Skip to content

Callbacks

Weights & Biases Logging callback. To use this callback pip install wandb. Any metric that contains epoch will be plotted with epoch and all the other metrics will be plotted against global_step which is total training steps. You can change the default axis by providing default_step_metric.

Parameters:

Name Type Description Default
log_model bool

Whether to upload model artifact to Wandb

False
code_file Optional[str]

path of the code you want to upload as artifact to Wandb

None
default_step_metric

Metrics will be plotted against the default_step_metric. Default value is global_step.

'global_step'
from gradsflow.callbacks import WandbCallback
from timm import create_model

cnn = create_model("resnet18", pretrained=False, num_classes=1)
model = Model(cnn)
model.compile()
cb = WandbCallback()
autodataset = None  # create your dataset
model.fit(autodataset, callbacks=cb)

on_epoch_end(self)

Called after each epoch

on_fit_start(self)

Called on each model.fit(...)

on_train_step_end(self, outputs=None, **_)

Called after training step


Tracks the carbon emissions produced by deep neural networks using CodeCarbon. To use this callback first install codecarbon using pip install codecarbon. For offline use, you must have to specify the country code.

Parameters:

Name Type Description Default
offline bool

whether to use internet connection or not. You will have to provide the country code country_iso_code for offline use.

False
**kwargs

passed directly to codecarbon class.

{}

on_fit_end(self)

Called after model.fit(...)


Comet Logging callback. This callback requires comet-ml to be pre-installed (pip install comet-ml). Automatically log your Experiment to Comet logging platform. You need to provide API key either by setting environment variable COMET_API_KEY or directly pass as an argument to the callback. Checkout the documentation for more examples.

Parameters:

Name Type Description Default
project_name str

Name of the Project

'awesome-project'
api_key Optional[str]

project API key

None
offline bool

log experiment offline

False

on_epoch_end(self)

Called after each epoch

on_fit_start(self)

Called on each model.fit(...)

on_train_epoch_start(self)

Called on start of training epoch

on_train_step_end(self, outputs=None, **_)

Called after training step

on_val_epoch_start(self)

Called on start of validation epoch

on_val_step_end(self, outputs=None, **_)

Called after validation step



Last update: January 14, 2022