simple digit recognition

[robertjankowski]

Demo of digits recognition task. I use pretrained model on MNIST dataset.

[robertjankowski]

Simple UI

[LaCuneta]

Thank you for the pull request! We appreciate it. We think this is a good example model for using NetLogo and Python together, and we have a few comments for you.

One item is on the Info Tab, which at the moment is mostly empty. It'd be good to fill in as much detail as you were able to. You can look at other example models in the NetLogo library for how we usually fill them out. In particular with this model, it'd be important to note the Python libraries used and what they're for.

There were also some code changes:

• When running a prediction, the output widget area retains the last prediction, which was confusing for me. You could use clear-output to make sure only the most recent prediction is presented, or switch to having a monitor widget for a prediction variable instead of an output area.
• Along those lines, you could have the prediction constantly updating as the drawing changes using a prediction reporter procedure, instead of need to hit a separate button for it. Any time the draw procedure changes the view, just run the prediction.
• One other change is to the draw procedure - you can either change the model view updates to be Continuous instead of On Ticks, or you can add the display command to the end of your the procedure (preferred). Either way, you should be able to remove the tick there, since we're not using ticks for anything in this model.
• Finally, you could switch to using an in-memory transfer of the view data instead of using a file as an intermediate store. Below is a sample of how that might look (warning, I have not tested it). It just gets the pcolor of each patch into a 2D list, which is converted into an array by np on the Python side:

to-report predict
  py:set images (list map (range max-pycor (min-pycor - 1) -1) [ row ->
    map (range min-pxcor (max-pxcor + 1)) [ col ->
      [ pcolor / 10 ] of patch row col
    ]
  ])
  report py:runresult "model.predict(np.array(images))[0]"
end

Thanks, again!

[robertjankowski]

Changes:

• automatically predict digit and display on monitor widget
• remote tick -> add display command
• not saving patch view to file
• add information in info tab

[robertjankowski]

Thanks for your review. I really appreciate it. I added monitors for handling probabilities from softmax layer and fixed other issues.

[robertjankowski]

[qiemem]

Oh, one more quick comment (feel free to ignore). You might consider swapping out the NN model for a fully convolutional one (so, no Dense layers). Right now, performance suffers when the digit is too far left or right. For instance, try drawing a straight line near the right or left side. It will register as a 4 or a 6 instead of a 1. A fully convolutional network should fix that. Maybe something like adding another conv layer or two on and then finish up with a global avg pooling layer with a softmax activation.

[robertjankowski]

I just added fixes. Referring to the model I think maybe in couple of days I would try to improve NN architecture. Many thanks for your help!

[LaCuneta]

This looks good to me. @qiemem, @robertjankowski: Any further changes, or are we good to merge?

[robertjankowski]

For me PR is ready to merge.