Grasshopper

algorithmic modeling for Rhino

Hi Mateusz,

Firstly thank you for putting this together, what a great way to get stuck into machine learning, through Grasshopper.

I am just having a play with the component at the moment and I am struggling to get good results.

In addition it seems it is not possible to have more than one output node, is this correct?

I am using the MNIST data as a test case and trying 10 output nodes, one for each digit. I tried using one node but this gave poor results.

Thank you in advance,

Sam

Views: 240

Replies to This Discussion

Apologies, I've just answered my own question... it is possible, I just hadn't matched the no. of outputs correctly.

Thanks,

Sam

...However I am still struggling to get reliable results. My set-up is based on the article here: http://neuralnetworksanddeeplearning.com/chap1.html

I have attached my GH file for reference.

Thanks,

Sam

File size was a problem with the large dataset internalised. Here's an image instead:

For anyone who finds themselves following my monologue I have managed to get this to work.

The key problem was that neural networks are very sensitive to their hyper parameters (e.g. learning rate, alpha value for the Sigmoid function).

I ended up reducing my training data set size (to speed up training) and then used an optimisation algorithm (e.g the Nelder-Mead plugin - shameless plug) to optimise these parameters. This method reasonably quickly gave me decent results (>90% accuracy on my validation set).

I have since found a slightly different approach here: http://www.grasshopper3d.com/group/owl/forum/topics/examples-discus...

RSS

About

Translate

Search

Photos

  • Add Photos
  • View All

© 2020   Created by Scott Davidson.   Powered by

Badges  |  Report an Issue  |  Terms of Service