build_embedding.py

def build_embedding(df, features, emb_dim = 10, name = 'embedding_layer'):
    '''
        Define the embedding neural network to encode features in a emb_dim-dimensional vector.
        
        Parameters
        ----------
            df : pandas DataFrame
                    dataframe containing input metadata
            
            features : list of str
                            list of categorical features (columns of df)
                                
            emb_dim    : int
                                vector size, dimension of the embedding space
                                Default : 10
            
            name    : str
                        name given to the embedding layer
                        Default : 'embedding_layer'
        
        Return
        ------
            network : Keras Model object
                        Partial network architecture modelling embeddings to be trained
    '''
    inputs = []
    concat = []

    cat_sizes = {}
    cat_embsizes = {}
    for cat in features:
        cat_sizes[cat] = df[cat].nunique()
        cat_embsizes[cat] = min(50, cat_sizes[cat]//2+1)
    
        x = Input((1,), name=cat)
        inputs.append(x)
        x = Embedding(cat_sizes[cat] + 1, cat_embsizes[cat], input_length=1)(x)
        x = Reshape((cat_embsizes[cat],))(x)
        concat.append(x)
    
    if len(concat) > 1:
        x = Concatenate()(concat)
    
    x = Dense(emb_dim, activation='relu')(x)
    
    return x, inputs