disable __setitem__ in static mode & add API paddle.static.setitem with dy2st strategy

[zoooo0820]

PR types

Breaking changes

PR changes

APIs

Description

Pcard-66985

Motivation

Currently, the __setitem__ in static mode is implemented by directly overwriting the result on the original Variable. This leads to although the forward result is correct, the calculation result is wrong in the backward stage （some are because result relies on the initial value of this Variable; some are because the graph in backward constructed according to this also has the problem of overwriting ).

# Dygraph is right
>>> a = paddle.full((2,3,4), 1.2, dtype='float32')
>>> a.stop_gradient=False
>>> v = paddle.ones((1,), dtype='float32')
>>> v.stop_gradient=False
>>> b = a * 2
>>> b[0,:, 2] = b[0, :, 2] * v
>>> l = b.mean()
>>> l.backward()
>>> b.sum()
Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=False,
       57.60000229)
>>> a.grad.sum()
Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=False,
       2.)
>>> v.grad
Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=False,
       [0.30000001])

# In static mode, the forward b is right, but grad of a and v are not
>>> paddle.enable_static()
>>> exe = paddle.static.Executor()
>>>
>>> train_program = paddle.static.Program()
>>> startup_program = paddle.static.Program()
>>> with paddle.static.program_guard(train_program, startup_program):
...     a = paddle.full((2,3,4), 1.2, dtype='float32')
...     a.stop_gradient=False
...     v = paddle.ones((1,), dtype='float32')
...     v.stop_gradient=False
...     b = a * 2
...     b[0, :, 2] = b[0, :, 2] * v
...     l = b.mean()
...     paddle.static.append_backward(l)
...     res = exe.run(fetch_list=[b, a.grad_name, v.grad_name])
...     print('b.sum() =', res[0].sum())
...     print('a.grad.sum() = ', res[1].sum())
...     print('v.grad.sum() = ', res[2].sum())
...
b.sum() = 57.600002
a.grad.sum() =  1.75
v.grad.sum() =  0.0

The essence of this problem is that the method of directly rewriting variables in static graphs violates the static single assignment form (SSA).

What this PR changed

• In Dynamic Mode:
  The behavior of __setitem__ is unchanged. Users can still write code like: x[0] = 10.

• In Static Mode
  __setitem__ is disabled. Users should use x = paddle.static.setitem(x, 0, 10) to replace x[0] = 10.
  API paddle.static.setitem will provide the variable assignment with indexing in static mode, but unlike __setitem__, it will return a new variable instead of overwriting the original variable.

>>> paddle.enable_static()
>>> a = paddle.ones((2,3,4))
>>> a[0] = 10
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.7/dist-packages/paddle/fluid/framework.py", line 2303, in __setitem__
    "In static mode, the __setitem__ (looks like: x[indices] = values) should not be used. Please use x = paddle.static.setitem(x, indices, values)"
RuntimeError: In static mode, the __setitem__ (looks like: x[indices] = values) should not be used. Please use x = paddle.static.setitem(x, indices, values)

• In Dynamic-to-Static (with @paddle.jit.to_static )
  As Fix setitem error in static mode #55282 , introducing a ParametersMap, if there are a new variable mapped to an old variable, the old one will be replaced by the new one.
  Users are not aware of this change in d2s, with just still using paddle.jit.to_static.

[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.

[paddle-bot]

❌ The PR is not created using PR's template. You can refer to this Demo.
Please use PR's template, it helps save our maintainers' time so that more developers get helped.

[paddle-ci-bot]

Sorry to inform you that 788f8ac's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[Aurelius84]

Suggested change

	return paddle.fluid.framework._setitem_impl_(self.obj, item, value)
	new_var = paddle.fluid.framework._setitem_impl_(self.obj, item, value)
	# NOTE(dev): Update __dict__ will not modify the id(), but only move the
	# pointed reference object to the new one.
	self.obj.__dict__.update(new_var.__dict__)

[Aurelius84]

LGTM

[jeff41404]

LGTM

[XiaoguangHu01]

LGTM