Difference between trainnet and trainNetwork

Hi, I am creating a deep neural network for a regression problem with tabular data as the inputs (6 features and 1 target).

I was wondering if there is any difference between the trainnet and trainNetwork functions? They seem to be quite similar so I am unsure of the pros and cons to using either function?

Answer 1

[Ben Hinkle]

Here are some of the differences between the two trainers:

• trainnet accepts a loss function input, which makes it easier to customize the loss if the standard regression loss is not enough.
• trainnet lets you track additional metrics while training.
• trainnet supports the LBFGS solver, which can converge quickly for small problems
• trainnet outputs a dlnetwork, while trainNetwork outputs a SeriesNetwork or DAGNetwork.
• trainNetwork accepts a table as input, while trainnet does not. This might make trainNetwork slightly easier to use when training data is in a table.

[Steve Herman]

Good afternoon,

I'm running "Deep Learning Toolbox Version 14.6 (R2023a)", and I don't seem to have the "trainnet" function available.

>> help trainnet

trainnet not found.

trainNetwork is available, but it does not suit my application as I want to have more control over, e.g. loss function.

Can you help?

Sh

[Cris LaPierre]

trainnet was introduced in R2023b. You must be using at least R2023b to access it.

[Onur Kilic]

Thank you, this is useful information. Should we expect performance improvements for trainnet in the near future? Currently, when training larger models, trainnet is much slower than trainNetwork and also gives 'out of memory' errors (GPU training) for model or minibatch sizes that trainNetwork would handle without issues.