[FEATURE] Replace SASRec TF with PyTorch version

[miguelgfierro]

Description

SASRec tests are disabled: https://github.com/recommenders-team/recommenders/blob/main/tests/ci/azureml_tests/test_groups.py#L410
We could replace the TF algo with https://github.com/microsoft/UniRec/blob/main/unirec/model/sequential/sasrec.py

Expected behavior with the suggested feature

Branch: https://github.com/recommenders-team/recommenders/tree/miguel/sasrec_unirec

Tasks:

• Create unit tests of the classes
• Make sure the original script runs
• Create a functional test with the minimal parts of the code training movielens

Other Comments

[miguelgfierro]

pytest -s tests/unit/recommenders/models/test_unirec_model.py::test_sasrec_train --disable-warnings

[miguelgfierro]

import cvxpy as cp
E   ModuleNotFoundError: No module named 'cvxpy'

solved with pip install cvxpy

another error:

FAILED tests/unit/recommenders/models/test_unirec_model.py::test_sasrec_train - ModuleNotFoundError: No module named 'feather'

solved by installing install feather-format

[miguelgfierro]

FIXED

 @pytest.mark.gpu
    def test_sasrec_train(base_config, unirec_config_path):
        # config = copy.deepcopy(base_config)
        # yaml_file = os.path.join(unirec_config_path, "model", "SASRec.yaml")
        # config.update(load_yaml(yaml_file))
    
        # model = SASRec(config)
        import copy
        import datetime
        from recommenders.models.unirec.main import main
    
        GLOBAL_CONF = {
            # "config_dir": f"{os.path.join(unirec_config_path, 'unirec', 'config')}",
            "config_dir": unirec_config_path,
            "exp_name": "pytest",
            "checkpoint_dir": f'{datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")}',
            "model": "",
            "dataloader": "SeqRecDataset",
            "dataset": "",
            "dataset_path": os.path.join(unirec_config_path, "tests/.temp/data"),
            "output_path": "",
            "learning_rate": 0.001,
            "dropout_prob": 0.0,
            "embedding_size": 32,
            "hidden_size": 32,
            "use_pre_item_emb": 0,
            "loss_type": "bce",
            "max_seq_len": 10,
            "has_user_bias": 1,
            "has_item_bias": 1,
            "epochs": 1,
            "early_stop": -1,
            "batch_size": 512,
            "n_sample_neg_train": 9,
            "valid_protocol": "one_vs_all",
            "test_protocol": "one_vs_all",
            "grad_clip_value": 0.1,
            "weight_decay": 1e-6,
            "history_mask_mode": "autoagressive",
            "user_history_filename": "user_history",
            "metrics": "['hit@5;10', 'ndcg@5;10']",
            "key_metric": "ndcg@5",
            "num_workers": 4,
            "num_workers_test": 0,
            "verbose": 2,
            "neg_by_pop_alpha": 0.0,
            "conv_size": 10,  # for ConvFormer-series
        }
        config = copy.deepcopy(GLOBAL_CONF)
        config["task"] = "train"
        config["dataset_path"] = os.path.join(config["dataset_path"], "ml-100k")
        config["dataset"] = "ml-100k"
        config["model"] = "SASRec"
        config["output_path"] = os.path.join(unirec_config_path, f"tests/.temp/output/")
>       result = main.run(config)

tests/unit/recommenders/models/test_unirec_model.py:146: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
recommenders/models/unirec/main/main.py:676: in run
    res = main(config, accelerator)
recommenders/models/unirec/main/main.py:357: in main
    user2history, user2history_time = get_user_history(
recommenders/models/unirec/main/main.py:137: in get_user_history
    user2history, user2history_time = general.load_user_history(
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

file_path = '/home/u/MS/recommenders/recommenders/models/unirec/config/tests/.temp/data/ml-100k', file_name = 'user_history', n_users = 940, format = 'user-item_seq', time_seq = 0

    def load_user_history(
        file_path, file_name, n_users=None, format="user-item", time_seq=0
    ):
        if os.path.exists(os.path.join(file_path, file_name + ".ftr")):
            df = pd.read_feather(os.path.join(file_path, file_name + ".ftr"))
        elif os.path.exists(os.path.join(file_path, file_name + ".pkl")):
            df = load_pkl_obj(os.path.join(file_path, file_name + ".pkl"))
        else:
>           raise NotImplementedError(
                "Unsupported user history file type: {0}".format(file_name)
            )
E           NotImplementedError: Unsupported user history file type: user_history

recommenders/models/unirec/utils/general.py:134: NotImplementedError
----------------------------------------------------------------------------------------------------- Captured log call -----------------------------------------------------------------------------------------------------
INFO     SASRec-pytest:logger.py:61 config={'gpu_id': 0, 'use_gpu': True, 'seed': 2022, 'state': 'INFO', 'verbose': 2, 'saved': True, 'use_tensorboard': False, 'use_wandb': False, 'init_method': 'normal', 'init_std': 0.02, 'init_mean': 0.0, 'scheduler': 'reduce', 'scheduler_factor': 0.1, 'time_seq': 0, 'seq_last': False, 'has_user_emb': False, 'has_user_bias': 1, 'has_item_bias': 1, 'use_features': False, 'use_text_emb': False, 'use_position_emb': True, 'load_pretrained_model': False, 'embedding_size': 32, 'hidden_size': 32, 'inner_size': 512, 'dropout_prob': 0.0, 'epochs': 1, 'batch_size': 512, 'learning_rate': 0.001, 'optimizer': 'adam', 'eval_step': 1, 'early_stop': -1, 'clip_grad_norm': None, 'weight_decay': 1e-06, 'num_workers': 4, 'persistent_workers': False, 'pin_memory': False, 'shuffle_train': False, 'use_pre_item_emb': 0, 'loss_type': 'bce', 'ccl_w': 150, 'ccl_m': 0.4, 'distance_type': 'dot', 'metrics': "['hit@5;10', 'ndcg@5;10']", 'key_metric': 'ndcg@5', 'test_protocol': 'one_vs_all', 'valid_protocol': 'one_vs_all', 'test_batch_size': 100, 'model': 'SASRec', 'dataloader': 'SeqRecDataset', 'max_seq_len': 10, 'history_mask_mode': 'autoagressive', 'tau': 1.0, 'enable_morec': 0, 'morec_objectives': ['fairness', 'alignment', 'revenue'], 'morec_objective_controller': 'PID', 'morec_ngroup': [10, 10, -1], 'morec_alpha': 0.1, 'morec_lambda': 0.2, 'morec_expect_loss': 0.2, 'morec_beta_min': 0.6, 'morec_beta_max': 1.3, 'morec_K_p': 0.01, 'morec_K_i': 0.001, 'morec_objective_weights': '[0.3,0.3,0.4]', 'n_layers': 2, 'n_heads': 16, 'hidden_dropout_prob': 0.5, 'attn_dropout_prob': 0.5, 'hidden_act': 'swish', 'layer_norm_eps': '1e-10', 'group_size': -1, 'n_items': 1017, 'n_neg_test_from_sampling': 0, 'n_neg_train_from_sampling': 0, 'n_neg_valid_from_sampling': 0, 'n_users': 940, 'test_file_format': 'user-item', 'train_file_format': 'user-item', 'user_history_file_format': 'user-item_seq', 'valid_file_format': 'user-item', 'base_model': 'GRU', 'freeze': 0, 'train_type': 'Base', 'config_dir': PosixPath('/home/u/MS/recommenders/recommenders/models/unirec/config'), 'exp_name': 'SASRec-pytest', 'checkpoint_dir': '2024-07-05_12-25-03', 'dataset': 'ml-100k', 'dataset_path': '/home/u/MS/recommenders/recommenders/models/unirec/config/tests/.temp/data/ml-100k', 'output_path': '/home/u/MS/recommenders/recommenders/models/unirec/config/tests/.temp/output/', 'n_sample_neg_train': 9, 'grad_clip_value': 0.1, 'user_history_filename': 'user_history', 'num_workers_test': 0, 'neg_by_pop_alpha': 0.0, 'conv_size': 10, 'task': 'train', 'cmd_args': {'base_model': 'GRU', 'freeze': 0, 'train_type': 'Base', 'config_dir': PosixPath('/home/u/MS/recommenders/recommenders/models/unirec/config'), 'exp_name': 'SASRec-pytest', 'checkpoint_dir': '2024-07-05_12-25-03', 'model': 'SASRec', 'dataloader': 'SeqRecDataset', 'dataset': 'ml-100k', 'dataset_path': '/home/u/MS/recommenders/recommenders/models/unirec/config/tests/.temp/data/ml-100k', 'output_path': '/home/u/MS/recommenders/recommenders/models/unirec/config/tests/.temp/output/', 'learning_rate': 0.001, 'dropout_prob': 0.0, 'embedding_size': 32, 'hidden_size': 32, 'use_pre_item_emb': 0, 'loss_type': 'bce', 'max_seq_len': 10, 'has_user_bias': 1, 'has_item_bias': 1, 'epochs': 1, 'early_stop': -1, 'batch_size': 512, 'n_sample_neg_train': 9, 'valid_protocol': 'one_vs_all', 'test_protocol': 'one_vs_all', 'grad_clip_value': 0.1, 'weight_decay': 1e-06, 'history_mask_mode': 'autoagressive', 'user_history_filename': 'user_history', 'metrics': "['hit@5;10', 'ndcg@5;10']", 'key_metric': 'ndcg@5', 'num_workers': 4, 'num_workers_test': 0, 'verbose': 2, 'neg_by_pop_alpha': 0.0, 'conv_size': 10, 'task': 'train', 'logger_time_str': '2024-07-05_122503', 'logger_rand': 91}, 'device': device(type='cpu'), 'logger_time_str': '2024-07-05_122503', 'logger_rand': 91}
INFO     SASRec-pytest:main.py:136 Loading user history from user_history ...
================================================================================================== short test summary info ==================================================================================================
FAILED tests/unit/recommenders/models/test_unirec_model.py::test_sasrec_train - NotImplementedError: Unsupported user history file type: user_history

If I download the original repo and run the tests, I get the same error:

TOL = 0.05
ABS_TOL = 0.05

GLOBAL_CONF = {
    "config_dir": f"{os.path.join(UNIREC_PATH, 'unirec', 'config')}",
    "exp_name": "pytest",
    "checkpoint_dir": f'{datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")}',
    "model": "",
    "dataloader": "SeqRecDataset",
    "dataset": "",
    "dataset_path": os.path.join(UNIREC_PATH, "tests/.temp/data"),
    "output_path": "",
    "learning_rate": 0.001,
    "dropout_prob": 0.0,
    "embedding_size": 32,
    "hidden_size": 32,
    "use_pre_item_emb": 0,
    "loss_type": "bce",
    "max_seq_len": 10,
    "has_user_bias": 1,
    "has_item_bias": 1,
    "epochs": 1,  # 3, # MIG
    "early_stop": -1,
    "batch_size": 512,
    "n_sample_neg_train": 9,
    "valid_protocol": "one_vs_all",
    "test_protocol": "one_vs_all",
    "grad_clip_value": 0.1,
    "weight_decay": 1e-6,
    "history_mask_mode": "autoagressive",
    "user_history_filename": "user_history",
    "metrics": "['hit@5;10', 'ndcg@5;10']",
    "key_metric": "ndcg@5",
    "num_workers": 4,
    "num_workers_test": 0,
    "verbose": 2,
    "neg_by_pop_alpha": 0.0,
    "conv_size": 10,  # for ConvFormer-series
}

# SEQ_MODELS = ["SVDPlusPlus", "FASTConvFormer", "ConvFormer", "SASRec", "AvgHist", "GRU", "AttHist"]  # Each test is ordered according to the list
# SEQ_MODELS = ["SASRec"]
SEQ_MODELS = ["SVDPlusPlus"]
LOSS_TYPES = ["bce", "bpr", "softmax", "ccl", "fullsoftmax"]
EXPECTED_METRICS = {
    "SVDPlusPlus": {"hit@5": 0.04792, "ndcg@5": 0.03394},
    "FASTConvFormer": {"hit@5": 0.05005, "ndcg@5": 0.03355},
    "ConvFormer": {"hit@5": 0.05005, "ndcg@5": 0.03538},
    "SASRec": {"hit@5": 0.04792, "ndcg@5": 0.03184},
    "AvgHist": {"hit@5": 0.05005, "ndcg@5": 0.03423},
    "GRU": {"hit@5": 0.04686, "ndcg@5": 0.03197},
    "AttHist": {"hit@5": 0.04686, "ndcg@5": 0.03221},
    "SASRec_bce": {"hit@5": 0.04792, "ndcg@5": 0.03184},
    "SASRec_bpr": {"hit@5": 0.04686, "ndcg@5": 0.03122},
    "SASRec_softmax": {"hit@5": 0.04686, "ndcg@5": 0.03066},
    "SASRec_ccl": {"hit@5": 0.02449, "ndcg@5": 0.01318},
    "SASRec_fullsoftmax": {"hit@5": 0.04792, "ndcg@5": 0.03155},
    "SASRec_with_text_emb": {"hit@5": 0.04686, "ndcg@5": 0.03219},
    "SASRec_with_max_len": {"hit@5": 0.04686, "ndcg@5": 0.03122},
}


# >>>>>> Test train pipeline of sequential models and check the performance
# Note: the test instance should be put in the first place because the model checkpoint files generated here are required in following tests
@pytest.mark.parametrize(
    "data, models, expected_values", [("ml-100k", SEQ_MODELS, EXPECTED_METRICS)]
)
def test_train_pipeline(data, models, expected_values):
    all_result = {}
    # finish all training first for following evaluation and infer test
    for model in models:
        config = copy.deepcopy(GLOBAL_CONF)
        config["task"] = "train"
        config["dataset_path"] = os.path.join(config["dataset_path"], data)
        config["dataset"] = data
        config["model"] = model
        config["output_path"] = os.path.join(
            UNIREC_PATH, f"tests/.temp/output/{data}/{model}"
        )
        result = main.run(config)
        all_result[model] = result

    # check the performance
    failed_models = []
    for model in models:
        exp_value = expected_values[model]
        result = all_result[model]
        for k, v in exp_value.items():
            if not result[k] == pytest.approx(v, rel=TOL, abs=ABS_TOL):
                failed_models.append(model)
                break
    assert (
        len(failed_models) == 0
    ), f"performance of [{', '.join(failed_models)}] not correct."


$ tests/test_model/test_seq_model_mig.py F                                                                               [100%]

========================================================================================================= FAILURES =========================================================================================================
__________________________________________________________________________________ test_train_pipeline[ml-100k-models0-expected_values0] ___________________________________________________________________________________

data = 'ml-100k', models = ['SVDPlusPlus']
expected_values = {'AttHist': {'hit@5': 0.04686, 'ndcg@5': 0.03221}, 'AvgHist': {'hit@5': 0.05005, 'ndcg@5': 0.03423}, 'ConvFormer': {'hit@5': 0.05005, 'ndcg@5': 0.03538}, 'FASTConvFormer': {'hit@5': 0.05005, 'ndcg@5': 0.03355}, ...}

    @pytest.mark.parametrize(
        "data, models, expected_values", [("ml-100k", SEQ_MODELS, EXPECTED_METRICS)]
    )
    def test_train_pipeline(data, models, expected_values):
        all_result = {}
        # finish all training first for following evaluation and infer test
        for model in models:
            config = copy.deepcopy(GLOBAL_CONF)
            config["task"] = "train"
            config["dataset_path"] = os.path.join(config["dataset_path"], data)
            config["dataset"] = data
            config["model"] = model
            config["output_path"] = os.path.join(
                UNIREC_PATH, f"tests/.temp/output/{data}/{model}"
            )
>           result = main.run(config)

tests/test_model/test_seq_model_mig.py:97: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
unirec/main/main.py:492: in run
    res = main(config, accelerator)
unirec/main/main.py:272: in main
    user2history, user2history_time = get_user_history(user2history, user2history_time, config, DATA_TRAIN_NAME)
unirec/main/main.py:116: in get_user_history
    user2history, user2history_time = general.load_user_history(file_path, _user_history_filename, config['n_users'], _user_history_data_format, config['time_seq'])
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

file_path = '/home/u/MS/UniRec/tests/.temp/data/ml-100k', file_name = 'user_history', n_users = 940, format = 'user-item_seq', time_seq = 0

    def load_user_history(file_path, file_name, n_users=None, format='user-item', time_seq=0):
        if os.path.exists(os.path.join(file_path, file_name + '.ftr')):
            df = pd.read_feather(os.path.join(file_path, file_name + '.ftr'))
        elif os.path.exists(os.path.join(file_path, file_name + '.pkl')):
            df = load_pkl_obj(os.path.join(file_path, file_name + '.pkl'))
        else:
>           raise NotImplementedError("Unsupported user history file type: {0}".format(file_name) )
E           NotImplementedError: Unsupported user history file type: user_history

unirec/utils/general.py:117: NotImplementedError

We need to download the dataset python download_split_ml100k.py and preprocess it sh preprocess_ml100k.sh before being able to run the training.

[miguelgfierro]

Work so far: staging...miguel/sasrec_unirec

Next step is to create a unit test called test_sasrec_train which should train sasrec with the minimum set of options on a dummy dataset. We should first make sure that the code with result = main.run(config) runs, and then, replace it with the minimum set of functions.

The steps should follow the structure of https://github.com/recommenders-team/recommenders/blob/main/examples/00_quick_start/sar_movielens.ipynb:

• Data loading
• split train and test iterators
• instantiate the model
• train the model

if we want, we can also do:

• evaluate
• get metrics

After this, we will create a notebook explaining an end 2 end case with a real dataset, and we will replace the TF notebook.

Tasks:

• Understand the extra dependencies I need to add: only accelerate and cvxpy. Both seem strong libraries.
• Understand python download_split_ml100k.py.
• Understand sh preprocess_ml100k.sh.
• Understand result = main.run(config)
• Replicate and adapt python download_split_ml100k.py.
• Replicate and adapt sh preprocess_ml100k.sh.
• Replicate and adapt result = main.run(config)