def test_fprop(self):
        """
        compare `fprop` results for cpu and gpu backends
        """
        r = []
        for i in xrange(self.N):
            batch_size, x_dim = self.rng.random_integers(3000, size=2)
            x = self.rng.rand(batch_size, x_dim).astype(np.float32)

            for nonlinearity in ['sigmoid', 'tanh', 'relu']:
                state = self.rng.get_state()
                quagga.processor_type = 'gpu'
                x_gpu = Connector(Matrix.from_npa(x))
                nonlinearity_block = NonlinearityBlock(x_gpu, nonlinearity)
                x_gpu.fprop()
                nonlinearity_block.fprop()
                output_gpu = nonlinearity_block.output.to_host()

                self.rng.set_state(state)
                quagga.processor_type = 'cpu'
                x_cpu = Connector(Matrix.from_npa(x))
                nonlinearity_block = NonlinearityBlock(x_cpu, nonlinearity)
                x_cpu.fprop()
                nonlinearity_block.fprop()
                output_cpu = nonlinearity_block.output.to_host()

                r.append(np.allclose(output_gpu, output_cpu))

        self.assertEqual(sum(r), len(r))

    def __init__(self, W, b, x, device_id=None):
        self.f_context = Context(device_id)
        device_id = self.f_context.device_id

        if W.bpropagable:
            self.W, self.dL_dW = W.register_usage(device_id, device_id)
        else:
            self.W = W.register_usage(device_id)
        if b:
            if b.bpropagable:
                self.b, self.dL_db = b.register_usage(device_id, device_id)
                self.ones = Matrix.empty(x.nrows, 1, self.b.dtype, device_id)
                self.ones.sync_fill(1.0)
            else:
                self.b = b.register_usage(device_id)
        if x.bpropagable:
            self.x, self.dL_dx = x.register_usage(device_id, device_id)
        else:
            self.x = x.register_usage(device_id)

        output = Matrix.empty(x.nrows, self.W.ncols, device_id=device_id)
        self.learning = hasattr(self, 'dL_dW') or hasattr(self, 'dL_db') or \
                        hasattr(self, 'dL_dx')
        if self.learning:
            self.b_context = Context(device_id)
            self.output = Connector(output, device_id)
        else:
            self.output = Connector(output)

    def __init__(self, x, nonlinearity, device_id=None):
        """


        """
        self.f_context = Context(device_id)
        device_id = self.f_context.device_id
        self.learning = x.bpropagable
        if self.learning:
            self.b_context = Context(device_id)
            self.x, self.dL_dx = x.register_usage(device_id, device_id)
            self._df_dpref = Matrix.empty_like(self.x, device_id)
        else:
            self.x = x.register_usage(device_id)
        output = Matrix.empty_like(x, device_id)
        self.output = Connector(output, device_id if self.learning else None)
        if nonlinearity == "sigmoid":
            self.f = self.x.sigmoid
        elif nonlinearity == "tanh":
            self.f = self.x.tanh
        elif nonlinearity == "relu":
            self.f = self.x.relu
        elif nonlinearity == "softmax":
            raise ValueError("For softmax nonlinearity use SoftmaxBlock!")
        else:
            raise ValueError("TODO!")
        self.training_mode = True

    def register_usage(self, fu_device_id, bo_device_id=None):
        """
        Register usage of connector's forward_matrix.

        :param fu_device_id: context in which `forward_matrix` will be used
        :param bo_device_id: context in which `backward_matrix`
                                    of the connector will be calculated
        """

        if not self.bpropagable and bo_device_id:
            raise ValueError("Nobody is going to use computation from backward step. "
                             "You mustn't register for backward propagate!")
        if fu_device_id != self._fo_device_id and fu_device_id not in self._f_matrices:
            self._f_matrices[fu_device_id] = Matrix.empty_like(self, fu_device_id)
            self.context[fu_device_id] = Context(fu_device_id)
        if bo_device_id is None:
            return self._f_matrices[fu_device_id]

        for device_id in [self._bu_device_id, bo_device_id]:
            if device_id not in self._b_matrices:
                self._b_matrices[device_id] = Matrix.empty_like(self, device_id)
                if device_id not in self.context:
                    self.context[device_id] = Context(device_id)
        if self._bu_device_id != bo_device_id and self._bu_device_id not in self._b_matrices_pool:
            self._b_matrices_pool[self._bu_device_id] = Matrix.empty_like(self, self._bu_device_id)
        return self._f_matrices[fu_device_id], self._b_matrices[bo_device_id]

    def test_theano_fprop_vector(self):
        r = []
        for _ in xrange(self.N):
            embd_dim = self.rng.random_integers(10000)
            batch_size, output_dim = self.rng.random_integers(2000, size=2)
            W = self.get_orthogonal_matrix(embd_dim, output_dim)
            row_idxs = self.rng.randint(embd_dim, size=(batch_size, 1)).astype(np.int32)

            quagga.processor_type = 'gpu'
            qrow_idxs = Connector(Matrix.from_npa(row_idxs))
            qW = Connector(Matrix.from_npa(W))
            row_slicing_block = RowSlicingBlock(qW, qrow_idxs)
            qW.fprop()
            qrow_idxs.fprop()
            row_slicing_block.fprop()
            q_output = row_slicing_block.output.to_host()

            trow_idxs = T.ivector()
            row_slicing_layer = RowSlicingLayer(W)
            t_output = row_slicing_layer.get_output_expr(trow_idxs)
            t_output = theano.function([trow_idxs], t_output)(row_idxs[:, 0])

            r.append(np.allclose(q_output, t_output))

        self.assertEqual(sum(r), len(r))

    def test_bprop(self):
        r = []
        for i in xrange(self.N):
            repeats = self.rng.random_integers(42)
            axis = self.rng.randint(2)
            input_dim, output_dim = self.rng.random_integers(2000, size=2)
            x = self.get_normal_matrix(input_dim, output_dim)
            input_dim = input_dim if axis else input_dim * repeats
            true_labels = self.rng.randint(output_dim, size=(input_dim, 1)).astype(np.int32)
            device_id = 0

            output = {}
            for processor_type in ['gpu', 'cpu']:
                quagga.processor_type = processor_type
                qx = Connector(Matrix.from_npa(x), device_id)
                qtrue_labels = Connector(Matrix.from_npa(true_labels))
                repeat_block = RepeatBlock(qx, repeats, axis)
                sce_block = SoftmaxCeBlock(repeat_block.output, qtrue_labels)
                qx.fprop()
                qtrue_labels.fprop()
                repeat_block.fprop()
                sce_block.fprop()
                sce_block.bprop()
                repeat_block.bprop()
                output[processor_type] = qx.backward_matrix.to_host()

            r.append(np.allclose(output['gpu'], output['cpu']))

        self.assertEqual(sum(r), len(r))

    def test_bprop_vector(self):
        r = []
        for _ in xrange(self.N):
            embd_dim = self.rng.random_integers(10000)
            batch_size, output_dim = self.rng.random_integers(2000, size=2)
            W = self.get_orthogonal_matrix(embd_dim, output_dim)
            row_idxs = self.rng.randint(embd_dim, size=(batch_size, 1)).astype(np.int32)
            true_labels = self.rng.randint(output_dim, size=(batch_size, 1)).astype(np.int32)
            device_id = 0

            output = {}
            for processor_type in ['gpu', 'cpu']:
                quagga.processor_type = processor_type
                qrow_idxs = Connector(Matrix.from_npa(row_idxs))
                qtrue_labels = Connector(Matrix.from_npa(true_labels))
                qW = Connector(Matrix.from_npa(W), device_id)
                row_slicing_block = RowSlicingBlock(qW, qrow_idxs)
                sce_block = SoftmaxCeBlock(row_slicing_block.output, qtrue_labels)
                qW.fprop()
                qrow_idxs.fprop()
                row_slicing_block.fprop()
                sce_block.fprop()
                sce_block.bprop()
                row_slicing_block.bprop()
                qW.add(Context(), qW.backward_matrix)
                output[processor_type] = qW.to_host()

            r.append(np.allclose(output['gpu'], output['cpu']))

        self.assertEqual(sum(r), len(r))

    def test_fprop(self):
        """
        compare `fprop` results for cpu and gpu backends
        """
        r = []
        for i in xrange(self.N):
            max_input_sequence_len = self.rng.random_integers(500)
            sequence_len = max_input_sequence_len if i == 0 else self.rng.random_integers(max_input_sequence_len)
            batch_size = self.rng.random_integers(512)
            dim = self.rng.random_integers(1500)
            x = [self.rng.rand(batch_size, dim).astype(dtype=np.float32) for _ in xrange(max_input_sequence_len)]

            state = self.rng.get_state()
            quagga.processor_type = 'gpu'
            x_gpu = List([Connector(Matrix.from_npa(e)) for e in x])
            smean_pooling_block_gpu = SequentialMeanPoolingBlock(x_gpu)
            x_gpu.set_length(sequence_len)
            smean_pooling_block_gpu.fprop()
            output_gpu = smean_pooling_block_gpu.output.to_host()

            self.rng.set_state(state)
            quagga.processor_type = 'cpu'
            x_cpu = List([Connector(Matrix.from_npa(e)) for e in x])
            smean_pooling_block_cpu = SequentialMeanPoolingBlock(x_cpu)
            x_cpu.set_length(sequence_len)
            smean_pooling_block_cpu.fprop()
            output_cpu = smean_pooling_block_cpu.output.to_host()

            r.append(np.allclose(output_gpu, output_cpu))

        self.assertEqual(sum(r), self.N)

    def test_theano_fprop_matrix(self):
        r = []
        for i in xrange(self.N):
            max_input_sequence_len = self.rng.random_integers(300)
            sequence_len = max_input_sequence_len if i == 0 else self.rng.random_integers(max_input_sequence_len)
            embd_dim = self.rng.random_integers(10000)
            batch_size = self.rng.random_integers(500)
            output_dim = self.rng.random_integers(2000)
            W = self.get_orthogonal_matrix(embd_dim, output_dim)
            row_idxs = self.rng.randint(embd_dim, size=(batch_size, max_input_sequence_len)).astype(np.int32)

            quagga.processor_type = 'gpu'
            qrow_idxs = Connector(Matrix.from_npa(row_idxs))
            qW = Connector(Matrix.from_npa(W))
            row_slicing_block = RowSlicingBlock(qW, qrow_idxs)
            qW.fprop()
            qrow_idxs.ncols = sequence_len
            qrow_idxs.fprop()
            row_slicing_block.fprop()
            q_output = row_slicing_block.output.to_host()

            th_row_idxs = T.imatrix()
            row_slicing_layer = RowSlicingLayer(W)
            toutput = row_slicing_layer.get_output_expr(th_row_idxs)
            th_output = theano.function([th_row_idxs], toutput)(row_idxs)

            for i in xrange(sequence_len):
                r.append(np.allclose(q_output[i], th_output[i]))

        self.assertEqual(sum(r), len(r))

    def test_theano_grad(self):
        quagga.processor_type = 'gpu'
        r = []
        for i in xrange(self.N):
            batch_size, dim = self.rng.random_integers(2000, size=2)
            y_hat = self.rng.randn(batch_size, dim).astype(dtype=np.float32)
            y = self.rng.randn(batch_size, dim).astype(dtype=np.float32)

            # Theano model
            th_y_hat, th_y = T.fmatrix(), T.fmatrix()
            loss = T.mean(T.sum((th_y_hat - th_y) ** 2, axis=1))
            get_theano_grads = theano.function([th_y_hat, th_y], T.grad(loss, wrt=th_y_hat))
            th_dL_dy_hat = get_theano_grads(y_hat, y)

            # quagga model
            context = Context()
            y_hat_gpu = Connector(Matrix.from_npa(y_hat), context, context)
            y_gpu = Connector(Matrix.from_npa(y))
            sigmoid_ce_block = SseBlock(y_hat_gpu, y_gpu)
            sigmoid_ce_block.fprop()
            sigmoid_ce_block.bprop()
            q_dL_dy_hat = y_hat_gpu.backward_matrix.to_host()

            r.append(np.allclose(th_dL_dy_hat, q_dL_dy_hat))

        self.assertEqual(sum(r), self.N)

    def test_bprop(self):
        """
        compare `bprop` results for cpu and gpu backends
        """
        r = []
        for i in xrange(self.N):
            batch_size, dim = self.rng.random_integers(2000, size=2)
            y_hat = self.rng.randn(batch_size, dim).astype(dtype=np.float32)
            y = self.rng.randn(batch_size, dim).astype(dtype=np.float32)

            quagga.processor_type = 'gpu'
            context = Context()
            y_hat_gpu = Connector(Matrix.from_npa(y_hat), context, context)
            y_gpu = Connector(Matrix.from_npa(y))
            sse_block = SseBlock(y_hat_gpu, y_gpu)
            sse_block.fprop()
            sse_block.bprop()
            dL_dy_hat_gpu = y_hat_gpu.backward_matrix.to_host()

            quagga.processor_type = 'cpu'
            context = Context()
            y_hat_cpu = Connector(Matrix.from_npa(y_hat), context, context)
            y_cpu = Connector(Matrix.from_npa(y))
            sse_block = SseBlock(y_hat_cpu, y_cpu)
            sse_block.fprop()
            sse_block.bprop()
            dL_dy_hat_cpu = y_hat_cpu.backward_matrix.to_host()

            r.append(np.allclose(dL_dy_hat_gpu, dL_dy_hat_cpu))

        self.assertEqual(sum(r), self.N)

    def test_fprop_matrix(self):
        """
        compare `fprop` results for cpu and gpu backends
        """
        r = []
        for i in xrange(self.N):
            max_input_sequence_len = self.rng.random_integers(300)
            sequence_len = max_input_sequence_len if i == 0 else self.rng.random_integers(max_input_sequence_len)
            embd_dim = self.rng.random_integers(10000)
            batch_size, output_dim = self.rng.random_integers(2000, size=2)
            W = self.get_orthogonal_matrix(embd_dim, output_dim)
            row_idxs = self.rng.randint(embd_dim, size=(batch_size, max_input_sequence_len)).astype(np.int32)

            output = {}
            for processor_type in ['gpu', 'cpu']:
                quagga.processor_type = processor_type
                qrow_idxs = Connector(Matrix.from_npa(row_idxs))
                qW = Connector(Matrix.from_npa(W))
                row_slicing_block = RowSlicingBlock(qW, qrow_idxs)
                qW.fprop()
                qrow_idxs.ncols = sequence_len
                qrow_idxs.fprop()
                row_slicing_block.fprop()
                output[processor_type] = row_slicing_block.output.to_host()

            for output_gpu, output_cpu in izip(output['gpu'], output['cpu']):
                r.append(np.allclose(output_gpu, output_cpu))

        self.assertEqual(sum(r), len(r))

    def test_bprop(self):
        """
        compare `fprop` results for cpu and gpu backends
        """

        r = []
        for i in xrange(self.N):
            max_input_sequence_len = self.rng.random_integers(500)
            sequence_len = max_input_sequence_len if i == 0 else self.rng.random_integers(max_input_sequence_len)
            batch_size = self.rng.random_integers(256)
            input_dim, hidden_dim = self.rng.random_integers(1500, size=2)
            x = [self.rng.randn(batch_size, input_dim).astype(np.float32) for _ in xrange(max_input_sequence_len)]
            true_labels = [self.rng.randint(hidden_dim, size=(batch_size, 1)).astype(np.int32) for _ in xrange(max_input_sequence_len)]
            W = self.get_orthogonal_matrix(input_dim, hidden_dim)
            b = self.rng.rand(1, hidden_dim).astype(np.float32)
            device_id = 0

            quagga_grads = {}
            for reverse in [False, True]:
                for with_bias in [False, True]:
                    for processor_type in ['gpu', 'cpu']:
                        quagga.processor_type = processor_type
                        qx = List([Connector(Matrix.from_npa(e), device_id) for e in x])
                        qtrue_labels = List([Connector(Matrix.from_npa(e)) for e in true_labels], len(qx))
                        qW = Connector(Matrix.from_npa(W), device_id)
                        qb = Connector(Matrix.from_npa(b), device_id) if with_bias else None
                        seq_dot_block = SequencerBlock(block_class=DotBlock,
                                                       params=[qW, qb],
                                                       sequences=[qx],
                                                       output_names=['output'],
                                                       reverse=reverse)
                        seq_sce_block = SequencerBlock(block_class=SoftmaxCeBlock,
                                                       params=[],
                                                       sequences=[seq_dot_block.output, qtrue_labels],
                                                       reverse=reverse)
                        qx.length = sequence_len
                        qx.fprop()
                        qtrue_labels.fprop()
                        qW.fprop()
                        if qb:
                            qb.fprop()
                        seq_dot_block.fprop()
                        seq_sce_block.fprop()
                        seq_sce_block.bprop()
                        seq_dot_block.bprop()
                        quagga_grads[processor_type] = [qW.backward_matrix.to_host()]
                        if with_bias:
                            quagga_grads[processor_type].append(qb.backward_matrix.to_host())
                        quagga_grads[processor_type].extend(e.backward_matrix.to_host() for e in qx)

                    for grad_gpu, grad_cpu in izip(quagga_grads['gpu'], quagga_grads['cpu']):
                        r.append(np.allclose(grad_gpu, grad_cpu, atol=1e-5))

        self.assertEqual(sum(r), len(r))

 def bprop(self):
     dL_doutput = self.output.backward_matrix
     self.dL_dpre_a.assign_dL_dpre_a(self.context, dL_doutput, self.a,
                                     self.matrices[:self.length])
     if hasattr(self, 'dL_dmatrices'):
         Matrix.add_attention_tile(self.context, dL_doutput, self.a,
                                   self.dL_dpre_a, self.u,
                                   self.dL_dmatrices[:self.length])
     if hasattr(self, 'dL_du'):
         self.dL_du.add_attention_derivative(self.context, self.dL_dpre_a,
                                             self.matrices[:self.length])

 def __init__(self, dropout_prob, x, seed=42, device_id=None):
     self.dropout_prob = dropout_prob
     self.f_context = Context(device_id)
     device_id = self.f_context.device_id
     self.generator = Matrix.get_random_generator(seed)
     if x.bpropagable:
         self.b_context = Context(device_id)
         self.x, self.dL_dx = x.register_usage(device_id, device_id)
     else:
         self.x = x.register_usage(device_id)
     self.output = Matrix.empty_like(self.x)
     self.output = Connector(self.output, device_id if x.bpropagable else None)
     self.training_mode = True

    def test_bprop(self):
        """
        compare `bprop` results for cpu and gpu backends
        """
        r = []
        for i in xrange(self.N):
            batch_size, x_dim, output_dim = self.rng.random_integers(2000, size=3)
            x = self.rng.rand(batch_size, x_dim).astype(np.float32)
            W = self.get_orthogonal_matrix(x_dim, output_dim)
            b = self.rng.rand(1, output_dim).astype(np.float32) if self.rng.randint(2) else None
            device_id = 0

            state = self.rng.get_state()
            quagga.processor_type = 'gpu'
            context = Context()
            x_gpu = Connector(Matrix.from_npa(x), device_id)
            W_gpu = Connector(Matrix.from_npa(W), device_id)
            b_gpu = Connector(Matrix.from_npa(b), device_id) if b is not None else b
            dot_block_gpu = DotBlock(W_gpu, b_gpu, x_gpu)
            x_gpu.fprop()
            W_gpu.fprop()
            if b_gpu:
                b_gpu.fprop()
            dot_block_gpu.fprop()
            _, dL_doutput = dot_block_gpu.output.register_usage(device_id, device_id)
            random_matrix = self.rng.rand(dL_doutput.nrows, dL_doutput.ncols)
            dL_doutput.assign(context, Matrix.from_npa(random_matrix, 'float'))
            dot_block_gpu.bprop()
            if b is not None:
                dL_db_gpu = b_gpu.backward_matrix.to_host()
            dL_dW_gpu = W_gpu.backward_matrix.to_host()
            dL_dx_gpu = x_gpu.backward_matrix.to_host()

            self.rng.set_state(state)
            quagga.processor_type = 'cpu'
            context = Context()
            x_cpu = Connector(Matrix.from_npa(x), device_id)
            W_cpu = Connector(Matrix.from_npa(W), device_id)
            b_cpu = Connector(Matrix.from_npa(b), device_id) if b is not None else b
            dot_block_cpu = DotBlock(W_cpu, b_cpu, x_cpu)
            x_cpu.fprop()
            W_cpu.fprop()
            if b_cpu:
                b_cpu.fprop()
            dot_block_cpu.fprop()
            _, dL_doutput = dot_block_cpu.output.register_usage(device_id, device_id)
            random_matrix = self.rng.rand(dL_doutput.nrows, dL_doutput.ncols)
            dL_doutput.assign(context, Matrix.from_npa(random_matrix, 'float'))
            dot_block_cpu.bprop()
            if b is not None:
                dL_db_cpu = b_cpu.backward_matrix.to_host()
            dL_dW_cpu = W_cpu.backward_matrix.to_host()
            dL_dx_cpu = x_cpu.backward_matrix.to_host()

            r.append(np.allclose(dL_dx_gpu, dL_dx_cpu, atol=1e-5))
            r.append(np.allclose(dL_dW_gpu, dL_dW_cpu, atol=1e-5))
            if b is not None:
                r.append(np.allclose(dL_db_gpu, dL_db_cpu, atol=1e-5))

        self.assertEqual(sum(r), len(r))

    def test_theano_grad(self):
        quagga.processor_type = 'gpu'
        r = []
        for i in xrange(self.N):
            for sparse in [True, False]:
                batch_size, dim = self.rng.random_integers(2000, size=2)
                if sparse:
                    true_labels = np.zeros((batch_size, dim), np.float32)
                    for k, j in enumerate(self.rng.randint(dim, size=batch_size)):
                        true_labels[k, j] = 1.0
                else:
                    true_labels = self.rng.randint(dim, size=(batch_size, 1)).astype(np.int32)
                x = self.rng.randn(batch_size, dim).astype(np.float32)
                mask = (self.rng.rand(batch_size, 1) < 0.8).astype(np.float32)
                device_id = 0
                for with_mask in [False, True]:
                    # Theano model
                    th_x = T.fmatrix()
                    th_mask = T.fcol()
                    th_true_labels = T.fmatrix() if sparse else T.ivector()
                    if with_mask:
                        probs = T.nnet.softmax(th_mask * th_x)
                    else:
                        probs = T.nnet.softmax(th_x)
                    loss = T.mean(T.nnet.categorical_crossentropy(probs, th_true_labels))
                    if with_mask:
                        get_theano_grads = theano.function([th_x, th_true_labels, th_mask], T.grad(loss, wrt=th_x))
                        th_dL_dx = get_theano_grads(x, true_labels if sparse else true_labels[:, 0], mask)
                    else:
                        get_theano_grads = theano.function([th_x, th_true_labels], T.grad(loss, wrt=th_x))
                        th_dL_dx = get_theano_grads(x, true_labels if sparse else true_labels[:, 0])

                    # quagga model
                    x_gpu = Connector(Matrix.from_npa(x), device_id)
                    true_labels_gpu = Connector(Matrix.from_npa(true_labels))
                    mask_gpu = Connector(Matrix.from_npa(mask)) if with_mask else None
                    softmax_ce_block = SoftmaxCeBlock(x_gpu, true_labels_gpu, mask_gpu)
                    x_gpu.fprop()
                    true_labels_gpu.fprop()
                    if with_mask:
                        mask_gpu.fprop()
                    softmax_ce_block.fprop()
                    softmax_ce_block.bprop()
                    q_dL_dx = x_gpu.backward_matrix.to_host()

                    r.append(np.allclose(th_dL_dx, q_dL_dx))

        self.assertEqual(sum(r), len(r))

    def test_bprop(self):
        r = []
        for i in xrange(self.N):
            matrices = []
            nrows = self.rng.random_integers(1, 3000)
            ncols = [0]
            col_slices = []
            device_ids = []
            for _ in xrange(self.rng.random_integers(1, 10)):
                _ncols = self.rng.random_integers(1, 2000)
                ncols.append(ncols[-1] + _ncols)
                if self.rng.choice([True, False]):
                    device_ids.append(0)
                    col_slices.append((ncols[-2], ncols[-1]))
                else:
                    device_ids.append(None)
                matrices.append(self.rng.rand(nrows, _ncols).astype(np.float32))
            true_labels = self.rng.randint(ncols[-1], size=(nrows, 1)).astype(np.int32)
            if not col_slices:
                r.append(True)
                continue

            output = {}
            for processor_type in ['gpu', 'cpu']:
                quagga.processor_type = processor_type
                qmatrices = [Connector(Matrix.from_npa(m), d_id) for m, d_id in izip(matrices, device_ids)]
                qtrue_labels = Connector(Matrix.from_npa(true_labels))
                hstack_block = HorizontalStackBlock(*qmatrices)
                sce_block = SoftmaxCeBlock(hstack_block.output, qtrue_labels)

                for m in qmatrices:
                    m.fprop()
                qtrue_labels.fprop()
                hstack_block.fprop()
                sce_block.fprop()
                sce_block.bprop()
                hstack_block.bprop()

                output[processor_type] = [m.backward_matrix.to_host()
                                          for m in qmatrices if m.bpropagable]

            for dL_dm_gpu, dL_dm_cpu in izip(output['gpu'], output['cpu']):
                if not np.allclose(dL_dm_gpu, dL_dm_cpu):
                    r.append(False)
                    break
            else:
                r.append(True)
        self.assertEqual(sum(r), self.N)

    def test_fprop(self):
        """
        compare `fprop` results for cpu and gpu backends
        """

        r = []
        for i in xrange(self.N):
            max_input_sequence_len = self.rng.random_integers(500)
            sequence_len = max_input_sequence_len if i == 0 else self.rng.random_integers(max_input_sequence_len)
            batch_size = self.rng.random_integers(256)
            input_dim, hidden_dim = self.rng.random_integers(1500, size=2)
            x = [self.rng.randn(batch_size, input_dim).astype(np.float32) for _ in xrange(max_input_sequence_len)]
            W = self.get_orthogonal_matrix(input_dim, hidden_dim)
            b = self.rng.rand(1, hidden_dim).astype(np.float32)

            from quagga.cuda import cudart
            cudart.cuda_set_device(1)

            qoutput = {}
            for reverse in [False, True]:
                for with_bias in [False, True]:
                    for processor_type in ['gpu', 'cpu']:
                        quagga.processor_type = processor_type
                        qx = List([Connector(Matrix.from_npa(e)) for e in x])
                        qW = Connector(Matrix.from_npa(W))
                        qb = Connector(Matrix.from_npa(b)) if with_bias else None
                        seq_dot_block = SequencerBlock(block_class=DotBlock,
                                                       params=[qW, qb],
                                                       sequences=[qx],
                                                       output_names=['output'],
                                                       reverse=reverse)
                        qx.length = sequence_len
                        qx.fprop()
                        qW.fprop()
                        if qb:
                            qb.fprop()
                        seq_dot_block.fprop()
                        qoutput[processor_type] = seq_dot_block.output.to_host()

                    for output_gpu, output_cpu in izip(qoutput['gpu'], qoutput['cpu']):
                        if not np.allclose(output_gpu, output_cpu, atol=1e-5):
                            r.append(False)
                            break
                    else:
                        r.append(True)

        self.assertEqual(sum(r), len(r))

    def test_theano_grad(self):
        quagga.processor_type = 'gpu'
        r = []
        for i in xrange(self.N):
            batch_size, dim = self.rng.random_integers(2000, size=2)
            true_labels = self.rng.randint(2, size=(batch_size, dim)).astype(dtype=np.float32)
            mask = (self.rng.rand(batch_size, 1) < 0.8).astype(np.float32)
            x = self.rng.randn(batch_size, dim).astype(dtype=np.float32)
            device_id = 0

            for with_mask in [False, True]:
                # Theano model
                th_x = T.fmatrix()
                th_mask = T.fmatrix()
                th_true_labels = T.fmatrix()
                if with_mask:
                    probs = T.nnet.sigmoid(theano.compile.ops.Rebroadcast((0, False), (1, True))(th_mask) * th_x)
                else:
                    probs = T.nnet.sigmoid(th_x)
                loss = - th_true_labels * T.log(probs) - \
                       (1.0 - th_true_labels) * T.log(1.0 - probs)
                loss = T.sum(loss, axis=1).mean()

                if with_mask:
                    get_theano_grads = theano.function([th_x, th_true_labels, th_mask], T.grad(loss, wrt=th_x))
                    th_dL_dx = get_theano_grads(x, true_labels, mask)
                else:
                    get_theano_grads = theano.function([th_x, th_true_labels], T.grad(loss, wrt=th_x))
                    th_dL_dx = get_theano_grads(x, true_labels)

                # quagga model
                x_gpu = Connector(Matrix.from_npa(x), device_id)
                true_labels_gpu = Connector(Matrix.from_npa(true_labels))
                mask_gpu = Connector(Matrix.from_npa(mask)) if with_mask else None
                sigmoid_ce_block = SigmoidCeBlock(x_gpu, true_labels_gpu, mask_gpu)
                x_gpu.fprop()
                true_labels_gpu.fprop()
                if with_mask:
                    mask_gpu.fprop()
                sigmoid_ce_block.fprop()
                sigmoid_ce_block.bprop()
                q_dL_dx = x_gpu.backward_matrix.to_host()

                r.append(np.allclose(th_dL_dx, q_dL_dx))

        self.assertEqual(sum(r), len(r))