GRAD CAM for test images

[SaraPiresNobrega]

Hello ! I am developing my master's degree dissertation and I would like to used GRAD CAM or other class activation maps to corroborate the test results of my object detection model (YOLOv4).

Using the information present in https://github.com/AlexeyAB/darknet/issues/5117 it is possible to do that ?

[AlexeyAB]

@SaraPiresNobrega Hi,

Yes, use this in cfg-file for Grad-CAM:

[net]
adversarial_lr=0.5
attention=1
burn_in=0
learning_rate=0.001

And remove all stopbackward= lines from cfg-file (if they exist).

And run training command with your pre-trained weights.

---

For example:

• Use this cfg-file (rename it from yolov4x-mish.cfg.txt to yolov4x-mish.cfg): yolov4x-mish.cfg.txt
• Download this weights: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4x-mish.weights
• You can un-comment these lines and recompile to save outputs to png-files:
  • https://github.com/AlexeyAB/darknet/blob/97ccee56283ad92118153f0cd1036d1ea07ba115/src/network_kernels.cu#L242-L246
  • https://github.com/AlexeyAB/darknet/blob/97ccee56283ad92118153f0cd1036d1ea07ba115/src/network_kernels.cu#L253-L254
• Run training: ./darknet detector train F:/MSCOCO/coco_f.data cfg/yolov4x-mish.cfg yolov4x-mish.weights

You will see such results:

Although it looks less beautiful and less understandable, it is closer to the truth than other Grad-CAM implementations.

[JeremyKeusters]

Thanks for this @AlexeyAB ! Very interesting to look at! Would it be easy to also have the labels displayed on these images?

[SaraPiresNobrega]

@AlexeyAB thank you very much ! This approach will be very helpful in my master's degree dissertation ! I thought that there was some way to visualize the GRAD CAM maps using the "test" command but I think that this way I can also do what I want.

[AlexeyAB]

Also try the latest commit, I fixed a bug recently.

[SaraPiresNobrega]

@AlexeyAB thank you very much !