search

eerwitt / home-value-deep-learning

Train a model which can estimate a home's value based on pictures.
http://cedarstreet.io/
5 stars 1 forks source link

CNN Architecture #1

Open hanneshapke opened 8 years ago

hanneshapke commented 8 years ago

Hi Erik,

I attended your PyDX presentation about deep learning and I finally got around to try out my own CNN development with Theano and Lasagne. I looked through your code examples and your Caffe again, but I couldnt find a description of the network architecture.

How did you decide on the number of layers and the filter size for your project? Could you please elaborate more on your CNN design decisions?

Thank you for your reply!

eerwitt commented 8 years ago

Hi @hanneshapke,

Excellent! Theano and Lasagne are excellent pieces of software.

Sure, the layers are described here for this example implementation.

They were a first version based on the LeNet MNIST Tutorial.

Many of my design decisions were based on exploratory analysis of the filters considering input and output. These are resources I find useful and continually go back to looking at. I hope they may help you as well.

  1. Berkeley deepViz
  2. CS231n Explanation
  3. Great Paper on Tuning Networks

Thank you, Erik

hanneshapke commented 8 years ago

Hi Erik,

Thank you for the explanation! Very appreciated.