Print network architecture in cpp jit

jit
1

Hi,

In python, we could use print (model) to show the network architecture, how about in cpp?

Say i hv loaded jit model with jit load with the name model, how could i do the similar in cpp.

Thanks,
Rgds,
CL

2

We don’t have any built-in utils for this, but you can do a simple version manually:

#include <torch/script.h>

#include <iostream>
#include <memory>

void tabs(size_t num) {
  for (size_t i = 0; i < num; i++) {
    std::cout << "\t";
  }
}

void print_modules(const torch::jit::script::Module& module, size_t level = 0) {
  std::cout << module.name().qualifiedName() << " (\n";
  for (const auto& module : module.get_modules()) {
    tabs(level + 1);
    print_modules(module.module, level + 1);
  }
  tabs(level);
  std::cout << ")\n";
}

int main(int argc, const char *argv[]) {
  torch::jit::script::Module container = torch::jit::load("m.pt");
  print_modules(container);
  return 0;
}
3

Hi Driazati,

Thanks for the codes, it helps. just some typo i think, the

print_modules(module.module, level + 1);

should be

print_modules(module, level + 1);

Appreciate your help. The references for these seems to be not much, could you share some good reference for this ? For e.g., the method (or the var) that can return the input size of a pre-trained network etc.

Thanks again.

Regards,
CL

4

You can find the generated docs here https://pytorch.org/cppdocs/api/namespace_torch__jit.html#namespace-torch-jit. Unfortunately they are pretty sparse, we have ongoing efforts to improve them.

As for getting the input size a network expects, we don’t have any utilities for that since TorchScript models aren’t specialized to specific shapes.

5

hi,

thanks again for the reply and link. yes i’ve visited the link before, might be too sparse for non hardcore programmer. anyway, thanks for your reply again.

rgds,
CL

6

Hi,

Hope you don’t mind, I created another topic to get more details from the jit model. I understand that there are no shape properties for the input, I was wondering whether we could print the each layers details as stated in :

1 Like
7

hi,

adding these lines

  for (const auto& parameter : module.get_parameters()) {
	tabs(level + 1);
	std::cout << parameter.name() << '\t';
	std::cout << parameter.value().toTensor().sizes() << '\n';
  }

shall give the summary for the network.

if anyone having better way to show the network summary for jit model in C++, please feel free to share.

Thanks.

rgds,
CL

1 Like
8

does anybody know how to do this using libtorch (iOs)? module.get_parameters() is not defined.