hi,
I want to use tensorRT to accelerate my model. I convert tensorflow pb file to uff file, and there is a option named Conv2DBackpropInput which uff parser can not process. I use plugin layer to implement it. The option has two inputs, one is the previous option’s result, and another one is the const weights, but I can not get the values of the const weights.
As showed below, “conv2d_transpose” is the option I use plugin layer to implement, “up4/up_filter” is the const weights,
“conv5/conv5_3/LeakRelu” is the previous option’s result. I can get values of “conv5/conv5_3/LeakRelu”, but can not get values of inputs: “up4/up_filter”. Can you help me, thanks.
hi, NVES,
The “conv2d_transpose” is a plugin layer I want to implement, it has two inputs: “up4/up_filter” and “conv5/conv5_3/LeakRelu”. I can get shape information of “conv5/conv5_3/LeakRelu” through API getOutputDimensions, but I can not get any information of “up4/up_filter”. “up4/up_filter” is a Const option, which I think could be treated as weights, but I can not get it.
configureWithFormat function in the plugin contains the input dims as one of the parameters so you can access the weight information from there.
If you are using the nvuffparser::IPluginFactoryExt with plugins, then the parser will call createPlugin() with the weights field. However in this case the parser assumes that the plugin inputs are ordered such that all const weight values are towards the end of the input list.