clearml-docs/docs/guides/frameworks/huggingface/transformers.md

title
Transformers

The Hugging Face Transformers example demonstrates how to integrate ClearML into your Transformer's Trainer code. The Hugging Face Trainer automatically uses the built-in ClearMLCallback if the clearml package is already installed, to log Transformers models, parameters, scalars, and more.

In the example, ClearML is installed and set up in the training environment. This way ClearML can log models, parameters, scalars, and more.

When the example runs, it creates a ClearML task called Trainer in the HuggingFace Transformers project. To change the task's name or project, use the CLEARML_PROJECT and CLEARML_TASK environment variables respectively.

For more information about integrating ClearML into your Transformers code, see Hugging Face Transformers.

WebApp

Hyperparameters

ClearML automatically captures all the Trainer parameters. Notice in the code example that only a few of the TrainingArguments are explicitly set:

training_args = TrainingArguments(
     output_dir="path/to/save/folder/",
     learning_rate=2e-5,
     per_device_train_batch_size=8,
     per_device_eval_batch_size=8,
     num_train_epochs=2,
)

ClearML captures all of the TrainingArguments passed to the Trainer.

View these parameters in the experiment's CONFIGURATION tab > Hyperparameters section.

Models

In order for ClearML to log the models created during training in this example, the CLEARML_LOG_MODEL environment variable is set to True.

ClearML automatically captures the model snapshots created by the Trainer, and saves them as artifacts. View the snapshots in the experiment's ARTIFACTS tab.

Scalars

ClearML automatically captures the Trainer's scalars, which can be viewed in the experiment's Scalars tab.