Can not load scripted model using torch::jit::load

1

I use torch.jit.script to save model and using C++ to load it：

#include <torch/script.h> 
using namespace std;

int main() {
    std::string model_path = "G:/modelscriptcuda.pt";
    
   try {
    torch::jit::script::Module module = torch::jit::load(model_path, torch::kCUDA);
} catch (const c10::Error& e) {
    std::cerr << "error loading the model: " << e.what() << std::endl;
    return -1;
}}

And I got this error：

error loading the model: [enforce fail at ..\..\caffe2\serialize\inline_container.cc:197] . file not found: modelscriptcuda/version
(no backtrace available)
2

How did you save the checkpoint? Based on the error message I would guess your checkpoint might be in an older format as the version file is missing in the archive.

3

I use torch.save to save my model and I trace it through:
torch.save(model.state_dict(), save_path)

` model.load_state_dict(torch.load(‘G:/chapter5nomissing.pt’))
model.eval()
model = model.float()

example_input = torch.rand(1,6, 850)
example_rawdata = torch.rand(1,6,1000)
example_index = torch.rand(1,2,30)
example_mask = torch.rand(1,1,6,1000)
print(example_input.type(),example_rawdata.type(),example_index.type(),example_mask.type())

traced_script_module = torch.jit.script(model, (example_input,example_rawdata,example_mask) )

traced_script_module.save('G:/modelscriptcuda.pt')
print("save")
load_scripted_model = torch.jit.load('G:/modelscriptcuda.pt')
print(load_scripted_model)`

And I found it is okay to load it in python

4

Which PyTorch and libtorch versions are you using as it still seems they mismatch due to the unexpected checkpoint format?

5

torch == 1.11.0+cu113
– Found torch: I:/Project/CNN-1DMEAE/libtorch/lib/torch.lib
– Found Torch: 1.4.0

6

Update both to the same version and rerun your code again.