[API Compatibility] Improve compat.nn.Linear init and param reset

[Enigmatisms]

PR Category

Operator Mechanism

PR Types

Improvements

Description

优化了

• [API Compatibility] Add paddle.compat.nn.Linear #76169

中引入的 compat.nn.Linear 的初始化和 reset 函数。reset 函数中，nn.Initializer.Uniform 与 nn.Initializer.KaimingUniform 的行为都被修改:

• 当 nn.initializer.XXX 被直接实例化然后用 forward 形式初始化参数时，原始的行为如下代码块所示：

>>> a = paddle.nn.Linear(3, 4).to('cpu')
>>> a
Linear(in_features=3, out_features=4, dtype=None)
>>> a.weight
Parameter containing:
Tensor(shape=[3, 4], dtype=float32, place=Place(cpu), stop_gradient=False,
       [[-0.45837218, -0.32479179, -0.47977936,  0.04010978],
        [-0.36938468,  0.12928823,  0.82120913, -0.60326314],
        [ 0.89723587, -0.30610302,  0.82338744, -0.04669143]])
>>> init = paddle.nn.initializer.XavierNormal()
>>> init(a.weight)
>>> a.weight
Parameter containing:
Tensor(shape=[3, 4], dtype=float32, place=Place(gpu:0), stop_gradient=False,
       [[ 0.06874870, -0.84757495,  1.06253850, -0.43288031],
        [ 0.54353887, -0.06623430, -0.98724484, -0.62573409],
        [ 0.52956545,  0.14853251, -0.45487842, -0.41272211]])

可以看到，place 从 cpu 变成了 gpu:0。这是因为 initializer 在动态图执行时，传入的 place 是 _current_expected_place()，而不是输入 tensor 的 place。而如果强行将所有 initializer 对应位置修改为 var.place 也会引起问题：如果 initializer 实例传入 ParamAttr 中，ParamAttr 辅助创建 parameters 时，tensor 的 place 是 undefined:0，此时必须使用 _current_expected_place()。

本 PR 只修改了与 compat.nn.Linear 有关的 initializer（kaiming_uniform 以及 uniform），其他 initializer 行为未变（个人觉得原始行为是不合理的，很显然，如果确实不合理，对应的initializer都需要修改，则需要一个新的PR完成对所有的 initializer 行为的‘修正’）。基于此，修正了 compat.nn.Linear 的初始化行为：不再手动进行 place 一致性检查以及 to 操作。

Pcard-89620

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zhwesky2010]

LGTM

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@94faddc). Learn more about missing BASE report.

Additional details and impacted files

@@             Coverage Diff             @@
##             develop    #76196   +/-   ##
===========================================
  Coverage           ?   100.00%           
===========================================
  Files              ?         1           
  Lines              ?         2           
  Branches           ?         0           
===========================================
  Hits               ?         2           
  Misses             ?         0           
  Partials           ?         0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.