Python MatrixMul.lmul_T方法代码示例

# 需要导入模块: from pylearn2.linear.matrixmul import MatrixMul [as 别名]
# 或者: from pylearn2.linear.matrixmul.MatrixMul import lmul_T [as 别名]

#.........这里部分代码省略.........
                    ('mean_x.min_u', v_mean.min())
                    ]:
                rval[prefix+key] = val

        return rval

    def sample(self, state_below = None, state_above = None,
            layer_above = None,
            theano_rng = None):

        if theano_rng is None:
            raise ValueError("theano_rng is required; it just defaults to None so that it may appear after layer_above / state_above in the list.")

        if state_above is not None:
            msg = layer_above.downward_message(state_above)
        else:
            msg = None

        if self.requires_reformat:
            state_below = self.input_space.format_as(state_below, self.desired_space)

        z = self.transformer.lmul(state_below) + self.b

        if msg != None:
            z = z + msg

        on_prob = T.nnet.sigmoid(2. * z)

        samples = theano_rng.binomial(p = on_prob, n=1, size=on_prob.shape, dtype=on_prob.dtype) * 2. - 1.

        return samples

    def downward_message(self, downward_state):
        rval = self.transformer.lmul_T(downward_state)

        if self.requires_reformat:
            rval = self.desired_space.format_as(rval, self.input_space)

        return rval

    def init_mf_state(self):
        raise NotImplementedError("This is just a copy-paste of BVMP")
        # work around theano bug with broadcasted vectors
        z = T.alloc(0., self.dbm.batch_size, self.detector_layer_dim).astype(self.b.dtype) + \
                self.b.dimshuffle('x', 0)
        rval = max_pool_channels(z = z,
                pool_size = self.pool_size)
        return rval

    def make_state(self, num_examples, numpy_rng):
        """ Returns a shared variable containing an actual state
           (not a mean field state) for this variable.
        """
        driver = numpy_rng.uniform(0.,1., (num_examples, self.dim))
        on_prob = sigmoid_numpy(2. * self.b.get_value())
        sample = 2. * (driver < on_prob) - 1.

        rval = sharedX(sample, name = 'v_sample_shared')

        return rval

    def make_symbolic_state(self, num_examples, theano_rng):
        mean = T.nnet.sigmoid(2. * self.b)
        rval = theano_rng.binomial(size=(num_examples, self.nvis), p=mean)
        rval = 2. * (rval) - 1.

# 需要导入模块: from pylearn2.linear.matrixmul import MatrixMul [as 别名]
# 或者: from pylearn2.linear.matrixmul.MatrixMul import lmul_T [as 别名]

#.........这里部分代码省略.........
            training examples and the second indexing data dimensions.

        Returns
        -------
        a : tensor_like or list of tensor_likes
            Theano symbolic (or list thereof) representing the input to each
            hidden unit for each training example.
        """

        if isinstance(v, tensor.Variable):
            return self.bias_hid + self.transformer.lmul(v)
        else:
            return [self.input_to_h_from_v(vis) for vis in v]

    def input_to_v_from_h(self, h):
        """
        Compute the affine function (linear map plus bias) that serves as
        input to the visible layer in an RBM.

        Parameters
        ----------
        h  : tensor_like or list of tensor_likes
            Theano symbolic (or list thereof) representing the one or several
            minibatches on the hidden units, with the first dimension indexing
            training examples and the second indexing data dimensions.

        Returns
        -------
        a : tensor_like or list of tensor_likes
            Theano symbolic (or list thereof) representing the input to each
            visible unit for each row of h.
        """
        if isinstance(h, tensor.Variable):
            return self.bias_vis + self.transformer.lmul_T(h)
        else:
            return [self.input_to_v_from_h(hid) for hid in h]

    def upward_pass(self, v):
        """
        wrapper around mean_h_given_v method.  Called when RBM is accessed
        by mlp.HiddenLayer.
        """
        return self.mean_h_given_v(v)

    def mean_h_given_v(self, v):
        """
        Compute the mean activation of the hidden units given visible unit
        configurations for a set of training examples.

        Parameters
        ----------
        v  : tensor_like or list of tensor_likes
            Theano symbolic (or list thereof) representing the hidden unit
            states for a batch (or several) of training examples, with the
            first dimension indexing training examples and the second indexing
            data dimensions.

        Returns
        -------
        h : tensor_like or list of tensor_likes
            Theano symbolic (or list thereof) representing the mean
            (deterministic) hidden unit activations given the visible units.
        """
        if isinstance(v, tensor.Variable):
            return nnet.sigmoid(self.input_to_h_from_v(v))
        else:

# 需要导入模块: from pylearn2.linear.matrixmul import MatrixMul [as 别名]
# 或者: from pylearn2.linear.matrixmul.MatrixMul import lmul_T [as 别名]

#.........这里部分代码省略.........

        for s, t, c, e in safe_zip(state, target, coeff, eps):
            assert all([isinstance(elem, float) for elem in [t, c, e]])
            if c == 0.:
                continue
            m = s.mean(axis=0)
            assert m.ndim == 1
            rval += T.maximum(abs(m-t)-e,0.).mean()*c

        return rval

    def sample(self, state_below = None, state_above = None,
            layer_above = None,
            theano_rng = None):

        if theano_rng is None:
            raise ValueError("theano_rng is required; it just defaults to None so that it may appear after layer_above / state_above in the list.")

        if state_above is not None:
            msg = layer_above.downward_message(state_above)
        else:
            msg = None

        if self.requires_reformat:
            state_below = self.input_space.format_as(state_below, self.desired_space)

        z = self.transformer.lmul(state_below) + self.b
        p, h, p_sample, h_sample = max_pool_channels(z,
                self.pool_size, msg, theano_rng)

        return p_sample, h_sample

    def downward_message(self, downward_state):
        rval = self.transformer.lmul_T(downward_state)

        if self.requires_reformat:
            rval = self.desired_space.format_as(rval, self.input_space)

        return rval

    def make_state(self, num_examples, numpy_rng):
        """ Returns a shared variable containing an actual state
           (not a mean field state) for this variable.
        """

        t1 = time.time()

        empty_input = self.h_space.get_origin_batch(num_examples)
        h_state = sharedX(empty_input)

        default_z = T.zeros_like(h_state) + self.b

        theano_rng = MRG_RandomStreams(numpy_rng.randint(2 ** 16))

        p_exp, h_exp, p_sample, h_sample = max_pool_channels(
                z = default_z,
                pool_size = self.pool_size,
                theano_rng = theano_rng)

        assert h_sample.dtype == default_z.dtype

        p_state = sharedX( self.output_space.get_origin_batch(
            num_examples))

        t2 = time.time()