Are you reproducing the results of paper in this project?

[jiafw]

Hello,

Now I'm working on plant diseases diagnosis. It's nice to see your reproduced code for PD2SE.

When training the PD2SEModel I find that the three accuracies are very low. I want to know what's your best result in your experiment.

If the code can't reproduce the paper, I will check the code in the project carefully and make further work based on your code. Hopefully we can work together to improve the code.

thanks jia

[vrtompki]

Hi Jia,

Sorry its been so long. My university purchased an enterprise account for students and I have been using that account exclusively until today. Unfortunately, my enterprise account does not allow me to link to this one. With that being said, I have since abandoned replicating the results of the paper since it is for a research project that required more flexibility and portability, and my motivation has changed. The model itself was just too big (too many parameters) for my personal machine at the time, and since my lab as a whole focuses on computational intelligence, resources for our performance hardware is first come, first serve.

If you're planning to use this model still, I can still work with you if you want, but you have probably already figured it out by now. In my opinion you could just as easily achieve the authors results with three smaller networks, each dedicated to one of the three tasks described in the paper, which is essentially what they're doing, except the three individual networks share the same input and first three layers. In fact the labeling/objective of the third task encompasses the other two tasks as the 45 class labels include plant, disease, and severity. I have achieved similar results using much smaller networks not mentioned in this repo.

I am sure their model works and the code I have can replicate their results with some minor tweaking if I knew their selection criteria and had their actual shuffle units. Just let me know.