ConvTranspose

Description

• Performs a transposed convolution operation using an input tensor and a filter to generate the output tensor.
• If the pads parameter is provided, the output shape is calculated as follows:

• output_shape can also be explicitly specified. In this case, pads are automatically generated using the following formula:
  

  • If auto_pads == SAME_UPPER

    

  • Otherwise,

    

Input

• X: input tensor from the upper layer. For 2D convolution, the shape is [N, C, H, W], where N indicates the batch size, C indicates the number of channels, H and W indicate the height and width. For more than two dimensions, the shape is [N, C, D, H, W] and the data type is float16.
• W: weight tensor. For 2D convolution, the shape is [C, M/group, k_H, k_W], where C indicates the number of channels, k_H and k_W indicate the height and width of the kernel, and M indicates the number of feature maps. For more than two dimensions, the weight shape is [C, M/group, k_D, k_H, k_W]. The number of channels in the output must be equal to W.shape[1] * group (assuming that the index of the shape array starts from 0). The data type can be float16.
• B (optional): 1D tensor of type float16, with shape [M].

Attribute

• auto_pad: string. The value can be NOTSET, SAME_UPPER, SAME_LOWER, or VALID. The default value is NOTSET, indicating that padding is explicitly used. SAME_UPPER or SAME_LOWER indicates that the input is padded. The shape of the output y is y_H = x_H * stride_H, y_W = x_W * stride_W. When auto_pad is set to VALID, padding is not used.
• dilations: a sequence of integers (defaulting to all ones) that represents the dilation rate for each axis of the filter.
• group: The data type is int. The default value is 1, indicating the number of input channel groups.
• kernel_shape: The data type is ints. The shape of the filter needs to be input, indicating the size of the convolution kernel. The two-dimensional convolution input is [k _H, k _W], and the three-dimensional convolution input is [k _D, k _H, k _W].
• output_padding: A sequence of integers (defaulting to an all-zero array) that specifies the additional padding applied to the output. For a 4D input tensor, it is a list of 4 integers; for a 5D input tensor, it is a list of 5 integers.
• output_shape: The data type is ints. The output shape is automatically calculated based on pad.
• pads: The data type is ints. The default value is an all-0 matrix, indicating the padding value specified for each axis.
• strides: The data type is ints. The default value is an all-1 matrix, indicating the stride value of each axis.

Output

y: output tensor of the convolution result. The data type is the same as that of the input.

Constraints

• The number of input channels [C] must be divisible by group. Constraints must be satisfied for the bias shape [M], the first dimension of the filter [M/group], and the first dimension of the output y [M].
• The auto_pad attribute supports SAME_UPPER, SAME_LOWER, and VALID in 2D scenarios, but does not support them in 3D scenarios.
• When the auto_pad attribute is set to SAME_UPPER, SAME_LOWER, or VALID in 2D scenarios, the pads attribute will be corrected if it is provided.

ONNX Opset Support

Opset v8/v9/v10/v11/v12/v13/v14/v15/v16/v17/v18