pytorch

4ab73c1f - [docs] Fix EmbeddingBag docs (#45763)

[docs] Fix EmbeddingBag docs (#45763) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/45763 **Summary** This commit updates the documentation for `EmbeddingBag` to say that for bags of constant length with no per-sample weights, the class is equivalent to `Embedding` followed by `torch.sum(dim=1)`. The current docs say `dim=0` and this is readily falsifiable. **Test Plan** 1) Tried `Embedding` + `sum` with `dim`=0,1 in interpreter and compared to `EmbeddingBag` ``` >>> import torch >>> weights = torch.nn.Parameter(torch.randn(10, 3)) >>> e = torch.nn.Embedding(10, 3) >>> eb = torch.nn.EmbeddingBag(10, 3, mode="sum") >>> e.weight = weights >>> eb.weight = weights # Use 2D inputs because we are trying to test the case in which bags have constant length >>> inputs = torch.LongTensor([[4,1,2,7],[5,6,0,3]]) >>> eb(inputs) tensor([[-2.5497, -0.1556, -0.5166], [ 2.2528, -0.3627, 2.5822]], grad_fn=<EmbeddingBagBackward>) >>> torch.sum(e(inputs), dim=0) tensor([[ 1.6181, -0.8739, 0.8168], [ 0.0295, 2.3274, 1.2558], [-0.7958, -0.4228, 0.5961], [-1.1487, -1.5490, -0.6031]], grad_fn=<SumBackward1>) >>> torch.sum(e(inputs), dim=1) tensor([[-2.5497, -0.1556, -0.5166], [ 2.2528, -0.3627, 2.5822]], grad_fn=<SumBackward1>) ``` So clearly `torch.sum` with `dim=0` is not correct here. 2) Built docs and viewed in browser. *Before* <img width="882" alt="Captura de Pantalla 2020-10-02 a la(s) 12 26 20 p m" src="https://user-images.githubusercontent.com/4392003/94963035-557be100-04ac-11eb-986c-088965ac3050.png"> *After* <img width="901" alt="Captura de Pantalla 2020-10-05 a la(s) 11 26 51 a m" src="https://user-images.githubusercontent.com/4392003/95117732-ea294d80-06fd-11eb-9d6b-9b4e6c805cd0.png"> **Fixes** This commit closes #43197. Test Plan: Imported from OSS Reviewed By: ansley Differential Revision: D24118206 Pulled By: SplitInfinity fbshipit-source-id: cd0d6b5db33e415d8e04ba04f2c7074dcecf3eee