TensorBoard Integration

OG-Learn provides built-in TensorBoard support for visualizing training progress.

Basic Setup

Enable TensorBoard Logging

from og_learn import OGModel

model = OGModel(
    hv='lightgbm',
    lv='mlp',
    tensorboard_dir='runs/experiment1',  # Log directory
    tensorboard_name='og_mlp',           # Run name
    eval_every_epochs=5                   # Log frequency
)

model.fit(
    X_train, y_train,
    density=density_train,
    X_valid=X_valid,  # Required for validation metrics
    y_valid=y_valid,
    epochs=100
)

Launch TensorBoard

from og_learn import launch_tensorboard

# Start TensorBoard server
tb_process = launch_tensorboard('runs/experiment1', open_browser=True)

Or from command line:

tensorboard --logdir=runs/experiment1

Then open http://localhost:6006 in your browser.

---

Comparing Multiple Models

from og_learn import OGModel, compare_models
from og_learn.presets import get_lv_model

# Clear old logs
import shutil
shutil.rmtree('runs/comparison', ignore_errors=True)

# Define models
models = {
    'MLP': get_lv_model('mlp', num_features=X_train.shape[1]),
    'OG_LightGBM_MLP': OGModel(hv='lightgbm', lv='mlp'),
    'OG_XGBoost_MLP': OGModel(hv='xgboost', lv='mlp'),
    'OG_CatBoost_ResNet': OGModel(hv='catboost', lv='resnet'),
}

# Run comparison with TensorBoard logging
results = compare_models(
    models,
    X_train, y_train,
    X_test, y_test,
    density=density_train,
    tensorboard_dir='runs/comparison',  # All models log here
    eval_every_epochs=5
)

# Launch TensorBoard to compare
launch_tensorboard('runs/comparison', open_browser=True)

---

What Gets Logged

Metric	Description
R2_train	Training R² score
R2_val	Validation R² score
Loss_train	Training loss (MSE)

---

Viewing in TensorBoard

Scalars Tab

View training curves:

• R² scores over epochs
• Loss over epochs
• Compare multiple runs side-by-side

Text Tab

View run configuration and parameters.

---

Tips

Clear Old Logs

Before running new experiments, clear old logs to avoid confusion:

import shutil
shutil.rmtree('runs/', ignore_errors=True)

Logging Frequency

Set eval_every_epochs=1 for detailed curves, or higher values (5-10) for faster training.

Validation Data

Always provide X_valid and y_valid to see validation metrics in TensorBoard.

Background Process

launch_tensorboard() runs in background. To stop:

tb_process.terminate()