Converting std::vector<Example> into an input tensor in C++

C++
1

TL;DR: What’s the best way in C++ to convert the batch-size-length list of std::vector returned by a dataloader into a tensor, where the first dimension is batch_size and the remaining dimensions are the inputs in the Example.data ?

I have a numerical dataset for learning sequence, where the input dataset and output is a Tensor of size N x SEQLEN x INPUT, and the output is N x SEQLEN x OUTPUT. In Python, I use a TensorDataset, passing both 3D tensors, without issue. However, it appears TensorDataset in C++ doesn’t (yet?) support a labelled dataset (libtorch 1.5).

So I created as custom dataset class and tried to use it with a dataloader (for batching, random, shuffling examples, etc…):

class CustomDataset : public torch::data::datasets::Dataset<CustomDataset> {
	using Example = torch::data::Example<torch::Tensor, torch::Tensor>;

	torch::Tensor X, Y;

	public:
		CustomDataset(const torch::Tensor& Xs, const torch::Tensor& Ys) : X(Xs), Y(Ys) {}
		Example get(size_t index) {
			return {X[index], Y[index]};
		}
		torch::optional<size_t> size() const {
    			return X.size(0);
  		}
};

However, when I create a dataloader using this dataset, in my training loop I get a std::vector. My network’s forward function expects a tensor of shape BATCH_SIZE x SEQLEN x INPUT.

What is the best way to convert this list of Examples into a tensor of the right shape? Or is there a better way to do this?

2

I’m able to get past this issue using this construct to manually turn the “list of examples” back into a new 3D tensor… but this seems MASSIVELY inefficient. Surely there’s a better way.

    		for (auto& batch : *data_loader) {
			    std::vector<torch::Tensor> Xs_list, Ys_list;
		        std::transform(batch.begin(), batch.end(), std::back_inserter(Xs_list), [](auto &e) {return e.data;});
		        std::transform(batch.begin(), batch.end(), std::back_inserter(Ys_list), [](auto &e) {return e.target;});
			    torch::Tensor X = torch::stack(Xs_list);
			    torch::Tensor Y = torch::stack(Ys_list);
      			optimizer.zero_grad();
      			torch::Tensor prediction = model->forward(X);	
      			torch::Tensor loss = torch::mse_loss(prediction, Y);
....

Edit: previous version of this post tried to use torch::from_blob(list.data(), ..., and was wrong. Changing that to torch::stack() seems to work a lot better.

3

How did you resolve this?
I have a tensor that returns from a module without a batch dimension:

torch::Tensor out_tensor = module.forward({input_tensor}).toTensor();
However its is 3 by 512 by 512 and this fails:
out_tensor = out_tensor.squeeze().detach().permute({1, 2, 0});

How can I add the batch dimension?

Thanks