Provide a few default args for numpy translation #20451
Conversation
pytorchbot
added
module: internals
umanwizard
added
the
module: numpy
|
Related to #2228 |
| static const std::unordered_map<std::string, std::vector<std::string>> numpy_compatibility_arg_names = { | ||
| {"dim", {"axis"}}, | ||
| {"keepdim", {"keepdims"}}, | ||
| {"input", {"x", "a", "x1"}}, |
There was a problem hiding this comment.
I'm a bit ambivalent about the input and other aliases. I'm not sure people use them as kwargs often enough to make it worthwhile. There's also a long tail of names that NumPy uses instead of "input", including "A", "arr", "ary", "m", "a1", "array".
There was a problem hiding this comment.
I'm a bit ambivalent...
I don't know how often people use them as kwargs. But I want to be as complete as possible.
There's also a long tail...
Yes, and I've also seen p, and who knows what else there is. I think we need to decide case-by-case, depending on how common they are, whether to include them in the default common args here, or annotate individual functions with alternate arg names (I will introduce a facility for doing that in a future PR).
There was a problem hiding this comment.
@umanwizard has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
|
@umanwizard merged this pull request in 72bb84c. |
Summary:
Add automatic translations for a few argument names that commonly differ between PyTorch and NumPy.
For now, they are as follows:
* `keepdim` -> `keepdims`
* `dim` -> `axis`
* `input` -> (any of `a`, `x`, `x1`)
* `other` -> `x2`
Basic examples:
```python
>>> t=torch.randn(10,10)
>>> torch.sum(x=t, axis=1)
tensor([ 0.5199, -0.3768, 4.3619, -0.9105, 1.1804, 1.0837, -0.9036, 0.2365,
1.1171, -0.0999])
```
```python
>>> torch.add(x1=5, x2=6)
tensor(11)
```
The additional overhead is zero when using traditional PyTorch argument names, and a few (usually 1) extra PyDict lookups when using NumPy argument names.
Pull Request resolved: pytorch#20451
Differential Revision: D15337521
Pulled By: umanwizard
fbshipit-source-id: 7a7d389786f4ccf5c86a14ecb2002c61730c51b5
Add automatic translations for a few argument names that commonly differ between PyTorch and NumPy.
For now, they are as follows:
keepdim->keepdimsdim->axisinput-> (any ofa,x,x1)other->x2Basic examples:
The additional overhead is zero when using traditional PyTorch argument names, and a few (usually 1) extra PyDict lookups when using NumPy argument names.