Skip to content

Latest commit

 

History

History
107 lines (70 loc) · 4.21 KB

README.md

File metadata and controls

107 lines (70 loc) · 4.21 KB
Raw

darknet README

Darknet pre-trained model

We tested some darknet pre-trained models to others, get more detail from this file

Models Keras CNTK
yolov2 √(final conv) √(final conv)
yolov3 √(boxed image)

##Download darknet pre-trained model

$ mmdownload -f darknet

Supported models: ['yolov3', 'yolov2']

$ mmdownload -f darknet -n yolov3 -o ./

Downloading file [./yolov3.cfg] from [https://github.com/pjreddie/darknet/blob/master/cfg/yolov3.cfg]
progress: 200.0 KB downloaded, 100%
Downloading file [./yolov3.weights] from [https://pjreddie.com/media/files/yolov3.weights]
progress: 242200.0 KB downloaded, 100%
Darknet Model yolov3 saved as [./yolov3.cfg] and [./yolov3.weights].

Step-by-step conversion (for debugging)

Convert model from darknet to IR (darknet -> IR)

You can use following bash command to convert the network architecture [./yolov3.cfg] with weights [./yolov3.weights] to IR architecture file [darknet_yolov3.pb], [darknet_yolov3.json] and IR weights file [darknet_yolov3.npy]

'darknetStart' is to decide the start buf offsize to parse the darknet weight file.

$ mmtoir -f darknet -n yolov3.cfg -w yolov3.weights -o darknet_yolov3 --darknetStart 0
.
.
.
weights buf size: 62001758
Warning: Graph Construct a self-loop node data. Ignored.
loaded weights buf size: 62001758
IR network structure is saved as [darknet_yolov3.json].
IR network structure is saved as [darknet_yolov3.pb].
IR weights are saved as [darknet_yolov3.npy].

Convert model from IR to Keras/CNTK code (IR -> Keras/CNTK)

You can use following bash command to convert the IR architecture file [darknet_yolov3.pb] and weights file [darknet_yolov3.npy] to Keras Python code file[keras_yolov3_converted.py]

$ mmtocode -f keras -n darknet_yolov3.pb -w darknet_yolov3.npy -d keras_yolov3_converted.py

Parse file [darknet_yolov3.pb] with binary format successfully.
Target network code snippet is saved as [keras_yolov3_converted.py].

Using Keras code to generate the final result image

You can use following bash command to generate the final result image [yolov3_detect.jpg] using Keras Python code file[keras_yolov3_converted.py]

$ python -m mmdnn.conversion.examples.keras.imagenet_test -n keras_yolov3_converted.py -w darknet_yolov3.npy -i mmdnn/conversion/examples/data/dog.jpg -s darknet -p yolov3 --detect test

Found 3 boxes for img
('dog 1.00', (68, 164), (175, 393))
('truck 0.86', (255, 68), (377, 126))
('bicycle 0.99', (62, 87), (311, 316))
Keras yolo model result file is saved as [test.jpg], generated by [keras_yolov3_converted.py] and [darknet_yolov3.npy].

Using Keras/CNTK code to get the 3 final convolution layer output file

52,26,13

The order of output conv layer is [(1, 52, 52, 256), (1, 26, 26, 256), (1, 13, 13, 256)]. You can use following bash command to generate the final convolution layer output file [finalconv_52.npy, finalconv_26.npy, finalconv_13.npy] using CNTK Python code file[cntk_yolov3_converted.py]

$ python -m mmdnn.conversion.examples.cntk.imagenet_test -n cntk_yolov3_converted.py -w darknet_yolov3.npy -i mmdnn/conversion/examples/data/dog.jpg -s darknet -p yolov3 --detect test

Found 3 outputs

The output of CNTK model file is saved as [finalconv_52.npy].
The output of CNTK model file is saved as [finalconv_26.npy].
The output of CNTK model file is saved as [finalconv_13.npy].
generated by [cntk_yolov3_converted.py], [darknet_yolov3.npy] and [mmdnn/conversion/examples/data/dog.jpg].

Limitation

  • Currently no image classification support caused by the different architecture

Acknowledgement

Thanks to pytorch-caffe-darknet-convert the code of parsing the config file, keras-yolo3 the code of generating detection result image from final conv.