Training Hyperparameters¶

Configure training parameters for optimal model performance.

Overview¶

Hyperparameters control how the model learns. The defaults work well for most cases, but tuning can improve results.

Core Parameters¶

Batch Size¶

Default: 32

Number of samples processed together:

Adjustment tips:

• Reduce if out of memory
• Increase for larger datasets
• Adjust learning rate proportionally

Max Epochs¶

Default: 100

Maximum training iterations:

Note: Training usually stops earlier due to early stopping.

Early Stopping Patience¶

Default: 8

Epochs without improvement before stopping:

Optimization Parameters¶

Learning Rate¶

Default: 0.0001

Step size for weight updates:

Uses OneCycleLR scheduler with warmup.

Dropout¶

Default: 0.5

Regularization to prevent overfitting:

Increase if validation loss >> training loss.

Weight Decay¶

Default: 0.01

L2 regularization on weights:

Loss Function¶

Focal Loss (Default)¶

Better for imbalanced datasets:

Alpha Guidelines¶

Gamma Guidelines¶

Augmentation During Training¶

Label Smoothing¶

Default: 0.1

Softens target labels:

Prevents overconfidence.

Mixup Alpha¶

Default: 0.5

Mixes training examples:

Creates virtual training samples.

Presets¶

Fast Training¶

{
    "batch_size": 64,
    "max_epochs": 50,
    "early_stopping_patience": 5,
    "learning_rate": 0.0002,
    "dropout": 0.4,
}

Balanced (Default)¶

{
    "batch_size": 32,
    "max_epochs": 100,
    "early_stopping_patience": 8,
    "learning_rate": 0.0001,
    "dropout": 0.5,
}

Maximum Accuracy¶

{
    "batch_size": 32,
    "max_epochs": 200,
    "early_stopping_patience": 15,
    "learning_rate": 0.00005,
    "dropout": 0.5,
}

Troubleshooting¶