SNPE can dynamically resize networks when they are loaded. Suppose that we have converted a model from Caffe2 with the following command:

snpe-caffe2-to-dlc -p predict_net.pb -e init_net.pb -i data 1,3,244,244 -d model.dlc

Then, we might use the following code to initialize the network to consume a batch of 1080p images:

#include <SNPE/SNPE.hpp>

auto container = zdl::DlContainer::IDlContainer::open("model.dlc");

zdl::DlSystem::TensorShapeMap inputShapeMap;
inputShapeMap.add("data", {3,1080,1440,3})

zdl::SNPE::SNPEBuilder builder(container.get());
builder.setInputDimensions(inputShapeMap);

auto snpe = builder.build();