The transpose of `conv3d`. This operation is sometimes called "deconvolution" after [Deconvolutional Networks](http://www.matthewzeiler.com/pubs/cvpr2010/cvpr2010.pdf), but is actually the transpose (gradient) of `conv2d` rather than an actual deconvolution. Args: input: A 5-D `Tenso
(input, # pylint: disable=redefined-builtin
filters,
output_shape,
strides,
padding="SAME",
data_format="NDHWC",
dilations=None,
name=None)
| 2575 | |
| 2576 | @tf_export("nn.conv3d_transpose", v1=[]) |
| 2577 | def conv3d_transpose_v2(input, # pylint: disable=redefined-builtin |
| 2578 | filters, |
| 2579 | output_shape, |
| 2580 | strides, |
| 2581 | padding="SAME", |
| 2582 | data_format="NDHWC", |
| 2583 | dilations=None, |
| 2584 | name=None): |
| 2585 | """The transpose of `conv3d`. |
| 2586 | |
| 2587 | This operation is sometimes called "deconvolution" after [Deconvolutional |
| 2588 | Networks](http://www.matthewzeiler.com/pubs/cvpr2010/cvpr2010.pdf), but is |
| 2589 | actually the transpose (gradient) of `conv2d` rather than an actual |
| 2590 | deconvolution. |
| 2591 | |
| 2592 | Args: |
| 2593 | input: A 5-D `Tensor` of type `float` and shape `[batch, height, width, |
| 2594 | in_channels]` for `NHWC` data format or `[batch, in_channels, height, |
| 2595 | width]` for `NCHW` data format. |
| 2596 | filters: A 5-D `Tensor` with the same type as `value` and shape `[height, |
| 2597 | width, output_channels, in_channels]`. `filter`'s `in_channels` dimension |
| 2598 | must match that of `value`. |
| 2599 | output_shape: A 1-D `Tensor` representing the output shape of the |
| 2600 | deconvolution op. |
| 2601 | strides: An int or list of `ints` that has length `1`, `3` or `5`. The |
| 2602 | stride of the sliding window for each dimension of `input`. If a single |
| 2603 | value is given it is replicated in the `D`, `H` and `W` dimension. By |
| 2604 | default the `N` and `C` dimensions are set to 0. The dimension order is |
| 2605 | determined by the value of `data_format`, see below for details. |
| 2606 | padding: A string, either `'VALID'` or `'SAME'`. The padding algorithm. See |
| 2607 | the "returns" section of `tf.nn.convolution` for details. |
| 2608 | data_format: A string. 'NDHWC' and 'NCDHW' are supported. |
| 2609 | dilations: An int or list of `ints` that has length `1`, `3` or `5`, |
| 2610 | defaults to 1. The dilation factor for each dimension of`input`. If a |
| 2611 | single value is given it is replicated in the `D`, `H` and `W` dimension. |
| 2612 | By default the `N` and `C` dimensions are set to 1. If set to k > 1, there |
| 2613 | will be k-1 skipped cells between each filter element on that dimension. |
| 2614 | The dimension order is determined by the value of `data_format`, see above |
| 2615 | for details. Dilations in the batch and depth dimensions if a 5-d tensor |
| 2616 | must be 1. |
| 2617 | name: Optional name for the returned tensor. |
| 2618 | |
| 2619 | Returns: |
| 2620 | A `Tensor` with the same type as `value`. |
| 2621 | """ |
| 2622 | with ops.name_scope(name, "conv3d_transpose", |
| 2623 | [input, filter, output_shape]) as name: |
| 2624 | if data_format is None: |
| 2625 | data_format = "NDHWC" |
| 2626 | channel_index = 1 if data_format.startswith("NC") else 4 |
| 2627 | |
| 2628 | strides = _get_sequence(strides, 3, channel_index, "strides") |
| 2629 | dilations = _get_sequence(dilations, 3, channel_index, "dilations") |
| 2630 | |
| 2631 | return gen_nn_ops.conv3d_backprop_input_v2( |
| 2632 | input_sizes=output_shape, |
| 2633 | filter=filters, |
| 2634 | out_backprop=input, |
no test coverage detected