📚 The doc issue
The doc of torch.gradient() shows its description as below:
|
edge_order (``int``, optional): 1 or 2, for `first-order |
|
<https://www.ams.org/journals/mcom/1988-51-184/S0025-5718-1988-0935077-0/S0025-5718-1988-0935077-0.pdf>`_ or |
|
`second-order <https://www.ams.org/journals/mcom/1988-51-184/S0025-5718-1988-0935077-0/S0025-5718-1988-0935077-0.pdf>`_ |
|
estimation of the boundary ("edge") values, respectively. |
For some limitations that usually happen below, which should be noted in docs:
|
if (dim.has_value()) { |
|
// The following function get called to check whether dim argument satisfies prerequisites. |
|
// The output of the function is not used for the computation of gradient. |
|
dim_list_to_bitset(dim.value(), self.dim()); |
|
for (const auto i : c10::irange(dim.value().size())) { |
|
TORCH_CHECK(self.size(dim.value()[i]) >= edge_order + 1, "torch.gradient expected each dimension size to be at least edge_order+1"); |
|
} |
|
} else { |
|
for (const auto i : c10::irange(self.dim())) { |
|
TORCH_CHECK(self.size(i) >= edge_order + 1, "torch.gradient expected each dimension size to be at least edge_order+1"); |
|
} |
|
} |
Here is a repro:
Repro
import torch
input_data = torch.randn(10, 2, 32, 32)
gradient = torch.gradient(input_data, spacing=1, dim=None, edge_order=2)
Output
RuntimeError: torch.gradient expected each dimension size to be at least edge_order+1
Thanks for noting!
Suggest a potential alternative/fix
- Fix the Doc of the description of
edge_order
