Fix Layer train eval setting failed in static mode

[chenwhql]

PR types

Bug fixes

PR changes

APIs

Describe

Fix Layer train eval setting failed in static mode

By our design, paddlle.nn.Layer also can be used in static graph mode, but Layer.train() and Layer.eval() can only be set in dygraph mode, if set in static mode, it willl throw like:

Traceback (most recent call last):
  File "train_mode.py", line 5, in <module>
    l.train()
  File "/usr/local/python2.7.15/lib/python2.7/site-packages/paddle/fluid/dygraph/layers.py", line 135, in train
    framework._dygraph_tracer().train_mode()
AttributeError: 'NoneType' object has no attribute 'train_mode'

related issue: #29534

This may cause some problem. For example, the behavior of paddle.nn.Dropoutin Layer.train() and Layer.eval() are diffferent, the 2.0 paddle.nn.Dropout implement as:

def forward(self, input):
        out = F.dropout(
            input,
            p=self.p,
            axis=self.axis,
            training=self.training,
            mode=self.mode,
            name=self.name)
        return out

It need self.training as argument of original dropout, if user want use paddle.nn.Dropout for test in static mode, it need to set Layer.eval() firstly.

So this PR fix this problem, only set framework._dygraph_tracer().train_mode() in dygraph mode.

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zhiqiu]

Actually, _dygraph_tracer().train_mode() can be removed.

[chenwhql]

agree, and need to revert dropout(is_test=None) to dropout(is_test=False) changed in #29064

[chenwhql]

for compatibility, removing it in next PR is better, this PR need to be chery-pick into 2.0 to fix this bug