-2

I am working on a word-based sign language recognition system. I'm new to deep learning. My experience is limited. I have 10 classes. The size of the data ( 1256, 32, 112, 112, 3). For video classification I use 3D CNN, which has been mentioned in articles as being better than lstm's. My model couldn't learn at first. Or it was overfitting. I made some improvements. Using batchnormalization instead of Dropout in feature learning layers, keeping the learning_rate high at first and then lowering it, lowering the total parameter of my model. These were the improvements I tried. And results have improved. But in my test results, while acc is 0.69, my loss value is 1.02, which is very high. What improvements should I make for better results in this regard, do you have any recommendations? Below is the information about my model and its outputs.

Model accuracy and model loss graphs is here

My Model is here

Huriye Feyza Kurt
  • 1
  • 1

1 Answers1

0

Is your model based on other popular model? You could try using a transfer learning model reusing low level layers. Also increasing your data and setting a different batch size can help

luismontanaresm
  • 34
  • 1