFE_ROOT/distribute/python:/usr/local/lib/python2.7/dist-packages/numpy/core/include/:$PYTHONPATH

To test the environment, execute the following command line:

$ python -c "import caffe; print caffe.__file__"

You will see commands as shown in the following figure:


The YOLOv3 Example

This sections details the flow described in 'The Processing Flow' section, using the YOLOv3 CNN as an example. The files are placed in the example_yolov3 folder, which is organized as shown below:

├── 0_model_darknet
│ ├── yolov3.cfg
│ └── yolov3.weights
├── 1_model_caffe
│ ├── v3.caffemodel
│ └── v3.prototxt
├── 2_model_for_quantize
│ ├── v3.caffemodel
│ ├── v3_example.prototxt
│ └── v3.prototxt
├── 3_model_after_quantize
│ ├── deploy.caffemodel
│ └── deploy.prototxt
├── 4_model_elf
│ ├── dpu_yolo.elf
│ ├── yolo_kernel_graph.jpg
│ └──
├── 5_file_for_test
│ ├── calib_data
│ ├── calib.txt
│ ├──
│ ├── coco.names
│ ├── image.txt
│ ├── test.jpg
│ ├── yolov3_caffe_result.txt
│ └── yolov3_darknet_result.txt
├── detection.jpg
├── results

Step 1: Darknet to Caffe Model Conversion

Use the following commands to launch the Darknet to Caffe conversion process:

$ cd example_yolov3
$ bash

The output is shown in the following figure: figure

The script converts the Darknet model(stored in the 0_model_darknet folder), to a Caffe model (stored in the 1_model_caffe folder), using the script.

The script performs the following actions to generate the two Caffe files, v3.prototxt and v3.caffemodel:

$ python ../ \
       0_model_darknet/yolov3.cfg        #path to Darknet cfg file \
       0_model_darknet/yolov3.weights    #path to Darknet weights file \
       1_model_caffe/v3.prototxt         #path to Caffe prototxt file \
       1_model_caffe/v3.caffemodel       #path to Caffe caffemodel file

Step 2: Test the Darknet and Caffe YOLOv3 models

The files that are required to test both the Darknet and Caffe models are placed in the 5_file_for_test folder. Ensure to delete all the *result*.txt files in the repository

Test Darknet

Execute the following commands to test Darknet:

$ cd example_yolov3
rm results/*
rm 5_file_for_test/yolov3_*_result.txt
$ bash

You will see the following: figure

The script uses the standard Darknet framework to get the detection result of the network, which is saved as yolov3_darknet_result.txt. The folder and file name in the script can be modified accordingly.

Note: The confidence threshold will determine the number of bounding boxes that will be the output of the final detection result. If the threshold needs to be changed in testing the Darknet model, modify line 419 of the detector.c file, clean and re-make Darknet, and run again.

Test Caffe

Execute the following commands to test Caffe:

$ cd example_yolov3
$ bash

You will see the following: figure

The script uses the functions added to the standard Caffe framework by DeePhi to get the detection result of the network after the conversion. The final detection result is saved in the 5_file_for_test folder as yolov3_caffe_result.txt. The path names, classes, and anchorCnt parameters in the script can be modified accordingly.

The following is the script:

$ ../caffe-master/build/examples/yolo/yolov3_detect.bin \
                                     1_model_caffe/v3.prototxt         #path to prototxt \
                                     1_model_caffe/v3.caffemodel       #path to caffemodel \
                                     5_file_for_test/image.txt         #image.txt specifies images for test \
                                     -confidence_threshold 0.005       #threshold for confidence \
                                     -classes 80                       #class num of network \
                                     -anchorCnt 3                      #anchor num of network \
                                     -out_file 5_file_for_test/yolov3_caffe_result.txt

If you plan to use the Darknet network model with different anchor parameters, modify line 134 of the yolov3_detect.cpp file according to anchor parameters in the cfg file. In the official YOLOv3 model, the anchor parameters of the Darknet model are the following:

anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,

Therefore, the corresponding values in yolov3_detect.cpp are:

float biases[18]= {10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326};

Note: To test the Caffe networks converted from YOLOv2/tiny YOLOv2 basic networks, use yolov2_detect.bin placed in the Caffe build examples folder (caffe-master/build/examples/yolo/), and modify the related parameters in yolov2_detect.cpp.

Note: To test the Caffe networks converted from YOLOv3/tiny YOLOv3 basic networks, use yolov3_detect.bin placed in the Caffe build examples folder (caffe-master/build/examples/yolo/) and modify the related parameters in yolov3_detect.cpp.

Detection Result Comparison

The Darknet detection results for the input image test.jpg extracted from the yolov3_darknet_result.txt file (you might get the same results but in a different order), are as follows:

test 0.006225 14.738144 25.408401  350.169128 324
test 0.006644 1.000000  176.401611 353.000000 484
test 0.012753 4.063614  33.500977  353.000000 500
test 0.097317 1.000000  108.329865 353.000000 490
test 0.105670 1.000000  108.329865 353.000000 490
test 0.209107 4.063614  33.500977  353.000000 500
test 0.247878 48.710381 237.655457 193.563782 373
test 0.996792 42.086151 16.412552  353.000000 485

The Caffe detection results for the example image extracted from the yolov3_caffe_result.txt file (you might get the same results but in a different order), are as follows:

test.jpg 0.00619627 14.7674 25.3904 350.165 324
test.jpg 0.00668068 1       176.519 353     484
test.jpg 0.0126261  4.07681 33.4886 353     500
test.jpg 0.0978761  1       108.419 353     490
test.jpg 0.105866   1       108.419 353     490
test.jpg 0.208174   4.07681 33.4886 353     500
test.jpg 0.245967   48.6832 237.693 193.577 373
test.jpg 0.996799   42.0844 16.3986 353     485

The difference of confidences and bonding box coordinates between Darknet and Caffe model is negligible.

Step 3: Quantize the Caffe Model

Use the following commands to quantize the Caffe model:

$ cd example_yolov3
$ cp 1_model_caffe/v3.caffemodel  ./2_model_for_quantize/
$ bash

You will see the following: figure


Note: It is normal to see Loss values of 0 in the calibration phase, because loss layers are not included in the converted Caffe model.

In order to quantize the converted Caffe network, copy the v3.prototxt and v3.caffemodel files from the 1_model_caffe folder to the 2_model_for_quantize folder. Then, modify the v3.prototxt file by:

  • Commenting out the first five lines
  • Adding an ImageData layer with the calibration images for the train phase as shown in the following fragment:
name: "Darknet2Caffe"
#####Comment following five lines generated by converter#####
#input: "data"
#input_dim: 1
#input_dim: 3
#input_dim: 608
#input_dim: 608
#####Change input data layer to ImageDate and modify root_folder/source before run DECENT#####
layer {
name: "data"
type: "ImageData"
top: "data"
top: "label"
include {
  phase: TRAIN
transform_param {
  mirror: false
  yolo_height:608    #change height according to Darknet model
  yolo_width:608     #change width  according to Darknet model
image_data_param {
  source:"/PATH_TO/5_file_for_test/calib.txt"         #change path accordingly
  root_folder:"/PATH_TO/5_file_for_test/calib_data/"  #change path accordingly
  batch_size: 1
  shuffle: false
##### No changes after below layers#####

For your convenience, the v3_example.prototxt file illustrates the changes you should do on the original v3.prototxt file.

The format of calib.txt used in the calibration phase of DNNDK decent is as follows:

#image_name                     fake_label_number
COCO_train2014_000000000009.jpg 1
COCO_train2014_000000000025.jpg 1
COCO_train2014_000000000030.jpg 1
COCO_train2014_000000000034.jpg 1
COCO_train2014_000000000036.jpg 1

Note: The label number is not used in the calibration process.

The 5_file_for_test/calib_data folder contains some images from the COCO dataset, to be used for the calibration process. The script performs the following actions:

#Assuming "decent" tool is already in the PATH
$ decent quantize -model 2_model_for_quantize/v3.prototxt        #path to prototxt \
                -weights 2_model_for_quantize/v3.caffemodel    #path to caffemodel \
                -gpu 0 \
                -sigmoided_layers layer81-conv,layer93-conv,layer105-conv \
                -output_dir 3_model_after_quantize \
                -method 1

In the YOLOv3 network, the conv layer before the yolo layer (that is, the output layer in Caffe model) will be quantized with the -sigmoided_layers flag for better accuracy. For more information, use the decent –help command or see the DNNDK User Guide UG1327.

Step 4: Compile the Quantized Model

Use the following commands to compile the ELF file:

$ cd example_yolov3
$ cp 3_model_after_quantize/ref_deploy.prototxt 3_model_after_quantize/deploy.prototxt
$ bash

You will see the following: figure

Modify the deploy.prototxt (generated in the previous step) in the 3_model_after_quantize folder as follows:

layer {
name: "data"
type: "Input"
top: "data"
#####Comment following five lines #####
#transform_param {
# mirror: false
# yolo_height: 608
# yolo_width: 608
# }
#####Nothing change to below layers#####
input_param {
shape {yolov3_deploy.tar.gz
dim: 1
dim: 3
dim: 608
dim: 608

Note: The ref_deploy.prototxt file already contains all the above described changes. So, just copy it and rename as deploy.prototxt.

The script uses dnnc to compile and generate the ELF file for the target ZCU102 board as follows:

#Assume the dnnc-dpu1.3.0 is installed in your $PATH

$ dnnc-dpu1.3.0 --prototxt=3_model_after_quantize/deploy.prototxt \
              --caffemodel=3_model_after_quantize/deploy.caffemodel \
              --dpu=4096FA \
              --cpu_arch=arm64 \
              --output_dir=4_model_elf \
              --net_name=yolo \
              --mode=normal \

For more information, use the dnnc-dpu1.3.0 –help command or see the DNNDK User Guide UG1327. For other platforms, specify --dpu and --cpu_arch accordingly.

Step 5: Deploy YOLOv3 on the ZCU102 Board

Use the following steps to deploy YOLOv3 on the ZCU102 board:

  1. Copy the dpu_yolo.elf file (generated by dnnc) from the 4_model_elf folder, to the model folder. Then, use the following commands to archive the yolov3_deploy folder.
$ cd yolo_convertor
$ cp example_yolov3/4_model_elf/dpu_yolo.elf yolov3_deploy/model/
$ tar -cvf yolov3_deploy.tar ./yolov3_deploy
$ gzip -v  yolov3_deploy.tar
  1. Assuming that the ZCU102 board is turned on and it is connected to a 4K external monitor via Display Port, and the usual communication is setup between target board and host Linux PC (that is, the microUSB-to-USB cable and an Ethernet Point2Point cable), open PuTTy with the following command:
$ sudo putty /dev/ttyUSB0 -serial -sercfg 115200,8,n,1,N
  1. Once the UART communication is established, execute the following command:
$ ifconfig eth0 netmask
  1. From the host PC execute the following command (depending on your PC, select either eth0 or eth1. In this case eth1 is selected)
$ sudo ifconfig eth1 netmask
  1. Establish the UART communication between the host and the target. What you have done so far, after booting the board, was to use the serial port to set the IP address of the board and ensure that the host PC and target board are in the same subnet. The following screenshot illustrates this procedure: figure

  2. From your host PC copy the yolov3_deploy.tar.gz file to the ZCU102 board using the ssh/scp commands:

$ scp yolov3_coco80_zcu102.tar root@
  1. Open the archive on the board and type the following command:
tar -xvf yolov3_deploy.tar.gz

After extraction, the files in the package are as follows:

├── coco_test.jpg
├── test.avi
├── Makefile
├── model
│ └── dpu_yolo.elf
└── src
└── utils.h
  1. Compile the code in the yolov3_deploy folder on ZCU102 with the make command, as follows:
make -j
  1. Use the following command to run the executable yolo:

#Test image ./yolo coco_test.jpg i

#Test video ./yolo v

You will see the following: figure figure

If you are using a Windows OS PC, you can use Teraterm with the same settings as Putty, as shown below: figure

In that case you proceed accordingly to set the IP addresses and use pscp.exe utility instead of scp, as shown in the following screenshot: figure figure

Note: For more information on setting up the ZCU102 communication, preparing to boot the SD card, connecting serial port/ethernet cables, and setting up the 4K display, see the DNNDK User Guide UG1327.

You can’t perform that action at this time.