cdgreenidge / leabra7 Goto Github PK

The current logging system is pretty slow, since we do some fancy SQL-style joining. It also has some bugs (for example, it reports 1x1 Tensors instead of floats for Layer unit-level attributes.)

To fix this, we need to consider two kinds of attributes: "whole" attributes, like layer.avg_act, and "part" attributes, like the net value of a layer's 3rd unit. This will allow us to get rid of the nasty joins.

Changes

ObservableMixin

Abstract attributes

• A set of "whole" attributes

• A set of "part" attributes

• Abstract methods

  • observe_parts_attr(self, attr: str) -> Dict[str, List[Any]]. Should return a dict that looks like {"unit": [0, 1], "act": [0.3, 0.5]}.

• Already implemented methods

  • observe_whole_attr(self, attr: str) -> Tuple[str, Any] (this is just a simple getattr)
  • validate_attr()

Logger

• New method to record whole attributes
• New method to record parts attributes
• Method to transform to Logs

New stuff

• A function to transform a list of whole observations (type Tuple[str, Any] into a dict).
• Logs class. This is just a named tuple with two attributes: whole and parts. Each contains a data frame with the respective logs. We will return this from net.logs() instead of a single data frame.

Other stuff:

• Clean up WholeObs and PartsObs comments
• Reorder logs.py to make sense

Useful link: https://grey.colorado.edu/emergent/index.php/Leabra#Learning

Needed for the Schapiro model.

https://github.com/jupyter/nbconvert

We should direct people to the project pages instead.

It doesn't document the exception raised.

Hey Noam,

I just realized that the README points users to the Hitchiker's Guide to Python for the Python installation instructions. Since we moved our packaging to Anaconda, we actually want them to install the Anaconda python distribution (https://conda.io/docs/user-guide/install/download.html)

Could you update the README (again, sorry!) to point them there?

Necessary for the Schapiro model

• Need to add separate theta tough, peak, and plus cycle events to capture theta oscillation.
• Need to be able to disable projections
• Need to be able to define projections with learning defined for different phases

KetzMorkondaOReilly13.pdf

Actually, we want to check it on spec validation.

Please mention and add a prominent link to your Arxiv paper in the readme of this repo. I believe it is https://arxiv.org/abs/1809.04166

Right now, we encode the "unit" number in the dataframe column title. This is annoying and should be fixed.

The link to the Hitchiker's Guide to Python in the README is broken.

Since you've implemented the average KWTA already, that code is no longer needed :)

Necessary for Schapiro Model

Currently, we just check of the object exists. We should (probably) also validate the object's type.