mindspore.nn.Conv2dThor

class mindspore.nn.Conv2dThor(in_channels, out_channels, kernel_size, stride=1, pad_mode='same', padding=0, dilation=1, group=1, has_bias=False, weight_init='normal', bias_init='zeros')[source]

2D convolution layer and saving the information needed for THOR.

Applies a 2D convolution over an input tensor which is typically of shape \((N, C_{in}, H_{in}, W_{in})\), where \(N\) is batch size, \(C_{in}\) is channel number, and \(H_{in}, W_{in})\) are height and width. And saves the information A and G in the 2D convolution layer needed for THOR. The detail can be seen in paper: https://www.aaai.org/AAAI21Papers/AAAI-6611.ChenM.pdf

For each batch of shape \((C_{in}, H_{in}, W_{in})\), the formula is defined as:

\[out_j = \sum_{i=0}^{C_{in} - 1} ccor(W_{ij}, X_i) + b_j,\]

where \(ccor\) is the cross-correlation operator, \(C_{in}\) is the input channel number, \(j\) ranges from \(0\) to \(C_{out} - 1\), \(W_{ij}\) corresponds to the \(i\)-th channel of the \(j\)-th filter and \(out_{j}\) corresponds to the \(j\)-th channel of the output. \(W_{ij}\) is a slice of kernel and it has shape \((\text{ks_h}, \text{ks_w})\), where \(\text{ks_h}\) and \(\text{ks_w}\) are the height and width of the convolution kernel. The full kernel has shape \((C_{out}, C_{in} // \text{group}, \text{ks_h}, \text{ks_w})\), where group is the group number to split the input x in the channel dimension.

If the ‘pad_mode’ is set to be “valid”, the output height and width will be \(\left \lfloor{1 + \frac{H_{in} + 2 \times \text{padding} - \text{ks_h} - (\text{ks_h} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor\) and \(\left \lfloor{1 + \frac{W_{in} + 2 \times \text{padding} - \text{ks_w} - (\text{ks_w} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor\) respectively.

Parameters

  • in_channels (int) – The number of the input channel \(C_{in}\).

  • out_channels (int) – The number of the output channel \(C_{out}\).

  • kernel_size (Union[int, tuple[int]]) – The data type is int or a tuple of 2 integers. Specifies the height and width of the 2D convolution window. Single int means that the value is not only the height, but also the width of the kernel. A tuple of 2 integers means the height and the width of the kernel respectively.

  • stride (Union[int, tuple[int]]) – The distance of kernel moving, an int number represents the height and width of movement, or a tuple of two int numbers that represent height and width of movement, respectively. Default: 1.

  • pad_mode (str) –

    Specifies padding mode. The optional values are “same”, “valid”, “pad”. Default: “same”.

    • same: Adopts the way of completion. The shape of the output will be the same as the x. The total number of padding will be calculated in horizontal and vertical directions and evenly distributed to top and bottom, left and right if possible. Otherwise, the last extra padding will be done from the bottom and the right side. If this mode is set, padding must be 0.

    • valid: Adopts the way of discarding. The possible largest height and width of output will be returned without padding. Extra pixels will be discarded. If this mode is set, padding must be 0.

    • pad: Implicit paddings on both sides of the input x. The number of padding will be padded to the input Tensor borders. padding must be greater than or equal to 0.

  • padding (Union[int, tuple[int]]) – Implicit paddings on both sides of the input x. If padding is an integer, the paddings of top, bottom, left and right are the same, equal to padding. If padding is a tuple with four integers, the paddings of top, bottom, left and right will be equal to padding[0], padding[1], padding[2], and padding[3] accordingly. Default: 0.

  • dilation (Union[int, tuple[int]]) – The data type is int or a tuple of 2 integers. Specifies the dilation rate to use for dilated convolution. If set to be \(k > 1\), there will be \(k - 1\) pixels skipped for each sampling location. Its value must be greater or equal to 1 and bounded by the height and width of the input x. Default: 1.

  • group (int) – Splits filter into groups, in_ channels and out_channels must be divisible by the number of groups. If the group is equal to in_channels and out_channels, this 2D convolution layer also can be called 2D depthwise convolution layer. Default: 1.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: False.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializes the convolution kernel. It can be a Tensor, a string, an Initializer or a number. When a string is specified, values from ‘TruncatedNormal’, ‘Normal’, ‘Uniform’, ‘HeUniform’ and ‘XavierUniform’ distributions as well as constant ‘One’ and ‘Zero’ distributions are possible. Alias ‘xavier_uniform’, ‘he_uniform’, ‘ones’ and ‘zeros’ are acceptable. Uppercase and lowercase are both acceptable. Refer to the values of Initializer for more details. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializes the bias vector. Possible Initializer and string are the same as ‘weight_init’. Refer to the values of Initializer for more details. Default: ‘zeros’.

Inputs:

  • x (Tensor) - Tensor of shape \((N, C_{in}, H_{in}, W_{in})\).

Outputs:

Tensor of shape \((N, C_{out}, H_{out}, W_{out})\).

Supported Platforms:

Ascend GPU

Examples

>>> net = nn.Conv2dThor(120, 240, 4, has_bias=False, weight_init='normal')
>>> x = Tensor(np.ones([1, 120, 1024, 640]), mindspore.float32)
>>> print(net(x).shape)
(1, 240, 1024, 640)
save_gradient(dout)[source]