Weights & Biases (W&B) Tutorial: TF2

1. Log-in the wandb in a command window by the following command: $ wandb login

2. Open the jupyter notebook

3. Initialize wandb as follows:

wandb.init(project="bayesian_optimization_test01",
           config={"lr": 0.001,
                  "layer_size": 64}
          )

In config you can specify all the hyper-parameters you wanna track.

4. Log

• all the metrics and loss values can be stored by the following code:

  from wandb.keras import WandbCallback
  ...
  model.fit(X_train, y_train,  validation_data=(X_test, y_test),
          callbacks=[WandbCallback()])

• you can log any data by wandb.log(..)

  wandb.log({"mse": mse})

Once the training begins, your W&B project page link will show up where you can keep track of the training status.

5. Visualize effects of the hyper-parameters w.r.t your metrics

charts to analyze the effects of the hyper-parameters

Note that these two charts can be added through Parallel coordinates and Parameter importance from the + dropdown menu.

 

References

• docs.wandb.com/quickstart
• docs.wandb.com/library/integrations/keras, docs.wandb.com/library/integrations/jupyter
• docs.wandb.com/library/init, docs.wandb.com/library/config, docs.wandb.com/library/log