fix(qat): Rescale input data for C++ application

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>

Author: narendasan

examples/int8/datasets/cifar10.cpp

@@ -50,7 +50,7 @@ std::pair<torch::Tensor, torch::Tensor> read_batch(const std::string& path) {
     labels.push_back(label);
     auto image_tensor =
         torch::from_blob(image.data(), {kImageChannels, kImageDim, kImageDim}, torch::TensorOptions().dtype(torch::kU8))
-            .to(torch::kF32);
+            .to(torch::kF32).div(255);
     images.push_back(image_tensor);
   }
 

examples/int8/ptq/main.cpp

@@ -140,4 +140,5 @@ int main(int argc, const char* argv[]) {
 
   auto trt_runtimes = benchmark_module(trt_mod, dims[0]);
   print_avg_std_dev("TRT quantized model", trt_runtimes, dims[0][0]);
+  trt_mod.save("/tmp/ptq_vgg16.trt.ts");
 }

examples/int8/qat/main.cpp

@@ -124,5 +124,6 @@ int main(int argc, const char* argv[]) {
 
   auto trt_runtimes = benchmark_module(trt_mod, dims[0]);
   print_avg_std_dev("TRT quantized model", trt_runtimes, dims[0][0]);
+  trt_mod.save("/tmp/qat_vgg16.trt.ts");
 }