issue with "train categorizers"

[vzimmern]

Thanks for a wonderful piece of programming. It has been absolutely transformative in our research project on gene therapy in mice. I was really struggling to get anywhere with DeepLabCut and your piece of programming has helped our project along immensely.

I have a question about the "training categorizer" step. I am on a Macintosh laptop in a Python virtual environment. LabGym installs correctly and runs smoothly until I get to this step.

Here is the setup:

And the error in Terminal is:

I am struggling to make sense of the error. I am happy to provide any other information needed in order to get an answer.

best, Vincent Zimmern, MS, MD

[yujiahu415]

Hi Vincent,

Thank you for the nice comments! This issue may be because too few examples in certain behavior category. But I need more information to be sure. So how many pairs of behavior examples in each behavior category (folder)? Can you also double check the behavior examples to see whether they are all in pairs (one animation + one pattern image, with the same name but different file extension)? Thanks!

[vzimmern]

There were 2 behavior types that had fewer than 100 pairs -- one only had 1 pair and the other had 21 pairs. What's the minimum number of behaviors that are needed?

Thanks again not only for such a nice piece of useful code (I'm raving about it to my colleagues) but also for such quick responses on Github.

[yujiahu415]

You need to at least have 10 pairs of examples for each category to enable the training. You can either add in more training examples or just delete that category if its frequency is also very low in your videos to analyze. Generally, the more training examples, the higher accuracy of the Categorizer.

[yujiahu415]

By the way, I saw that you included body parts in pattern images with the STD of 0. You probably could try higher STD, like 50. Higher STD typically reduces unnecessary details of the body parts and may enhance the learning efficiency. But since you already generated and sorted the examples with STD as 0, you may first try Categorizer with different complexity and input size, if the training accuracy is not satisfying.

[vzimmern]

Thanks for the quick response and great advice. I will try to find a different video that has more of the missing behavior (the one with only 1 pair). Can't thank you enough for this excellent code.