Why my model .network work not good?

so i make some .network model with my own dataset (number 0-9) and i use the model architecture from ‘smile_train_test.prototxt’
and then i train and test with accuracy 95+%, i quantization and convert it to .network model and deploy to OpenMV H7
but… when i try to use it. it can’t detect anything that i have train before
what should i do?
my model use 64x64 resolution to train and what is initialize of my camera that i should to setup?
thank you for your answer

i edit my model follow this
1.Snap my dataset from openmv about 5 pictures per class (0-9) with 80x60 resolution
2.Augmentation own dataset into 400 pictures per class about (4,000 of all classes)
3.Train my model using smile model architecture from your guide
4.Test my model and got accuracy about 90% up
5.Quantization and convert model to .network and deploy to openMV H7
6.I setup my main.py with ‘lenet_search_whole_window’ example and adapt it for my .network model like set framesize …
7.I test it with real dataset but it work so bad such as
i try to show number 0 but it tell me number 8 with 1.00 confidence!
i don’t know what’s wrong? what should i do? can you help me to solve this problem? i try to setup like this Re-loadable CNN on the OpenMV Cam M7/H7 with CMSIS-NN - YouTube and try to use with lenet.network it’s not work well too!

thank you very much

Maybe it’s overfitted, you need more training images. The smile network is trained on 18,000 images.

but testing set have work i’m confuse

I don’t know what could be wrong with your network, but the lenet.network example should work fine. Maybe it’s your hand writing ? Try it on printed digits.

I tested lenet.network, as you can see it’s working fine (Note: I had to lower the threshold for these numbers)