[FEA] Support feeding pre-trained embeddings to TF4Rec model with high-level api

[rnyak]

🚀 Feature request

Currently we do not have out of the box support for adding pre-trained embeddings to embedding layer, and ability to freeze them, and train a TF4Rec model. We have embedding_initializer but we never tested if it works accurately and as expected. May be we can create in PyTorch a class like TensorInitializer (TF) as we did in Merlin Models and expose the embedding initializer and trainable args to the user.

We need to

• Expose definition of embeddings module in the input blocks: TabularFeatures and TabularSequenceFeatures
• Support feeding pre-trained embeddings to TF4Rec model with high-level api (users should be add them to the embedding layer, and freeze them, i.e., set trainable=False (TF Api) or requires_grad=False (PyTorch API))
• create an example notebook for showcasing that functionality

Motivation

This is a FEA coming from our customers and users.

[karlhigley]

Is this related to or part of NVIDIA-Merlin/Merlin#211?

[rnyak]

Is this related to or part of NVIDIA-Merlin/Merlin#211?

@karlhigley more related to NVIDIA-Merlin/Merlin#471. Not sure about the link to 211.

[gabrielspmoreira]

When the embedding table are not huge and fit GPU memory, the new PretrainedEmbeddingsInitializer ( #572 ) can be used to initialize the embedding matrix with pre-trained embeddings and set them to trainable or not.