Implementation of getSerializationSize and serialize

[LinHungShi]

I am trying to write the engine to a plain file. It seems two related functions haven't been implemented yet. Is there any tutorial demonstrating how to implement these two functions?

[Alexey-Kamenev]

Not sure I understand the question. Are you asking in general how serialization should be implemented for plugins in TensorRT? Or specifically in Stereo DNN plugins? If former - then please refer to TensorRT documentation and samples. If latter - take a look at some of our plugins, for example this.

[LinHungShi]

Hi @Alexey-Kamenev thanks for your reply. Since there're no weights in ELU layer, I am wondering how to serialize weight. To be more specific, do all layer variables need serialization? For example, there are dependencies between kernel_weights and kernel_weights_d_ in Conv3DPlugin, is it necessary to serialize both of them?

[Alexey-Kamenev]

In general, you should serialize all variables which are required to restore the plugin state correctly during deserialization (i.e. when deserialization ctor is called). That includes weights and biases in case of conv3d plugin. kernel_weights_d_ is just a device copy of kernel_weights_ so you only need to serialize one of them - kernel_weights_. Some variables which can be re-created do not require serialization, like various descriptors, auxiliary tensors etc.

[LinHungShi]

Hi, your answer really saves my time. Do you mean I just need to serialize those important variables and recreate less important variables during deserialization?