Hyperparameters

Tune your model's training settings

This feature is coming soon.

A hyperparameter is a configuration that is external to the model and can be tweaked and tuned to optimize a model's performance.

Hyperparameters can be configured when creating a new model version. Open the optional SmartML Hyperparameters section at the bottom of Add New Model to adjust these advanced settings for your training run.

If no hyperparameters are set, these defaults are used:

{
"aug": "Sequential",
"args": [
[{
"aug": "CropAndPad",
"kwargs": {
"percent": {
"tuple": [-0.1, 0.0]
},
"keep_size": false,
"sample_independently": false
}
},
{
"aug": "GammaContrast",
"args": [{
"tuple": [0.8, 1.2]
}]
},
{
"aug": "HorizontalFlip",
"args": [0.5]
},
{
"aug": "VerticalFlip",
"args": [0.5]
}
]
]
}

Learning rate

Learning rate strategy defaults to Adaptive. At a high level, this strategy can be summarized this way: start with a high learning rate, then lower it when you’ve gone long enough without seeing a noticeable decrease in training loss. Once you’ve gone even longer without seeing a noticeable decrease, abort training.

The Baseline is set to R101 FPN 3x by default and has a base learning rate of 0.02.

Baseline options include:

  • R50 FPN 3X

  • R101 FPN 3X

  • X101 FPN 3X

Resolution

You can choose from an Inference Resize Strategy of "None" or "Fixed". "Fixed" provides Aspect Ratio options to select from:

  • 4:3

  • 16:9

  • 1:1.77

  • custom width and height

Data Augmentation

Read more about Data Augmentation with SmartML and the options available: