returnn.frontend._utils#

Internal utils

returnn.frontend._utils.get_backend_from_tensors(*args)[source]#
Parameters:

args –

Returns:

frontend, fallback to global frontend

returnn.frontend._utils.get_dtype_name(x: T | Tensor[T] | int | float) → str[source]#
Parameters:

x – tensor

Returns:

dtype of tensor, as string

returnn.frontend._utils.is_int(x: T | Tensor[T] | int | float) → bool[source]#
Parameters:

x –

Returns:

whether the dtype is int

returnn.frontend._utils.bin_op_out_template(backend: Type[Backend], a: Tensor[T] | int | float | number, b: Tensor[T] | int | float | number, *, name: str, copy_sparse_dim: bool = True, allow_broadcast_all_sources: bool | None = None, dim_order: Sequence[Dim] | None = None, allow_scalar: bool = True) → Tuple[Tensor[T], T, T][source]#

make template for output tensor of binary op

Parameters:

  • backend –

  • a –

  • b –

  • name – for returned Tensor. no other functionality

  • copy_sparse_dim –

  • allow_broadcast_all_sources – if True, it is allowed that neither a nor b has all dims of the result. Not needed when out_dims is specified explicitly.

  • dim_order – defines the order of the resulting dims. if None, it is automatically inferred from a and b. Not all the dims of a and b need to be specified here, and there could also be other dims in the dim_order.

  • allow_scalar – if True, it is allowed that a or b is a scalar, and then no broadcast dims are added. This can be relevant to allow things like x * 2, where x in on GPU, and then PyTorch allows 2 to stay on CPU.

Returns:

out, a_raw, b_raw

returnn.frontend._utils.res_feature_dim(a: Tensor, b: Tensor) → Dim | None[source]#
Parameters:

  • a –

  • b –

Returns:

feature dim if consistent or None

returnn.frontend._utils.res_sparse_dim(a: Tensor, b: Tensor) → Dim | None[source]#
Parameters:

  • a –

  • b –

Returns:

sparse dim if consistent or None

returnn.frontend._utils.strided_slice_raw_key(tensor: Tensor, axis: Dim | Sequence[Dim] | None = None, key: int | float | complex | number | ndarray | bool | str | Tensor | slice | Sequence[int | float | complex | number | ndarray | bool | str | Tensor | slice] | None = None, key_dim: Dim | Sequence[Dim] | None = None) → Tuple[slice | int | T | Sequence[None | slice | int | T], Tuple[Dim, ...]][source]#

Given an axis and a key, return a raw key that can be used to index into a raw tensor, as in raw_tensor[key]. The tensor is needed to infer the raw axis index.

Returns:

raw item key, resulting dims of reduced tensor