Regarding training of a neural network

#1

can anyone plzz look at my google colab file and tell me where i am wrong as i was trying to train a deep learning model for predicting steering angle of a car. The val_loss decreased from 0.74 to 0.22 but now it is not decreasing and the model is not performing well.i kept a learning rate of 0.00001 and the model architecture is taken from nvidia self driving car paper CNN(2016). Can u plzz take a look at my custom data generators and tell me if i am wrong there and tell me the reason why my model is not performing well and how it can be improved from here

Here is the link to my google colab file

https://colab.research.google.com/drive/1mHo8a8DMvo2sDKHr7RzhBQESMFHTEr5a?usp=sharing

#2

hey @amankharb ,
There are a lot things in this you can try on.
Apart from those i told you on whatsapp.
You can try

  1. Augmentation.
  2. A modified version of Learning Rate Scheduler is implementing it in a cyclic order. Means not always decreasing it , but based on performance and epochs , you need to increase it too.
  3. Using correct Batch Size.
  4. Exploring different pre trained Models , fine tuning them.
  5. Using Correct optimizer
  6. Regularization , only if needed.
  7. Search about more papers on the same topic and try to implement them , after confirming there results that they are better then yours or not.

I hope this helps You .
Thank You :slightly_smiling_face:.