• 标题：Momentum优化器的rescale_grad和grad_clip参数没生效
• 版本、环境信息：

1）PaddlePaddle版本：2.2.0
2）CPU：i5
3）GPU：3060 cuda 11.2
4）系统环境：windows10，Python 3.6

• 模型信息

convs2s

• 复现信息：运行下面的代码，分别修改clip_norm和rescale_grad，比较前后的参数、梯度、梯度范数。
• 问题描述：没有报错，我发现加和不加该参数，最后得到的参数和梯度没发生改变。他们在哪里起作用了呢？

`
import paddle
import numpy as np
import random
from paddle.fluid.dygraph.parallel import ParallelEnv
nranks = ParallelEnv().nranks
seed=1
np.random.seed(seed)
paddle.seed(seed)
random.seed(seed)
random.seed(seed)

def ClipGradByNorm(tensor, clip_norm=0.1):
''' 按tensor的l2范数缩放 '''
norm = ((tensor2).sum())(1 / 2)
c = clip_norm / max(norm, clip_norm)
tensor = tensor * c
return tensor

def calc_global_norm(tensor_groups):
l2_norm_fn = lambda tensor: ((tensor2).sum())(1 / 2)
tensor_norms = [l2_norm_fn(tensor) for tensor in tensor_groups]
global_norm = sum([norm**2 for norm in tensor_norms])**(1 / 2)
return global_norm

def ClipGradByGlobalNorm(tensor_groups, clip_norm=0.1):
''' 按所有tensor的l2范数缩放 '''
global_norm = calc_global_norm(tensor_groups)
c = clip_norm / max(global_norm, clip_norm)
tensor_groups = [tensor * c for tensor in tensor_groups]
return tensor_groups

bsz=256
epoch=10
##################### 分别修改下面两个 ######################
clip_norm=0.1
rescale_grad=1
###########################################################
model = paddle.nn.Linear(10, 10)
clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=clip_norm)
optim = paddle.optimizer.Momentum(learning_rate=0.1,parameters=model.parameters(),rescale_grad=rescale_grad,grad_clip=clip)
for i in range(5):
x = paddle.randn([bsz, 10])
out = model(x)
loss = paddle.mean(out)

back prop gradients

loss.backward()

update params by gradients

optim.step()
param_groups=optim._param_groups
grad_groups=[param.grad for param in param_groups]
print('param_groups',param_groups[0][0].numpy())
print('grad_groups',grad_groups[0][0].numpy())
global_norm=calc_global_norm(grad_groups)
print('global_norm',global_norm.numpy())
'''
clip_norm=0.1
param_groups [-0.5496709 -0.4565471 0.10905335 0.42701575 0.5106469 -0.34190077 0.01443232 -0.11369913 -0.26169622 0.26478162]
grad_groups [0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471]
*global_norm [1.5874481]

clip_norm=0.2
param_groups [-0.5516438 -0.45852 0.10708045 0.42504287 0.50867397 -0.34387365 0.01245942 -0.11567203 -0.2636691 0.26280874]
grad_groups [0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471 0.01216471]
*global_norm [1.5874481]

'''
`