PluginV2

Adds a node with custom functionality to TensorRT.

See also

Extending TensorRT with Custom Layers

Attributes

Plugin specific attributes.

Inputs

Plugin specific inputs.

Outputs

Plugin specific outputs.

Data Types

Plugin specific data types.

Shape Information

Plugin specific shapes.

DLA Restrictions

Plugin specific restrictions.

Examples

Plugin

def init_gelu_plugin(bias):
    for plugin_creator in trt.get_plugin_registry().plugin_creator_list:
        if plugin_creator.name == "ustomGeluPluginDynamic":
            gelu_data_type = trt.PluginField("type_id", np.array([0], dtype=np.int32), type=trt.PluginFieldType.INT8)
            gelu_bias = trt.PluginField("bias", bias, type=trt.PluginFieldType.FLOAT32)
            field_collection = trt.PluginFieldCollection([gelu_data_type, gelu_bias])
            return plugin_creator.create_plugin(name="CustomGeluPluginDynamic", field_collection=field_collection)
    
    raise Exception('CustomGeluPluginDynamic Plugin not found')

bias = np.array([[[1, -0.5, 0, 0, 0.7]]], dtype=np.float32)

input_layer = network.add_input(name="input_layer", dtype=trt.float32, shape=(1, 1, 5))
lrelu = network.add_plugin_v2(inputs=[input_layer], plugin=init_gelu_plugin(bias))
network.mark_output(lrelu.get_output(0))

inputs[input_layer.name] = np.array(
    [[
        [3.0, -4.3, 22.8, -2.97, 143.2],
    ]]
)
outputs[lrelu.get_output(0).name] = lrelu.get_output(0).shape

input_with_bias = inputs[input_layer.name] + bias
expected_output = 0.5 * input_with_bias * (1 + np.tanh(np.sqrt(2 / math.pi) * (input_with_bias + 0.044715 * np.power(input_with_bias, 3))))

expected[lrelu.get_output(0).name] = expected_output

C++ API

For more information about the C++ IPluginV2Layer operator, refer to the C++ IPluginV2Layer documentation.

Python API

For more information about the Python IPluginV2Layer operator, refer to the Python IPluginV2Layer documentation.