search

ceccocats / tkDNN

Deep neural network library and toolkit to do high performace inference on NVIDIA jetson platforms
GNU General Public License v2.0
718 stars 209 forks source link

How does input_bins and output_bins affect on custom data? #152

Closed Goru1890 closed 3 years ago

Goru1890 commented 3 years ago

I wrote a custom yolo4.cpp for my own dataset and trained weights, do I have to customize also the paths in input_bins and output_bins? For example, I don't have the layer161_out.bin in my exported weights.

I also noticed that the ap is lower in a tkDNN demo, using a jetson tx2, than in a darknet demo, using a gtx 2080.

mive93 commented 3 years ago

Hi @Goru1890

generally with the yolo family just check in the layer folder (with the weights) the weights that start with g, those are the output layers. For instance, with Yolov4 you will have g139, g150 and g161, that correspond to output bins layer139_out.bin, layer150_out.bin, layer161_out.bin. For other models, it depends, could be more complex, but the point is that it correspond to the number of the output layers.

What do you mean by lower ap? Which ap are you mentioning? MAP(0.5:0.95)? AP50? AP75? Darknet does not compute the AP50, we compute all of them.

mive93 commented 3 years ago

Closing for now, feel free to reopen.