Manually set Variable data in C++

C++
1

I’m trying to do the equivalent of layer.weights.data = tensor using the PyTorch C++ frontend.

I’ve tried a few things:

*module.named_parameters()["layer"].data<float>() = *vector.data();

This segmentation faults.

module.named_parameters()["layer"].set_data(torch::from_blob(vector.data(), tensor_size));

This gives !is_variable() ASSERT FAILED.

module.named_parameters()["layer"] = torch::from_blob(vector.data(), tensor_size);

This apparently doesn’t do anything. The weights afterwards don’t change.

Is there a way to do this in C++?

1 Like
2

It’s a bit awkward, but I found you can do it with memcpy.

memcpy(module.named_parameters()["layer"].data<float>(), vector.data(), element_count);
3

Hey @Omegastick,

how about doing this instead

IntList sizes = net->named_parameters()["layer"].sizes();
torch::Tensor tensor = torch::from_blob(data.data(), sizes); // where data is an std::vector<float>

net->named_parameters()["layer"].set_data(tensor);
4

That gives:

terminate called after throwing an instance of 'c10::Error'
  what():  !is_variable() ASSERT FAILED at /home/px046/dev/pytorch/c10/core/TensorImpl.h:758, please report a bug to PyTorch. (set_sizes_and_strides at /home/px046/dev/pytorch/c10/core/TensorImpl.h:758)
5

I found an error in my original memcpy code. You need to multiply the number of elements by the size of the data type:

memcpy(module.named_parameters()["layer"].data<float>(), vector.data(), element_count * sizeof(float));
6

Hm, the memcpy approach I wrote above only works for small tensors. It works for 5x5 tensors but on 64x64 ones it stops working part-way through the tensor.

EDIT:
Okay, I found the proper way to do it. You need a NoGradGuard and Tensor::copy():

NoGradGuard guard;
module.named_parameters()["layer"].copy_(torch::from_blob(vector.data(), tensor_size));
1 Like
7

take care that your guard is only locally defined

8

Hey @Omegastick,

what I dont like about your solution is that it has the potential to break other peoples code. Your solution would at least require to put it in parenthesis, e.g.

{
    NoGradGuard guard;
    module.named_parameters()["layer"].copy_(torch::from_blob(vector.data(), tensor_size));
}

You could either do this or this, as mentioned above

torch::Tensor tensor = torch::from_blob(vector.data(), tensor_size);
net->named_parameters()["layer"].set_data(tensor);
9

that works for me! thanks!

10

Hi Martin,

Thanks for the suggested code. I just have a quick question, can from_blob create a constant tensor by taking a constant array as input? For example, given a constant reference of a float array data_array, and try to do the following:
``’
torch::Tensor tensor = torch::from_blob(data_array, tensor_size);

Is there any way to make from_blob (or any other alternative method) works in this case? Thanks!
1 Like
11

Hi @nkg114mc,

you may want to checkout the tensor basics

Cheers

Martin

12

Thank you very much for your suggestion, mhubii!

13

Maybe using const_cast?