[Dy2Stat] Fix Error when generating train_program in eval mode

[Aurelius84]

PR types

Function optimization

PR changes

Others

Describe

Fix Error when generating train_program in eval mode. See following example code:

class Net(nn.Layer):
    def __init__(self, in_channels, hidden_size):
        super(Net, self).__init__()
        self.lstm = nn.LSTM(
            in_channels, hidden_size, direction='bidirectional', num_layers=2)

    @paddle.jit.to_static
    def forward(self, x):
        x, _ = self.lstm(x)
        return x

net = Net(12, 2)
net.eval() # switch into eval mode firstly
# Here, we will get infer_program defaultly after converting dygraph layers into static.
# When initializing ParitalLayer, a train_program will be generated by `append_backward`
# from converted program(here is infer_program) with `is_test=True`, which will raise a error.
net = paddle.jit.to_static(
    net, input_spec=[paddle.static.InputSpec(shape=[-1, 10, 12])])
paddle.jit.save(net, 'simple_lstm')

Solution:

1. Add change_is_test_status to make sure all is_test=True before append_backward
2. Introduce lazy initialization mechanism to create program while only we need it.

[paddle-bot-old]

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

[zhhsplendid]

LGTM