What to Look for During Training and Testing
In addition to debugging errors with your models, there are a number of statistics to watch during training and testing. These can be viewed in real-time via PerceptiLabs' Statistics and Test views respectively.
Below are common things to look for while training and testing different types of models.
Tip: Also be sure to check out our Machine Learning Minute videos on YouTube where we provide brief overviews of modeling and debugging.
Common Training Issues
Classification Training
Gradients: gradients of 0 indicate that the model is not training and gradients that grow infinitely mean that the model is changing too much. Note that usually the gradient in the last layer is sufficient to watch.
Loss: if loss increases during validation, this means the model is overfitting.
Predictions per class: there should be one color per class, otherwise the model is mixing classes together. The model may need to be made more complex to rectify this.
Segmentation Training
Gradients: if gradients die (decrease to nothing) quick, try normalizing the input using batch normalization and/or try a different output activation function.
Prediction: if the predictions seem to contain a lot of noise then the model has not been trained enough yet or needs to be made more complex. If you are only predicting one's (i.e., a white image) follow the tips in the previous point.
General Training
CPU/GPU usage: if resources aren't being utilized used as much as they should be, try larger batch sizes (this can be adjusted in the training settings).
Common Testing Issues
Classification Testing
Confusion matrix: if the confusion matrix shows that classes are mixed up, try building a more complex model.
Table: look closely at Precision and Recall. Also if the test accuracy is a lot worse than it is for training and validation, the model is likely overfitting.
Segmentation Testing
Image: see Segmentation Training above.
Table: if test Dice is a lot worse than it is training and validation, the model is likely overfitting.
Last updated