Google I/O is a wrap! Catch up on TensorFlow sessions View sessions

Model Cards in TFX

The ModelCardGenerator TFX pipeline component generates model cards.

For the detailed model card format, see the Model Card API.

For more general information about TFX, please see the TFX User Guide.

Configuring the ModelCardGenerator Component

The ModelCardGenerator takes dataset statistics, model evaluation, and a pushed model to automatically populate parts of a model card.

Model card fields can also be explicitly populated with a JSON string (this can be generated using the json module, see Example below). If a field is populated both by TFX and JSON, the JSON value will overwrite the TFX value.

The ModelCardGenerator writes model card documents to the model_card/ directory of its artifact output. It uses a default HTML model card template, which is used to generate model_card.html. Custom templates can also be used; each template input must be accompanied by a file name output in the template_io arg.


from model_card_toolkit import ModelCardGenerator
import json

model_card_fields = {
  'model_details': {
    'name': 'my_model',
    'owners': 'Google',
    'version': 'v0.1'
  'considerations': {
    'limitations': 'This is a demo model.'
mc_gen = ModelCardGenerator(
        ('html/default_template.html.jinja', 'model_card.html'),
        ('md/', '')

More details are available in the ModelCardGenerator API reference.

See our end-to-end demo for a full working example.