feat: Add XAI SDK integration to TensorFlow models with LIT integration

[taiseiak]

Add automatic addition of feature attribution for TensorFlow 2 models in the LIT integration on Vertex Notebooks. Detects for Vertex Notebooks by looking for the same environment variable to check for Vertex Notebooks as the LIT library does.

Fixes b/210943910 🦕

go/local-explanations-lit-xai-notebook

[Ark-kun]

In the future, would it be possible to add the automatic addition of feature attribution for TensorFlow 2 models imported using Model.upload_tensorflow_saved_model?

[taiseiak]

Where is Models.upload_tensorflow_model? I cannot find it.

[Ark-kun]

Documentation: https://cloud.google.com/python/docs/reference/aiplatform/latest/google.cloud.aiplatform.Model#google_cloud_aiplatform_Model_upload_tensorflow_saved_model

Code:

python-aiplatform/google/cloud/aiplatform/models.py

Line 2857 in 8df74a2

def upload_tensorflow_saved_model(

[taiseiak]

Documentation: https://cloud.google.com/python/docs/reference/aiplatform/latest/google.cloud.aiplatform.Model#google_cloud_aiplatform_Model_upload_tensorflow_saved_model

Code:

python-aiplatform/google/cloud/aiplatform/models.py

Line 2857 in 8df74a2

def upload_tensorflow_saved_model(

I think this is possible, but I don't know if we should do this automatically. We charge users for using explanations on uploaded models. I think having a parameter that sets up explanation metadata or having a function that utilizes the XAI SDK to create metadata would be better. Would those suffice?

[Ark-kun]

Documentation: https://cloud.google.com/python/docs/reference/aiplatform/latest/google.cloud.aiplatform.Model#google_cloud_aiplatform_Model_upload_tensorflow_saved_model
Code:

python-aiplatform/google/cloud/aiplatform/models.py

Line 2857 in 8df74a2

def upload_tensorflow_saved_model(

I think this is possible, but I don't know if we should do this automatically. We charge users for using explanations on uploaded models. I think having a parameter that sets up explanation metadata or having a function that utilizes the XAI SDK to create metadata would be better. Would those suffice?

I have couple of questions:

1. Are users charged extra when XAI metadata is added to a model?
2. Is there an easy way for a user to add explanation metadata to a model that would fit most users?

[taiseiak]

Documentation: https://cloud.google.com/python/docs/reference/aiplatform/latest/google.cloud.aiplatform.Model#google_cloud_aiplatform_Model_upload_tensorflow_saved_model
Code:

python-aiplatform/google/cloud/aiplatform/models.py

Line 2857 in 8df74a2

def upload_tensorflow_saved_model(

I think this is possible, but I don't know if we should do this automatically. We charge users for using explanations on uploaded models. I think having a parameter that sets up explanation metadata or having a function that utilizes the XAI SDK to create metadata would be better. Would those suffice?

I have couple of questions:

1. Are users charged extra when XAI metadata is added to a model?
2. Is there an easy way for a user to add explanation metadata to a model that would fit most users?

1. Currently, no. However, this will change in the future. When a user deploys a model to an endpoint with explanation metadata, they will be charged for the explainer GKE pod for that endpoint. Users are also charged extra for using the explain call also as it takes computation time to calculate explanations.
2. The most easy way right now is to use the XAI SDK (https://github.com/GoogleCloudPlatform/explainable_ai_sdk). This SDK has many convenience methods for generating metadata. This is what I used to generate metadata for this PR.