layers.py 文件源码

def __init__(self, rng, name, is_train, x, n_in, n_out, W=None, b=None, activation=ReLU, p=0.5):
        """p is the probability of NOT dropping out a unit"""
        self.name = name
        self.x = x
        bound = np.sqrt(6./(n_in+n_out))
        if W is None:
            W_values = np.asarray(
                    rng.uniform(
                        low=-bound,
                        high=bound,
                        size=(n_in, n_out)
                        ),
                    dtype=theano.config.floatX)
            if activation == theano.tensor.nnet.sigmoid:
                W_values *= 4
            W = theano.shared(value=W_values, name='W', borrow=True)

        if b is None:
            # b_values = np.zeros((n_out,), dtype=theano.config.floatX)
            b_values = np.ones((n_out,), dtype=theano.config.floatX) * np.cast[theano.config.floatX](bound)
            b = theano.shared(value=b_values, name='b', borrow=True)

        self.W = W
        self.b = b

        lin_output= T.dot(x, self.W) + self.b
        output = (
                lin_output if activation is None
                else activation(lin_output))

        def drop(x, rng=rng, p=p):
            """p is the probability of NOT dropping out a unit"""
            srng = RandomStreams(rng.randint(999999))
            mask = srng.binomial(n=1, p=p, size=x.shape, dtype=theano.config.floatX)
            return x * mask

        train_output = drop(np.cast[theano.config.floatX](1./p) * output)

        self.output = T.switch(T.neq(is_train, 0), train_output, output)

        self.params = [self.W, self.b]