How to modifly code for custom yolo?

[ashwinnair14]

files_to_test.zip
Hi @duangenquan @ichigoi7e ,

I have trained a custom yolo model with only 4 classes and which has a different structure than yolo.
I have attached the cfg and weights and also the NCS graph to test. Tested with the person image from darknet.

When I modify the python code to set self.classes = 4 I dont get any outputs on the script(ObjectWrapper.py).

But when I run the same model and image on the original darknet I get this output.

Is there an easy way to change the C++ or the python code ? or do I need to rewrite it?

Thanks

layer     filters    size              input                output
   conv     12  3 x 3 / 2   416 x 416 x   3   ->   208 x 208 x  12
..
..
.
.
.
.
   conv     45  1 x 1 / 1    13 x  13 x 100   ->    13 x  13 x  45
   detection
mask_scale: Using default '1.000000'
Loading weights from 8.weights...Done!
data/person.jpg: Predicted in 0.001252 seconds.
person: 100%
motorbike: 82%

[necusjz]

Please modify the C++ code ./src/Region.cpp as well, refer to Update parameters section in README.

[ashwinnair14]

@ichigoi7e Thanks but I was unable to understand the paramater
I guess self.blockwd = 12 is the the lastConv layer lenght - 1?

Whats the paramater targetBlockwd ?
Cause I dont get expected outputs no matter what I try ?
for the person image I was 100% score on the darknet platform.

[necusjz]

@ashwinnair14
The exact values of these parameters are related to your code on the darknet platform.
For example, how many grids do you divide a picture into?
Please read your previous code carefully.

[duangenquan]

@ashwinnair14 , blockwd and targetBlockwd are used to rectify detection results a little bit, because the converted caffe model from yolo generates smaller blocks sometimes.
please set blockwd as the block numbers running on stick, and targetBlockwd from darknet models.

BTW, you can also tune other parameters in caffe cfg, such as pad, to generate the same block dimensions as darknet models.