libtorch jit forward error

[moodeerf]

Discussed in #154

Hi, I had encounted the same problem.
When I call the 'forward' function with one parameters (like this forward(inputs)), the code can be compiled correctly, but it can't run with the error 'what(): forward() is missing value for argument 'sr'.'.
However, When I call the 'forward' function with two parameters (like this forward(inputs,sr)), the code can't be compiled correctly with the error 'candidate expects 1 argument, 2 provided'.
I doubt that the jit model was not exported correctly, please check it at your convenience. By the way, I can call the 'forward' function with one parameters (like this forward(inputs)) with the old jit model.
On the other hand, the function 'reset_states' of the jit model can't be used in C++ code, can you provide a name so we can use it like this 'model.get_method('name')()' or 'model.run_method('name')'

[snakers4]

Hi,

Not sure that we will be opening old models, but we are planning a major update some time soon.
This may be a good time for debugging with libtorch as well.
Maybe this is an issue of libtorch itself, idk.

with the old jit model.

If remember correctly, the old jit models had a large chunk size, did not have an SR parameter, did not have internal state.

On the other hand, the function 'reset_states' of the jit model can't be used in C++ code, can you provide a name so we can use it like this 'model.get_method('name')()' or 'model.run_method('name')'

It is the case for all C++ models, only .forward() can be invoked, or is it a particular error with this model?

[moodeerf]

Hi,

Not sure that we will be opening old models, but we are planning a major update some time soon. This may be a good time for debugging with libtorch as well. Maybe this is an issue of libtorch itself, idk.

with the old jit model.

If remember correctly, the old jit models had a large chunk size, did not have an SR parameter, did not have internal state.

On the other hand, the function 'reset_states' of the jit model can't be used in C++ code, can you provide a name so we can use it like this 'model.get_method('name')()' or 'model.run_method('name')'

It is the case for all C++ models, only .forward() can be invoked, or is it a particular error with this model?

Em...,
Firstly I had used other jit models，.forward() can be invoked correctly , so most likely not the issue of libtorch itself, idk.
Secondly, If 'reset_states' is able to be invoked in C++ code, 'reset_states' need be defined in Python code follow the rules of torch.jit.export https://pytorch.org/docs/stable/jit.html
For example:
the function ' right_context' is defined in line:549 of the file
https://github.com/wenet-e2e/wenet/blob/main/wenet/transformer/asr_model.py
and the function can be invoked in the line:58 of the file https://github.com/wenet-e2e/wenet/blob/main/runtime/core/decoder/torch_asr_model.cc
So,the 'reset_states' should be defined like above.

[snakers4]

Firstly I had used other jit models，.forward() can be invoked correctly , so most likely not the issue of libtorch itself, idk.

With multiple input variables?

[moodeerf]

I will check it .

[moodeerf]

Firstly I had used other jit models，.forward() can be invoked correctly , so most likely not the issue of libtorch itself, idk.

With multiple input variables?
Hi,
Finally i solved the problem like this .
std::vector<torch::jit::IValue> inputs;
inputs.push_back(torch::from_blob(buffer,{chunk_size},torch::kFloat32));
int sr = 16000;
inputs.push_back(sr);
at::Tensor output = module.forward(inputs).toTensor();