Created by: Aurelius84

PR types

New features

PR changes

APIs

Describe

support usage: to_static(model)

Code Example:

Before this PR，if we want decorate a forward() from model, we shall write code like（not decorator mode）:

model.forward = to_static(model.forward, input_spec=...)

It's not easy-to-use. In this PR，we support decorate model directly and convert call for foward and all inner function.

class SimpleNet(Layer):
    def __init__(self, ...):
        .......

    def forward(self, x):
        ......
        return out

net = SimpleNet()
x = paddle.to_tensor(np.ones(4, 10))

# case 1:  decorate model and feed data
net = to_static(net)
out = net(x)
print(out)

# case 2: decorate model with InputSpec
net = to_static(net, input_spec=[InputSpec([None, 8, 10])])
self.assertListEqual(list(input_shape), [-1, 8, 10])


# case 3: redecorate with new InputSpec
net = declarative(net, input_spec=[InputSpec([None, 16, 10])])
input_shape = net.forward.inputs[0].shape
self.assertListEqual(list(input_shape), [-1, 16, 10])