Wolfram Function Repository

Instant-use add-on functions for the Wolfram Language

Function Repository Resource:

 TrainingStateLayer

Source Notebook

Create a NetGraph that returns the current NetEvalutionMode as 0 or 1

Contributed by: Pierre-André Brousseau

ResourceFunction["TrainingStateLayer"][]

provides a NetGraph that returns the current NetEvalutionMode.

Details and Options

During training, ResourceFunction["TrainingStateLayer"] returns 1. Otherwise, it returns 0.
Any network can have online data augmentation by simply multiplying by this layer.
If NetEvaluationMode→"Train", this graph returns 1. If NetEvaluationMode→"Test", this graph returns 0.

Examples

Basic Examples

Create the training state layer:

In[1]:=
ts = ResourceFunction["TrainingStateLayer"][]
Out[1]=

Simple output:

In[2]:=
ts[]
Out[2]=
In[3]:=
ts[NetEvaluationMode -> "Train"]
Out[3]=
In[4]:=
ts[NetEvaluationMode -> "Test"]
Out[4]=

Use it as part of a network for data augmentation. This net adds noise as data augmentation during training:

In[5]:=
augmentationNet = NetGraph@FunctionLayer[Block[{ra, ts}, ( ra = RandomArrayLayer[UniformDistribution[{-1, 1}], "Output" -> {4}][]; ts = ResourceFunction["TrainingStateLayer"][][]; #Input + (ra*ts) )] &]
Out[5]=

Apply an input to the net, no noise is added:

In[6]:=
augmentationNet[<|"Input" -> {1, 2, 3, 4}|>]
Out[6]=

When in training mode, noise is added each time the net is used:

In[7]:=
Table[ augmentationNet[<|"Input" -> {1, 2, 3, 4}|>, NetEvaluationMode -> "Train"] , 4]
Out[7]=

Test mode does not add noise:

In[8]:=
Table[ augmentationNet[<|"Input" -> {1, 2, 3, 4}|>, NetEvaluationMode -> "Test"] , 4]
Out[8]=

Publisher

Pierre-Andre Brousseau

Requirements

Wolfram Language 13.0 (December 2021) or above

Version History

  • 1.0.0 – 07 June 2024

Related Resources

Author Notes

-

License Information