Questions about plugin Layer of priorbox in SSD

Autonomous Vehicles DRIVE AGX Xavier DRIVE AGX Xavier General
1

I’m trying to write a priorbox layer in SSD for driveworks plugin, and the priorbox layer need to use the PriorBoxParameters from model.prototxt. However, in driveworks, the constructor of a plugin layer seems directly load from weights, as follows,

dwStatus _dwDNNPlugin_initializeFromWeights(_dwDNNPluginHandle_t* handle, const char*,
                                            const dwDNNPluginWeights* weights, int32_t numWeights)
{
    std::unique_ptr<FCPlugin> fcPlugin(new FCPlugin(weights, numWeights, 10));
    *handle = reinterpret_cast<_dwDNNPluginHandle_t>(fcPlugin.release());
    return DW_SUCCESS;
}

How can I pulgin the priorbox layer using this API ?

2

Dear huchenyang,
Did you check your model working well with TensorRT first? Including plugin layer code in DW is similar to including plugin layer in TensorRT. Please check TensorRT sample_plugin and dnn_plugin in DW as a reference.
_dwDNNPlugin_initialize() function can read plugin information byte by byte as shown in sample.

3

Yes, I just write the driveworks plugin class with reference to tensorRT.
However, the priorBOX layer is constructed by the deploy file under caffe model in tensorRT. The API of Driveworks need weights to construct the plugin layer. It’s totally different. PriorBox doesn’t need weights.
So I have no idea how to modify it from tensorRT plugin class to driveworks plugin class.

4

Dear huchnyang,
Can you keep the constructor with weights as an empty block and check.

5

Hi huchnyang,

Have you tried our suggestions to keep the constructor with weights as an empty block and check?
Any result can be shared?