Python caffe2.python.core.CreateOperator() Examples

def _run_op_test(self, X, I, check_grad=False):
        with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
            op = core.CreateOperator('BatchPermutation', ['X', 'I'], ['Y'])
            workspace.FeedBlob('X', X)
            workspace.FeedBlob('I', I)
        workspace.RunOperatorOnce(op)
        Y = workspace.FetchBlob('Y')

        if check_grad:
            gc = gradient_checker.GradientChecker(
                stepsize=0.1,
                threshold=0.001,
                device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
            )

            res, grad, grad_estimated = gc.CheckSimple(op, [X, I], 0, [0])
            self.assertTrue(res, 'Grad check failed')

        Y_ref = X[I]
        np.testing.assert_allclose(Y, Y_ref, rtol=1e-5, atol=1e-08) 

def _run_op_test(self, X, I, check_grad=False):
        with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
            op = core.CreateOperator('BatchPermutation', ['X', 'I'], ['Y'])
            workspace.FeedBlob('X', X)
            workspace.FeedBlob('I', I)
        workspace.RunOperatorOnce(op)
        Y = workspace.FetchBlob('Y')

        if check_grad:
            gc = gradient_checker.GradientChecker(
                stepsize=0.1,
                threshold=0.001,
                device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
            )

            res, grad, grad_estimated = gc.CheckSimple(op, [X, I], 0, [0])
            self.assertTrue(res, 'Grad check failed')

        Y_ref = X[I]
        np.testing.assert_allclose(Y, Y_ref, rtol=1e-5, atol=1e-08) 

def _CorrectMomentum(self, correction):
        """The MomentumSGDUpdate op implements the update V as

            V := mu * V + lr * grad,

        where mu is the momentum factor, lr is the learning rate, and grad is
        the stochastic gradient. Since V is not defined independently of the
        learning rate (as it should ideally be), when the learning rate is
        changed we should scale the update history V in order to make it
        compatible in scale with lr * grad.
        """
        logger.info(
            'Scaling update history by {:.6f} (new lr / old lr)'.
            format(correction))
        for i in range(cfg.NUM_GPUS):
            with c2_utils.CudaScope(i):
                for param in self.TrainableParams(gpu_id=i):
                    op = core.CreateOperator(
                        'Scale', [param + '_momentum'], [param + '_momentum'],
                        scale=correction)
                    workspace.RunOperatorOnce(op) 

def _CorrectMomentum(self, correction):
        """The MomentumSGDUpdate op implements the update V as

            V := mu * V + lr * grad,

        where mu is the momentum factor, lr is the learning rate, and grad is
        the stochastic gradient. Since V is not defined independently of the
        learning rate (as it should ideally be), when the learning rate is
        changed we should scale the update history V in order to make it
        compatible in scale with lr * grad.
        """
        logger.info(
            'Scaling update history by {:.6f} (new lr / old lr)'.
            format(correction))
        for i in range(cfg.NUM_GPUS):
            with c2_utils.CudaScope(i):
                for param in self.TrainableParams(gpu_id=i):
                    op = core.CreateOperator(
                        'Scale', [param + '_momentum'], [param + '_momentum'],
                        scale=correction)
                    workspace.RunOperatorOnce(op) 

def ConvertTensorProtosToInitNet(net_params, input_name):
    """Takes the net_params returned from TranslateModel, and wrap it as an
    init net that contain GivenTensorFill.

    This is a very simple feature that only works with float tensors, and is
    only intended to be used in an environment where you want a single
    initialization file - for more complex cases, use a db to store the
    parameters.
    """
    init_net = caffe2_pb2.NetDef()
    for tensor in net_params.protos:
        if len(tensor.float_data) == 0:
            raise RuntimeError(
                "Only float tensors are supported in this util.")
        op = core.CreateOperator(
            "GivenTensorFill", [], [tensor.name],
            arg=[
                utils.MakeArgument("shape", list(tensor.dims)),
                utils.MakeArgument("values", tensor.float_data)])
        init_net.op.extend([op])
    init_net.op.extend([core.CreateOperator("ConstantFill", [], [input_name], shape=[1])])
    return init_net 

def enqueue_blobs(self, gpu_id, blob_names, blobs):
        """Put a mini-batch on a BlobsQueue."""
        assert len(blob_names) == len(blobs)
        t = time.time()
        dev = c2_utils.CudaDevice(gpu_id)
        queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
        blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
        for (blob_name, blob) in zip(blob_names, blobs):
            workspace.FeedBlob(blob_name, blob, device_option=dev)
        logger.debug(
            'enqueue_blobs {}: workspace.FeedBlob: {}'.
            format(gpu_id, time.time() - t)
        )
        t = time.time()
        op = core.CreateOperator(
            'SafeEnqueueBlobs', [queue_name] + blob_names,
            blob_names + [queue_name + '_enqueue_status'],
            device_option=dev
        )
        workspace.RunOperatorOnce(op)
        logger.debug(
            'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
            format(gpu_id, time.time() - t)
        ) 

def run_add5_and_add5gradient_op(device):
    # clear the workspace before running the operator
    workspace.ResetWorkspace()
    add5 = core.CreateOperator("Add5",
                               ["X"],
                               ["Y"],
                               device_option=device)
    print("==> Running Add5 op:")
    workspace.FeedBlob("X", (np.random.rand(5, 5)), device_option=device)
    print("Input of Add5: ", workspace.FetchBlob("X"))
    workspace.RunOperatorOnce(add5)
    print("Output of Add5: ", workspace.FetchBlob("Y"))

    print("\n\n==> Running Add5Gradient op:")
    print("Input of Add5Gradient: ", workspace.FetchBlob("Y"))
    add5gradient = core.CreateOperator("Add5Gradient",
                                       ["Y"],
                                       ["Z"],
                                       device_option=device)
    workspace.RunOperatorOnce(add5gradient)
    print("Output of Add5Gradient: ", workspace.FetchBlob("Z")) 

def enqueue_blobs(self, gpu_id, blob_names, blobs):
        """Put a mini-batch on a BlobsQueue."""
        assert len(blob_names) == len(blobs)
        t = time.time()
        dev = c2_utils.CudaDevice(gpu_id)
        queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
        blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
        for (blob_name, blob) in zip(blob_names, blobs):
            workspace.FeedBlob(blob_name, blob, device_option=dev)
        logger.debug(
            'enqueue_blobs {}: workspace.FeedBlob: {}'.
            format(gpu_id, time.time() - t)
        )
        t = time.time()
        op = core.CreateOperator(
            'SafeEnqueueBlobs', [queue_name] + blob_names,
            blob_names + [queue_name + '_enqueue_status'],
            device_option=dev
        )
        workspace.RunOperatorOnce(op)
        logger.debug(
            'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
            format(gpu_id, time.time() - t)
        ) 

def _run_op_test(self, X, I, check_grad=False):
        with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
            op = core.CreateOperator('BatchPermutation', ['X', 'I'], ['Y'])
            workspace.FeedBlob('X', X)
            workspace.FeedBlob('I', I)
        workspace.RunOperatorOnce(op)
        Y = workspace.FetchBlob('Y')

        if check_grad:
            gc = gradient_checker.GradientChecker(
                stepsize=0.1,
                threshold=0.001,
                device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
            )

            res, grad, grad_estimated = gc.CheckSimple(op, [X, I], 0, [0])
            self.assertTrue(res, 'Grad check failed')

        Y_ref = X[I]
        np.testing.assert_allclose(Y, Y_ref, rtol=1e-5, atol=1e-08) 

def enqueue_blobs(self, gpu_id, blob_names, blobs):
        """Put a mini-batch on a BlobsQueue."""
        assert len(blob_names) == len(blobs)
        t = time.time()
        dev = c2_utils.CudaDevice(gpu_id)
        queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
        blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
        for (blob_name, blob) in zip(blob_names, blobs):
            workspace.FeedBlob(blob_name, blob, device_option=dev)
        logger.debug(
            'enqueue_blobs {}: workspace.FeedBlob: {}'.
            format(gpu_id, time.time() - t)
        )
        t = time.time()
        op = core.CreateOperator(
            'SafeEnqueueBlobs', [queue_name] + blob_names,
            blob_names + [queue_name + '_enqueue_status'],
            device_option=dev
        )
        workspace.RunOperatorOnce(op)
        logger.debug(
            'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
            format(gpu_id, time.time() - t)
        ) 

def enqueue_blobs(
        self,
        gpu_id,
        enqueue_blobs_names,
        blob_values,
    ):
        enqueue_blobs_names = [
            'gpu_{}/{}'.format(
                gpu_id, enqueue_blob_name
            ) for enqueue_blob_name in enqueue_blobs_names
        ]

        deviceOption = core.DeviceOption(caffe2_pb2.CUDA, gpu_id)
        for (blob_name, blob) in zip(enqueue_blobs_names, blob_values):
            workspace.FeedBlob(blob_name, blob, device_option=deviceOption)

        queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
        workspace.RunOperatorOnce(
            core.CreateOperator(
                'SafeEnqueueBlobs',
                [queue_name] + enqueue_blobs_names,
                enqueue_blobs_names + [queue_name + '_enqueue_status'],
                device_option=deviceOption,
            )
        ) 

def enqueue_blobs(self, gpu_id, blob_names, blobs):
        """Put a mini-batch on a BlobsQueue."""
        assert len(blob_names) == len(blobs)
        t = time.time()
        dev = c2_utils.CudaDevice(gpu_id)
        queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
        blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
        for (blob_name, blob) in zip(blob_names, blobs):
            workspace.FeedBlob(blob_name, blob, device_option=dev)
        logger.debug(
            'enqueue_blobs {}: workspace.FeedBlob: {}'.
            format(gpu_id, time.time() - t)
        )
        t = time.time()
        op = core.CreateOperator(
            'SafeEnqueueBlobs', [queue_name] + blob_names,
            blob_names + [queue_name + '_enqueue_status'],
            device_option=dev
        )
        workspace.RunOperatorOnce(op)
        logger.debug(
            'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
            format(gpu_id, time.time() - t)
        ) 

def _run_op_test(self, X, I, check_grad=False):
        with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
            op = core.CreateOperator('BatchPermutation', ['X', 'I'], ['Y'])
            workspace.FeedBlob('X', X)
            workspace.FeedBlob('I', I)
        workspace.RunOperatorOnce(op)
        Y = workspace.FetchBlob('Y')

        if check_grad:
            gc = gradient_checker.GradientChecker(
                stepsize=0.1,
                threshold=0.001,
                device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
            )

            res, grad, grad_estimated = gc.CheckSimple(op, [X, I], 0, [0])
            self.assertTrue(res, 'Grad check failed')

        Y_ref = X[I]
        np.testing.assert_allclose(Y, Y_ref, rtol=1e-5, atol=1e-08) 

def _run_op_test(self, X, I, check_grad=False):
        with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
            op = core.CreateOperator('BatchPermutation', ['X', 'I'], ['Y'])
            workspace.FeedBlob('X', X)
            workspace.FeedBlob('I', I)
        workspace.RunOperatorOnce(op)
        Y = workspace.FetchBlob('Y')

        if check_grad:
            gc = gradient_checker.GradientChecker(
                stepsize=0.1,
                threshold=0.001,
                device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
            )

            res, grad, grad_estimated = gc.CheckSimple(op, [X, I], 0, [0])
            self.assertTrue(res, 'Grad check failed')

        Y_ref = X[I]
        np.testing.assert_allclose(Y, Y_ref, rtol=1e-5, atol=1e-08) 

def test_add5(self, N, C, gc, dc):
        # set th seed
        np.random.seed(101)
        # TODO: test double, int and int64
        data = np.random.rand(N, C).astype(np.float32)
        op = core.CreateOperator("Add5", ["data"], ["output"])

        # device check
        self.assertDeviceChecks(dc, op, [data], [0])

        # gradient check
        self.assertGradientChecks(gc, op, [data], 0, [0])

        # reference check
        def ref_add5(input):
            return [input + 5]
        self.assertReferenceChecks(gc, op, [data], ref_add5) 

def _run_test(self, A, B, check_grad=False):
        with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
            op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
            workspace.FeedBlob('A', A)
            workspace.FeedBlob('B', B)
        workspace.RunOperatorOnce(op)
        C = workspace.FetchBlob('C')

        if check_grad:
            gc = gradient_checker.GradientChecker(
                stepsize=0.005,
                threshold=0.005,
                device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
            )

            res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
            self.assertTrue(res, 'Grad check failed')

        dims = C.shape
        C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
        np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08) 

def _create_const_fill_op_from_c2_int8_tensor(name, int8_tensor):
    assert type(int8_tensor) == workspace.Int8Tensor
    kTypeNameMapper = {
        np.dtype("int32"): "Int8GivenIntTensorFill",
        np.dtype("uint8"): "Int8GivenTensorFill",
    }

    tensor = int8_tensor.data
    assert tensor.dtype in [np.dtype("uint8"), np.dtype("int32")]
    values = tensor.tobytes() if tensor.dtype == np.dtype("uint8") else tensor

    return core.CreateOperator(
        kTypeNameMapper[tensor.dtype],
        [],
        [name],
        values=values,
        shape=tensor.shape,
        Y_scale=int8_tensor.scale,
        Y_zero_point=int8_tensor.zero_point,
    ) 

def _create_const_fill_op_from_numpy(name, tensor, device_option=None):
    assert type(tensor) == np.ndarray
    kTypeNameMapper = {
        np.dtype("float32"): "GivenTensorFill",
        np.dtype("int32"): "GivenTensorIntFill",
        np.dtype("int64"): "GivenTensorInt64Fill",
        np.dtype("uint8"): "GivenTensorStringFill",
    }

    args_dict = {}
    if tensor.dtype == np.dtype("uint8"):
        args_dict.update({"values": [str(tensor.data)], "shape": [1]})
    else:
        args_dict.update({"values": tensor, "shape": tensor.shape})

    if device_option is not None:
        args_dict["device_option"] = device_option

    return core.CreateOperator(kTypeNameMapper[tensor.dtype], [], [name], **args_dict) 

def _CorrectMomentum(self, correction):
        """The MomentumSGDUpdate op implements the update V as

            V := mu * V + lr * grad,

        where mu is the momentum factor, lr is the learning rate, and grad is
        the stochastic gradient. Since V is not defined independently of the
        learning rate (as it should ideally be), when the learning rate is
        changed we should scale the update history V in order to make it
        compatible in scale with lr * grad.
        """
        logger.info(
            'Scaling update history by {:.6f} (new lr / old lr)'.
            format(correction))
        for i in range(cfg.NUM_GPUS):
            with c2_utils.CudaScope(i):
                for param in self.TrainableParams(gpu_id=i):
                    op = core.CreateOperator(
                        'Scale', [param + '_momentum'], [param + '_momentum'],
                        scale=correction)
                    workspace.RunOperatorOnce(op) 

def enqueue_blobs(self, gpu_id, blob_names, blobs):
        """Put a mini-batch on a BlobsQueue."""
        assert len(blob_names) == len(blobs)
        t = time.time()
        dev = c2_utils.CudaDevice(gpu_id)
        queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
        blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
        for (blob_name, blob) in zip(blob_names, blobs):
            workspace.FeedBlob(blob_name, blob, device_option=dev)
        logger.debug(
            'enqueue_blobs {}: workspace.FeedBlob: {}'.
            format(gpu_id, time.time() - t)
        )
        t = time.time()
        op = core.CreateOperator(
            'SafeEnqueueBlobs', [queue_name] + blob_names,
            blob_names + [queue_name + '_enqueue_status'],
            device_option=dev
        )
        workspace.RunOperatorOnce(op)
        logger.debug(
            'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
            format(gpu_id, time.time() - t)
        ) 

def _run_op_test(self, X, I, check_grad=False):
        with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
            op = core.CreateOperator('BatchPermutation', ['X', 'I'], ['Y'])
            workspace.FeedBlob('X', X)
            workspace.FeedBlob('I', I)
        workspace.RunOperatorOnce(op)
        Y = workspace.FetchBlob('Y')

        if check_grad:
            gc = gradient_checker.GradientChecker(
                stepsize=0.1,
                threshold=0.001,
                device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
            )

            res, grad, grad_estimated = gc.CheckSimple(op, [X, I], 0, [0])
            self.assertTrue(res, 'Grad check failed')

        Y_ref = X[I]
        np.testing.assert_allclose(Y, Y_ref, rtol=1e-5, atol=1e-08) 

def _run_test(self, A, B, check_grad=False):
        with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA, 0)):
            op = core.CreateOperator('SpatialNarrowAs', ['A', 'B'], ['C'])
            workspace.FeedBlob('A', A)
            workspace.FeedBlob('B', B)
        workspace.RunOperatorOnce(op)
        C = workspace.FetchBlob('C')

        if check_grad:
            gc = gradient_checker.GradientChecker(
                stepsize=0.005,
                threshold=0.005,
                device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
            )

            res, grad, grad_estimated = gc.CheckSimple(op, [A, B], 0, [0])
            self.assertTrue(res, 'Grad check failed')

        dims = C.shape
        C_ref = A[:dims[0], :dims[1], :dims[2], :dims[3]]
        np.testing.assert_allclose(C, C_ref, rtol=1e-5, atol=1e-08) 

def enqueue_blobs(self, gpu_id, blob_names, blobs):
        """Put a mini-batch on a BlobsQueue."""
        assert len(blob_names) == len(blobs)
        t = time.time()
        dev = c2_utils.CudaDevice(gpu_id)
        queue_name = 'gpu_{}/{}'.format(gpu_id, self._blobs_queue_name)
        blob_names = ['gpu_{}/{}'.format(gpu_id, b) for b in blob_names]
        for (blob_name, blob) in zip(blob_names, blobs):
            workspace.FeedBlob(blob_name, blob, device_option=dev)
        logger.debug(
            'enqueue_blobs {}: workspace.FeedBlob: {}'.
            format(gpu_id, time.time() - t)
        )
        t = time.time()
        op = core.CreateOperator(
            'SafeEnqueueBlobs', [queue_name] + blob_names,
            blob_names + [queue_name + '_enqueue_status'],
            device_option=dev
        )
        workspace.RunOperatorOnce(op)
        logger.debug(
            'enqueue_blobs {}: workspace.RunOperatorOnce: {}'.
            format(gpu_id, time.time() - t)
        ) 

def _CorrectMomentum(self, correction):
        """The MomentumSGDUpdate op implements the update V as

            V := mu * V + lr * grad,

        where mu is the momentum factor, lr is the learning rate, and grad is
        the stochastic gradient. Since V is not defined independently of the
        learning rate (as it should ideally be), when the learning rate is
        changed we should scale the update history V in order to make it
        compatible in scale with lr * grad.
        """
        logger.info(
            'Scaling update history by {:.6f} (new lr / old lr)'.
            format(correction))
        for i in range(cfg.NUM_GPUS):
            with c2_utils.CudaScope(i):
                for param in self.TrainableParams(gpu_id=i):
                    op = core.CreateOperator(
                        'Scale', [param + '_momentum'], [param + '_momentum'],
                        scale=correction)
                    workspace.RunOperatorOnce(op) 

def _CorrectMomentum(self, correction):
        """The MomentumSGDUpdate op implements the update V as

            V := mu * V + lr * grad,

        where mu is the momentum factor, lr is the learning rate, and grad is
        the stochastic gradient. Since V is not defined independently of the
        learning rate (as it should ideally be), when the learning rate is
        changed we should scale the update history V in order to make it
        compatible in scale with lr * grad.
        """
        logger.info(
            'Scaling update history by {:.6f} (new lr / old lr)'.
            format(correction))
        for i in range(cfg.NUM_GPUS):
            with c2_utils.CudaScope(i):
                for param in self.TrainableParams(gpu_id=i):
                    op = core.CreateOperator(
                        'Scale', [param + '_momentum'], [param + '_momentum'],
                        scale=correction)
                    workspace.RunOperatorOnce(op) 

def test_forward_and_gradient(self):
        Y = np.random.randn(128, 4 * 21).astype(np.float32)
        Y_hat = np.random.randn(128, 4 * 21).astype(np.float32)
        inside_weights = np.random.randn(128, 4 * 21).astype(np.float32)
        inside_weights[inside_weights < 0] = 0
        outside_weights = np.random.randn(128, 4 * 21).astype(np.float32)
        outside_weights[outside_weights < 0] = 0
        scale = np.random.random()
        beta = np.random.random()

        op = core.CreateOperator(
            'SmoothL1Loss', ['Y_hat', 'Y', 'inside_weights', 'outside_weights'],
            ['loss'],
            scale=scale,
            beta=beta
        )

        gc = gradient_checker.GradientChecker(
            stepsize=0.005,
            threshold=0.005,
            device_option=core.DeviceOption(caffe2_pb2.CUDA, 0)
        )

        res, grad, grad_estimated = gc.CheckSimple(
            op, [Y_hat, Y, inside_weights, outside_weights], 0, [0]
        )

        self.assertTrue(
            grad.shape == grad_estimated.shape,
            'Fail check: grad.shape != grad_estimated.shape'
        )

        # To inspect the gradient and estimated gradient:
        # np.set_printoptions(precision=3, suppress=True)
        # print('grad:')
        # print(grad)
        # print('grad_estimated:')
        # print(grad_estimated)

        self.assertTrue(res) 

def close_blobs_queues(self):
        """Close a BlobsQueue."""
        for gpu_id in range(self._num_gpus):
            with core.NameScope('gpu_{}'.format(gpu_id)):
                workspace.RunOperatorOnce(
                    core.CreateOperator(
                        'CloseBlobsQueue', [self._blobs_queue_name], []
                    )
                ) 

def create_blobs_queues(self):
        """Create one BlobsQueue for each GPU to hold mini-batches."""
        for gpu_id in range(self._num_gpus):
            with c2_utils.GpuNameScope(gpu_id):
                workspace.RunOperatorOnce(
                    core.CreateOperator(
                        'CreateBlobsQueue', [], [self._blobs_queue_name],
                        num_blobs=len(self.get_output_names()),
                        capacity=self._blobs_queue_capacity
                    )
                )
        return self.create_enqueue_blobs() 

def close_blobs_queues(self):
        """Close a BlobsQueue."""
        for gpu_id in range(self._num_gpus):
            with core.NameScope('gpu_{}'.format(gpu_id)):
                workspace.RunOperatorOnce(
                    core.CreateOperator(
                        'CloseBlobsQueue', [self._blobs_queue_name], []
                    )
                ) 

def add_bbox_ops(args, net, blobs):
    new_ops = []
    new_external_outputs = []

    # Operators for bboxes
    op_box = core.CreateOperator(
        "BBoxTransform",
        ["rpn_rois", "bbox_pred", "im_info"],
        ["pred_bbox"],
        weights=cfg.MODEL.BBOX_REG_WEIGHTS,
        apply_scale=False,
        correct_transform_coords=True,
    )
    new_ops.extend([op_box])

    blob_prob = "cls_prob"
    blob_box = "pred_bbox"
    op_nms = core.CreateOperator(
        "BoxWithNMSLimit",
        [blob_prob, blob_box],
        ["score_nms", "bbox_nms", "class_nms"],
        arg=[
            putils.MakeArgument("score_thresh", cfg.TEST.SCORE_THRESH),
            putils.MakeArgument("nms", cfg.TEST.NMS),
            putils.MakeArgument("detections_per_im", cfg.TEST.DETECTIONS_PER_IM),
            putils.MakeArgument("soft_nms_enabled", cfg.TEST.SOFT_NMS.ENABLED),
            putils.MakeArgument("soft_nms_method", cfg.TEST.SOFT_NMS.METHOD),
            putils.MakeArgument("soft_nms_sigma", cfg.TEST.SOFT_NMS.SIGMA),
        ],
    )
    new_ops.extend([op_nms])
    new_external_outputs.extend(["score_nms", "bbox_nms", "class_nms"])

    net.Proto().op.extend(new_ops)
    net.Proto().external_output.extend(new_external_outputs)