conv3d_transpose

paddle.static.nn. conv3d_transpose ( input, num_filters, output_size=None, filter_size=None, padding=0, stride=1, dilation=1, groups=None, param_attr=None, bias_attr=None, use_cudnn=True, act=None, name=None, data_format='NCDHW' ) [source]

The convolution3D transpose layer calculates the output based on the input, filter, and dilations, strides, paddings. Input(Input) and output(Output) are in NCDHW or NDHWC format. Where N is batch size, C is the number of channels, D is the depth of the feature, H is the height of the feature, and W is the width of the feature. Parameters(dilations, strides, paddings) are two elements. These two elements represent height and width, respectively. The details of convolution transpose layer, please refer to the following explanation and references therein. If bias attribution and activation type are provided, bias is added to the output of the convolution, and the corresponding activation function is applied to the final result.

For each input \(X\), the equation is:

\[Out = \sigma (W \ast X + b)\]

In the above equation:

  • \(X\): Input value, a Tensor with NCDHW or NDHWC format.

  • \(W\): Filter value, a Tensor with MCDHW format.

  • \(\ast\): Convolution operation.

  • \(b\): Bias value, a 2-D Tensor with shape [M, 1].

  • \(\sigma\): Activation function.

  • \(Out\): Output value, the shape of \(Out\) and \(X\) may be different.

Example

  • Input:

    Input shape: \((N, C_{in}, D_{in}, H_{in}, W_{in})\)

    Filter shape: \((C_{in}, C_{out}, D_f, H_f, W_f)\)

  • Output:

    Output shape: \((N, C_{out}, D_{out}, H_{out}, W_{out})\)

Where

\[\begin{split}D^\prime_{out} &= (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\ H^\prime_{out} &= (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\ W^\prime_{out} &= (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1 \\ D_{out} &\in [ D^\prime_{out}, D^\prime_{out} + strides[0] ] \\ H_{out} &\in [ H^\prime_{out}, H^\prime_{out} + strides[1] ] \\ W_{out} &\in [ W^\prime_{out}, W^\prime_{out} + strides[2] ]\end{split}\]

If padding = “SAME”:

\[\begin{split}D^\prime_{out} &= \frac{(D_{in} + stride[0] - 1)}{stride[0]} \\ H^\prime_{out} &= \frac{(H_{in} + stride[1] - 1)}{stride[1]} \\ W^\prime_{out} &= \frac{(H_{in} + stride[2] - 1)}{stride[2]}\end{split}\]

If padding = “VALID”:

\[\begin{split}D^\prime_{out} &= (D_{in} - 1) * strides[0] + dilations[0] * (D_f - 1) + 1 \\ H^\prime_{out} &= (H_{in} - 1) * strides[1] + dilations[1] * (H_f - 1) + 1 \\ W^\prime_{out} &= (W_{in} − 1) * strides[2] + dilations[2] * (W_f − 1) + 1\end{split}\]

If output_size is None, \(D_{out} = D^\prime_{out}, :math:\); else, the specified output_size_depth (the depth of the output feature layer) \(D_{out}\) must between \(D^\prime_{out}\) and \(D^\prime_{out} + strides[0]\)), the specified output_size_height (the height of the output feature layer) \(H_{out}\) must between \(H^\prime_{out}\) and \(H^\prime_{out} + strides[1]\)), and the specified output_size_width (the width of the output feature layer) \(W_{out}\) must between \(W^\prime_{out}\) and \(W^\prime_{out} + strides[2]\)).

Since transposed convolution can be treated as the inverse of convolution, and since different sized input feature layers may correspond to the same sized output feature layer according to the input-output formula for convolution, the size of the output feature layer for a fixed sized input feature layer is not unique to transposed convolution.

If output_size is specified, conv3d_transpose can compute the kernel size automatically.

Parameters

  • input (Tensor) – The input is 5-D Tensor with shape [N, C, D, H, W] or [N, D, H, W, C], the data type of input is float32 or float64.

  • num_filters (int) – The number of the filter. It is as same as the output image channel.

  • output_size (int|tuple, optional) – The output image size. If output size is a tuple, it must contain three integers, (image_depth, image_height, image_width). This parameter only works when filter_size is None. If output_size and filter_size are specified at the same time, They should follow the formula above. Default: None. Output_size and filter_size should not be None at the same time.

  • filter_size (int|tuple, optional) – The filter size. If filter_size is a tuple, it must contain three integers, (filter_size_depth, filter_size_height, filter_size_width). Otherwise, filter_size_depth = filter_size_height = filter_size_width = filter_size. None if use output size to calculate filter_size. Default: None. filter_size and output_size should not be None at the same time.

  • padding (int|list|str|tuple, optional) – The padding size. The padding argument effectively adds dilation * (kernel - 1) amount of zero-padding on both sides of input. If padding is a string, either ‘VALID’ or ‘SAME’ supported, which is the padding algorithm. If padding is a tuple or list, it could be in three forms: [pad_depth, pad_height, pad_width] or [pad_depth_front, pad_depth_back, pad_height_top, pad_height_bottom, pad_width_left, pad_width_right], and when data_format is ‘NCDHW’, padding can be in the form [[0,0], [0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]. when data_format is ‘NDHWC’, padding can be in the form [[0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]. Default: padding = 0.

  • stride (int|tuple, optional) – The stride size. It means the stride in transposed convolution. If stride is a tuple, it must contain three integers, (stride_depth, stride_height, stride_width). Otherwise, stride_depth = stride_height = stride_width = stride. Default: stride = 1.

  • dilation (int|tuple, optional) – The dilation size. It means the spacing between the kernel points. If dilation is a tuple, it must contain three integers, (dilation_depth, dilation_height, dilation_width). Otherwise, dilation_depth = dilation_height = dilation_width = dilation. Default: dilation = 1.

  • groups (int, optional) – The groups number of the Conv3d transpose layer. Inspired by grouped convolution in Alex Krizhevsky’s Deep CNN paper, in which when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels. Default: groups=1

  • param_attr (ParamAttr, optional) – The parameter attribute for learnable parameters/weights of conv3d_transpose. If it is set to None or one attribute of ParamAttr, conv3d_transpose will create ParamAttr as param_attr. If the Initializer of the param_attr is not set, the parameter is initialized with Xavier. Default: None.

  • bias_attr (ParamAttr|bool, optional) – The parameter attribute for the bias of conv3d_transpose. If it is set to False, no bias will be added to the output units. If it is set to None or one attribute of ParamAttr, conv3d_transpose will create ParamAttr as bias_attr. If the Initializer of the bias_attr is not set, the bias is initialized zero. Default: None.

  • use_cudnn (bool, optional) – Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True

  • act (str, optional) – Activation type, if it is set to None, activation is not appended. Default: None.

  • name (str, optional) – For detailed information, please refer to Name. Usually name is no need to set and None by default.

  • data_format (str, optional) – Specify the data format of the input, and the data format of the output will be consistent with that of the input. An optional string from: “NCHW”, “NHWC”. The default is “NCHW”. When it is “NCHW”, the data is stored in the order of: [batch_size, input_channels, input_height, input_width].

Returns

A Tensor representing the conv3d_transpose, whose data type is the same with input and shape is (num_batches, channels, out_d, out_h, out_w) or (num_batches, out_d, out_h, out_w, channels). If act is None, the tensor variable storing the transposed convolution result, and if act is not None, the tensor variable storing transposed convolution and non-linearity activation result.

Examples

>>> import paddle
>>> import numpy as np

>>> paddle.seed(1107)
>>> np.random.seed(1107)
>>> paddle.enable_static()
>>> data = paddle.static.data(name='data', shape=[None, 3, 12, 32, 32], dtype='float32')
>>> param_attr = paddle.framework.ParamAttr(name='conv3d.weight', initializer=paddle.nn.initializer.XavierNormal(), learning_rate=0.001)
>>> res = paddle.static.nn.conv3d_transpose(input=data, num_filters=2, filter_size=3, act="relu", param_attr=param_attr)
>>> place = paddle.CPUPlace()
>>> exe = paddle.static.Executor(place)
>>> exe.run(paddle.static.default_startup_program())
>>> x = np.random.rand(1, 3, 12, 32, 32).astype("float32")
>>> output = exe.run(feed={"data": x}, fetch_list=[res.mean()])
>>> print(output)
[array(0.5148856, dtype=float32)]