Using a sparse coo matrix with Ray Tune

1

Hello,

I’m using an algorithm that is called lightfm and it takes sparse matrices as input and it does some conversions to scipy coo matrices behind the scenes. So, I don’t have control over this conversion part.

When trying to do hyperparameter optimization, I get the following error

AttributeError: 'ray._raylet.ObjectRef' object has no attribute 'tocoo'

How can I solve this issue ? Here is the full code

from ray import air, tune, put
from ray.tune.search.optuna import OptunaSearch

# train interactions, test_interations & dataset_item_features are sparse matrices
ray_train_interactions = put(train_interactions)
ray_test_interactions = put(test_interactions)
ray_dataset_item_features = put(dataset_item_features)


def train_lightfm(config):
    
    model = LightFM(no_components=config["components"], 
                learning_rate=config["lr"],
                learning_schedule="adagrad",
                item_alpha=config["item_alpha"],
                loss="warp")
    
    for i in range(30):
        model = model.fit_partial(ray_train_interactions,
                        epochs=1, 
                        item_features=ray_dataset_item_features,
                        num_threads=2)
        maptk = precision_at_k(model, ray_test_interactions, item_features = ray_dataset_item_features, k=5).mean()
        out = -maptk
        
        # Handle some weird numerical going on
        if np.abs(out + 1) < 0.01 or out < -1.0:
            out = 0
        
        tune.report(out) 


search_space = {
    "lr": tune.loguniform(1e-5, 5e-1),
    "components": tune.uniform(10, 100),
    "item_alpha": tune.loguniform(1e-4, 5e-1)
}

tuner = tune.Tuner(
    train_lightfm,
    param_space=search_space,
    tune_config=tune.TuneConfig(
        search_alg=optuna_search,
    ),
)

results = tuner.fit()

Thanks

2

Maybe you should ask this question in Ray AIR cateogry…

cc: @kai @matthewdeng