I would like to create a model that I can implement into an app that can detect plant diseases. To do that I have to train my own object detection tf model. And so my problem is the average precision is always low no matter what I try to do.
I am using this dataset: https://universe.roboflow.com/graduation-project-2023/plants-diseases-detection-and-classification/dataset/12 I used their formatting to get it to Tensorflow object detection. I set up my Tensorflow object detection on cmd:
This is what the inside of my Tensorflow folder looks like where inside pretrain-models I'm using efficientdet_d1_coco17_tpu-32 and inside "models" I have my pipeline.config which has these changed.
Pipeline.config:
Now I have tried to run 100k steps and it gave me an average precision of 0.137.
So I decided to up it up to 250k steps and it gave me an average precision of 0.245 and it seemed like upping the steps was what I needed to do so i left it on 1m steps and it came back with 0.276 and this is the screenshot of the other stats:
And it's like I hit a wall and I have no clue what am I supposed to do any more because I have tried changing dataset before and it seemed like it got worse and continuing the training on past checkpoints only seems to lower the average precision overall.