Tensor API

Overview

FTorch provides a torch_tensor derived type, which exposes the functionality of the torch::Tensor C++ class. The interface is designed to be familiar to Fortran programmers, whilst retaining strong similarity with torch::Tensor and the torch.Tensor Python class.

Under the hood, the torch_tensor type holds a pointer to a torch::Tensor object in C++ (implemented using c_ptr from the iso_c_binding intrinsic module). This allows us to avoid unnecessary data copies between C++ and Fortran.

Procedures

Constructors

We provide several subroutines for constructing torch_tensor objects. These include:

• torch_tensor_empty, which allocates memory for the torch_tensor, but does not set any values.
• torch_tensor_zeros, which creates a torch_tensor whose values are uniformly zero.
• torch_tensor_ones, which creates a torch_tensor whose values are uniformly one.
• torch_tensor_from_array, which allows the user to create a torch_tensor with the same rank, shape, and data type as a given Fortran array. Note that the data is not copied - the tensor data points to the Fortran array, meaning the array must have been declared with the target property. The array will continue to be pointed to even when operations are applied to the tensor, so this subroutine can be used 'in advance' to set up an array for outputting data. torch_tensor_from_array may be called with or without the layout argument - an array which specifies the order in which indices should be looped over. The default layout is [1,2,...,n] implies that data will be read into the same indices by Torch. (See the transposing user guide page for more details.

It is compulsory to call one of the constructors before interacting with it in any of the ways described in the following. Each of the constructors sets the pointer attribute of the torch_tensor; without this being set, most of the other operations are meaningless.

Tensor interrogation

We provide several subroutines for interrogating torch_tensor objects. These include:

• torch_tensor_get_rank, which determines the rank (i.e., dimensionality) of the tensor.
• torch_tensor_get_shape, which determines the shape (i.e., extent in each dimension) of the tensor.
• torch_tensor_get_dtype, which determines the data type of the tensor in terms of the enums torch_kInt8, torch_kFloat32, etc.
• torch_tensor_get_device_type, which determines the device type that the tensor resides on in terms of the enums torch_kCPU, torch_kCUDA, torch_kXPU, etc.
• torch_tensor_get_device_index, which determines the index of the device that the tensor resides on as an integer. For a CPU device, this index should be set to -1 (the default). For GPU devices, the index should be non-negative (defaulting to 0).

Procedures for interrogation are implemented as methods as well as stand-alone procedures. For example, tensor%get_rank can be used in place of torch_tensor_get_rank, omitting the first argument (which would be the tensor itself). The naming pattern is similar for the other methods (simply drop the preceding torch_tensor_).

Tensor deallocation

We provide a subroutine for deallocating the memory associated with a torch_tensor object: torch_tensor_delete. An interface is provided such that this can also be applied to arrays of tensors. Calling this subroutine manually is optional as it is called as a destructor when the torch_tensor goes out of scope anyway.

Tensor manipulation

We provide the following subroutines for manipulating the data values associated with a torch_tensor object:

• torch_tensor_zero (aliased as class method torch_tensor%zero), which sets all the data entries associated with a tensor to zero.

Operator overloading

Mathematical operators involving Tensors are overloaded, so that we can compute expressions involving outputs from one or more ML models.

Whilst it's possible to import such functionality with a bare

use ftorch

statement, the best practice is to import specifically the operators that you wish to use. Note that the assignment operator = has a slightly different notation:

use ftorch, only: assignment(=), operator(+), operator(-), operator(*), &
  operator(/), operator(**)

Overloaded assignment operator

Particular care should be taken with the overloaded assignment operator. Whenever you execute code involving torch_tensors on each side of an equals sign, the overloaded assignment operator should be triggered. As such, if you aren't using the bare use ftorch import then you should ensure you specify use ftorch, only: assignment(=) (as well as any other module members you require).

For a straightforward assignment of two torch_tensors a and b,

b = a

the overloaded assignment operator is called once.

For overloaded operators the situation is more complex. Consider the overloaded addition operator (the same applies for the rest). When we execute the line

c = a + b

the addition is evaluated first. It is implemented as a Fortran function and its return value is an intermediate tensor. The setup is such that this is created using torch_tensor_empty under the hood (inheriting all the properties of the tensors being added¹). Following this, the intermediate tensor is assigned to c. Finally, the finalizer for torch_tensor is called for the intermediate tensor because it goes out of scope.

Similarly as above, in the case where you have some function func that returns a torch_tensor, an intermediate torch_tensor will be created, assigned, and destroyed because the call will have the form

a = func()

Other operators acting on tensors

We have also exposed the operators for taking the sum or mean over the entries in a tensor, which can be achieved with the subroutines torch_tensor_sum and in torch_tensor_mean, respectively.

Examples

For a concrete example of how to construct, interrogate, manipulate, and delete Torch tensors, see the tensor manipulation worked example.

For an example of how to compute mathematical expressions involving Torch tensors, see the autograd worked example.