nn.Embedding/F.embedding support scale_grad_by_freq 易用性提升
#70090
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
| @@ -1714,6 +1714,8 @@ class Embedding(Layer): | |||
| The local word vector needs to be transformed into numpy format, and the shape of local word | |||
| vector should be consistent with :attr:`num_embeddings` . Then :ref:`api_paddle_nn_initializer_Assign` | |||
| is used to load custom or pre-trained word vectors. See code example for details. | |||
| scale_grad_by_freq(bool, optional): Indicating whether to scale the gradients by the inverse frequency of the | |||
There was a problem hiding this comment.
这个是新增参数，且是最后面的的，按道理直接修改原kernel不会不兼容吧
There was a problem hiding this comment.
原kernel的配置是动静不一致的，情况比较复杂，反向kernel要修改的话涉及很多个文件，我不确定会不会改错了。前几天试着改了一下但是最后在pir模式中 scale_grad_by_freq 没法正确传参。
paddle\phi\ops\yaml\inconsistent\dygraph_backward.yaml
paddle\phi\ops\yaml\inconsistent\static_backward.yaml
paddle\phi\ops\yaml\op_compat.yaml
There was a problem hiding this comment.
原kernel的配置是动静不一致的，情况比较复杂，反向kernel要修改的话涉及很多个文件，我不确定会不会改错了。前几天试着改了一下但是最后在pir模式中 scale_grad_by_freq 没法正确传参。
paddle\phi\ops\yaml\inconsistent\dygraph_backward.yaml
paddle\phi\ops\yaml\inconsistent\static_backward.yaml
paddle\phi\ops\yaml\op_compat.yaml
kernel命名为scale_grad_by_frea是不规范的
There was a problem hiding this comment.
https://github.com/PaddlePaddle/Paddle/pull/70124/files
我不知道是不是哪里配置的不对
There was a problem hiding this comment.
https://github.com/PaddlePaddle/Paddle/pull/70124/files 我不知道是不是哪里配置的不对
这个kernel的命名优化一下吧
PR Category
User Experience
PR Types
Improvements
Description
为避免不兼容修改，新增 embedding_with_scaled_gradient op，前向复用 embedding_kernel，反向新增 embedding_with_scaled_gradient_grad_kernel。