add new API paddle.nn.initializer.Dirac

[zhwesky2010]

PR types

New features

PR changes

APIs

Describe

paddle.nn.initializer.Dirac

Initialize the 3D/4D/5D Tensor with Dirac delta function. It can reserve the feature of convolution layer input, which means that as many channels are reserved as possible.

attr = paddle.ParamAttr(initializer=paddle.nn.initializer.Dirac())
conv = paddle.nn.Conv1D(3, 2, 3, weight_attr=attr)
# conv.weight:
# Tensor(shape=[2, 3, 3], dtype=float32, place=CPUPlace, stop_gradient=False,
#       [[[0., 1., 0.],
#         [0., 0., 0.],
#         [0., 0., 0.]],
# 
#        [[0., 0., 0.],
#         [0., 1., 0.],
#         [0., 0., 0.]]])

input = paddle.rand([8, 3, 10])
output = conv(input)
input.shape  # [8, 3, 10]
output.shape # [8, 2, 8]
paddle.equal_all(output, input[:, 0:2, 1:9])
# [True]
# It means output is almost the same with input, 2 channels are reserved.

[paddle-bot-old]

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

[jeff41404]

lgtm

[TCChenlong]

LGTM

[TCChenlong]

will return 1.0？

[zhwesky2010]

This API returns a float number.