search

MichalDanielDobrzanski / DeepLearningPython

neuralnetworksanddeeplearning.com integrated scripts for Python 3.5.2 and Theano with CUDA support
MIT License
2.79k stars 1.27k forks source link

Custom BackPropagation & Neural Network architecture for a Regression problem #22

Closed diliprk closed 3 years ago

diliprk commented 5 years ago

Hi Michal, Thank you for putting up Nielsen's code in Python 3.5. I am trying to build a backpropagation algorithm for a custom physics guided loss function where a neural network works in tandem with a physics-based simulator for a regression problem. Since heavy customizations are required, libraries like Keras are not suitable for the job and I have to build the neural network from scratch. I read the first two chapters of Nielsen's book and found the explanation and code quite suitable to use as a starting template. My clarifications are: 1.) Can I start with network.py or should I use the improved versions (network2.py & network3.py as a starting base? 2.) Which activation function should I use? Can I continue using Sigmoid as in the code or should I implement ReLuor another activation function for regression predictions? 3.) What parts of the code I should change for this regression problem so that the network architecture accepts an np.array of 7 input features and 1 output column, because right now it does not accept multidimensional array in the training data?

MichalDanielDobrzanski commented 3 years ago

Hey @diliprk. It is up to you to find the best neural network architecture. This is all about it. You must validate your architectures against your data set.

I am closing this issue.