Add new deep learning functionality by ncguilbeault · Pull Request #78 · bonsai-rx/machinelearning

ncguilbeault · 2025-12-12T10:05:25Z

This PR adds new functionality into the Bonsai.ML.Torch package to allow creation of deep learning modules directly in Bonsai. Operators were added to the Bonsai.ML.Torch.NeuralNets namespace to support a wide range of deep learning functionality, including allowing the direct creation of torch modules (convolutional layers, pooling layers, padding layers, recurrent neural networks, transformers, etc), and existing operators were refactored to support this new functionality in addition to previous functionality.

The PR in #77 should be merged first before this one.

…harp modules

…unctionality

…ing the training mode

…ule operators

…rrectly display element name and properties of module

…ng a property

…s not accepting input

… use pytorch naming conventions

…ule classes

…odule-specific parameters

…e state

…ss for feeding sequential container

…les`

glopesdev

Massive effort, thanks for putting this all together! Lots of comments below mainly on naming and organization for us to think about.

glopesdev · 2026-03-12T09:32:15Z

src/Bonsai.ML.Torch/NeuralNets/ActivationFunction/Celu.cs

+/// See <see href="https://pytorch.org/docs/stable/generated/torch.nn.CELU.html"/> for more information.
+/// </remarks>
+[Description("Creates a continuously differentiable exponential linear unit (CELU) activation function.")]
+[DisplayName("CELU")]


We have on occasion changed the display name for dynamic operators, but usually I try to avoid doing it for casing reasons, since on XML the actual class name will persist, and may create confusion between diffs and the visual editor.

I do agree this is a strange acronym, I looked elsewhere and couldn't find a better name, but would still keep the C# naming conventions.

glopesdev · 2026-03-12T09:32:34Z

src/Bonsai.ML.Torch/NeuralNets/ActivationFunction/Elu.cs

+/// See <see href="https://pytorch.org/docs/stable/generated/torch.nn.ELU.html"/> for more information.
+/// </remarks>
+[Description("Creates an exponential linear unit (ELU) activation function.")]
+[DisplayName("ELU")]


Same comment as above.

glopesdev · 2026-03-12T09:32:43Z

src/Bonsai.ML.Torch/NeuralNets/ActivationFunction/Gelu.cs

+/// See <see href="https://pytorch.org/docs/stable/generated/torch.nn.GELU.html"/> for more information.
+/// </remarks>
+[Description("Creates a gaussian error linear unit (GELU) activation function.")]
+[DisplayName("GELU")]


Same as above.

glopesdev · 2026-03-12T09:32:54Z

src/Bonsai.ML.Torch/NeuralNets/ActivationFunction/Glu.cs

+/// </remarks>
+[Description("Creates a gated linear unit (GLU) module.")]
+[DisplayName("GLU")]
+public class Glu


Same as above. Alternatively, we could emulate our decision in the Torch package and simply expand all acronyms into their full long names to favor readability, as these acronyms are impenetrable to cursory reading, and easily confused with each other.

glopesdev · 2026-03-12T09:35:27Z

src/Bonsai.ML.Torch/NeuralNets/ActivationFunction/LeakyRelu.cs

+/// See <see href="https://pytorch.org/docs/stable/generated/torch.nn.LeakyReLU.html"/> for more information.
+/// </remarks>
+[Description("Creates a leaky rectified linear unit (LeakyReLU) activation function.")]
+[DisplayName("LeakyReLU")]


I can see that the casing in torch is all over the place, so maybe we just default to expanding long-names and make everything clearer?

glopesdev · 2026-03-12T10:20:15Z

src/Bonsai.ML.Torch/NeuralNets/SaveModuleStateBuilder.cs

+/// <summary>
+/// Represents an operator that saves a module's state to a file.
+/// </summary>
+[Combinator]


Same as above.

glopesdev · 2026-03-12T10:20:31Z

src/Bonsai.ML.Torch/NeuralNets/ShuffleModuleBuilder.cs

+/// </summary>
+[XmlInclude(typeof(Shuffle.ChannelShuffle))]
+[DefaultProperty(nameof(ShuffleModule))]
+[Combinator]


Same as above.

glopesdev · 2026-03-12T10:20:40Z

src/Bonsai.ML.Torch/NeuralNets/SparseModuleBuilder.cs

+[XmlInclude(typeof(Sparse.EmbeddingBagFromPretrained))]
+[XmlInclude(typeof(Sparse.EmbeddingFromPretrained))]
+[DefaultProperty(nameof(SparseModule))]
+[Combinator]


Same as above.

glopesdev · 2026-03-12T10:20:54Z

src/Bonsai.ML.Torch/NeuralNets/TransformerModuleBuilder.cs

+[XmlInclude(typeof(Transformer.TransformerEncoder))]
+[XmlInclude(typeof(Transformer.TransformerEncoderLayer))]
+[DefaultProperty(nameof(TransformerModule))]
+[Combinator]


Same as above.

glopesdev · 2026-03-12T10:21:03Z

src/Bonsai.ML.Torch/NeuralNets/VisionModuleBuilder.cs

+[XmlInclude(typeof(Vision.PixelUnshuffle))]
+[XmlInclude(typeof(Vision.Upsample))]
+[DefaultProperty(nameof(VisionModule))]
+[Combinator]


Same as above.

ncguilbeault force-pushed the dev/torch-neuralnets-additions branch 2 times, most recently from 2df9d1a to e9ad97a Compare December 16, 2025 11:24

ncguilbeault force-pushed the dev/torch-neuralnets-additions branch 2 times, most recently from d4c9318 to 5369308 Compare December 19, 2025 14:01

ncguilbeault force-pushed the dev/torch-neuralnets-additions branch from 5369308 to a4400fc Compare January 13, 2026 11:39

ncguilbeault added 25 commits January 15, 2026 13:58

Added autogenerated classes which create observable streams of TorchS…

912049d

…harp modules

Added learning rate schedulers

7261825

Updated with improved naming

eefefea

Adding changes which include renaming files and improving doc strings

7aecc5d

Added XmlIgnore attribute to device property on modules

20d4cea

Added operator to explicitly allow setting the modules training mode

96e5611

Updated forward method to handle both Module and ScriptModule types

bc1e4bb

Testing inference expression builder

7457b0b

Organized modules into categories that better describe the module's f…

a7d3a62

…unctionality

Added correctly typed overloads to loading/saving modules and to sett…

8c0a1b2

…ing the training mode

Added a type converter for converting module operator classes

f11a427

Added an abstract class for module combinator builders

6c4fe96

Added a convolution module builder class and modified convolution mod…

349b4f5

…ule operators

Added module combinator property descriptor and updated builder to co…

d84c277

…rrectly display element name and properties of module

Added operator to collect module parameters

de16d5d

Updated convolution module builder

1b3fa2a

Refactored sequential container to only support inputs instead of usi…

a5f5d80

…ng a property

Added pooling module builder

32ad837

Updated modules to support module combinator builder pattern

2d2a1a5

Fixed issue with TransformerDecoder and TransformerEncoder module…

0a22aaa

…s not accepting input

Refactored modules into ActivationFunction namespace and changed to…

b3da5e7

… use pytorch naming conventions

Refactored module builders to use name of builder

8d3b49f

Refactored forward method into a builder class to support derived mod…

cfd257e

…ule classes

Refactored LoadModule into builder class to support derived modules

5c8904f

Updated modules to return derived class which allows easy access to m…

e9f4d09

…odule-specific parameters

ncguilbeault force-pushed the dev/torch-neuralnets-additions branch from a4400fc to e9f4d09 Compare January 19, 2026 14:57

ncguilbeault added 5 commits January 29, 2026 22:43

Refactored save module into a builder class and renamed to save modul…

aa457e5

…e state

Added operator to merge derived modules into a common module base cla…

18d8107

…ss for feeding sequential container

Updated base module builder class to set argument range

291d6a1

Updated to support correct sink behavior

f2cc9bf

Fixed sequential container to correctly receive input from `MergeModu…

a11b3de

…les`

ncguilbeault requested a review from glopesdev March 11, 2026 10:43

glopesdev changed the title ~~Addition of new deep learning functionality~~ Add new deep learning functionality Mar 12, 2026

glopesdev reviewed Mar 12, 2026

View reviewed changes

glopesdev added this to the v0.5.0 milestone Mar 12, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add new deep learning functionality#78

Add new deep learning functionality#78
ncguilbeault wants to merge 30 commits intobonsai-rx:mainfrom
ncguilbeault:dev/torch-neuralnets-additions

ncguilbeault commented Dec 12, 2025

Uh oh!

glopesdev left a comment

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

glopesdev Mar 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

ncguilbeault commented Dec 12, 2025

Uh oh!

glopesdev left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants