| description |
|---|
How to integrate a Keras script to log metrics to W&B |
Use the Keras callback to automatically save all the metrics and the loss values tracked in model.fit.
{% code title="example.py" %}
import wandb
from wandb.keras import WandbCallback
wandb.init(config={"hyper": "parameter"})
# Magic
model.fit(X_train, y_train, validation_data=(X_test, y_test),
callbacks=[WandbCallback()]){% endcode %}
See our example projects for a complete script example.
Keras WandbCallback() class supports a number of options:
| Keyword argument | Default | Description |
|---|---|---|
| monitor | val_loss | The training metric used to measure performance for saving the best model. i.e. val_loss |
| mode | auto | 'min', 'max', or 'auto': How to compare the training metric specified in monitor between steps |
| save_weights_only | False | only save the weights instead of the entire model |
| save_model | True | save the model if it's improved at each step |
| log_weights | False | log the values of each layers parameters at each epoch |
| log_gradients | False | log the gradients of each layers parametres at each epcoh |
| training_data | None | tuple (X,y) needed for calculating gradients |
| data_type | None | the type of data we're saving, currently only "image" is supported |
| labels | None | only used if data_type is specified, list of labels to convert numeric output to if you are building classifier. (supports binary classification) |
| predictions | 36 | the number of predictions to make if data_type is specified. Max is 100. |
| generator | None | if using data augmentation and data_type you can specify a generator to make predictions with. |
If you're setting use_multiprocessing=True and seeing the error Error('You must call wandb.init() before wandb.config.batch_size') then try this:
- In the Sequence class init, add:
wandb.init(group='...') - In your main program, make sure you're using
if __name__ == "__main__":and then put the rest of your script logic inside that.