Awesome
OpenVINO-YOLO-Automatic-Generation
This repos parses yolov3/yolov3-spp/yolov4/yolov4-tiny/yolov4-tiny-3l's . cfg file to automatically generate tensorflow1.x code
As long as you have the .cfg file，you can convert it to .pb and OpenVINO. This means that code can be generated automatically even if the model(.cfg and .weights) is changed
Environment
python3.x (test on python3.6.5)
OpenVINO2020R4 :https://docs.openvinotoolkit.org/latest/index.html
Win or Ubuntu
Tensorflow 1.12.0
How to use
You could use changed yolov3.cfg ,yolov3-spp.cfg,yolov4.cfg And i will use original yolov3.cfg and yolov4.cfg as a demonstration. And i will show all the details to help you understand this repos!
yolov3-spp
For example: yolov3-spp
python parse_config.py --cfg cfg/yolov3-spp.cfg
you will get the output like this:
inputs = _conv2d_fixed_padding(inputs, 32, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 64, 3,strides=2)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 32, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 64, 3,strides=1)
inputs = inputs + shortcut
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=2)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=2)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
route36=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=2)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
route61=inputs
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=2)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = inputs + shortcut
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
route77=inputs
maxpool78=slim.max_pool2d(inputs, 5, 1, 'SAME')
inputs=route77
maxpool80=slim.max_pool2d(inputs, 9, 1, 'SAME')
inputs=route77
maxpool82=slim.max_pool2d(inputs, 13, 1, 'SAME')
inputs=tf.concat([maxpool82,maxpool80,maxpool78,route77],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
route86=inputs
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
detect_1 = _detection_layer(inputs, num_classes, _ANCHORS[6:9], img_size, data_format)
detect_1 = tf.identity(detect_1, name='detect_1')
inputs=route86
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _upsample(inputs,route61.get_shape().as_list(),data_format)
route92=inputs
inputs=tf.concat([route92,route61],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
route98=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
detect_2 = _detection_layer(inputs, num_classes, _ANCHORS[3:6], img_size, data_format)
detect_2 = tf.identity(detect_2, name='detect_2')
inputs=route98
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _upsample(inputs,route36.get_shape().as_list(),data_format)
route104=inputs
inputs=tf.concat([route104,route36],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
detect_3 = _detection_layer(inputs, num_classes, _ANCHORS[0:3], img_size, data_format)
detect_3 = tf.identity(detect_3, name='detect_3')
detections = tf.concat([detect_1, detect_2, detect_3], axis=1)
detections = tf.identity(detections, name='detections')
copy all the output and paste here: https://github.com/TNTWEN/OpenVINO-YOLO-Automatic-generation/blob/5f58e072c6a065ae8593e2be78f9e92163819637/yolo_v3.py#L173
And i upload yolo_v3.py here demo/yolo_v3.py. I hope it will help you understand easily
Then we can convert .weights -> .pb -> OpenVINO
#windows default OpenVINO path
python convert_weights_pb.py --class_names cfg/coco.names --weights_file yolov3-spp.weights --data_format NHWC
"C:\Program Files (x86)\IntelSWTools\openvino\bin\setupvars.bat"
python "C:\Program Files (x86)\IntelSWTools\openvino_2020.4.287\deployment_tools\model_optimizer\mo.py" --input_model frozen_darknet_yolov3_model.pb --transformations_config yolov3spp.json --batch 1 --reverse_input_channels
python object_detection_demo_yolov3_async.py -i cam -m frozen_darknet_yolov3_model.xml -d CPU
yolov4
For example: yolov4
cd yolov4parse
python parse_config.py --cfg cfg/yolov4.cfg
the output is
inputs = _conv2d_fixed_padding(inputs, 32, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 64, 3,strides=2)
route1=inputs
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
route2=inputs
inputs=route1
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 32, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 64, 3,strides=1)
inputs = inputs + shortcut
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
route8=inputs
inputs=tf.concat([route8,route2],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=2)
route11=inputs
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
route12=inputs
inputs=route11
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 64, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 64, 3,strides=1)
inputs = inputs + shortcut
inputs = _conv2d_fixed_padding(inputs, 64, 1,strides=1)
route21=inputs
inputs=tf.concat([route21,route12],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=2)
route24=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
route25=inputs
inputs=route24
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 3,strides=1)
inputs = inputs + shortcut
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
route52=inputs
inputs=tf.concat([route52,route25],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
route54=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=2)
route55=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
route56=inputs
inputs=route55
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = inputs + shortcut
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
route83=inputs
inputs=tf.concat([route83,route56],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
route85=inputs
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=2)
route86=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
route87=inputs
inputs=route86
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
shortcut=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = inputs + shortcut
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
route102=inputs
inputs=tf.concat([route102,route87],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 1024, 1,strides=1)
************************************
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
route107=inputs
maxpool108=slim.max_pool2d(inputs, 5, 1, 'SAME')
inputs=route107
maxpool110=slim.max_pool2d(inputs, 9, 1, 'SAME')
inputs=route107
maxpool112=slim.max_pool2d(inputs, 13, 1, 'SAME')
inputs=tf.concat([maxpool112,maxpool110,maxpool108,route107],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
route116=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _upsample(inputs,route85.get_shape().as_list(),data_format)
route118=inputs
inputs=route85
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
route120=inputs
inputs=tf.concat([route120,route118],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
route126=inputs
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _upsample(inputs,route54.get_shape().as_list(),data_format)
route128=inputs
inputs=route54
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
route130=inputs
inputs=tf.concat([route130,route128],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 128, 1,strides=1)
route136=inputs
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=1)
detect_1 = _detection_layer(inputs, num_classes, _ANCHORS[0:3], img_size, data_format)
detect_1 = tf.identity(detect_1, name='detect_1')
inputs=route136
inputs = _conv2d_fixed_padding(inputs, 256, 3,strides=2)
route141=inputs
inputs=tf.concat([route141,route126],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 256, 1,strides=1)
route147=inputs
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=1)
detect_2 = _detection_layer(inputs, num_classes, _ANCHORS[3:6], img_size, data_format)
detect_2 = tf.identity(detect_2, name='detect_2')
inputs=route147
inputs = _conv2d_fixed_padding(inputs, 512, 3,strides=2)
route152=inputs
inputs=tf.concat([route152,route116],axis=1 if data_format == 'NCHW' else 3)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
inputs = _conv2d_fixed_padding(inputs, 512, 1,strides=1)
inputs = _conv2d_fixed_padding(inputs, 1024, 3,strides=1)
detect_3 = _detection_layer(inputs, num_classes, _ANCHORS[6:9], img_size, data_format)
detect_3 = tf.identity(detect_3, name='detect_3')
detections = tf.concat([detect_1, detect_2, detect_3], axis=1)
detections = tf.identity(detections, name='detections')
this line ************************************ divides the output into two parts
that means the code above this line uses mish activation
copy the code above this line and paste here:
https://github.com/TNTWEN/OpenVINO-YOLO-Automatic-generation/blob/0714d18acf97a9ab6f6539599b6b8f242004d0be/yolov4parse/yolo_v4.py#L170
the code below this line uses leaky activation
copy the code bellow this line and paste here:
https://github.com/TNTWEN/OpenVINO-YOLO-Automatic-generation/blob/0714d18acf97a9ab6f6539599b6b8f242004d0be/yolov4parse/yolo_v4.py#L186
And i upload yolo_v4.py here demo/yolo_v4.py. I hope it will help you understand easily
Then we can convert .weights -> .pb -> OpenVINO
#windows default OpenVINO path
python convert_weights_pb.py --class_names cfg/coco.names --weights_file yolov4.weights --data_format NHWC
"C:\Program Files (x86)\IntelSWTools\openvino\bin\setupvars.bat"
python "C:\Program Files (x86)\IntelSWTools\openvino_2020.4.287\deployment_tools\model_optimizer\mo.py" --input_model frozen_darknet_yolov4_model.pb --transformations_config yolov4.json --batch 1 --reverse_input_channels
python object_detection_demo_yolov3_async.py -i cam -m frozen_darknet_yolov4_model.xml -d CPU
We could find the result is the same as https://github.com/TNTWEN/OpenVINO-YOLOV4
