Module tfrec.utils.model_utils

Functions

def cross_validate(model, X, y, folds=5, epochs=5, batch_size=32, callbacks=None, shuffle=False, random_state=None)

Performs K-Fold cross validation on data.

Parameters

model : tf.keras.Model

A compiled model with appropriate metrics to keep track of.

X : numpy.ndarray

Numpy array of shape (m, 2) of user-item pairs.

y : numpy.ndarray

Numpy array of shape (m,) of target rating.

folds : integer, optional

Number of splits for K Fold (default is 5).

epochs : integer, optional

Epochs to fit the model (default is 10).

batch_size : integer, optional

Batch_size to be passed in while fitting (default is 32).

callbacks : list, optional

List of callbacks to be passed in while fitting (default None).

shuffle : boolean, optional

Shuffle before splitting into batches or not (default False).

random_state : integer, optional

Seed variable, only useful if shuffle = True (default None).

Returns

numpy.ndarray

Numpy array of shape (folds, 1(loss) + no_of_metrics).

Contains validation scores evaluated for each fold.

Expand source code

def cross_validate(model, X, y, folds=5, epochs=5, batch_size=32, callbacks=None, shuffle=False, random_state=None):
    """Performs K-Fold cross validation on data.

    Parameters
    ----------
    model : tf.keras.Model
        A compiled model with appropriate metrics to keep track of.
    X : numpy.ndarray
        Numpy array of shape `(m, 2)` of user-item pairs.
    y : numpy.ndarray
        Numpy array of shape `(m,)` of target rating.
    folds : integer, optional
        Number of splits for K Fold (default is `5`).
    epochs : integer, optional
        Epochs to fit the model (default is `10`).
    batch_size : integer, optional
        Batch_size to be passed in while fitting (default is `32`).
    callbacks : list, optional
        List of callbacks to be passed in while fitting (default `None`).
    shuffle : boolean, optional
        Shuffle before splitting into batches or not (default `False`).
    random_state : integer, optional
        Seed variable, only useful if `shuffle = True` (default `None`).

    Returns
    -------
    numpy.ndarray
        Numpy array of shape `(folds, 1(loss) + no_of_metrics)`.

        Contains validation scores evaluated for each fold.
    """

    # Initalize KFold
    kfolds = KFold(n_splits=folds, random_state=random_state, shuffle=shuffle)
    all_metrics = []

    # To build the model
    if type(model).__name__ == 'SVDpp':
        model.implicit_feedback(X[:10, :])
    model(X[:10, :])

    # Workaround to reset weights after each fold fit
    weights = model.get_weights()
    i = 1

    for train, val in kfolds.split(X, y):

        # Gather implicit feedback if model is SVD++
        if type(model).__name__ == 'SVDpp':
            model.implicit_feedback(X[train])

        print(f'\nFitting on Fold {i}')
        # Train and evaluate metrics
        history = model.fit(
            X[train], y[train], batch_size=batch_size, epochs=epochs, callbacks=callbacks)
        print(f'\nEvaluating on Fold {i}')
        fold_score = history.model.evaluate(X[val], y[val])
        all_metrics.append(fold_score)

        if folds != i:
            model.set_weights(weights)

        i += 1

    all_metrics = np.array(all_metrics)

    for i, metric in enumerate(model.metrics_names):
        print(f'Mean {metric.capitalize()} : {np.mean(all_metrics.T[i])}')

    return all_metrics, model

def preprocess_and_split(df, shuffle=True, split=True, test_size=0.2, random_state=None)

Preprocesses and Splits the dataset if necessary.

Encodes & Decodes User and Item, each from 0 to len(user) and len(item).

Splits into Train and Test and shuffles(if shuffle = True).

Returns Dataset as numpy arrays and Encoding and Decodings of users and items.

Parameters

df : pandas.core.frame.DataFrame

Dataframe of shape (m,3) with columns in the order - user, item, rating.

shuffle : boolean, optional

Boolean to determine should the dataset be shuffled after encoding (default is True).

split : boolean, optional

Boolean to determine if the dataset should be split into train and test set (default is True).

test_size : float, optional

Size of test split (default is 0.2).

random_state : integer, optional

Seed variable, useful for reproducing results (default is None).

Returns

dataset, user_item_encodings : tuple

Dataset contains tuple of train and test variables if shuffle = True.

dataset

(x_train, y_train), (x_test, y_test).

user_item_encodings

(user_to_encoded, encoded_to_user,item_to_encoded, encoded_to_item).

Expand source code

def preprocess_and_split(df, shuffle=True, split=True, test_size=0.2, random_state=None):
    """Preprocesses and Splits the dataset if necessary.

    **Encodes** & **Decodes** User and Item, each from `0 to len(user) and len(item)`.

    Splits into **Train** and **Test** and shuffles(if `shuffle = True`).

    Returns Dataset as numpy arrays and **Encoding** and **Decodings** of users and items.

    Parameters
    ----------
    df : pandas.core.frame.DataFrame
        Dataframe of shape `(m,3)` with columns in the order - `user`, `item`, `rating`.
    shuffle : boolean, optional
        Boolean to determine should the dataset be shuffled after encoding (default is `True`).
    split : boolean, optional
        Boolean to determine if the dataset should be split into train and test set (default is `True`).
    test_size : float, optional
        Size of test split (default is `0.2`).
    random_state : integer, optional
        Seed variable, useful for reproducing results (default is `None`).

    Returns
    -------
    dataset, user_item_encodings : tuple
        Dataset contains tuple of train and test variables if `shuffle = True`.

        **dataset**

            (x_train, y_train), (x_test, y_test).

        **user_item_encodings**

            (user_to_encoded, encoded_to_user,item_to_encoded, encoded_to_item).
    """

    # Get unique ids
    user_ids = df.iloc[:, 0].unique().tolist()
    item_ids = df.iloc[:, 1].unique().tolist()

    # Create encoding and decoding maps for users
    user_to_encoded = {user_id: encoded_id for encoded_id,
                       user_id in enumerate(user_ids)}
    encoded_to_user = {encoded_id: user_id for encoded_id,
                       user_id in enumerate(user_ids)}

    # Create encoding and decoding maps for items
    item_to_encoded = {item_id: encoded_id for encoded_id,
                       item_id in enumerate(item_ids)}
    encoded_to_item = {encoded_id: item_id for encoded_id,
                       item_id in enumerate(item_ids)}

    user_item_encodings = (user_to_encoded, encoded_to_user,
                           item_to_encoded, encoded_to_item)

    # Encode Dataset
    df_encoded = pd.DataFrame([], columns=['user', 'item', 'rating'])
    df_encoded['user'] = df.iloc[:, 0].map(user_to_encoded)
    df_encoded['item'] = df.iloc[:, 1].map(item_to_encoded)
    df_encoded['rating'] = df.iloc[:, 2].values.astype(np.float32)

    # Shuffle if shuffle = True
    if shuffle:
        df_encoded = df_encoded.sample(frac=1, random_state=None)

    # Dataframe -> Numpy Arrays
    x = df_encoded[['user', 'item']].values
    y = df_encoded[['rating']].values
    dataset = x, y

    # Split if split = True
    if split:
        x_train, x_test, y_train, y_test = train_test_split(
            x, y, test_size=test_size, random_state=random_state
        )
        dataset = (x_train, y_train), (x_test, y_test)

    return dataset, user_item_encodings