Python theano.tensor.ones_like() Examples

def test_gpualloc_output_to_gpu():
    a_val = numpy.asarray(numpy.random.rand(4, 5), dtype='float32')
    a = tcn.shared_constructor(a_val)

    b = T.fscalar()
    f = theano.function([b], T.ones_like(a) + b, mode=mode_without_gpu)
    f_gpu = theano.function([b], B.gpu_from_host(T.ones_like(a)) + b,
                            mode=mode_with_gpu)

    f(2)
    f_gpu(2)

    assert sum([node.op == T.alloc for node in f.maker.fgraph.toposort()]) == 1
    assert sum([node.op == B.gpu_alloc
                for node in f_gpu.maker.fgraph.toposort()]) == 1

    assert numpy.allclose(numpy.ones(a.get_value(borrow=True).shape) + 9,
                          f_gpu(9))
    assert numpy.allclose(f(5), f_gpu(5)) 

def sp_ones_like(x):
    """
    Construct a sparse matrix of ones with the same sparsity pattern.

    Parameters
    ----------
    x
        Sparse matrix to take the sparsity pattern.

    Returns
    -------
    A sparse matrix
        The same as `x` with data changed for ones.

    """
    # TODO: don't restrict to CSM formats
    data, indices, indptr, shape = csm_properties(x)
    return CSM(format=x.format)(tensor.ones_like(data), indices, indptr, shape) 

def normalize_batch_in_training(x, gamma, beta,
                                reduction_axes, epsilon=1e-3):
    """Computes mean and std for batch then apply batch_normalization on batch.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_train is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_train'):
        return _old_normalize_batch_in_training(x, gamma, beta, reduction_axes, epsilon)

    if gamma is None:
        if beta is None:
            gamma = ones_like(x)
        else:
            gamma = ones_like(beta)
    if beta is None:
        if gamma is None:
            beta = zeros_like(x)
        beta = zeros_like(gamma)

    normed, mean, stdinv = T.nnet.bn.batch_normalization_train(
        x, gamma, beta, reduction_axes, epsilon)

    return normed, mean, T.inv(stdinv ** 2) 

def test_int32_dtype(self):
        # Reported on the theano-user mailing-list:
        # https://groups.google.com/d/msg/theano-users/MT9ui8LtTsY/rwatwEF9zWAJ
        size = 9
        intX = 'int32'

        C = tensor.matrix('C', dtype=intX)
        I = tensor.matrix('I', dtype=intX)

        fI = I.flatten()
        data = tensor.ones_like(fI)
        indptr = tensor.arange(data.shape[0] + 1, dtype='int32')

        m1 = sparse.CSR(data, fI, indptr, (8, size))
        m2 = sparse.dot(m1, C)
        y = m2.reshape(shape=(2, 4, 9), ndim=3)

        f = theano.function(inputs=[I, C], outputs=y)
        i = numpy.asarray([[4, 3, 7, 7], [2, 8, 4, 5]], dtype=intX)
        a = numpy.asarray(numpy.random.randint(0, 100, (size, size)),
                          dtype=intX)
        f(i, a) 

def batch_normalization(x, mean, var, beta, gamma, epsilon=1e-3):
    """Apply batch normalization on x given mean, var, beta and gamma.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_test is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_test'):
        return _old_batch_normalization(x, mean, var, beta, gamma, epsilon)

    if gamma is None:
        gamma = ones_like(var)
    if beta is None:
        beta = zeros_like(mean)

    if mean.ndim == 1:
        # based on TensorFlow's default: normalize along rightmost dimension
        reduction_axes = list(range(x.ndim - 1))
    else:
        reduction_axes = [i for i in range(x.ndim) if mean.broadcastable[i]]

    return T.nnet.bn.batch_normalization_test(
        x, gamma, beta, mean, var, reduction_axes, epsilon)


# TODO remove this function when Theano without
# T.nnet.bn.batch_normalization_train is deprecated 

def test_structured_add_s_v(self):
        sp_types = {'csc': sp.csc_matrix,
                    'csr': sp.csr_matrix}

        for format in ['csr', 'csc']:
            for dtype in ['float32', 'float64']:
                x = theano.sparse.SparseType(format, dtype=dtype)()
                y = tensor.vector(dtype=dtype)
                f = theano.function([x, y], structured_add_s_v(x, y))

                spmat = sp_types[format](random_lil((4, 3), dtype, 3))
                spones = spmat.copy()
                spones.data = numpy.ones_like(spones.data)
                mat = numpy.asarray(numpy.random.rand(3), dtype=dtype)

                out = f(spmat, mat)

                utt.assert_allclose(as_ndarray(spones.multiply(spmat + mat)),
                                    out.toarray()) 

def batch_normalization(x, mean, var, beta, gamma, epsilon=1e-3):
    """Apply batch normalization on x given mean, var, beta and gamma.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_test is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_test'):
        return _old_batch_normalization(x, mean, var, beta, gamma, epsilon)

    if gamma is None:
        gamma = ones_like(var)
    if beta is None:
        beta = zeros_like(mean)

    if mean.ndim == 1:
        # based on TensorFlow's default: normalize along rightmost dimension
        reduction_axes = list(range(x.ndim - 1))
    else:
        reduction_axes = [i for i in range(x.ndim) if mean.broadcastable[i]]

    return T.nnet.bn.batch_normalization_test(
        x, gamma, beta, mean, var, reduction_axes, epsilon)


# TODO remove this function when Theano without
# T.nnet.bn.batch_normalization_train is deprecated 

def make_node(self, acts, labels, input_lengths):
    # Unless specified, assume all sequences have full sequence length, i.e. acts_.shape[0]
    if input_lengths == None:
      input_lengths = T.cast(acts.shape[0], dtype="int32") * T.ones_like(acts[0,:,0], dtype=np.int32)

    # acts.shape = [seqLen, batchN, outputUnit]
    if acts.dtype != "float32":
      raise Exception("acts must be float32 instead of %s" % acts.dtype)
    # labels.shape = [batchN, labelLen]
    if labels.dtype != "int32":
      raise Exception("labels must be int32 instead of %s" % labels.dtype)
    # input_lengths.shape = [batchN]
    if input_lengths.dtype != "int32":
      raise Exception("input_lengths must be int32 instead of %s" % input_lengths.dtype)

    applyNode = theano.Apply(self, inputs=[acts, input_lengths, labels], outputs=[self.costs, self.gradients])

    # Return only the cost. Gradient will be returned by grad()
    self.default_output = 0 

    return applyNode 

def normalize_batch_in_training(x, gamma, beta,
                                reduction_axes, epsilon=1e-3):
    """Computes mean and std for batch then apply batch_normalization on batch.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_train is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_train'):
        return _old_normalize_batch_in_training(x, gamma, beta, reduction_axes, epsilon)

    if gamma is None:
        if beta is None:
            gamma = ones_like(x)
        else:
            gamma = ones_like(beta)
    if beta is None:
        if gamma is None:
            beta = zeros_like(x)
        beta = zeros_like(gamma)

    normed, mean, stdinv = T.nnet.bn.batch_normalization_train(
        x, gamma, beta, reduction_axes, epsilon)

    return normed, mean, T.inv(stdinv ** 2) 

def batch_normalization(x, mean, var, beta, gamma, epsilon=1e-3):
    """Apply batch normalization on x given mean, var, beta and gamma.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_test is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_test'):
        return _old_batch_normalization(x, mean, var, beta, gamma, epsilon)

    if gamma is None:
        gamma = ones_like(var)
    if beta is None:
        beta = zeros_like(mean)

    if mean.ndim == 1:
        # based on TensorFlow's default: normalize along rightmost dimension
        reduction_axes = list(range(x.ndim - 1))
    else:
        reduction_axes = [i for i in range(x.ndim) if mean.broadcastable[i]]

    return T.nnet.bn.batch_normalization_test(
        x, gamma, beta, mean, var, reduction_axes, epsilon)


# TODO remove this function when Theano without
# T.nnet.bn.batch_normalization_train is deprecated 

def test_structured_add_s_v(self):
        sp_types = {'csc': sp.csc_matrix,
                    'csr': sp.csr_matrix}

        for format in ['csr', 'csc']:
            for dtype in ['float32', 'float64']:
                x = theano.sparse.SparseType(format, dtype=dtype)()
                y = tensor.vector(dtype=dtype)
                f = theano.function([x, y], structured_add_s_v(x, y))

                spmat = sp_types[format](random_lil((4, 3), dtype, 3))
                spones = spmat.copy()
                spones.data = numpy.ones_like(spones.data)
                mat = numpy.asarray(numpy.random.rand(3), dtype=dtype)

                out = f(spmat, mat)

                utt.assert_allclose(as_ndarray(spones.multiply(spmat + mat)),
                                    out.toarray()) 

def normalize_batch_in_training(x, gamma, beta,
                                reduction_axes, epsilon=1e-3):
    """Computes mean and std for batch then apply batch_normalization on batch.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_train is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_train'):
        return _old_normalize_batch_in_training(x, gamma, beta, reduction_axes, epsilon)

    if gamma is None:
        if beta is None:
            gamma = ones_like(x)
        else:
            gamma = ones_like(beta)
    if beta is None:
        if gamma is None:
            beta = zeros_like(x)
        beta = zeros_like(gamma)

    normed, mean, stdinv = T.nnet.bn.batch_normalization_train(
        x, gamma, beta, reduction_axes, epsilon)

    return normed, mean, T.inv(stdinv ** 2) 

def batch_normalization(x, mean, var, beta, gamma, epsilon=1e-3):
    """Apply batch normalization on x given mean, var, beta and gamma.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_test is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_test'):
        return _old_batch_normalization(x, mean, var, beta, gamma, epsilon)

    if gamma is None:
        gamma = ones_like(var)
    if beta is None:
        beta = zeros_like(mean)

    if mean.ndim == 1:
        # based on TensorFlow's default: normalize along rightmost dimension
        reduction_axes = list(range(x.ndim - 1))
    else:
        reduction_axes = [i for i in range(x.ndim) if mean.broadcastable[i]]

    return T.nnet.bn.batch_normalization_test(
        x, gamma, beta, mean, var, reduction_axes, epsilon)


# TODO remove this function when Theano without
# T.nnet.bn.batch_normalization_train is deprecated 

def sp_ones_like(x):
    """
    Construct a sparse matrix of ones with the same sparsity pattern.

    Parameters
    ----------
    x
        Sparse matrix to take the sparsity pattern.

    Returns
    -------
    A sparse matrix
        The same as `x` with data changed for ones.

    """
    # TODO: don't restrict to CSM formats
    data, indices, indptr, shape = csm_properties(x)
    return CSM(format=x.format)(tensor.ones_like(data), indices, indptr, shape) 

def test_int32_dtype(self):
        # Reported on the theano-user mailing-list:
        # https://groups.google.com/d/msg/theano-users/MT9ui8LtTsY/rwatwEF9zWAJ
        size = 9
        intX = 'int32'

        C = tensor.matrix('C', dtype=intX)
        I = tensor.matrix('I', dtype=intX)

        fI = I.flatten()
        data = tensor.ones_like(fI)
        indptr = tensor.arange(data.shape[0] + 1, dtype='int32')

        m1 = sparse.CSR(data, fI, indptr, (8, size))
        m2 = sparse.dot(m1, C)
        y = m2.reshape(shape=(2, 4, 9), ndim=3)

        f = theano.function(inputs=[I, C], outputs=y)
        i = numpy.asarray([[4, 3, 7, 7], [2, 8, 4, 5]], dtype=intX)
        a = numpy.asarray(numpy.random.randint(0, 100, (size, size)),
                          dtype=intX)
        f(i, a) 

def test_gpujoin_gpualloc():
    a = T.fmatrix('a')
    a_val = numpy.asarray(numpy.random.rand(4, 5), dtype='float32')
    b = T.fmatrix('b')
    b_val = numpy.asarray(numpy.random.rand(3, 5), dtype='float32')

    f = theano.function([a, b], T.join(0, T.zeros_like(a), T.ones_like(b)) + 4,
                        mode=mode_without_gpu)
    f_gpu = theano.function([a, b], T.join(0, T.zeros_like(a), T.ones_like(b)),
                            mode=mode_with_gpu)
    f_gpu2 = theano.function([a, b], T.join(0, T.zeros_like(a),
                                           T.ones_like(b)) + 4,
                             mode=mode_with_gpu)

    assert sum([node.op == T.alloc for node in f.maker.fgraph.toposort()]) == 2
    assert sum([node.op == T.join for node in f.maker.fgraph.toposort()]) == 1
    assert sum([isinstance(node.op, B.GpuAlloc)
                for node in f_gpu.maker.fgraph.toposort()]) == 2
    assert sum([node.op == B.gpu_join
                for node in f_gpu.maker.fgraph.toposort()]) == 1
    assert sum([isinstance(node.op, B.GpuAlloc)
                for node in f_gpu2.maker.fgraph.toposort()]) == 2
    assert sum([node.op == B.gpu_join
                for node in f_gpu2.maker.fgraph.toposort()]) == 1
    assert numpy.allclose(f(a_val, b_val), f_gpu2(a_val, b_val)) 

def normalize_batch_in_training(x, gamma, beta,
                                reduction_axes, epsilon=1e-3):
    """Computes mean and std for batch then apply batch_normalization on batch.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_train is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_train'):
        return _old_normalize_batch_in_training(x, gamma, beta, reduction_axes, epsilon)

    if gamma is None:
        if beta is None:
            gamma = ones_like(x)
        else:
            gamma = ones_like(beta)
    if beta is None:
        if gamma is None:
            beta = zeros_like(x)
        beta = zeros_like(gamma)

    normed, mean, stdinv = T.nnet.bn.batch_normalization_train(
        x, gamma, beta, reduction_axes, epsilon)

    return normed, mean, T.inv(stdinv ** 2) 

def compute_kernel(lls, lsf, x, z):

    ls = T.exp(lls)
    sf = T.exp(lsf)

    if x.ndim == 1:
        x = x[ None, : ]

    if z.ndim == 1:
        z = z[ None, : ]

    lsre = T.outer(T.ones_like(x[ :, 0 ]), ls)

    r2 = T.outer(T.sum(x * x / lsre, 1), T.ones_like(z[ : , 0 : 1 ])) - np.float32(2) * \
        T.dot(x / lsre, T.transpose(z)) + T.dot(np.float32(1.0) / lsre, T.transpose(z)**2)

    k = sf * T.exp(-np.float32(0.5) * r2)

    return k 

def compute_psi1(lls, lsf, xmean, xvar, z):

    if xmean.ndim == 1:
        xmean = xmean[ None, : ]

    ls = T.exp(lls)
    sf = T.exp(lsf)
    lspxvar = ls + xvar
    constterm1 = ls / lspxvar
    constterm2 = T.prod(T.sqrt(constterm1), 1)
    r2_psi1 = T.outer(T.sum(xmean * xmean / lspxvar, 1), T.ones_like(z[ : , 0 : 1 ])) \
        - np.float32(2) * T.dot(xmean / lspxvar, T.transpose(z)) + \
        T.dot(np.float32(1.0) / lspxvar, T.transpose(z)**2)
    psi1 = sf * T.outer(constterm2, T.ones_like(z[ : , 0 : 1 ])) * T.exp(-np.float32(0.5) * r2_psi1)

    return psi1 

def compute_log_ei(self, x, incumbent):

        Kzz = compute_kernel(self.lls, self.lsf, self.z, self.z) + T.eye(self.z.shape[ 0 ]) * self.jitter * T.exp(self.lsf)
        KzzInv = T.nlinalg.MatrixInversePSD()(Kzz)
        LLt = T.dot(self.LParamPost, T.transpose(self.LParamPost))
        covCavityInv = KzzInv + LLt * casting(self.n_points - self.set_for_training) / casting(self.n_points)
        covCavity = T.nlinalg.MatrixInversePSD()(covCavityInv)
        meanCavity = T.dot(covCavity, casting(self.n_points - self.set_for_training) / casting(self.n_points) * self.mParamPost)
        KzzInvcovCavity = T.dot(KzzInv, covCavity)
        KzzInvmeanCavity = T.dot(KzzInv, meanCavity)
        Kxz = compute_kernel(self.lls, self.lsf, x, self.z)
        B = T.dot(KzzInvcovCavity, KzzInv) - KzzInv 
        v_out = T.exp(self.lsf) + T.dot(Kxz * T.dot(Kxz, B), T.ones_like(self.z[ : , 0 : 1 ])) # + T.exp(self.lvar_noise)
        m_out = T.dot(Kxz, KzzInvmeanCavity)
        s = (incumbent - m_out) / T.sqrt(v_out)

        log_ei = T.log((incumbent - m_out) * ratio(s) + T.sqrt(v_out)) + log_n_pdf(s)

        return log_ei 

def compute_kernel(lls, lsf, x, z):

    ls = T.exp(lls)
    sf = T.exp(lsf)

    if x.ndim == 1:
        x = x[ None, : ]

    if z.ndim == 1:
        z = z[ None, : ]

    lsre = T.outer(T.ones_like(x[ :, 0 ]), ls)

    r2 = T.outer(T.sum(x * x / lsre, 1), T.ones_like(z[ : , 0 : 1 ])) - np.float32(2) * \
        T.dot(x / lsre, T.transpose(z)) + T.dot(np.float32(1.0) / lsre, T.transpose(z)**2)

    k = sf * T.exp(-np.float32(0.5) * r2)

    return k 

def compute_psi1(lls, lsf, xmean, xvar, z):

    if xmean.ndim == 1:
        xmean = xmean[ None, : ]

    ls = T.exp(lls)
    sf = T.exp(lsf)
    lspxvar = ls + xvar
    constterm1 = ls / lspxvar
    constterm2 = T.prod(T.sqrt(constterm1), 1)
    r2_psi1 = T.outer(T.sum(xmean * xmean / lspxvar, 1), T.ones_like(z[ : , 0 : 1 ])) \
        - np.float32(2) * T.dot(xmean / lspxvar, T.transpose(z)) + \
        T.dot(np.float32(1.0) / lspxvar, T.transpose(z)**2)
    psi1 = sf * T.outer(constterm2, T.ones_like(z[ : , 0 : 1 ])) * T.exp(-np.float32(0.5) * r2_psi1)

    return psi1 

def compute_log_ei(self, x, incumbent):

        Kzz = compute_kernel(self.lls, self.lsf, self.z, self.z) + T.eye(self.z.shape[ 0 ]) * self.jitter * T.exp(self.lsf)
        KzzInv = T.nlinalg.MatrixInversePSD()(Kzz)
        LLt = T.dot(self.LParamPost, T.transpose(self.LParamPost))
        covCavityInv = KzzInv + LLt * casting(self.n_points - self.set_for_training) / casting(self.n_points)
        covCavity = T.nlinalg.MatrixInversePSD()(covCavityInv)
        meanCavity = T.dot(covCavity, casting(self.n_points - self.set_for_training) / casting(self.n_points) * self.mParamPost)
        KzzInvcovCavity = T.dot(KzzInv, covCavity)
        KzzInvmeanCavity = T.dot(KzzInv, meanCavity)
        Kxz = compute_kernel(self.lls, self.lsf, x, self.z)
        B = T.dot(KzzInvcovCavity, KzzInv) - KzzInv 
        v_out = T.exp(self.lsf) + T.dot(Kxz * T.dot(Kxz, B), T.ones_like(self.z[ : , 0 : 1 ])) # + T.exp(self.lvar_noise)
        m_out = T.dot(Kxz, KzzInvmeanCavity)
        s = (incumbent - m_out) / T.sqrt(v_out)

        log_ei = T.log((incumbent - m_out) * ratio(s) + T.sqrt(v_out)) + log_n_pdf(s)

        return log_ei 

def apply(self, inputs, states, cells, mask=None):
        def slice_last(x, no):
            return x[:, no * self.dim: (no + 1) * self.dim]

        activation = tensor.dot(states, self.W_state) + inputs
        in_gate = self.gate_activation.apply(
            slice_last(activation, 0))
        pre = slice_last(activation, 1)
        forget_gate = self.gate_activation.apply(
            pre + self.bias * tensor.ones_like(pre))
        next_cells = (
            forget_gate * cells +
            in_gate * self.activation.apply(slice_last(activation, 2)))
        out_gate = self.gate_activation.apply(
            slice_last(activation, 3))
        next_states = out_gate * self.activation.apply(next_cells)

        if mask:
            next_states = (mask[:, None] * next_states +
                           (1 - mask[:, None]) * states)
            next_cells = (mask[:, None] * next_cells +
                          (1 - mask[:, None]) * cells)

        return next_states, next_cells 

def batch_normalization(x, mean, var, beta, gamma, axis=-1, epsilon=1e-3):
    """Apply batch normalization on x given mean, var, beta and gamma.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_test is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_test'):
        return _old_batch_normalization(x, mean, var, beta, gamma, epsilon)

    if gamma is None:
        gamma = ones_like(var)
    if beta is None:
        beta = zeros_like(mean)

    if mean.ndim == 1:
        # based on TensorFlow's default: normalize along rightmost dimension
        reduction_axes = list(range(x.ndim - 1))
    else:
        reduction_axes = [i for i in range(x.ndim) if mean.broadcastable[i]]

    return T.nnet.bn.batch_normalization_test(
        x, gamma, beta, mean, var, reduction_axes, epsilon)


# TODO remove this function when Theano without
# T.nnet.bn.batch_normalization_train is deprecated 

def _compute_training_statistics(self, input_):
        axes = (0,) + tuple((i + 1) for i, b in
                            enumerate(self.population_mean.broadcastable)
                            if b)
        mean = input_.mean(axis=axes, keepdims=True)
        assert mean.broadcastable[1:] == self.population_mean.broadcastable
        add_role(mean, BATCH_NORM_MINIBATCH_ESTIMATE)
        if self.mean_only:
            stdev = tensor.ones_like(mean)
        else:
            stdev = tensor.sqrt(tensor.var(input_, axis=axes, keepdims=True) +
                                numpy.cast[theano.config.floatX](self.epsilon))
            assert (stdev.broadcastable[1:] ==
                    self.population_stdev.broadcastable)
            add_role(stdev, BATCH_NORM_MINIBATCH_ESTIMATE)
        return mean, stdev 

def normalize_batch_in_training(x, gamma, beta,
                                reduction_axes, epsilon=1e-3):
    """Computes mean and std for batch then apply batch_normalization on batch.
    """
    # TODO remove this if statement when Theano without
    # T.nnet.bn.batch_normalization_train is deprecated
    if not hasattr(T.nnet.bn, 'batch_normalization_train'):
        return _old_normalize_batch_in_training(x, gamma, beta, reduction_axes, epsilon)

    if gamma is None:
        if beta is None:
            gamma = ones_like(x)
        else:
            gamma = ones_like(beta)
    if beta is None:
        if gamma is None:
            beta = zeros_like(x)
        beta = zeros_like(gamma)

    normed, mean, stdinv = T.nnet.bn.batch_normalization_train(
        x, gamma, beta, reduction_axes, epsilon)

    return normed, mean, T.inv(stdinv ** 2) 

def test_gpujoin_gpualloc():
    a = T.fmatrix('a')
    a_val = numpy.asarray(numpy.random.rand(4, 5), dtype='float32')
    b = T.fmatrix('b')
    b_val = numpy.asarray(numpy.random.rand(3, 5), dtype='float32')

    f = theano.function([a, b], T.join(0, T.zeros_like(a), T.ones_like(b)) + 4,
                        mode=mode_without_gpu)
    f_gpu = theano.function([a, b], T.join(0, T.zeros_like(a), T.ones_like(b)),
                            mode=mode_with_gpu)
    f_gpu2 = theano.function([a, b], T.join(0, T.zeros_like(a),
                                            T.ones_like(b)) + 4,
                             mode=mode_with_gpu)
    assert sum([node.op == T.alloc for node in f.maker.fgraph.toposort()]) == 2
    assert sum([node.op == T.join for node in f.maker.fgraph.toposort()]) == 1
    assert sum([isinstance(node.op, GpuAlloc)
                for node in f_gpu.maker.fgraph.toposort()]) == 2
    assert sum([node.op == gpu_join
                for node in f_gpu.maker.fgraph.toposort()]) == 1
    assert sum([isinstance(node.op, GpuAlloc)
                for node in f_gpu2.maker.fgraph.toposort()]) == 2
    assert sum([node.op == gpu_join
                for node in f_gpu2.maker.fgraph.toposort()]) == 1
    assert numpy.allclose(f(a_val, b_val), f_gpu2(a_val, b_val)) 

def test_gpualloc_output_to_gpu():
    a_val = numpy.asarray(numpy.random.rand(4, 5), dtype='float32')
    a = tcn.shared_constructor(a_val)

    b = T.fscalar()
    f = theano.function([b], T.ones_like(a) + b, mode=mode_without_gpu)
    f_gpu = theano.function([b], B.gpu_from_host(T.ones_like(a)) + b,
                            mode=mode_with_gpu)

    f(2)
    f_gpu(2)

    assert sum([node.op == T.alloc for node in f.maker.fgraph.toposort()]) == 1
    assert sum([node.op == B.gpu_alloc
                for node in f_gpu.maker.fgraph.toposort()]) == 1

    assert numpy.allclose(numpy.ones(a.get_value(borrow=True).shape) + 9,
                          f_gpu(9))
    assert numpy.allclose(f(5), f_gpu(5)) 

def ones_like(x, dtype=None, name=None):
    return T.ones_like(x, dtype=dtype)