Created by: chenwhql

PR types

Function optimization

PR changes

APIs

Describe

Use structured name as key in loaded state_dict, instead use parameter.name

original example:

import paddle
import paddle.nn as nn

IMAGE_SIZE = 784
CLASS_NUM = 10

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)

    @paddle.jit.to_static
    def forward(self, x):
        return self._linear(x)

# enable dygraph mode
paddle.disable_static() 

# create network
layer = LinearNet()

# load
model_path = "linear.example.model"
state_dict, _ = paddle.load(model_path)

# inference
layer.set_state_dict(state_dict, use_structured_name=False)
layer.eval()
x = paddle.randn([1, IMAGE_SIZE], 'float32')
pred = layer(x)

new example change:

# inference
layer.set_state_dict(state_dict)
layer.eval()
x = paddle.randn([1, IMAGE_SIZE], 'float32')
pred = layer(x)

Pay attention to the use of Layer.set_state_dict, no need to configure use_structured_name=False