Migrate SLi-Rec from TensorFlow to PyTorch

[miguelgfierro]

Description

Migrates the SLi-Rec sequential recommender from TensorFlow (TF-1.x graph mode) to PyTorch, as part of the TF→PyTorch migration tracked in #2302.

The port is a self-contained nn.Module mirroring the existing PyTorch graphrec/lightgcn.py and the lightgcn_deep_dive notebook: all hyper-parameters are explicit arguments — no prepare_hparams/yaml/HParams object. Architecture goes on the constructor, training knobs on fit, evaluation knobs on run_eval; it reuses only the framework-agnostic metric functions in deeprec_utils. The TensorFlow SLi-Rec is removed and replaced by this port; the five other TF sequential models (asvd, caser, gru, nextitnet, sum) and the shared TF base are untouched.

model = SLI_RECModel(
    user_vocab=user_vocab, item_vocab=item_vocab, cate_vocab=cate_vocab,
    item_embedding_dim=32, cate_embedding_dim=8, user_embedding_dim=16,
    hidden_size=40, attention_size=40, max_seq_length=50,
    layer_sizes=[100, 64], att_fcn_layer_sizes=[80, 40], dropout=[0.3, 0.3], seed=SEED,
)
model.fit(train_file, valid_file, epochs=10, batch_size=400, learning_rate=0.001,
          train_num_ngs=4, valid_num_ngs=4, embed_l2=0.0, layer_l2=0.0)
model.run_eval(test_file, num_ngs=9)   # -> {"auc": ..., "group_auc": ..., ...}

New files

• recommenders/models/deeprec/models/sequential/pytorch/rnn_cell_pytorch.py — Time4LSTMCell + a dynamic_rnn-equivalent scan driver.
• recommenders/models/deeprec/models/sequential/pytorch/sequential_base_pytorch.py — shared embeddings, unmasked ASVD Attention, FcnNet MLP head, softmax pairwise loss, unique-embedding regularization, and the fit/run_eval/predict lifecycle.
• recommenders/models/deeprec/models/sequential/pytorch/sli_rec_pytorch.py — SLI_RECModel (forward == the TF _build_seq_graph) + masked AttentionFcn.
• recommenders/models/deeprec/io/sequential_dataset_pytorch.py — faithful port of SequentialIterator (parsing, time features, padding, in-batch negatives).
• tests/unit/recommenders/models/test_sli_rec_pytorch.py — 12 plain-function unit tests.
• examples/00_quick_start/slirec_amazon.ipynb — a dedicated PyTorch SLi-Rec quickstart that defines parameters as plain Python and passes them explicitly (no prepare_hparams), so the functional test still exercises SLi-Rec end-to-end.

Removed (TensorFlow SLi-Rec)

• recommenders/models/deeprec/models/sequential/sli_rec.py and its TF-only time-aware cell file rnn_cell_implement.py (used exclusively by TF SLi-Rec).
• The TF SLi-Rec unit (test_slirec_component_definition) and smoke (test_model_slirec) tests and their test_groups.yml entries.
• docs/models.rst: the RNN Cells section; the SLIRec section now points at the PyTorch modules.

Changed

• tests/conftest.py — the slirec_quickstart notebook fixture points at the new slirec_amazon.ipynb; README links updated.
• examples/00_quick_start/sequential_recsys_amazondataset.ipynb — kept as the shared multi-model quickstart, but its default active model is switched from the removed TF SLi-Rec to A2SVD (same IJCAI'19 paper), with a pointer to the new SLi-Rec notebook. The other TF sequential models are unaffected.

Validation

Component-level parity (weights copied TF→PyTorch, on the real slirec data):

• Time4LSTMCell + scan: single-step m/c diff ~1e-7; full padded-sequence rnn_outputs diff ~6e-8; padded steps exactly zero.
• SequentialDataset: parser, eval-batch arrays, and train-batch arrays (incl. in-batch negative sampling under a shared RNG seed) all bit-identical to the TF iterator.

End-to-end training (10 epochs, batch 400, seed 42, embed_l2=layer_l2=0 — the functional-test config):

	untrained	epoch 1 (valid)	epoch 10 (valid)	test
PyTorch AUC	0.4802	0.5071	0.7614	0.7355
TF AUC	0.4857	0.4975	0.7369	0.7174

The functional test asserts auc == pytest.approx(0.7183, rel=0.1, abs=0.05) → accepts [0.6465, 0.7901]. The PyTorch test AUC 0.7355 passes comfortably and exceeds both the 0.7183 target and the TF reference (group_auc 0.7201 vs TF 0.7073). Unit tests: 12/12 pass on CPU. The untrained baseline AUC (0.4802) is bit-identical to the earlier hparams-based version, confirming the explicit-params refactor preserves initialization and behavior.

Related Issues

• [FEATURE] Migrate TensorFlow algorithms to PyTorch #2302

Checklist

• I have not rewritten tests or code to pass CI without addressing the underlying issue.
• I ran black to format the code.
• I have added tests (unit tests + the existing functional notebook test now exercises the PyTorch model).
• The five other TF sequential models are untouched (byte-identical to staging); their tests still collect.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[miguelgfierro]

@anargyri I updated the code, please take a look

[anargyri]

Are these methods used anywhere?