Bilinear

class paddle.nn.initializer. Bilinear [source]

This initializer can be used in transposed convolution operator to act as upsampling. Users can upsample a feature map with shape of (B, C, H, W) by any integer factor.

Returns: Bilinear initializer instance objects.

Examples

           >>> import math

>>> import paddle
>>> import paddle.nn as nn
>>> from paddle.regularizer import L2Decay

>>> factor = 2
>>> C = 2
>>> B = 8
>>> H = W = 32
>>> w_attr = paddle.ParamAttr(
...     learning_rate=0.0,
...     regularizer=L2Decay(0.0),
...     initializer=nn.initializer.Bilinear(),
... )
>>> data = paddle.rand([B, 3, H, W], dtype='float32')
>>> conv_up = nn.Conv2DTranspose(
...     3,
...     out_channels=C,
...     kernel_size=2 * factor - factor % 2,
...     padding=int(math.ceil((factor - 1) / 2.0)),
...     stride=factor,
...     weight_attr=w_attr,
...     bias_attr=False,
... )
>>> x = conv_up(data)

          

Where, out_channels=C and groups=C means this is channel-wise transposed convolution. The filter shape will be (C, 1, K, K) where K is kernel_size, This initializer will set a (K, K) interpolation kernel for every channel of the filter identically. The resulting shape of the output feature map will be (B, C, factor * H, factor * W). Note that the learning rate and the weight decay are set to 0 in order to keep coefficient values of bilinear interpolation unchanged during training.

forward ( var: Tensor, block: Optional[Block] = None ) → paddle.Tensor | None forward¶

Initialize the input tensor with Bilinear initialization.

Parameters

var (Tensor) – Tensor that needs to be initialized.
block (Block|None, optional) – The block in which initialization ops should be added. Used in static graph only, default None.

Returns

The initialization op