VPU Devices

This chapter provides information on the OpenVINO™ Runtime plugins that enable inference of deep learning models on the supported VPU devices:

  • Intel® Neural Compute Stick 2 powered by the Intel® Movidius™ Myriad™ X — Supported by the MYRIAD Plugin

  • Intel® Vision Accelerator Design with Intel® Movidius™ VPUs — Supported by the HDDL Plugin

Note

With the OpenVINO™ 2020.4 release, Intel® Movidius™ Neural Compute Stick powered by the Intel® Movidius™ Myriad™ 2 is no longer supported.

Supported Networks

Caffe :

  • AlexNet

  • CaffeNet

  • GoogleNet (Inception) v1, v2, v4

  • VGG family (VGG16, VGG19)

  • SqueezeNet v1.0, v1.1

  • ResNet v1 family (18***, 50, 101, 152)

  • MobileNet (mobilenet-v1-1.0-224, mobilenet-v2)

  • Inception ResNet v2

  • DenseNet family (121,161,169,201)

  • SSD-300, SSD-512, SSD-MobileNet, SSD-GoogleNet, SSD-SqueezeNet

TensorFlow :

  • AlexNet

  • Inception v1, v2, v3, v4

  • Inception ResNet v2

  • MobileNet v1, v2

  • ResNet v1 family (50, 101, 152)

  • ResNet v2 family (50, 101, 152)

  • SqueezeNet v1.0, v1.1

  • VGG family (VGG16, VGG19)

  • Yolo family (yolo-v2, yolo-v3, tiny-yolo-v1, tiny-yolo-v2, tiny-yolo-v3)

  • faster_rcnn_inception_v2, faster_rcnn_resnet101

  • ssd_mobilenet_v1

  • DeepLab-v3+

Apache MXNet :

  • AlexNet and CaffeNet

  • DenseNet family (121,161,169,201)

  • SqueezeNet v1.1

  • MobileNet v1, v2

  • NiN

  • ResNet v1 (101, 152)

  • ResNet v2 (101)

  • SqueezeNet v1.1

  • VGG family (VGG16, VGG19)

  • SSD-Inception-v3, SSD-MobileNet, SSD-ResNet-50, SSD-300

*** Network is tested on Intel Neural Compute Stick 2 with BatchNormalization fusion optimization disabled during Model Optimizer import

Optimizations

VPU plugins support layer fusion and decomposition.

Layer Fusion

Fusing Rules

Certain layers can be merged into ‘convolution’, ‘ReLU’, and ‘Eltwise’ layers according to the patterns below:

  • Convolution

    • Convolution + ReLU → Convolution

    • Convolution + Clamp → Convolution

    • Convolution + LeakyReLU → Convolution

    • Convolution (3x3, stride=1, padding=1) + Pooling (2x2, stride=2, padding=0) → Convolution

  • Pooling + ReLU → Pooling

  • FullyConnected + ReLU → FullyConnected

  • Eltwise

    • Eltwise + ReLU → Eltwise

    • Eltwise + LeakyReLU → Eltwise

    • Eltwise + Clamp → Eltwise

Joining Rules

Note

Application of these rules depends on tensor sizes and available resources.

Layers can be joined only when the two conditions below are met:

  • Layers are located on topologically independent branches.

  • Layers can be executed simultaneously on the same hardware units.

Decomposition Rules

  • Convolution and Pooling layers are tiled, resulting in the following pattern:

    • A Split layer that splits tensors into tiles

    • A set of tiles, optionally with service layers like Copy

    • Depending on a tiling scheme, a Concatenation or Sum layer that joins all resulting tensors into one and restores the full blob that contains the result of a tiled operation

    Names of tiled layers contain the @soc=M/N part, where M is the tile number and N is the number of tiles:

    _images/yolo_tiny_v1.png

Note

Nominal layers, such as Shrink and Expand, are not executed.

Note

VPU plugins can add extra layers like Copy.

VPU Common Configuration Parameters

VPU plugins support the configuration parameters listed below. The parameters are passed as std::map<std::string, std::string> on InferenceEngine::Core::LoadNetwork or InferenceEngine::Core::SetConfig. When specifying key values as raw strings (when using Python API), omit the KEY_ prefix.

Parameter Name

Parameter Values

Default

Description

KEY_VPU_HW_STAGES_OPTIMIZATION

YES / NO

YES

Turn on HW stages usage Applicable for Intel Movidius Myriad X and Intel Vision Accelerator Design devices only.

KEY_VPU_COMPUTE_LAYOUT

VPU_AUTO , VPU_NCHW , VPU_NHWC

VPU_AUTO

Specify internal input and output layouts for network layers.

KEY_VPU_PRINT_RECEIVE_TENSOR_TIME

YES / NO

NO

Add device-side time spent waiting for input to PerformanceCounts. See the Data Transfer Pipelining section for details.

KEY_VPU_IGNORE_IR_STATISTIC

YES / NO

NO

VPU plugin could use statistic present in IR in order to try to improve calculations precision. This option is enabled to exclude the statistic.

KEY_VPU_CUSTOM_LAYERS

path to XML file

empty string

This option allows passing XML file with custom layers binding. If a layer is present in such file, it will be used during inference even if the layer is natively supported.

Data Transfer Pipelining

MYRIAD plugin tries to pipeline data transfer to/from a device with computations. While one infer request is executed, the data for the next infer request can be uploaded to a device in parallel. The same applies to result downloading.

KEY_VPU_PRINT_RECEIVE_TENSOR_TIME configuration parameter can be used to check the efficiency of current pipelining. The new record in performance counters will show the time that device spent waiting for input before starting the inference. In a perfect pipeline this time should be near zero, which means that the data was already transferred when new inference started.

Troubleshooting

When running inference with the VPU plugin: “[VPU] Cannot convert layer <layer_name> due to unsupported layer type <layer_type>”

This means that the topology has a layer unsupported by the target VPU plugin. To resolve this issue, the custom layer can be implemented for the target device, using the OpenVINO™ Extensibility mechanism. To quickly get a working prototype, use the heterogeneous scenario with the default fallback policy (see the Heterogeneous execution section). Use the HETERO mode with a fallback device that supports this layer, for example, CPU: HETERO:MYRIAD,CPU. For a list of VPU-supported layers, see the Supported Layers section of the Supported Devices page.

Known Layers Limitations

  • ScaleShift layer is supported for zero value of broadcast attribute only.

  • CTCGreedyDecoder layer works with the ctc_merge_repeated attribute equal to 1.

  • DetectionOutput layer works with zero values of interpolate_orientation and num_orient_classes parameters only.

  • MVN layer uses fixed value for eps parameters (1e-9).

  • Normalize layer uses fixed value for eps parameters (1e-9) and is supported for zero value of across_spatial only.

  • Pad layer works only with 4D tensors.