Python caffe2.python.core.Net() Examples

def _run_speed_test(self, iters=5, N=1024):
        """This function provides an example of how to benchmark custom
        operators using the Caffe2 'prof_dag' network execution type. Please
        note that for 'prof_dag' to work, Caffe2 must be compiled with profiling
        support using the `-DUSE_PROF=ON` option passed to `cmake` when building
        Caffe2.
        """
        net = core.Net('test')
        net.Proto().type = 'prof_dag'
        net.Proto().num_workers = 2
        Y = net.BatchPermutation(['X', 'I'], 'Y')
        Y_flat = net.FlattenToVec([Y], 'Y_flat')
        loss = net.AveragedLoss([Y_flat], 'loss')
        net.AddGradientOperators([loss])
        workspace.CreateNet(net)

        X = np.random.randn(N, 256, 14, 14)
        for _i in range(iters):
            I = np.random.permutation(N)
            workspace.FeedBlob('X', X.astype(np.float32))
            workspace.FeedBlob('I', I.astype(np.int32))
            workspace.RunNet(net.Proto().name)
            np.testing.assert_allclose(
                workspace.FetchBlob('Y'), X[I], rtol=1e-5, atol=1e-08
            ) 

def test_constant():
    workspace.ResetWorkspace()

    shape = [10, 10]
    val = random.random()
    net = core.Net("net")
    net.ConstantFill([], ["Y"], shape=shape, value=val, run_once=0, name="Y")

    # Execute via Caffe2
    workspace.RunNetOnce(net)

    # Import caffe2 network into ngraph
    importer = C2Importer()
    importer.parse_net_def(net.Proto(), verbose=False)

    # Get handle
    f_ng = importer.get_op_handle("Y")

    # Execute
    with ExecutorFactory() as ex:
        f_result = ex.executor(f_ng)()

        # compare Caffe2 and ngraph results
        assert(np.ma.allequal(f_result, workspace.FetchBlob("Y")))
        assert(np.isclose(f_result[0][0], val, atol=1e-6, rtol=0)) 

def _run_speed_test(self, iters=5, N=1024):
        """This function provides an example of how to benchmark custom
        operators using the Caffe2 'prof_dag' network execution type. Please
        note that for 'prof_dag' to work, Caffe2 must be compiled with profiling
        support using the `-DUSE_PROF=ON` option passed to `cmake` when building
        Caffe2.
        """
        net = core.Net('test')
        net.Proto().type = 'prof_dag'
        net.Proto().num_workers = 2
        Y = net.BatchPermutation(['X', 'I'], 'Y')
        Y_flat = net.FlattenToVec([Y], 'Y_flat')
        loss = net.AveragedLoss([Y_flat], 'loss')
        net.AddGradientOperators([loss])
        workspace.CreateNet(net)

        X = np.random.randn(N, 256, 14, 14)
        for _i in range(iters):
            I = np.random.permutation(N)
            workspace.FeedBlob('X', X.astype(np.float32))
            workspace.FeedBlob('I', I.astype(np.int32))
            workspace.RunNet(net.Proto().name)
            np.testing.assert_allclose(
                workspace.FetchBlob('Y'), X[I], rtol=1e-5, atol=1e-08
            ) 

def _run_speed_test(self, iters=5, N=1024):
        """This function provides an example of how to benchmark custom
        operators using the Caffe2 'prof_dag' network execution type. Please
        note that for 'prof_dag' to work, Caffe2 must be compiled with profiling
        support using the `-DUSE_PROF=ON` option passed to `cmake` when building
        Caffe2.
        """
        net = core.Net('test')
        net.Proto().type = 'prof_dag'
        net.Proto().num_workers = 2
        Y = net.BatchPermutation(['X', 'I'], 'Y')
        Y_flat = net.FlattenToVec([Y], 'Y_flat')
        loss = net.AveragedLoss([Y_flat], 'loss')
        net.AddGradientOperators([loss])
        workspace.CreateNet(net)

        X = np.random.randn(N, 256, 14, 14)
        for _i in range(iters):
            I = np.random.permutation(N)
            workspace.FeedBlob('X', X.astype(np.float32))
            workspace.FeedBlob('I', I.astype(np.int32))
            workspace.RunNet(net.Proto().name)
            np.testing.assert_allclose(
                workspace.FetchBlob('Y'), X[I], rtol=1e-5, atol=1e-08
            ) 

def get_net(data_loader, name):
    logger = logging.getLogger(__name__)
    blob_names = data_loader.get_output_names()
    net = core.Net(name)
    net.type = 'dag'
    for gpu_id in range(cfg.NUM_GPUS):
        with core.NameScope('gpu_{}'.format(gpu_id)):
            with core.DeviceScope(muji.OnGPU(gpu_id)):
                for blob_name in blob_names:
                    blob = core.ScopedName(blob_name)
                    workspace.CreateBlob(blob)
                net.DequeueBlobs(
                    data_loader._blobs_queue_name, blob_names)
    logger.info("Protobuf:\n" + str(net.Proto()))

    return net 

def get_net(data_loader, name):
    logger = logging.getLogger(__name__)
    blob_names = data_loader.get_output_names()
    net = core.Net(name)
    net.type = 'dag'
    for gpu_id in range(cfg.NUM_GPUS):
        with core.NameScope('gpu_{}'.format(gpu_id)):
            with core.DeviceScope(muji.OnGPU(gpu_id)):
                for blob_name in blob_names:
                    blob = core.ScopedName(blob_name)
                    workspace.CreateBlob(blob)
                net.DequeueBlobs(
                    data_loader._blobs_queue_name, blob_names)
    logger.info("Protobuf:\n" + str(net.Proto()))

    return net 

def _run_speed_test(self, iters=5, N=1024):
        """This function provides an example of how to benchmark custom
        operators using the Caffe2 'prof_dag' network execution type. Please
        note that for 'prof_dag' to work, Caffe2 must be compiled with profiling
        support using the `-DUSE_PROF=ON` option passed to `cmake` when building
        Caffe2.
        """
        net = core.Net('test')
        net.Proto().type = 'prof_dag'
        net.Proto().num_workers = 2
        Y = net.BatchPermutation(['X', 'I'], 'Y')
        Y_flat = net.FlattenToVec([Y], 'Y_flat')
        loss = net.AveragedLoss([Y_flat], 'loss')
        net.AddGradientOperators([loss])
        workspace.CreateNet(net)

        X = np.random.randn(N, 256, 14, 14)
        for _i in range(iters):
            I = np.random.permutation(N)
            workspace.FeedBlob('X', X.astype(np.float32))
            workspace.FeedBlob('I', I.astype(np.int32))
            workspace.RunNet(net.Proto().name)
            np.testing.assert_allclose(
                workspace.FetchBlob('Y'), X[I], rtol=1e-5, atol=1e-08
            ) 

def write_graph(self, model_or_nets_or_protos=None, **kwargs):
        '''Write graph to the summary.'''
        if isinstance(model_or_nets_or_protos, cnn.CNNModelHelper):
            current_graph, track_blob_names = model_to_graph(model_or_nets_or_protos, **kwargs)
        elif isinstance(model_or_nets_or_protos, list):
            if isinstance(model_or_nets_or_protos[0], core.Net):
                current_graph, track_blob_names = nets_to_graph(model_or_nets_or_protos, **kwargs)
            elif isinstance(model_or_nets_or_protos[0], caffe2_pb2.NetDef):
                current_graph, track_blob_names = protos_to_graph(model_or_nets_or_protos, **kwargs)
            else:
                raise NotImplementedError
        else:
            raise NotImplementedError
        self._file_writer.add_graph(current_graph)
        self._track_blob_names = track_blob_names
        # Once the graph is built, one can just map the blobs
        self.check_names()
        self.sort_out_names() 

def load_model_pb(net_file, init_file=None, is_run_init=True, is_create_net=True):
    net = core.Net("net")
    if net_file is not None:
        net.Proto().ParseFromString(open(net_file, "rb").read())

    if init_file is None:
        fn, ext = os.path.splitext(net_file)
        init_file = fn + "_init" + ext

    init_net = caffe2_pb2.NetDef()
    init_net.ParseFromString(open(init_file, "rb").read())

    if is_run_init:
        workspace.RunNetOnce(init_net)
        create_blobs_if_not_existed(net.external_inputs)
        if net.Proto().name == "":
            net.Proto().name = "net"
    if is_create_net:
        workspace.CreateNet(net)

    return (net, init_net) 

def _run_speed_test(self, iters=5, N=1024):
        """This function provides an example of how to benchmark custom
        operators using the Caffe2 'prof_dag' network execution type. Please
        note that for 'prof_dag' to work, Caffe2 must be compiled with profiling
        support using the `-DUSE_PROF=ON` option passed to `cmake` when building
        Caffe2.
        """
        net = core.Net('test')
        net.Proto().type = 'prof_dag'
        net.Proto().num_workers = 2
        Y = net.BatchPermutation(['X', 'I'], 'Y')
        Y_flat = net.FlattenToVec([Y], 'Y_flat')
        loss = net.AveragedLoss([Y_flat], 'loss')
        net.AddGradientOperators([loss])
        workspace.CreateNet(net)

        X = np.random.randn(N, 256, 14, 14)
        for _i in range(iters):
            I = np.random.permutation(N)
            workspace.FeedBlob('X', X.astype(np.float32))
            workspace.FeedBlob('I', I.astype(np.int32))
            workspace.RunNet(net.Proto().name)
            np.testing.assert_allclose(
                workspace.FetchBlob('Y'), X[I], rtol=1e-5, atol=1e-08
            ) 

def get_net(data_loader, name):
    logger = logging.getLogger(__name__)
    blob_names = data_loader.get_output_names()
    net = core.Net(name)
    net.type = 'dag'
    for gpu_id in range(cfg.NUM_GPUS):
        with core.NameScope('gpu_{}'.format(gpu_id)):
            with core.DeviceScope(muji.OnGPU(gpu_id)):
                for blob_name in blob_names:
                    blob = core.ScopedName(blob_name)
                    workspace.CreateBlob(blob)
                net.DequeueBlobs(
                    data_loader._blobs_queue_name, blob_names)
    logger.info("Protobuf:\n" + str(net.Proto()))

    return net 

def get_net(data_loader, name):
    logger = logging.getLogger(__name__)
    blob_names = data_loader.get_output_names()
    net = core.Net(name)
    net.type = 'dag'
    for gpu_id in range(cfg.NUM_GPUS):
        with core.NameScope('gpu_{}'.format(gpu_id)):
            with core.DeviceScope(muji.OnGPU(gpu_id)):
                for blob_name in blob_names:
                    blob = core.ScopedName(blob_name)
                    workspace.CreateBlob(blob)
                net.DequeueBlobs(
                    data_loader._blobs_queue_name, blob_names)
    logger.info("Protobuf:\n" + str(net.Proto()))

    return net 

def test_SquaredL2Distance():
    workspace.ResetWorkspace()
    shape = (10, 10)

    net = core.Net("net")
    Y = net.GivenTensorFill([], "Y", shape=shape, values=np.random.uniform(-1, 1, shape))
    T = net.GivenTensorFill([], "T", shape=shape, values=np.random.uniform(-1, 1, shape))
    net.SquaredL2Distance([Y, T], "dist")

    # Execute via Caffe2
    workspace.RunNetOnce(net)

    # Import caffe2 network into ngraph
    importer = C2Importer()
    importer.parse_net_def(net.Proto(), verbose=False)

    # Get handle
    f_ng = importer.get_op_handle("dist")

    # Execute
    with ExecutorFactory() as ex:
        f_result = ex.executor(f_ng)()

        assert(np.allclose(f_result, workspace.FetchBlob("dist"), equal_nan=False)) 

def _run_speed_test(self, iters=5, N=1024):
        """This function provides an example of how to benchmark custom
        operators using the Caffe2 'prof_dag' network execution type. Please
        note that for 'prof_dag' to work, Caffe2 must be compiled with profiling
        support using the `-DUSE_PROF=ON` option passed to `cmake` when building
        Caffe2.
        """
        net = core.Net('test')
        net.Proto().type = 'prof_dag'
        net.Proto().num_workers = 2
        Y = net.BatchPermutation(['X', 'I'], 'Y')
        Y_flat = net.FlattenToVec([Y], 'Y_flat')
        loss = net.AveragedLoss([Y_flat], 'loss')
        net.AddGradientOperators([loss])
        workspace.CreateNet(net)

        X = np.random.randn(N, 256, 14, 14)
        for _i in range(iters):
            I = np.random.permutation(N)
            workspace.FeedBlob('X', X.astype(np.float32))
            workspace.FeedBlob('I', I.astype(np.int32))
            workspace.RunNet(net.Proto().name)
            np.testing.assert_allclose(
                workspace.FetchBlob('Y'), X[I], rtol=1e-5, atol=1e-08
            ) 

def get_net(data_loader, name):
    logger = logging.getLogger(__name__)
    blob_names = data_loader.get_output_names()
    net = core.Net(name)
    net.type = 'dag'
    for gpu_id in range(cfg.NUM_GPUS):
        with core.NameScope('gpu_{}'.format(gpu_id)):
            with core.DeviceScope(muji.OnGPU(gpu_id)):
                for blob_name in blob_names:
                    blob = core.ScopedName(blob_name)
                    workspace.CreateBlob(blob)
                net.DequeueBlobs(
                    data_loader._blobs_queue_name, blob_names)
    logger.info("Protobuf:\n" + str(net.Proto()))

    return net 

def test_AveragedLoss():
    workspace.ResetWorkspace()
    shape = (32,)

    net = core.Net("net")
    X = net.GivenTensorFill([], "Y", shape=shape, values=np.random.uniform(-1, 1, shape))
    X.AveragedLoss([], ["loss"])

    # Execute via Caffe2
    workspace.RunNetOnce(net)

    # Import caffe2 network into ngraph
    importer = C2Importer()
    importer.parse_net_def(net.Proto(), verbose=False)

    # Get handle
    f_ng = importer.get_op_handle("loss")

    # Execute
    with ExecutorFactory() as ex:
        f_result = ex.executor(f_ng)()

        assert(np.allclose(f_result, workspace.FetchBlob("loss"), equal_nan=False)) 

def _try_get_shapes(nets):
    '''
    Get missing shapes for all blobs contained in the nets.

    Args:
        nets: List of core.Net to extract blob shape information from.

    Returns:
        Dictionary containing blob name to shape/dimensions mapping. The net
            is a computation graph that is composed of operators, and the
            operators have input and output blobs, each with their own dims.
    '''
    try:
        # Note: this will inspect the workspace for better or worse.
        # We don't care about the types, only the shapes
        shapes, _ = workspace.InferShapesAndTypes(nets)
        return shapes
    except Exception as e:
        logging.warning('Failed to compute shapes: %s', e)
        return {} 

def model_to_graph_def(model, **kwargs):
    '''
    Convert a Caffe2 model to a Tensorflow graph. This function extracts
    'param_init_net' and 'net' from the model and passes it to nets_to_graph()
    for further processing.

    Args:
        model (cnn.CNNModelHelper, model_helper.ModelHelper): The model to
            extract the nets (instances of core.Net) from.

    Returns:
        Call to nets_to_graph_def() with extracted 'param_init_net', 'net' and
            **kwargs. See _operators_to_graph_def for detailed **kwargs.
    '''
    nets = [model.param_init_net, model.net]
    return nets_to_graph_def(nets, **kwargs) 

def nets_to_graph_def(nets, shapes=None, **kwargs):
    '''
    Convert a set of Caffe2 nets to a Tensorflow graph.

    Args:
        nets: List of core.Nets. core.Net is a wrapper around a NetDef protobuf.
            The corresponding protobuf can be extracted using .Proto().
        shapes: Dictionary mapping blob names to their shapes/dimensions.

    Returns:
        Call to protos_to_graph_def() with the extracted NetDef protobufs and
            **kwargs. See _operators_to_graph_def for detailed **kwargs.
    '''
    # if shapes is None:
    #     shapes = _try_get_shapes(nets)
    # _try_get_shapes(nets) depends on workspace.InferShapesAndTypes(nets),
    # which is currently broken (segfault). We omit the shapes for now.
    shapes = {}
    nets = [copy.deepcopy(net.Proto()) for net in nets]
    shapes = copy.deepcopy(shapes)
    return protos_to_graph_def(nets, shapes, **kwargs) 

def get_net(data_loader, name):
    logger = logging.getLogger(__name__)
    blob_names = data_loader.get_output_names()
    net = core.Net(name)
    net.type = 'dag'
    for gpu_id in range(cfg.NUM_GPUS):
        with core.NameScope('gpu_{}'.format(gpu_id)):
            with core.DeviceScope(muji.OnGPU(gpu_id)):
                for blob_name in blob_names:
                    blob = core.ScopedName(blob_name)
                    workspace.CreateBlob(blob)
                net.DequeueBlobs(
                    data_loader._blobs_queue_name, blob_names)
    logger.info("Protobuf:\n" + str(net.Proto()))

    return net 

def _try_get_shapes(nets):
    '''
    Get missing shapes for all blobs contained in the nets.

    Args:
        nets: List of core.Net to extract blob shape information from.

    Returns:
        Dictionary containing blob name to shape/dimensions mapping. The net
            is a computation graph that is composed of operators, and the
            operators have input and output blobs, each with their own dims.
    '''
    try:
        # Note: this will inspect the workspace for better or worse.
        # We don't care about the types, only the shapes
        shapes, _ = workspace.InferShapesAndTypes(nets)
        return shapes
    except Exception as e:
        logging.warning('Failed to compute shapes: %s', e)
        return {} 

def nets_to_graph_def(nets, shapes=None, **kwargs):
    '''
    Convert a set of Caffe2 nets to a Tensorflow graph.

    Args:
        nets: List of core.Nets. core.Net is a wrapper around a NetDef protobuf.
            The corresponding protobuf can be extracted using .Proto().
        shapes: Dictionary mapping blob names to their shapes/dimensions.

    Returns:
        Call to protos_to_graph_def() with the extracted NetDef protobufs and
            **kwargs. See _operators_to_graph_def for detailed **kwargs.
    '''
    # if shapes is None:
    #     shapes = _try_get_shapes(nets)
    # _try_get_shapes(nets) depends on workspace.InferShapesAndTypes(nets),
    # which is currently broken (segfault). We omit the shapes for now.
    shapes = {}
    nets = [copy.deepcopy(net.Proto()) for net in nets]
    shapes = copy.deepcopy(shapes)
    return protos_to_graph_def(nets, shapes, **kwargs) 

def _run_speed_test(self, iters=5, N=1024):
        """This function provides an example of how to benchmark custom
        operators using the Caffe2 'prof_dag' network execution type. Please
        note that for 'prof_dag' to work, Caffe2 must be compiled with profiling
        support using the `-DUSE_PROF=ON` option passed to `cmake` when building
        Caffe2.
        """
        net = core.Net('test')
        net.Proto().type = 'prof_dag'
        net.Proto().num_workers = 2
        Y = net.BatchPermutation(['X', 'I'], 'Y')
        Y_flat = net.FlattenToVec([Y], 'Y_flat')
        loss = net.AveragedLoss([Y_flat], 'loss')
        net.AddGradientOperators([loss])
        workspace.CreateNet(net)

        X = np.random.randn(N, 256, 14, 14)
        for _i in range(iters):
            I = np.random.permutation(N)
            workspace.FeedBlob('X', X.astype(np.float32))
            workspace.FeedBlob('I', I.astype(np.int32))
            workspace.RunNet(net.Proto().name)
            np.testing.assert_allclose(
                workspace.FetchBlob('Y'), X[I], rtol=1e-5, atol=1e-08
            ) 

def load_model_pb(net_file, init_file=None, is_run_init=True, is_create_net=True):
    net = core.Net("net")
    if net_file is not None:
        net.Proto().ParseFromString(open(net_file, "rb").read())

    if init_file is None:
        fn, ext = os.path.splitext(net_file)
        init_file = fn + "_init" + ext

    init_net = caffe2_pb2.NetDef()
    init_net.ParseFromString(open(init_file, "rb").read())

    if is_run_init:
        workspace.RunNetOnce(init_net)
        create_blobs_if_not_existed(net.external_inputs)
        if net.Proto().name == "":
            net.Proto().name = "net"
    if is_create_net:
        workspace.CreateNet(net)

    return (net, init_net) 

def __init__(self, predict_net, init_net, *, convert_outputs=None):
        """
        Args:
            predict_net, init_net (core.Net): caffe2 nets
            convert_outptus (callable): a function that converts caffe2
                outputs to the same format of the original pytorch model.
                By default, use the one defined in the caffe2 meta_arch.
        """
        super().__init__()
        self.protobuf_model = ProtobufModel(predict_net, init_net)
        self.size_divisibility = get_pb_arg_vali(predict_net, "size_divisibility", 0)
        self.device = get_pb_arg_vals(predict_net, "device", b"cpu").decode("ascii")

        if convert_outputs is None:
            meta_arch = get_pb_arg_vals(predict_net, "meta_architecture", b"GeneralizedRCNN")
            meta_arch = META_ARCH_CAFFE2_EXPORT_TYPE_MAP[meta_arch.decode("ascii")]
            self._convert_outputs = meta_arch.get_outputs_converter(predict_net, init_net)
        else:
            self._convert_outputs = convert_outputs 

def _run_speed_test(self, iters=5, N=1024):
        """This function provides an example of how to benchmark custom
        operators using the Caffe2 'prof_dag' network execution type. Please
        note that for 'prof_dag' to work, Caffe2 must be compiled with profiling
        support using the `-DUSE_PROF=ON` option passed to `cmake` when building
        Caffe2.
        """
        net = core.Net('test')
        net.Proto().type = 'prof_dag'
        net.Proto().num_workers = 2
        Y = net.BatchPermutation(['X', 'I'], 'Y')
        Y_flat = net.FlattenToVec([Y], 'Y_flat')
        loss = net.AveragedLoss([Y_flat], 'loss')
        net.AddGradientOperators([loss])
        workspace.CreateNet(net)

        X = np.random.randn(N, 256, 14, 14)
        for _i in range(iters):
            I = np.random.permutation(N)
            workspace.FeedBlob('X', X.astype(np.float32))
            workspace.FeedBlob('I', I.astype(np.int32))
            workspace.RunNet(net.Proto().name)
            np.testing.assert_allclose(
                workspace.FetchBlob('Y'), X[I], rtol=1e-5, atol=1e-08
            ) 

def get_net(data_loader, name):
    logger = logging.getLogger(__name__)
    blob_names = data_loader.get_output_names()
    net = core.Net(name)
    net.type = 'dag'
    for gpu_id in range(cfg.NUM_GPUS):
        with core.NameScope('gpu_{}'.format(gpu_id)):
            with core.DeviceScope(muji.OnGPU(gpu_id)):
                for blob_name in blob_names:
                    blob = core.ScopedName(blob_name)
                    workspace.CreateBlob(blob)
                net.DequeueBlobs(
                    data_loader._blobs_queue_name, blob_names)
    logger.info("Protobuf:\n" + str(net.Proto()))

    return net 

def __init__(self, predict_net, init_net):
        logger.info("Initializing ProtobufModel ...")
        super().__init__()
        assert isinstance(predict_net, caffe2_pb2.NetDef)
        assert isinstance(init_net, caffe2_pb2.NetDef)
        self.ws_name = "__ws_tmp__"
        self.net = core.Net(predict_net)

        with ScopedWS(self.ws_name, is_reset=True, is_cleanup=False) as ws:
            ws.RunNetOnce(init_net)
            for blob in self.net.Proto().external_input:
                if blob not in ws.Blobs():
                    ws.CreateBlob(blob)
            ws.CreateNet(self.net)

        self._error_msgs = set() 

def _run_speed_test(self, iters=5, N=1024):
        """This function provides an example of how to benchmark custom
        operators using the Caffe2 'prof_dag' network execution type. Please
        note that for 'prof_dag' to work, Caffe2 must be compiled with profiling
        support using the `-DUSE_PROF=ON` option passed to `cmake` when building
        Caffe2.
        """
        net = core.Net('test')
        net.Proto().type = 'prof_dag'
        net.Proto().num_workers = 2
        Y = net.BatchPermutation(['X', 'I'], 'Y')
        Y_flat = net.FlattenToVec([Y], 'Y_flat')
        loss = net.AveragedLoss([Y_flat], 'loss')
        net.AddGradientOperators([loss])
        workspace.CreateNet(net)

        X = np.random.randn(N, 256, 14, 14)
        for _i in range(iters):
            I = np.random.permutation(N)
            workspace.FeedBlob('X', X.astype(np.float32))
            workspace.FeedBlob('I', I.astype(np.int32))
            workspace.RunNet(net.Proto().name)
            np.testing.assert_allclose(
                workspace.FetchBlob('Y'), X[I], rtol=1e-5, atol=1e-08
            ) 

def get_net(data_loader, name):
    logger = logging.getLogger(__name__)
    blob_names = data_loader.get_output_names()
    net = core.Net(name)
    net.type = 'dag'
    for gpu_id in range(cfg.NUM_GPUS):
        with core.NameScope('gpu_{}'.format(gpu_id)):
            with core.DeviceScope(muji.OnGPU(gpu_id)):
                for blob_name in blob_names:
                    blob = core.ScopedName(blob_name)
                    workspace.CreateBlob(blob)
                net.DequeueBlobs(
                    data_loader._blobs_queue_name, blob_names)
    logger.info("Protobuf:\n" + str(net.Proto()))

    return net