What is the objective function to optimize?

[dreamquark-ai]

What is the objective function to optimize?

[Optimox]

hello @xywust2014,

There are two distinct metrics to consider:

• the loss function that your model is trained to minimize
• the early stopping metric, which is used to select the best epoch and stop the training when no further improvement has been seen.

Currently you can easily change the loss function by using loss_fn parameter during the fit, you can pick any loss function from pytorch https://pytorch.org/docs/stable/nn.html or even create your own if you like. By default TabNetClassifier uses cross entropy while TabnetRegressor uses Mean Squarred Error.

About the early stopping metric, you can't change it easily with the current implementation, this will be improved in the future. For binary classification the early stop metric (which is the metric that the training displays) is AUC. For multiclass classfication it's the accuracy. For regression it's the mean squarred error.

You can easily access your training loss and early stop metrics by calling clf.history['train']['loss'] or clf.history['train']['metric'] or with the same command for the valid set clf.history['valid']['metric']

Hope this helps!

[xywust2014]

hello @xywust2014,

There are two distinct metrics to consider:

• the loss function that your model is trained to minimize
• the early stopping metric, which is used to select the best epoch and stop the training when no further improvement has been seen.

Currently you can easily change the loss function by using loss_fn parameter during the fit, you can pick any loss function from pytorch https://pytorch.org/docs/stable/nn.html or even create your own if you like. By default TabNetClassifier uses cross entropy while TabnetRegressor uses Mean Squarred Error.

About the early stopping metric, you can't change it easily with the current implementation, this will be improved in the future. For binary classification the early stop metric (which is the metric that the training displays) is AUC. For multiclass classfication it's the accuracy. For regression it's the mean squarred error.

You can easily access your training loss and early stop metrics by calling clf.history['train']['loss'] or clf.history['train']['metric'] or with the same command for the valid set clf.history['valid']['metric']

Hope this helps!

Thank you! Optimox. This helps a lot.