change prepend_op to append_op in initializer

[zhiqiu]

PR types

Breaking changes

PR changes

Others

Describe

change prepend_op to append_op in initializer to make dygraph and static graph consistent

Given the following code, with the same random SEED=123, the same parameter initialized in static graph and dynamic graph is different.

import paddle
paddle.set_device('cpu')
SEED = 123

weight_attr=paddle.framework.ParamAttr(name="linear_weight", learning_rate=1.0,
        trainable=False, regularizer=None, initializer=paddle.nn.initializer.TruncatedNormal(mean=0.0, std=2.0))
bias_attr=paddle.framework.ParamAttr(name="linear_bias", learning_rate=1.0,
        trainable=False, regularizer=None, initializer=paddle.nn.initializer.TruncatedNormal(mean=0.0, std=2.0))

if True:
    print('Test dynamic graph')
    paddle.seed(SEED)
    linear = paddle.nn.Linear(1, 1, weight_attr=weight_attr, bias_attr=bias_attr)

    print('linear.weight: ', linear.weight.numpy())
    print('linear.bias: ', linear.bias.numpy())

if True:
    print('Test static graph')
    paddle.enable_static()

    exe = paddle.static.Executor(paddle.CPUPlace())
    paddle.seed(SEED)

    linear = paddle.nn.Linear(1, 1, weight_attr=weight_attr, bias_attr=bias_attr)

    res = exe.run(paddle.static.default_startup_program(), fetch_list=['linear_weight', 'linear_bias'])

    print('linear.weight: ', res[0])
    print('linear.bias: ', res[1])

• Before

Test dynamic graph
linear.weight:  [[1.594699]]
linear.bias:  [3.8483276]
Test static graph
linear.weight:  [[3.8483276]]
linear.bias:  [1.594699]

• After

Test dynamic graph
linear.weight:  [[1.594699]]
linear.bias:  [3.8483276]
Test static graph
linear.weight:  [[1.594699]]
linear.bias:  [3.8483276]

[paddle-bot-old]

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

[Aurelius84]

LGTM