Class Dense

mla/neuralnet/layers/basic.py:53–89

class Dense(Layer, ParamMixin):
    def __init__(self, output_dim, parameters=None):
        """A fully connected layer.

        Parameters
        ----------
        output_dim : int
        """
        self._params = parameters
        self.output_dim = output_dim
        self.last_input = None

        if parameters is None:
            self._params = Parameters()

    def setup(self, x_shape):
        self._params.setup_weights((x_shape[1], self.output_dim))

    def forward_pass(self, X):
        self.last_input = X
        return self.weight(X)

    def weight(self, X):
        W = np.dot(X, self._params["W"])
        return W + self._params["b"]

    def backward_pass(self, delta):
        dW = np.dot(self.last_input.T, delta)
        db = np.sum(delta, axis=0)

        # Update gradient values
        self._params.update_grad("W", dW)
        self._params.update_grad("b", db)
        return np.dot(delta, self._params["W"].T)

    def shape(self, x_shape):
        return x_shape[0], self.output_dim


class Activation(Layer):