Darknet: Visualize how YOLO make prediction in forward pass

[phongnhhn92]

Hi guys, I have a quick question. For example, I already finished training my model and now I want to do to evaluation and investigate how did YOLO works at each layer (i mean how did the input image transform between layers). Is there any possible way to do that ?

[AlexeyAB]

@phongnhhn92 Hi,
Can you refer to to some example in other framework, how do it best?
This is not a trivial to show image with 1024 channels instead of 3 channels (RGB) on this layer: https://github.com/AlexeyAB/darknet/blob/53b69a21426d3593c8bf8d337c906731886310d9/cfg/yolo-voc.2.0.cfg#L217

[sivagnanamn]

@phongnhhn92 Did you try visualizing the features?

[iraadit]

It has been done in the openFrameworks port of Darknet/YOLO
Here is the link : ofxDarknet (Gene Kogan version
And here in the code : https://github.com/mrzl/ofxDarknet/blob/6b858584a8530ba96f9fbf577f5e56bccfb128e2/libs/darknet/include/darknet.c#L342

Twitter post announcing it : https://twitter.com/genekogan/status/852111806218752000

[sivagnanamn]

@iraadit Thank you!
I checked that repo sometime back, unfortunately it doesn't work on Linux currently. mrzl/ofxDarknet#32. And I'm looking for something in the lines of features & activation visualization. I found this GZHermit/darknet-visualization_script & you've already raised an issue GZHermit/darknet-visualization_script#2 to provide better README. :+1:

[phongnhhn92]

@iraadit its a nice repo. I will try it today.
However, I trained all my models in Linux so I doubt it will work in Windows.
In the past, I tried to load my model (trained on Linux) on a Windows machine but it doesnt work.
I will try it again haha :v

[iraadit]

@sivagnanamn
Yes, I hope @GZHermit will answer :smile:

@phongnhhn92
If it has been trained with exactly the same file than the one it is running with, it should work (I think)

[agirbau]

@sivagnanamn I was checking the repo https://github.com/GZHermit/darknet-visualization_script and I realized that it's not about visualizing the activation of the weights on inference, I think that it's more about don't having to type the darknet parameters by using the cmd.

[AlexeyAB]

You can use this solution: https://github.com/AlexeyAB/darknet/issues/571#issuecomment-378546396