I have tried to enhance the accuracy by changing many factors like the type of architecture, dropout rate, and number of layers. but nothing seems to help to increase the accuracy and decrease the loss. what should I do?
Project ID: Project ID 219803
Context/Use case: I need to classify is the plant is healthy or not and deploy this model to esp32 cam.
What accuracy are you getting, and what are you aiming to achieve?
How many images have you used to train the model?
What is your no. of cycles?
Have you decreased the learning rate?
Hi @aqila
I can see you used the EON Tuner to order by accuracy. Great! However you may need to go through a data quality process on your dataset to remove the blurry images, and any that are not obvious for you from the image taken that the plant is either healthy or unhealthy. This could be causing problems in building your model, and in addition to the steps mentioned by Lauren.
Try gathering the images again, all in focus and at approximately consistent lighting. 500+ images is a good start but our example which simply identifies a plant has 200. Have a look online for a plant health dataset to get an idea of the quantities you need. " This dataset consists of about 87K rgb images of healthy and diseased crop leaves which is categorized into 38 different classes."
If you still face accuracy on the model what you can do is try to build for a linux based device instead of esp32 and see if you can get better accuracy. You may end up hitting device limitations for this particular use case as it is a complex problem to solve. We have an example of uploading a pretrained model in the BYOM doc: Bring your own model (BYOM) - Edge Impulse Documentation
Please keep us posted on your progress if you try to train it from scratch, it would be great to see the improvements you can make on this with a smaller dataset using our platform.
Best
Eoin