Bilinear

class paddle.nn.initializer.Bilinear ( ) )

该接口为参数初始化函数,用于转置卷积函数中,对输入进行上采样。用户通过任意整型因子放大shape为(B,C,H,W)的特征图。

返回:对象

用法如下:

代码示例:

  1. import math
  2. import paddle
  3. import paddle.nn as nn
  4. from paddle.regularizer import L2Decay
  5. factor = 2
  6. C = 2
  7. B = 8
  8. H = W = 32
  9. w_attr = paddle.ParamAttr(learning_rate=0.,
  10. regularizer=L2Decay(0.),
  11. initializer=nn.initializer.Bilinear())
  12. data = paddle.rand([B, 3, H, W], dtype='float32')
  13. conv_up = nn.Conv2DTranspose(3,
  14. out_channels=C,
  15. kernel_size=2 * factor - factor % 2,
  16. padding=int(math.ceil((factor - 1) / 2.)),
  17. stride=factor,
  18. weight_attr=w_attr,
  19. bias_attr=False)
  20. x = conv_up(data)

上述代码实现的是将输入x(shape=[-1, 4, H, W])经过转置卷积得到shape=[-1, C, H*factor, W*factor]的输出,out_channels = C和groups = C 表示这是按通道转置的卷积函数,输出通道为C,转置卷积的groups为C。滤波器shape为(C,1,K,K),K为kernel_size。该初始化函数为滤波器的每个通道设置(K,K)插值核。输出特征图的最终输出shape为(B,C,factor*H,factor*W)。注意学习率和权重衰减设为0,以便在训练过程中双线性插值的系数值保持不变