This reference explains how logging and visualization work in BojAI, focusing on the Logger and Visualizer classes. These tools allow you to track your model’s training and evaluation performance over time.

Logging in trainer.py

Logging is only available inside trainer.py and is done using the Logger singleton class. You can access it inside the trainer class like this:

self.logger

What You Can Log

You can log:

How to Log

# Logging during training
self.logger.log(epoch=epoch, loss=loss, train=score, valid=val_score)

# Logging final evaluation score
self.logger.log(eval_score=final_score)

⚠️ You must always pass an epoch when logging training, validation, or loss.
Do not log eval_score in the same call as training/validation metrics.
This prevents optimization bias.

Visualization with Visualizer

BojAI’s Visualizer reads the log from the Logger class and plots it using Matplotlib.

It supports four kinds of plots:

In the Train Stage

You can access:

• Loss vs Epoch
• Training vs Validation

In the Deploy Stage

You can access:

• All four options below, if evaluation data is uploaded
• Only the first two, if evaluation data is not uploaded

The 4 Visualization Options

1. Loss vs Epoch

Plots how the training loss evolves over time.

visualizer.plot_loss()

Used to diagnose underfitting, overfitting, or unstable training.

2. Training vs Validation

Plots training vs validation performance per epoch.

visualizer.plot_validation_vs_training()

Helps you compare generalization quality and identify divergence.

3. Mean Train vs Eval (Deploy stage only)

Compares the average training score with the final evaluation score.

visualizer.plot_train_vs_eval()

Useful to spot overfitting if eval is much worse than train.

4. Mean Validation vs Eval (Deploy stage only)

Compares the average validation score with the final evaluation score.

visualizer.plot_valid_vs_eval()

Helps understand whether the validation set was representative of the evaluation set.