linearcorex.py 文件源码

def __init__(self, n_hidden=10, max_iter=10000, tol=1e-5, anneal=True, missing_values=None,
                 discourage_overlap=True, gaussianize='standard', gpu=False,
                 verbose=False, seed=None):
        self.m = n_hidden  # Number of latent factors to learn
        self.max_iter = max_iter  # Number of iterations to try
        self.tol = tol  # Threshold for convergence
        self.anneal = anneal
        self.eps = 0  # If anneal is True, it's adjusted during optimization to avoid local minima
        self.missing_values = missing_values

        self.discourage_overlap = discourage_overlap  # Whether or not to discourage overlapping latent factors
        self.gaussianize = gaussianize  # Preprocess data: 'standard' scales to zero mean and unit variance
        self.gpu = gpu  # Enable GPU support for some large matrix multiplications.
        if self.gpu:
            cm.cublas_init()

        self.yscale = 1.  # Can be arbitrary, but sets the scale of Y
        np.random.seed(seed)  # Set seed for deterministic results
        self.verbose = verbose
        if verbose:
            np.set_printoptions(precision=3, suppress=True, linewidth=160)
            print(('Linear CorEx with {:d} latent factors'.format(n_hidden)))

        # Initialize these when we fit on data
        self.n_samples, self.nv = 0, 0  # Number of samples/variables in input data
        self.ws = np.zeros((0, 0))  # m by nv array of weights
        self.moments = {}  # Dictionary of moments
        self.theta = None  # Parameters for preprocessing each variable
        self.history = {}  # Keep track of values for each iteration
        self.last_update = 0  # Used for momentum methods