The data i got is not structured, of course it has different classes in different folders, but it is not split into train and test. So I use a bit of python code to split the data into train and test, then saved the zip and downloaded the zip file, now I'll be using the zip file.
On 19th June 2025, i have used tinyVGG architecture, used in CNNExplainer website, for my use case of rose leaf disease prediction.
Again Trained on 22nd June- for 50 epochs with weight parameter for CELoss it converged with 79 percent test accuracy and 0.447 test_loss.
On 21st June 2025, I have implemented MobileNetV1 architecture from scratch- now need to use a tiny version of this for my use case.
The tiny version of MobileNetV1- is replicated with initial 3 blocks and removed 5x blocks and again last 2 blocks so size increases from 64 to 1024. The test accuracy was low, model was poorly performing. So a full architecture of MobileNetV1 was used now even though the converged test accuracy is over 80 percent, the test_loss was more than tinyVGG.
This may be due to reduced parameter count- which reduces representational power- especially on small feature maps.
This makes sense: VGG has more parameters and stability, MobileNetV1 is light but needs careful tuning.