tf.nn.atrous_conv2d_transpose(
value,
filters,
output_shape,
rate,
padding,
name=None
)
Defined in tensorflow/python/ops/nn_ops.py
.
The transpose of atrous_conv2d
.
This operation is sometimes called "deconvolution" after Deconvolutional
Networks, but is
actually the transpose (gradient) of atrous_conv2d
rather than an actual
deconvolution.
Args:
value
: A 4-DTensor
of typefloat
. It needs to be in the defaultNHWC
format. Its shape is[batch, in_height, in_width, in_channels]
.filters
: A 4-DTensor
with the same type asvalue
and shape[filter_height, filter_width, out_channels, in_channels]
.filters
'in_channels
dimension must match that ofvalue
. Atrous convolution is equivalent to standard convolution with upsampled filters with effective heightfilter_height + (filter_height - 1) * (rate - 1)
and effective widthfilter_width + (filter_width - 1) * (rate - 1)
, produced by insertingrate - 1
zeros along consecutive elements across thefilters
' spatial dimensions.output_shape
: A 1-DTensor
of shape representing the output shape of the deconvolution op.rate
: A positive int32. The stride with which we sample input values across theheight
andwidth
dimensions. Equivalently, the rate by which we upsample the filter values by inserting zeros across theheight
andwidth
dimensions. In the literature, the same parameter is sometimes calledinput stride
ordilation
.padding
: A string, either'VALID'
or'SAME'
. The padding algorithm.name
: Optional name for the returned tensor.
Returns:
A Tensor
with the same type as value
.
Raises:
ValueError
: If input/output depth does not matchfilters
' shape, or if padding is other than'VALID'
or'SAME'
, or if therate
is less than one, or if the output_shape is not a tensor with 4 elements.