AI engineering glossary · Deep learning core
What is evaluation mode?
Evaluation mode (also called Inference mode, model.eval): A framework setting that switches layers whose behaviour differs between training and testing to their test-time form: dropout is disabled and batch normalisation uses stored running statistics. Forgetting to set it gives noisy, inconsistent predictions.
Evaluation mode: an example
In PyTorch, model.eval() turns off dropout before validating.
Related terms
More Deep learning core terms
- Exploding gradient
- Exponential moving average of weights
- Fan-in and fan-out
- Feedforward neural network
- Flat minima
- Focal loss
- Forward pass
- Fully connected layer
Where you learn it
Evaluation mode is part of the AI Engineering Program, which teaches it with notes, a hands-on lab and an exam. The first lesson is free.