.rst .pdf

repository

Jump to

Code matchup

This page describes how the concepts discussed in the thesis are represented in this repository project.

Concept	Example	In code
Model size	16x16	16x16
	30x30x3	30x30x3
	etc.
Connections	1-to-9	@9
	1-to-25	@25
	1-to-49	@49
	Fully connected	@Inf
Branches	Direct connections only	DirectOnly
	Nearest neighbors	NearestNeighbor
	Next-to-nearest neighbors	NextToNN
	First ring	NearestNeighborOnly
	Second ring	NextToNNOnly
	Indirect connections only	IndirectOnly
	Grating strength 1.0	Kappa(1.0)
Dataset	MNIST	MNIST
	Fashion-MNIST	FASHION
Batch size	Batch size of 64	b64
	Batch size of 256	b256
	etc.
Training trial	Trial 1	t1
	Trial Name	tName
	etc.

For example, a weights file with the name weights_16x16x6@9-uniform.NextToNNOnly.FASHION.b64.t2.pt corresponds to the second trial of a model of height 16, width 16, and depth 6, with 1-to-9 connections and second ring branches (specifically from the uniform module), trained on the Fashion-MNIST dataset with a batch size of 64.

Training BCN models

Models are represented by bcn.BCN instances. The number of connections and the type of branches are specified with bcn.Connections and bcn.branches.Branches objects, respectively.

The dataset, batch size, and similar parameters are defined using the BCN.train method with a bcn.TrainingScheme object that represents the training scheme.

Setting up a model

Every model has a width and depth, and these are specified with the two positional arguments of bcn.BCN. For instance, the following code creates a 28x28x3 model with default (1-to-9) connections and default (direct-only) branches.

>>> from bcn import BCN
>>> model = BCN(28, 3)

To specify a different type of connections, such as 1-to-25 (Connections.ONE_TO_25), specify so by supplying the connections parameter with a Connections option.

>>> from bcn import Connections
>>> model = BCN(28, 3, connections=Connections.ONE_TO_25)

And to specify a different type of branches, such as uniform second ring branches (branches.uniform.NextToNNOnly), specify so by supplying the branches parameter with a branches.Branches object.

>>> from bcn.branches.uniform import NextToNNOnly
>>> nnn = NextToNNOnly()
>>> model = BCN(28, 3, branches=nnn)

Check out the bcn.BCN reference for other customizable parameters.

Training the model

Next we need to set up the training scheme. Like, what dataset will we train on? What’s the batch size? This information is specified with a bcn.TrainingScheme object. It takes only keyworded arguments. Use the width argument in case the dataset should be scaled to a size other than 28 (the original size of MNIST and Fashion-MNIST).

>>> from bcn import TrainingScheme, Dataset
>>> scheme = TrainingScheme(
...    dataset=Dataset.MNIST,
...    batch_size=64,
... )

Then set the scheme with the BCN.train method. Results can be saved using the save_path parameter. This method further accepts some extra information that are encoded in the results. See the documentation to learn what it accepts.

>>> model.train(
...    scheme=scheme,
...    save_path="./results/",
...    tag="Special test model",
... )

Finally, to train for, say, 100 epochs, use BCN.run_epochs:

>>> model.run_epochs(100)

Welcome to BCN’s documentation! BCN reference