[FEATURE] Migrate TensorFlow algorithms to PyTorch

[miguelgfierro]

Description

Curent TF algos:

 Here are the TensorFlow-based models in the repo:

  ┌────────────┬──────────────────────────────────────────────────────────────────────────────┐
  │ Subpackage │                                    Models                                    │
  ├────────────┼──────────────────────────────────────────────────────────────────────────────┤
  │ deeprec    │ XDeepFM, DKN, DKN_Item2Item, LightGCN, GRU, Caser, A2SVD, NextItNet, SLI_REC │
  ├────────────┼──────────────────────────────────────────────────────────────────────────────┤
  │ newsrec    │ NAML, NRMS, NPA, LSTUR                                                       │
  ├────────────┼──────────────────────────────────────────────────────────────────────────────┤
  │ ncf        │ Neural Collaborative Filtering                                               │
  ├────────────┼──────────────────────────────────────────────────────────────────────────────┤
  │ vae        │ MultinomialVAE, StandardVAE                                                  │
  ├────────────┼──────────────────────────────────────────────────────────────────────────────┤
  │ rbm        │ Restricted Boltzmann Machine                                                 │
  └────────────┴──────────────────────────────────────────────────────────────────────────────┘

  That's 5 subpackages with ~17 model implementations total. The deeprec and newsrec packages are the largest, each with their own base class hierarchy. Most use
  TF 1.x compatibility mode (tf.compat.v1), while newsrec uses Keras.

Expected behavior with the suggested feature

Working in PyTorch

Willingness to contribute

• Yes, I can contribute for this issue independently.
• Yes, I can contribute for this issue with guidance from Recommenders community.
• No, I cannot contribute at this time.

Other Comments

[miguelgfierro]

Related to:

• [BUG] tensorflow-estimator is removed from tensorflow 2.16.1 #2072
• [ASK] Stopped the support of new Tensorflow versions >= 2.16 and migrate to PyTorch #2073

[miguelgfierro]

@ds-wook these are the algos we have to migrate.

We think that VAEs or the RBM could be easy. Let us know if you want to take any of those? Then @anargyri and I will take some others

[ds-wook]

@miguelgfierro

I’ll take the VAE (MultinomialVAE, StandardVAE), Neural Collaborative Filtering, and XDeepFM.

I’ve previously worked with PyTorch-based implementations/libraries for these types of models, so I should be able to handle the migration.

[miguelgfierro]

the xDeepFM requires to migrate all deeprec package which is quite large. Perhaps is better to start with VAE, do a PR then NCF, do another PR, and then we can review what to do with deeprec

[ds-wook]

Sounds good, I’ll start with VAE and proceed step by step. Thanks for the opportunity to work on this!

[miguelgfierro]

I talked to @anargyri, he will take RBM and I will take NCF. And then we can work on the other ones. Does that sound good?

[ds-wook]

Good! Then I’ll take the VAE and start working on it first.