Can not export the transform graph from tf.Transform together with model definition using keras in TensorFlow 2.0.

[brightcoder01]

We cannot export the transform graph from tf.Transform and the graph from Keras Model together into a single SavedModel.

Background Let's focus on these three stages in the end-to-end machine learning lifecycle of TFX:
Transform stage: tf.Transform component transforms the raw data into the data used to train a machine learning model. It also exports the transform logic in the SavedModel format. We name it saved_model_trans.
Training stage: TensorFlow uses the preprocessed data to train a model. After completing the training job, it exports the model graph together with the transform graph (from saved_model_trans) as one SavedModel. We name it saved_model_final.
Serving stage: TensorFlow Serving can load saved_model_final to provide inference service. The schemas of the inference request and the raw data of the transform stage are exactly the same.

Current state: how to export transform graph and model graph together to a SavedModel?
√ Estimator (Work Well):
From tf.transform official tutorial, at the stage of exporting model, it will call estimator.export_saved_model(exported_model_dir, serving_input_fn) to complete the model exporting work. Inside its implementation, it calls serving_input_fn to load the transform graph from SavedModel at first, and then calls estimator's model_fn to generate the model graph, combines these two graph into one graph and finally exports it into one SavedModel. Please check the code snippet.

× Keras (Can not):
For TF2.0, we define a model using keras and exports it by calling tf.saved.saved_model. This SavedModel only contains the model definition including feature columns and NN structure. tf.saved.saved_model api doesn't have the parameter serving_input_fn just like estimator and it lacks the flexibility to combine transform graph and model graph together for inference. It will break the integration between tf.Transform and TensorFlow 2.0.

We want to improve tf.saved.saved_model api in Keras to support this function. So that TensorFlow 2.0 can work well with tf.Transform.

[brightcoder01]

Hi, @zoyahav @rmothukuru , could you please share more information about this bug?

[schmidt-jake]

As a hack, I implemented a custom Keras layer that applies the tf-transform function (saved_model_trans) on my raw inputs. This ensures the preprocessing function is transmitted to the saved_model_final for serving. But it's very hacky and also doesn't allow preprocessing transformations on the raw targets (ie mapping strings to integers for classification tasks). Can we expect that tf-transform will support keras models?

[mrcolo]

Any progress on this issue?

[zoyahav]

Yes, tf.transform currently supports Keras, though we're actively working on improving this support. There's currently the TFT layer which should be used when using Keras: https://www.tensorflow.org/tfx/transform/api_docs/python/tft/TFTransformOutput#transform_features_layer

Example usage: https://github.com/tensorflow/transform/blob/master/examples/census_example_v2.py#L344-L356 Note that the layer should be saved to the model as an attribute for resource tracking purposes.

[mrcolo]

Thanks, super helpful!

[rmothukuru]

@brightcoder01, Can you please let us know if this comment has resolved your issue? Thanks!

[rmothukuru]

Automatically closing due to lack of recent activity. Please update the issue when new information becomes available, and we will reopen the issue. Thanks!