update ops docs
This commit is contained in:
parent
9ad01103ad
commit
2d1c355b86
|
|
@ -5,9 +5,8 @@ mindspore.ops.FractionalAvgPool
|
|||
|
||||
在输入上执行分数平均池化。
|
||||
|
||||
分数平均池化类似于常规平均池化。在常规平均池中，可以通过获取集合中较小的 `N x N` 子部分（通常为2x2）的平均值来缩小输入集的大小，
|
||||
|
||||
并尝试将集合减少N的因子，其中N是整数。分数平均池化，意味着整体缩减比N不必是整数。在每个池化区域中，执行均值运算。
|
||||
分数平均池化类似于常规平均池化。在常规平均池化中，通过取集合较小的 `N x N` 子部分的平均值（通常为2x2）来缩小输入集的大小，目标是将集合缩小 `N` 倍，其中 `N` 为整数。
|
||||
但分数平均池化具有额外的灵活性，允许总缩小比率 `N` 为非整数值。
|
||||
|
||||
.. warning::
|
||||
`pooling_ratio` 当前只支持行和列轴，并行大于1.0，第一个和最后一个元素必须为1.0，因为我们不允许对batch和通道轴进行池化。
|
||||
|
|
@ -15,14 +14,14 @@ mindspore.ops.FractionalAvgPool
|
|||
参数：
|
||||
- **pooling_ratio** (list(float)) - 决定了输出的shape，数据类型是floats的列表，长度大于等于4。其值为每个维度的池化比率，目前仅支持行和列维度，
|
||||
应该大于等于0。第一个和最后一个元素必须为1.0，不支持对batch和通道轴进行池化。
|
||||
- **pseudo_random** (bool，可选) - 当设置为True时，以伪随机方式生成池序列，为False时，以随机方式生成池序列。默认为False。
|
||||
查看文章 `Fractional Max-Pooling <https://arxiv.org/pdf/1412.6071>`_ 以了解伪随机和随机之间的差异。
|
||||
- **pseudo_random** (bool，可选) - 控制序列生成机制是随机或伪随机。当设置为True时，以伪随机方式生成池序列，为False时，以随机方式生成池序列。默认值：False。
|
||||
参考 Benjamin Graham 的论文 `Fractional Max-Pooling <https://arxiv.org/pdf/1412.6071>`_ 以了解伪随机和随机池化之间的差异。
|
||||
- **overlapping** (bool，可选) - 当设置为True时，表示两个单元格都使用相邻池化单元边界的值进行池化，
|
||||
设置为False时，表示值不进行重复使用。默认为False。
|
||||
设置为False时，表示值不进行重复使用。默认值：False。
|
||||
- **deterministic** (bool，可选) - 当设置为True时，将在计算图中的FractionalAvgPool节点上进行迭代时使用固定池区域。
|
||||
主要用于单元测试，使FractionalAvgPool具有确定性。当设置为False时，将不使用固定池区域。默认为False。
|
||||
- **seed** (int，可选) - 如果seed或seed2被设置为非零，则随机数生成器由给定的seed生成，否则，它由随机种子生成。默认为0。
|
||||
- **seed2** (int，可选) - 第二个seed，以避免发生seed碰撞。默认为0。
|
||||
主要用于单元测试，使FractionalAvgPool具有确定性。当设置为False时，将不使用固定池区域。默认值：False。
|
||||
- **seed** (int，可选) - 如果seed或seed2被设置为非零，则随机数生成器由给定的seed生成，否则，它由随机种子生成。默认值：0。
|
||||
- **seed2** (int，可选) - 第二个seed，以避免发生seed碰撞。默认值：0。
|
||||
|
||||
输入：
|
||||
- **x** (Tensor) - 数据类型必须为：float32、float64、int32、int64。shape为 :math:`(N, H_{in}, W_{in}, C_{in})` 。
|
||||
|
|
|
|||
|
|
@ -5,24 +5,23 @@ mindspore.ops.FractionalMaxPool
|
|||
|
||||
在输入上执行分数最大池化。
|
||||
|
||||
分数最大池化类似于常规最大池化。在常规最大池中，通过获取集合中较小N×N子部分的最大值(通常为2x2)来缩小输入集的大小，并尝试将集合减少N倍，其中N是整数。
|
||||
分数最大池化类似于常规最大池化。在常规最大池化中，通过取集合较小的 `N x N` 的子部分的最大值（通常为2x2）来缩小输入集的大小，目标是将集合缩小 `N` 倍，其中 `N` 为整数。
|
||||
但分数最大池化具有额外的灵活性，允许总缩小比率 `N` 为非整数值。
|
||||
|
||||
分数最大池化意味着整体缩减比率N不必是整数。池区域的大小是随机生成的，但是相当均匀。
|
||||
|
||||
.. warning::
|
||||
`pooling_ratio` 当前只支持行和列轴，并行大于1.0，第一个和最后一个元素必须为1.0，因为我们不允许对batch和通道轴进行池化。
|
||||
|
||||
参数：
|
||||
- **pooling_ratio** (list(float)) - 决定了输出的shape，floats列表，长度大于等于4。对每个轴的value应该大于等于0，目前仅支持行和列维度。
|
||||
第一个和最后一个元素必须为1.0，因为我们不允许对batch和通道轴进行池化。
|
||||
- **pseudo_random** (bool，可选) - 当设置为True时，以伪随机方式生成池序列,否则以随机方式生成池序列。默认为False。
|
||||
查看文章 `Fractional Max-Pooling <https://arxiv.org/pdf/1412.6071>`_ 以了解伪随机和随机之间的差异。
|
||||
- **overlapping** (bool，可选) - 当设置为True时，表示池化时，两个单元格都使用相邻池化单元边界的值，
|
||||
设置为False时，表示值不进行重复使用。默认为False。
|
||||
- **pooling_ratio** (list(float)) - 决定了输出的shape，floats列表，长度大于等于4。对每个轴的value不能小于零0，目前仅支持行和列维度。
|
||||
- **pseudo_random** (bool，可选) - 控制序列生成机制是随机或伪随机。当设置为True时，以伪随机方式生成池序列，为False时，以随机方式生成池序列。默认值：False。
|
||||
参考 Benjamin Graham 的论文 `Fractional Max-Pooling <https://arxiv.org/pdf/1412.6071>`_ 以了解伪随机和随机池化之间的差异。
|
||||
- **overlapping** (bool，可选) - 当设置为True时，表示两个单元格都使用相邻池化单元边界的值进行池化，
|
||||
设置为False时，表示值不进行重复使用。默认值：False。
|
||||
- **deterministic** (bool，可选) - 当设置为True时，将在计算图中的FractionalMaxPool节点上进行迭代时使用固定池区域。
|
||||
主要用于单元测试，使FractionalMaxPool具有确定性。当设置为False时，将不使用固定池区域。默认为False。
|
||||
- **seed** (int，可选) - 如果seed或seed2被设置为非零，则随机数生成器由给定的seed生成，否则，它由随机种子生成。默认为0。
|
||||
- **seed2** (int，可选) - 第二个seed，以避免发生seed碰撞。默认为0。
|
||||
主要用于单元测试，使FractionalMaxPool具有确定性。当设置为False时，将不使用固定池区域。默认值：False。
|
||||
- **seed** (int，可选) - 如果seed或seed2被设置为非零，则随机数生成器由给定的seed生成，否则，它由随机种子生成。默认值：0。
|
||||
- **seed2** (int，可选) - 第二个seed，以避免发生seed碰撞。默认值：0。
|
||||
|
||||
输入：
|
||||
- **x** (Tensor) - 数据类型必须为：float32、float64、int32、int64。shape为： :math:`(N, H_{in}, W_{in}, C_{in})` 。
|
||||
|
|
|
|||
|
|
@ -3,11 +3,11 @@ mindspore.ops.Geqrf
|
|||
|
||||
.. py:class:: mindspore.ops.Geqrf
|
||||
|
||||
用于计算QR分解的低级函数。此函数返回两个Tensor（y，tau）。
|
||||
将矩阵分解为正交矩阵 `Q` 和上三角矩阵 `R` 的乘积。该过程称为QR分解： :math:`A = QR` 。
|
||||
|
||||
计算 `x` 的QR分解。 `Q` 和 `R` 矩阵都存储在同一个输出Tensor `y` 中。
|
||||
`Q` 和 `R` 矩阵都存储在同一个输出Tensor `y` 中。 `R` 的元素存储在对角线及上方。隐式定义矩阵 `Q` 的基本反射器（或户主向量）存储在对角线下方。
|
||||
|
||||
`R` 的元素存储在对角线及上方。隐式定义矩阵 `Q` 的基本反射器（或户主向量）存储在对角线下方。
|
||||
此函数返回两个Tensor（ `y`, `tau` ）。
|
||||
|
||||
输入：
|
||||
- **x** (Tensor) - shape为 :math:`(*, m, n)` ，输入矩阵维度必须为大于等于两维，支持dtype为float32、float64、complex64、complex128。
|
||||
|
|
|
|||
|
|
@ -3,7 +3,7 @@ mindspore.ops.HSVToRGB
|
|||
|
||||
.. py:class:: mindspore.ops.HSVToRGB
|
||||
|
||||
将一个或多个图像从HSV转换为RGB。图像的格式应为：NHWC。
|
||||
将一个或多个图像从HSV颜色空间转换为RGB颜色空间，其中每个像素的HSV值转换为其对应的RGB值。此函数仅适用于输入像素值在[0,1]范围内的情况。图像的格式应为：NHWC。
|
||||
|
||||
输入：
|
||||
- **x** (Tensor) - 输入的图像必须是shape为 :math:`[batch, image\_height, image\_width, channel]` 的四维Tensor。
|
||||
|
|
|
|||
|
|
@ -3,6 +3,6 @@ mindspore.ops.MatrixBandPart
|
|||
|
||||
.. py:class:: mindspore.ops.MatrixBandPart
|
||||
|
||||
复制一个Tensor，将每个最内层矩阵中中心带之外的所有值都设置为零。
|
||||
提取一个Tensor中每个矩阵的中心带，中心带之外的所有值都设置为零。
|
||||
|
||||
更多参考详见 :func:`mindspore.ops.matrix_band_part` 。
|
||||
|
|
|
|||
|
|
@ -3,6 +3,6 @@ mindspore.ops.MatrixDiagV3
|
|||
|
||||
.. py:class:: mindspore.ops.MatrixDiagV3
|
||||
|
||||
返回一个batch的对角Tensor，其具有给定的对角线值。
|
||||
将给定的输入Tensor构造为对角矩阵或一批对角矩阵。
|
||||
|
||||
更多参考详见 :func:`mindspore.ops.matrix_diag` 。
|
||||
|
|
|
|||
|
|
@ -3,38 +3,40 @@ mindspore.ops.MatrixSetDiagV3
|
|||
|
||||
.. py:class:: mindspore.ops.MatrixSetDiagV3(align="RIGHT_LEFT")
|
||||
|
||||
返回具有新的对角线值的批处理矩阵Tensor。
|
||||
给定输入 `x` 和对角线 `diagonal` ，此操作返回与 `x` 具有相同shape和值的Tensor，但返回的Tensor除开最内层矩阵的对角线，
|
||||
这些值将被对角线中的值覆盖。如果某些对角线比 `max_diag_len` 短，则需要被填充，其中 `max_diag_len` 指对角线的最长长度。
|
||||
返回一批具有新的对角线值的矩阵的Tensor。
|
||||
给定输入 `x` 和对角线 `diagonal` ，此操作返回与 `x` 具有相同shape和值的Tensor。但返回的Tensor除了最内层矩阵的对角线，
|
||||
这些值将被对角线中的值覆盖。
|
||||
|
||||
如果某些对角线比 `max_diag_len` 短，则需要被填充，其中 `max_diag_len` 指对角线的最长长度。
|
||||
`diagonal` 的维度 :math:`shape[-2]` 必须等于对角线个数 `num_diags` ， :math:`num\_diags = k[1] - k[0] + 1`，
|
||||
`diagonal` 的维度 :math:`shape[-1]` 必须等于最长对角线值 `max_diag_len` ，
|
||||
:math:`max\_diag\_len = min(x.shape[-2] + min(k[1], 0), x.shape[-1] + min(-k[0], 0))` 。
|
||||
设 `x` 具有 `r + 1` 维 :math:`[I, J, ..., L, M, N]` 。
|
||||
当 `k` 是整数或 :math:`k[0] == k[1]` 时，对角线 `diagonal` 的shape为 :math:`[I, J, ..., L, max\_diag\_len]` 。
|
||||
否则，其shape为 :math:`[I, J, ... L, num\_diags, max\_diag\_len]` 。
|
||||
|
||||
设 `x` 是一个n维Tensor，shape为： :math:`(d_1, d_2, ..., d_{n-2}, d_{n-1}, d_n)` 。
|
||||
当 `k` 是一个整数或 :math:`k[0] == k[1]` 时， `diagonal` 为n-1维Tensor，shape为 :math:`(d_1, d_2, ..., d_{n-2}, max\_diag\_len)` 。
|
||||
否则， `diagonal` 与 `x` 维度一致，其shape为 :math:`(d_1, d_2, ..., d_{n-2}, num\_diags, max\_diag\_len)` 。
|
||||
|
||||
参数：
|
||||
- **align** (str，可选) - 字符串，指定超对角线和次对角线的对齐方式。
|
||||
- **align** (str，可选) - 可选字符串，指定超对角线和次对角线的对齐方式。
|
||||
可选值："RIGHT_LEFT"、"LEFT_RIGHT"、"LEFT_LEFT"、"RIGHT_RIGHT"。
|
||||
默认值："RIGHT_LEFT"。
|
||||
|
||||
- "RIGHT_LEFT"表示将超对角线与右侧对齐（左侧填充行），将次对角线与左侧对齐（右侧填充行）。
|
||||
- "LEFT_RIGHT"表示将超对角线与左侧对齐（右侧填充行），将次对角线与右侧对齐（左侧填充行）。
|
||||
- "LEFT_LEFT"表示将超对角线与左侧对齐（右侧填充行），将次对角线与左侧对齐（右侧填充行）。
|
||||
- "RIGHT_RIGHT"表示将超对角线与右侧对齐（左侧填充行），将次对角线与右侧对齐（左侧填充行）。
|
||||
- "LEFT_LEFT"表示将超对角线和次对角线均与左侧对齐（右侧填充行）。
|
||||
- "RIGHT_RIGHT"表示将超对角线与次对角线均右侧对齐（左侧填充行）。
|
||||
|
||||
输入：
|
||||
- **x** (Tensor) - Tensor，其维度为 `r+1` 需要满足 `r >=1` 。
|
||||
- **x** (Tensor) - n维Tensor，其中 :math:`n >= 2` 。
|
||||
- **diagonal** (Tensor) - 输入对角线Tensor，具有与 `x` 相同的数据类型。
|
||||
当 `k` 是整数或 :math:`k[0] == k[1]` 时，其为维度 `r` ，否则，其维度为 `r + 1` 。
|
||||
当 `k` 是整数或 :math:`k[0] == k[1]` 时，其为维度 `n-1` ，否则，其维度为 `n` 。
|
||||
- **k** (Tensor) - int32类型的Tensor。对角线偏移量。正值表示超对角线，0表示主对角线，负值表示次对角线。
|
||||
`k` 可以是单个整数（对于单个对角线）或一对整数，分别指定矩阵带的上界和下界，且 `k[0]` 不得大于 `k[1]` 。
|
||||
其值必须在 :math:`(-x.shape[-2], x.shape[-1])` 中。采用图模式时，输入 `k` 必须是常量Tensor。
|
||||
|
||||
输出：
|
||||
Tensor，与 `x` 的类型相同。
|
||||
设 `x` 有 `r+1` 维 :math:`[I， J， ...， M， N]` 。
|
||||
则，输出Tensor的维度为 `r+1` 维 :math:`[I, J, ..., L, M, N]` ，与输入 `x` 相同。
|
||||
Tensor，数据类型和shape与 `x` 相同。
|
||||
|
||||
|
||||
异常：
|
||||
- **TypeError** - 若任一输入不是Tensor。
|
||||
|
|
|
|||
|
|
@ -5,7 +5,7 @@ mindspore.ops.MaxUnpool2D
|
|||
|
||||
MaxPool2D的逆过程。
|
||||
|
||||
MaxUnpool2D在计算过程中，保留最大值位置的元素，并将非最大值位置元素设置为0。
|
||||
由于MaxPool2D会丢失非最大值，因此它不是完全可逆的。MaxUnpool2D将MaxPool2D的输出作为输入，包括最大值的索引，并计算部分逆，其中所有非最大值都被设置为零。
|
||||
例如，输入的shape为 :math:`(N, C, H_{in}, W_{in})` ，输出的shape为 :math:`(N, C, H_{out}, W_{out})` ，
|
||||
则该操作如下式所示：
|
||||
|
||||
|
|
|
|||
|
|
@ -3,6 +3,6 @@ mindspore.ops.MultiMarginLoss
|
|||
|
||||
.. py:class:: mindspore.ops.MultiMarginLoss(p=1, margin=1.0, reduction="mean")
|
||||
|
||||
创建一个标准，用于优化输入和输出之间的多类分类铰链损失（基于边距的损失）。
|
||||
创建一个损失函数，用于优化输入和输出之间的多类分类 hinge 损失（基于边界的损失）。
|
||||
|
||||
更多细节请参考 :func:`mindspore.ops.multi_margin_loss` 。
|
||||
|
|
|
|||
|
|
@ -3,9 +3,7 @@ mindspore.ops.MultilabelMarginLoss
|
|||
|
||||
.. py:class:: mindspore.ops.MultilabelMarginLoss(reduction='mean')
|
||||
|
||||
二维卷积层。
|
||||
|
||||
创建一个标准，用于优化输入 :math:`x` （一个2D小批量Tensor）
|
||||
和输出 :math:`y` （一个目标类别索引的2DTensor）之间的多类分类铰链损失（基于边距的损失）。
|
||||
创建一个损失函数，用于最小化多分类任务的基于边际的损失。
|
||||
它以一个2D mini-batch Tensor :math:`x` 作为输入，以包含目标类索引的2D Tensor :math:`y` 作为输出。
|
||||
|
||||
更多细节请参考 :func:`mindspore.ops.multilabel_margin_loss` 。
|
||||
|
|
|
|||
|
|
@ -3,10 +3,11 @@ mindspore.ops.NonMaxSuppressionV3
|
|||
|
||||
.. py:class:: mindspore.ops.NonMaxSuppressionV3
|
||||
|
||||
贪婪选取一组按score降序排列后的边界框。
|
||||
按得分降序排列后采用贪婪策略地选择一组边界框，并剪枝掉与先前选择的框具有高重叠交并比(IOU)的框。
|
||||
得分低于 `score_threshold` 的边界框将被删除。
|
||||
|
||||
.. warning::
|
||||
如果 `max_output_size` 小于0，将使用0代替。
|
||||
如果 `max_output_size` 小于0，其值将置为0。
|
||||
|
||||
.. note::
|
||||
- 此算法与原点在坐标系中的位置无关。
|
||||
|
|
@ -14,10 +15,10 @@ mindspore.ops.NonMaxSuppressionV3
|
|||
|
||||
输入：
|
||||
- **boxes** (Tensor) - 二维Tensor，shape为 :math:`(num\_boxes, 4)` 。
|
||||
- **scores** (Tensor) - 一维Tensor，其shape为 :math:`(num\_boxes)` 。表示对应每一行每个方框的score值， `scores` 和 `boxes` 的num_boxes必须相等。支持的数据类型为float32。
|
||||
- **scores** (Tensor) - 一个shape为 :math:`(num\_boxes)` 的一维Tensor，表示每个边框（也就是 `boxes` Tensor的每一行）对应的单个分数。 `scores` 中的分数数量必须与 `boxes` 中的边框的数量相等。支持的数据类型为float32。
|
||||
- **max_output_size** (Union[Tensor, Number.Int]) - 选取最大的边框数，必须大于等于0，数据类型为int32。
|
||||
- **iou_threshold** (Union[Tensor, Number.Float]) - 边框重叠值阈值，重叠值大于此值说明重叠过大，其值必须大于等于0，小于等于1。支持的数据类型为float32。
|
||||
- **score_threshold** (Union[Tensor, Number.Float]) - 移除边框阈值，边框score值大于此值则移除相应边框。支持的数据类型为float32。
|
||||
- **iou_threshold** (Union[Tensor, Number.Float]) - 边框重叠值阈值，重叠值大于此值说明重叠过大。数据类型为float32，值必须在[0, 1]范围内。
|
||||
- **score_threshold** (Union[Tensor, Number.Float]) - 移除边框阈值，score值低于此值则边框被移除。支持的数据类型为float32。
|
||||
|
||||
输出：
|
||||
一维Tensor，表示被选中边框的index，其shape为 :math:`(M)` ，其中M <= `max_output_size` 。
|
||||
|
|
|
|||
|
|
@ -3,15 +3,16 @@ mindspore.ops.NonMaxSuppressionWithOverlaps
|
|||
|
||||
.. py:class:: mindspore.ops.NonMaxSuppressionWithOverlaps
|
||||
|
||||
贪婪选取一组按score降序排列后的边界框。
|
||||
按照得分从高到低贪心地选择一组边界框，同时移除与之前选定的边界框高度重叠的边界框
|
||||
得分低于 `score_threshold` 的边界框将被删除。可以定义自定义重叠标准（例如，IoU，IoA等），重叠值以N-by-N矩阵形式提供。
|
||||
|
||||
.. note::
|
||||
- 此算法与原点在坐标系中的位置无关。
|
||||
- 对于坐标系的正交变换和平移，该算法不受影响；因此坐标系的平移变换后算法会选择相同的框。
|
||||
- 对于坐标系的正交变换和平移，该算法不受影响，因此坐标系的平移变换后算法会选择相同的框。
|
||||
|
||||
输入：
|
||||
- **overlaps** (Tensor) - 二维Tensor，其shape为 :math:`(num\_boxes, num\_boxes)` ，表示n乘n的边框重叠值。支持的数据类型为float16、float32和float64。
|
||||
- **scores** (Tensor) - 一维Tensor，其shape为 :math:`(num\_boxes)` 。表示对应每一行每个方框的score值， `scores` 和 `overlaps` 的num_boxes必须相等。数据类型与 `overlaps` 一致。
|
||||
- **scores** (Tensor) - 一个shape为 :math:`(num\_boxes)` 的一维Tensor，表示每个边框（也就是 `boxes` Tensor的每一行）对应的单个分数。 `scores` 中的分数数量必须与 `boxes` 中的边框的数量相等。支持的数据类型为float32。
|
||||
- **max_output_size** (Union[Tensor, Number.Int]) - 选取最大的边框数，必须大于等于0，数据类型为int32。
|
||||
- **overlap_threshold** (Union[Tensor, Number.Float]) - 边框重叠值阈值，重叠值大于此值说明重叠过大。支持的数据类型为float16、float32和float64。
|
||||
- **score_threshold** (Union[Tensor, Number.Float]) - 移除边框阈值，边框score值大于此值则移除相应边框。数据类型与 `overlap_threshold` 一致。
|
||||
|
|
|
|||
|
|
@ -1250,7 +1250,7 @@ class Size(PrimitiveWithInfer):
|
|||
|
||||
class MatrixDiagV3(Primitive):
|
||||
"""
|
||||
Returns a batched diagonal tensor with given batched diagonal values.
|
||||
Constructs a diagonal matrix or a batch of diagonal matrices from a given input Tensor.
|
||||
|
||||
Refer to :func:`mindspore.ops.matrix_diag` for more details.
|
||||
|
||||
|
|
@ -1319,47 +1319,58 @@ class MatrixDiagPartV3(Primitive):
|
|||
|
||||
class MatrixSetDiagV3(Primitive):
|
||||
r"""
|
||||
Returns a batched matrix tensor with new batched diagonal values.
|
||||
Given x and diagonal, this operation returns a tensor with the same shape and values as x, except for the specified
|
||||
diagonals of the innermost matrices. These will be overwritten by the values in diagonal. Some diagonals are shorter
|
||||
than `max_diag_len` and need to be padded, where `max_diag_len` is the longest diagonal value.
|
||||
Returns a matrix Tensor with updated diagonal values across batches.
|
||||
It takes an Tensor `x` and `diagonal` as input and returns a Tensor of
|
||||
the same shape and values as `x`. But the specified diagonal values in
|
||||
the innermost matrices will be replaced by the values in the `diagonal`.
|
||||
|
||||
Diagonals shorter than `max_diag_len` need to be padded, where `max_diag_len` is the
|
||||
longest diagonal value.
|
||||
The diagonal.shape[-2] must be equal to num_diags calculated by :math:`k[1] - k[0] + 1`.
|
||||
The diagonal.shape[-1] must be
|
||||
equal to the longest diagonal value `max_diag_len` calculated
|
||||
by :math:`min(x.shape[-2] + min(k[1], 0), x.shape[-1] + min(-k[0], 0))` .
|
||||
Let x have r + 1 dimensions [I, J, ..., L, M, N].
|
||||
The diagonal tensor has rank r with shape :math:`[I, J, ..., L, max\_diag\_len]`
|
||||
when k is an integer or :math:`k[0] == k[1]` . Otherwise, it has rank r + 1
|
||||
with shape :math:`[I, J, ..., L, num\_diags, max\_diag\_len]` .
|
||||
|
||||
Assume `x` is an n-D Tensor with shape :math:`(d_1, d_2, ..., d_{n-2}, d_{n-1}, d_n)`.
|
||||
If `k` is an integer or :math:`k[0] == k[1]`, `diagonal` is an (n-1)-D Tensor with
|
||||
shape :math:`(d_1, d_2, ..., d_{n-2}, max\_diag\_len)`
|
||||
Otherwise, it has the same rank as `x`
|
||||
with shape :math:`(d_1, d_2, ..., d_{n-2}, num\_diags, max\_diag\_len)`.
|
||||
|
||||
Args:
|
||||
align (str, optional): An optional string from: "RIGHT_LEFT", "LEFT_RIGHT", "LEFT_LEFT", "RIGHT_RIGHT".
|
||||
Align is a string specifying how superdiagonals and subdiagonals should be aligned, respectively.
|
||||
align (str, optional): specifies how superdiagonals and subdiagonals should be aligned.
|
||||
Supported values:"RIGHT_LEFT", "LEFT_RIGHT", "LEFT_LEFT", "RIGHT_RIGHT".
|
||||
Default: "RIGHT_LEFT".
|
||||
|
||||
- "RIGHT_LEFT" aligns superdiagonals to the right (left-pads the row) and subdiagonals to the left
|
||||
(right-pads the row).
|
||||
- "LEFT_RIGHT" aligns superdiagonals to the left (right-pads the row) and subdiagonals to the right
|
||||
(left-pads the row).
|
||||
- "LEFT_LEFT" aligns superdiagonals to the left (right-pads the row) and subdiagonals to the left
|
||||
(right-pads the row).
|
||||
- "RIGHT_RIGHT" aligns superdiagonals to the right (left-pads the row) and subdiagonals to the right
|
||||
(left-pads the row).
|
||||
- When set to "RIGHT_LEFT", the alignment of superdiagonals will be towards the right side
|
||||
(padding the row on the left), while subdiagonals will be towards the left side
|
||||
(padding the row on the right)
|
||||
- When set to "LEFT_RIGHT", the alignment of superdiagonals will be towards the left side
|
||||
(padding the row on the right), while subdiagonals will be towards the right side
|
||||
(padding the row on the left)
|
||||
- When set to "LEFT_LEFT", the alignment of both superdiagonals and subdiagonals will be towards
|
||||
the left side(padding the row on the right).
|
||||
- When set to "RIGHT_RIGHT", the alignment of both superdiagonals and subdiagonals will be towards
|
||||
the right side(padding the row on the left).
|
||||
|
||||
Inputs:
|
||||
- **x** (Tensor) - Rank r + 1, where r >= 1.
|
||||
- **diagonal** (Tensor) - A Tensor. Have the same dtype as x. Rank r when k is an integer or k[0] == k[1].
|
||||
Otherwise, it has rank r + 1.
|
||||
- **k** (Tensor) - A Tensor of type int32. Diagonal offset(s). Positive value means superdiagonal, 0 refers to
|
||||
the main diagonal, and negative value means subdiagonals. k can be a single integer (for a single diagonal) or
|
||||
a pair of integers specifying the low and high ends of a matrix band. `k[0]` must not be larger than `k[1]` .
|
||||
The value of `k` has restructions, meaning value of k must be in (-x.shape[-2], x.shape[-1]).
|
||||
Input k must be const Tensor when taking Graph mode.
|
||||
- **x** (Tensor) - A n-D Tensor, where :math:`n >= 2`.
|
||||
- **diagonal** (Tensor) - A Tensor with the same dtype as `x`. Its rank depends on `k`.
|
||||
If `k` is an integer or :math:`k[0] == k[1]`, its dimension is :math:`n-1`.
|
||||
Otherwise, it has dimension :math:`n`.
|
||||
- **k** (Tensor) - Tensor type int32, used for diagonal offset(s).
|
||||
`k` can either be a single integer, which represents a single diagonal,
|
||||
or a pair of integers that specify the low and high ends of a matrix band.
|
||||
In this case(这里是指什么情况下？), `k[0]` should not be greater than `k[1]`.
|
||||
The value of `k` has restructions, which means that value of `k` must be in range (-x.shape[-2], x.shape[-1]).
|
||||
Input `k` must be const Tensor when taking Graph mode.
|
||||
|
||||
- `k > 0` refers to a superdiagonal.
|
||||
- `k = 0` refers to the main diagonal.
|
||||
- `k < 0` refers to subdiagonals.
|
||||
|
||||
Outputs:
|
||||
Tensor. The same type as x.
|
||||
Let x has r+1 dimensions :math:`[I, J, ..., L, M, N]` .
|
||||
The output is a tensor of rank r+1 with dimensions :math:`[I, J, ..., L, M, N]` , the same as input x.
|
||||
Tensor. The same type and shape as `x`.
|
||||
|
||||
Raises:
|
||||
TypeError: If any input is not Tensor.
|
||||
|
|
@ -1410,7 +1421,8 @@ class MatrixSetDiagV3(Primitive):
|
|||
|
||||
class MatrixBandPart(Primitive):
|
||||
r"""
|
||||
Copy a tensor setting everything outside a central band in each innermost matrix to zero.
|
||||
Extracts the central diagonal band of each matrix in a tensor, with all values outside
|
||||
the central band set to zero.
|
||||
|
||||
Refer to :func:`mindspore.ops.matrix_band_part` for more details.
|
||||
|
||||
|
|
@ -3618,7 +3630,7 @@ class DiagPart(PrimitiveWithCheck):
|
|||
|
||||
class Mvlgamma(Primitive):
|
||||
r"""
|
||||
Computes the multivariate log-gamma function with dimension `p` element-wise.
|
||||
Calculates the multivariate log-gamma function element-wise for a given dimension `p`.
|
||||
|
||||
Refer to :func:`mindspore.ops.mvlgamma` for more details.
|
||||
|
||||
|
|
@ -6560,13 +6572,14 @@ class TensorScatterDiv(_TensorScatterOp):
|
|||
|
||||
|
||||
class ListDiff(Primitive):
|
||||
r"""Computes the difference between two lists of numbers.
|
||||
r"""
|
||||
This function calculates the disparity between two numerical lists.
|
||||
|
||||
Given a list `x` and a list `y`, this operation returns a list `out` that
|
||||
represents all values that are in `x` but not in `y`. The returned list `out`
|
||||
is sorted in the same order that the numbers appear in `x` (duplicates are
|
||||
preserved). This operation also returns a list `idx` that represents the
|
||||
position of each `out` element in `x`. In other words:
|
||||
It generates a list of all elements that are present in list `x` but not in list `y`.
|
||||
The output list `out` retains the same order as the original `x` including duplicate elements.
|
||||
|
||||
Additionally, this class outputs a list `idx` that identifies the position of each element
|
||||
in `out` within the original `x`. That is to say:
|
||||
:code:`out[i] = x[idx[i]] for i in [0, 1, ..., len(out) - 1]` .
|
||||
|
||||
Args:
|
||||
|
|
@ -7146,9 +7159,10 @@ class RightShift(Primitive):
|
|||
|
||||
class LogSpace(Primitive):
|
||||
r"""
|
||||
Returns a one-dimensional tensor of size steps whose values are evenly
|
||||
spaced from :math:`base^{start}` to :math:`base^{end}` , inclusive,
|
||||
on a logarithmic scale with base.
|
||||
Generates a 1-D Tensor with a length of steps. The tensor's
|
||||
values are uniformly distributed on a logarithmic scale, ranging from
|
||||
:math:`base^{start}` to :math:`base^{end}`, including both endpoints.
|
||||
The logarithmic scale is based on the specified `base`.
|
||||
|
||||
.. math::
|
||||
\begin{aligned}
|
||||
|
|
@ -7304,8 +7318,8 @@ class Tril(Primitive):
|
|||
|
||||
class IndexFill(Primitive):
|
||||
"""
|
||||
Fills the elements under the dim dimension of the input Tensor with the input value
|
||||
by selecting the indices in the order given in index.
|
||||
Fills the specified elements of the input Tensor with a given value,
|
||||
using the indices specified in the input index array.
|
||||
|
||||
Refer to :func:`mindspore.ops.index_fill` for more details.
|
||||
|
||||
|
|
|
|||
|
|
@ -352,33 +352,36 @@ class CropAndResize(Primitive):
|
|||
|
||||
class NonMaxSuppressionV3(Primitive):
|
||||
r"""
|
||||
Greedily selects a subset of bounding boxes in descending order of score.
|
||||
Selects a subset of bounding boxes in a greedy manner, based on their descending score.
|
||||
It removes boxes that have high intersection-over-union (IOU) overlap with previously
|
||||
selected boxes, and eliminates boxes with scores lower than a given threshold.
|
||||
|
||||
.. warning::
|
||||
When input `max_output_size` is negative, it will be treated as 0.
|
||||
|
||||
Note:
|
||||
- This algorithm is agnostic to where the origin is in the coordinate system.
|
||||
- This algorithm is invariant to orthogonal transformations and translations of the coordinate system,
|
||||
thus translating or reflections of the coordinate system result in the same boxes being
|
||||
selected by the algorithm.
|
||||
- This algorithm does not depend on the location of the origin in the coordinate system.
|
||||
- This algorithm remains unaffected by orthogonal transformations and translations of
|
||||
the coordinate system, which means that translating or reflecting the coordinate system
|
||||
will result in the same boxes being chosen by the algorithm.
|
||||
|
||||
Inputs:
|
||||
- **boxes** (Tensor) - A 2-D Tensor of shape :math:`(num\_boxes, 4)`.
|
||||
- **scores** (Tensor) - A 1-D Tensor of shape :math:`(num\_boxes)` representing a single score
|
||||
corresponding to each box (each row of boxes), the num_boxes of `scores` must be equal to
|
||||
the num_boxes of `boxes`.
|
||||
- **scores** (Tensor) - A 1-D Tensor of shape :math:`(num\_boxes)` where each element represents a
|
||||
single score associated with each box (i.e., each row of the `boxes` Tensor).
|
||||
It is required that the number of scores in `scores` must be equal to the number of boxes in `boxes`.
|
||||
The supported data type is float32.
|
||||
- **max_output_size** (Union[Tensor, Number.Int]) - A scalar integer Tensor representing the maximum
|
||||
number of boxes to be selected by non max suppression.
|
||||
- **iou_threshold** (Union[Tensor, Number.Float]) - A 0-D float tensor representing the threshold for
|
||||
deciding whether boxes overlap too much with respect to IOU, and `iou_threshold` must be equal or greater
|
||||
than 0 and be equal or smaller than 1.
|
||||
- **score_threshold** (Union[Tensor, Number.Float]) - A 0-D float tensor representing the threshold for
|
||||
deciding when to remove boxes based on score.
|
||||
number of boxes to be selected by non max suppression. The supported data type is int32.
|
||||
- **iou_threshold** (Union[Tensor, Number.Float]) - A scalar float Tensor represents the threshold
|
||||
used for determining if the intersection over union (IOU) between boxes is too high.
|
||||
Data type of `iou_threshold` is float32 and must be in range [0, 1].
|
||||
- **score_threshold** (Union[Tensor, Number.Float]) - A scalar float Tensor represents the threshold for
|
||||
determining when to remove boxes based on score. The supported data type is float32.
|
||||
|
||||
Outputs:
|
||||
A 1-D integer Tensor of shape [M] representing the selected indices from the boxes tensor,
|
||||
where M <= max_output_size.
|
||||
where M <= `max_output_size`.
|
||||
|
||||
Raises:
|
||||
TypeError: If the dtype of `boxes` and `scores` are different.
|
||||
|
|
@ -418,33 +421,40 @@ class NonMaxSuppressionV3(Primitive):
|
|||
|
||||
class NonMaxSuppressionWithOverlaps(Primitive):
|
||||
r"""
|
||||
Greedily selects a subset of bounding boxes in descending order of score.
|
||||
Selects a subset of bounding boxes in a greedy manner by prioritizing those with higher
|
||||
scores and removing those with high overlaps with previously selected boxes.
|
||||
Boxes with scores lower than the score threshold are also removed.
|
||||
The overlap values between boxes are represented as an N-by-N square matrix,
|
||||
which can be customized to define different overlap criteria such as intersection
|
||||
over union or intersection over area.
|
||||
|
||||
|
||||
Note:
|
||||
- This algorithm is agnostic to where the origin is in the coordinate system.
|
||||
- This algorithm is invariant to orthogonal transformations and translations of the coordinate system;
|
||||
thus translating or reflections of the coordinate system result in the same boxes being
|
||||
selected by the algorithm.
|
||||
- This algorithm does not depend on the location of the origin in the coordinate system.
|
||||
- This algorithm remains unaffected by orthogonal transformations and translations of
|
||||
the coordinate system, which means that translating or reflecting the coordinate system
|
||||
will result in the same boxes being chosen by the algorithm.
|
||||
|
||||
Inputs:
|
||||
- **overlaps** (Tensor) - A 2-D Tensor of shape :math:`(num\_boxes, num\_boxes)`,
|
||||
representing the n-by-n box overlap values. Types allowed:float16, float32 and float64.
|
||||
- **scores** (Tensor) - A 1-D Tensor of shape :math:`(num\_boxes)` representing a single score
|
||||
corresponding to each box (each row of boxes), the num_boxes of `scores` must be equal to
|
||||
the num_boxes of `overlaps`. It has the same dtype as `overlaps`.
|
||||
- **scores** (Tensor) - A 1-D Tensor of shape :math:`(num\_boxes)` where each element represents a
|
||||
single score associated with each box (i.e., each row of the `boxes` Tensor).
|
||||
It is required that the number of scores in `scores` must be equal to the number of boxes in `boxes`.
|
||||
The supported data type is float32.
|
||||
- **max_output_size** (Union[Tensor, Number.Int]) - A scalar integer Tensor representing the maximum
|
||||
number of boxes to be selected by non max suppression, and max_output_size must be equal to or greater
|
||||
than 0.
|
||||
Types allowed:int32.
|
||||
- **overlap_threshold** (Union[Tensor, Number.Float]) - A 0-D float Tensor representing the threshold for
|
||||
deciding whether boxes overlap too much.
|
||||
- **overlap_threshold** (Union[Tensor, Number.Float]) - A scalar value, represented by a 0-D float Tensor,
|
||||
which is used as a threshold to determine if two boxes overlap too much.
|
||||
Types allowed:float16, float32 and float64.
|
||||
- **score_threshold** (Union[Tensor, Number.Float]) - A 0-D float Tensor representing the threshold for
|
||||
deciding when to remove boxes based on score. It has the same dtype as `overlap_threshold`.
|
||||
|
||||
Outputs:
|
||||
A 1-D integer Tensor of shape :math:`(M)` representing the selected indices from the boxes Tensor,
|
||||
where M <= max_output_size. Its data type is int32.
|
||||
A 1-D integer Tensor of shape :math:`(M)` representing the selected indices from the `boxes` Tensor,
|
||||
where M <= `max_output_size`. Its data type is int32.
|
||||
|
||||
Raises:
|
||||
TypeError: If the dtype of `overlaps` , `scores` `overlap_threshold` and `score_threshold`
|
||||
|
|
@ -488,10 +498,9 @@ class NonMaxSuppressionWithOverlaps(Primitive):
|
|||
|
||||
class HSVToRGB(Primitive):
|
||||
r"""
|
||||
Convert one or more images from HSV to RGB.
|
||||
Outputs a tensor of the same shape as the images tensor,
|
||||
containing the HSV value of the pixels. The output is only
|
||||
well defined if the value in images are in [0,1].
|
||||
Transform one single or a batch of images from HSV to RGB color space.
|
||||
Each pixel's HSV value is converted to its corresponding RGB value.
|
||||
Note that the function is only well-defined for input pixel values in the range [0, 1].
|
||||
|
||||
Inputs:
|
||||
- **x** (Tensor) - The input image must be a 4-D tensor of shape
|
||||
|
|
|
|||
|
|
@ -23,13 +23,14 @@ from mindspore.ops.primitive import prim_attr_register
|
|||
|
||||
class Geqrf(Primitive):
|
||||
r"""
|
||||
Geqrf is a low-level function for computing QR decompositions. This function returns two tensors
|
||||
(y, tau).
|
||||
Decomposes a matrix into the product of an orthogonal matrix `Q` and an upper triangular matrix `R`.
|
||||
The process is called QR decomposition: :math:`A = QR`.
|
||||
|
||||
Both `Q` and `R` matrices are stored in the same output tensor `y`.
|
||||
The elements of `R` are stored on and above the diagonal, whereas elementary reflectors
|
||||
(or Householder vectors) implicitly defining matrix `Q` are stored below the diagonal.
|
||||
|
||||
Computes a QR decomposition of `x`. Both `Q` and `R` matrices are stored in the same output tensor `y`.
|
||||
The elements of `R` are stored on and above the diagonal.
|
||||
Elementary reflectors (or Householder vectors) implicitly defining matrix `Q` are stored below the diagonal.
|
||||
This function returns two tensors (`y`, `tau`).
|
||||
|
||||
|
||||
Inputs:
|
||||
|
|
|
|||
|
|
@ -2818,7 +2818,7 @@ class Hypot(Primitive):
|
|||
|
||||
class Heaviside(Primitive):
|
||||
r"""
|
||||
Computes the Heaviside step function for each element in input.
|
||||
Applies the Heaviside step function for input `x` element-wise.
|
||||
|
||||
.. math::
|
||||
\text { heaviside }(\text { x, values })=\left\{\begin{array}{ll}
|
||||
|
|
@ -5486,9 +5486,9 @@ class MatrixInverse(Primitive):
|
|||
|
||||
class MatrixPower(Primitive):
|
||||
"""
|
||||
Computes the n-th power of a batch of square matrices.
|
||||
If n = 0, it returns a batch of identity matrices. If n is negative, it
|
||||
returns the inverse of each matrix (if invertible) raised to the power of abs(n).
|
||||
Calculates the n-th power of a batch of square matrices.
|
||||
When n equals 0, it returns a group of identity matrices. If n is negative,
|
||||
it computes the inverse of each matrix (if possible) raised to the power of abs(n).
|
||||
|
||||
Args:
|
||||
n (int) : The exponent, a required int.
|
||||
|
|
@ -5530,7 +5530,7 @@ class MatrixPower(Primitive):
|
|||
|
||||
class MatrixDeterminant(Primitive):
|
||||
"""
|
||||
Computes the determinant of one or more square matrices.
|
||||
Calculates the value of the determinant for one or more square matrices.
|
||||
|
||||
Refer to :func:`mindspore.ops.det` for more details.
|
||||
|
||||
|
|
@ -5554,7 +5554,7 @@ class MatrixDeterminant(Primitive):
|
|||
|
||||
class LogMatrixDeterminant(Primitive):
|
||||
"""
|
||||
Computes the sign and the log of the absolute value of the determinant of one or more square matrices.
|
||||
Calculates the sign and logarithm of the determinant of one or more square matrices.
|
||||
|
||||
Refer to :func:`mindspore.ops.slogdet` for more details.
|
||||
|
||||
|
|
@ -6109,7 +6109,8 @@ class Igammac(Primitive):
|
|||
|
||||
class IsClose(Primitive):
|
||||
r"""
|
||||
Returns a boolean Tensor where two tensors are element-wise equal within a tolerance.
|
||||
Returns a tensor of Boolean values indicating whether two input tensors
|
||||
are element-wise equal within a given tolerance.
|
||||
|
||||
Refer to :func:`mindspore.ops.isclose` for more details.
|
||||
|
||||
|
|
|
|||
|
|
@ -2001,11 +2001,13 @@ class MaxPool3D(Primitive):
|
|||
|
||||
class MaxUnpool2D(Primitive):
|
||||
r"""
|
||||
Computes the inverse of MaxPool2D.
|
||||
Calculates the partial inverse of MaxPool2D operation.
|
||||
|
||||
MaxUnpool2D keeps the maximal value and set all position of non-maximal values to zero.
|
||||
Typically the input is of shape :math:`(N, C, H_{in}, W_{in})` , the output is of
|
||||
shape :math:`(N, C, H_{out}, W_{out})` , the operation is as follows:
|
||||
Since MaxPool2D loses non-maximal values, it is not fully invertible.
|
||||
Therefore, MaxUnpool2D takes the output of MaxPool2D, including the indices of
|
||||
the maximal values, and computes a partial inverse where all non-maximal values are set to zero.
|
||||
Typically the input is of shape :math:`(N, C, H_{in}, W_{in})` ,
|
||||
the output is of shape :math:`(N, C, H_{out}, W_{out})` , the operation is as follows:
|
||||
|
||||
.. math::
|
||||
\begin{array}{ll} \\
|
||||
|
|
@ -2970,8 +2972,9 @@ class SmoothL1Loss(Primitive):
|
|||
|
||||
class MultiMarginLoss(Primitive):
|
||||
r"""
|
||||
Creates a criterion that optimizes a multi-class classification hinge loss (margin-based loss)
|
||||
between input and output.
|
||||
Creates a loss function that minimizes the margin-based loss or hinge loss
|
||||
for multi-class classification tasks.
|
||||
The loss is calculated by comparing the input and output of the function.
|
||||
|
||||
Refer to :func:`mindspore.ops.multi_margin_loss` for more details.
|
||||
|
||||
|
|
@ -8808,11 +8811,10 @@ class ApplyKerasMomentum(Primitive):
|
|||
|
||||
class MultilabelMarginLoss(Primitive):
|
||||
r"""
|
||||
MultilabelMarginLoss operation.
|
||||
|
||||
Creates a criterion that optimizes a multi-class multi-classification
|
||||
hinge loss (margin-based loss) between input :math:`x` (a 2D mini-batch `Tensor`)
|
||||
and output :math:`y` (which is a 2D `Tensor` of target class indices).
|
||||
Creates a loss criterion that minimizes a margin-based loss for multi-class
|
||||
classification tasks.
|
||||
It takes a 2D mini-batch Tensor :math:`x` as input and a 2D
|
||||
Tensor :math:`y` containing target class indices as output.
|
||||
|
||||
Refer to :func:`mindspore.ops.multilabel_margin_loss` for more details.
|
||||
|
||||
|
|
@ -8976,35 +8978,40 @@ class FractionalMaxPool(Primitive):
|
|||
r"""
|
||||
Performs fractional max pooling on the input.
|
||||
|
||||
Fractional max pooling is similar to regular max pooling, In regular max pooling, you downsize an
|
||||
input set by taking the maximum value of smaller N x N subsections of the set (often 2x2), and try
|
||||
to reduce the set by a factor of N, where N is an integer. Fractional max pooling, means that the
|
||||
overall reduction ratio N does not have to be an integer.
|
||||
The sizes of the pooling regions are generated randomly but are fairly uniform.
|
||||
Fractional max pooling is similar to regular max pooling, but with the added flexibility of
|
||||
allowing the overall reduction ratio `N` to be a non-integer value. In regular max pooling,
|
||||
an input set is reduced in size by taking the maximum value of `N x N` (usually 2x2)
|
||||
subsections of the set, with the goal of reducing the set by a factor of `N`, where `N` is an integer.
|
||||
|
||||
In contrast, fractional max pooling uses randomly generated pool sizes that are fairly uniform in size.
|
||||
|
||||
.. warning::
|
||||
"pooling_ratio", currently only supports row and col dimension and should be >= 1.0, the first
|
||||
and last elements must be 1.0 because we don't allow pooling on batch and channels dimensions.
|
||||
and last elements must be 1.0 because pooling on batch and channels dimensions is not allowed.
|
||||
|
||||
Args:
|
||||
pooling_ratio (list(float)): Decide the shape of output, is a list of floats that has length >= 4.
|
||||
Pooling ratio for each dimension of value should be >=0, currently only support for row and col
|
||||
dimension. The first and last elements must be 1.0 because we don't allow pooling on batch and
|
||||
channels dimensions.
|
||||
pseudo_random(bool, optional): When set to True, generates the pooling
|
||||
sequence in a pseudo random fashion, otherwise, in a random fashion.
|
||||
Check paper Benjamin Graham, Fractional Max-Pooling for difference between pseudo_random and
|
||||
random. Defaults to False.
|
||||
overlapping(bool, optional): When set to True, it means when pooling,
|
||||
the values at the boundary of adjacent pooling cells are used by both cells.
|
||||
When set to False, the values are not reused. Defaults to False.
|
||||
deterministic(bool, optional): When set to True, a fixed pooling region
|
||||
will be used when iterating over a FractionalMaxPool node in the computation graph. Mainly
|
||||
used in unit test to make FractionalMaxPool deterministic. When set to False,
|
||||
fixed pool regions will not be used. Defaults to False.
|
||||
seed(int, optional): If either seed or seed2 are set to be non-zero, the random number generator is
|
||||
seeded by the given seed. Otherwise, it is seeded by a random seed. Defaults to 0.
|
||||
seed2(int, optional): An second seed to avoid seed collision. Defaults to 0.
|
||||
pooling_ratio (list(float)): Decide the shape of output, is a list of float numbers has length >= 4.
|
||||
Pooling ratio for each dimension of value should not be less than 0, currently only support
|
||||
for row and col dimension.
|
||||
pseudo_random(bool, optional): Generate the pooling sequence either randomly or pseudo-randomly.
|
||||
If the pseudo_random parameter is set to True, the sequence will be generated in a
|
||||
pseudo-random fashion, otherwise it will be generated randomly.
|
||||
Refer to `Fractional Max-Pooling <https://arxiv.org/pdf/1412.6071>`_
|
||||
by Benjamin Graham to understand the distinction between the two.
|
||||
Default: False.
|
||||
overlapping(bool, optional): When set to True, the values at the boundary of adjacent pooling cells
|
||||
will be shared by both cells during pooling process. When set to False, the values are not reused.
|
||||
Default: False.
|
||||
deterministic(bool, optional): If deterministic is set to True, a fixed pooling region will be used
|
||||
in the computation graph, ensuring that the FractionalMaxPool is deterministic.
|
||||
This is often used in unit tests. When set to False, fixed pool regions will not be used.
|
||||
Default: False.
|
||||
seed(int, optional): If either seed or seed2 are set to a non-zero value, the random number
|
||||
generator will be seeded using the specified seed. If neither seed nor seed2 are set,
|
||||
the generator will be seeded by a random seed.
|
||||
Default: 0.
|
||||
seed2(int, optional): The second seed to avoid seed collision.
|
||||
Default: 0.
|
||||
|
||||
Inputs:
|
||||
- **x** (Tensor) -The data type must be one of the following types: float32, float64, int32, int64.
|
||||
|
|
@ -9148,11 +9155,10 @@ class FractionalAvgPool(Primitive):
|
|||
r"""
|
||||
Performs fractional avg pooling on the input.
|
||||
|
||||
Fractional avg pooling is similar to regular avg pooling, In regular avg pooling, you downsize an
|
||||
input set by taking the avgrage value of smaller N x N subsections of the set (often 2x2), and try
|
||||
to reduce the set by a factor of N, where N is an integer. Fractional avg pooling, means that the
|
||||
overall reduction ratio N does not have to be an integer. In each pooling region, a mean operation
|
||||
is performed.
|
||||
Fractional avg pooling is similar to regular avg pooling, but with the added flexibility of
|
||||
allowing the overall reduction ratio `N` to be a non-integer value. In regular avg pooling,
|
||||
an input set is reduced in size by taking the average value of `N x N` (usually 2x2)
|
||||
subsections of the set, with the goal of reducing the set by a factor of `N`, where `N` is an integer.
|
||||
|
||||
.. warning::
|
||||
"pooling_ratio", currently only supports row and col dimension and should be >= 1.0, the first
|
||||
|
|
@ -9163,20 +9169,25 @@ class FractionalAvgPool(Primitive):
|
|||
Pooling ratio for each dimension of value should be >=0, currently only support for row and col
|
||||
dimension. The first and last elements must be 1.0 because we don't allow pooling on batch and
|
||||
channels dimensions.
|
||||
pseudo_random(bool, optional): When set to True, generates the pooling
|
||||
sequence in a pseudorandom fashion, otherwise, in a random fashion.
|
||||
Check paper Benjamin Graham, Fractional Max-Pooling for difference between pseudo_random and
|
||||
random. Defaults to False.
|
||||
overlapping(bool, optional): When set to True, it means when pooling,
|
||||
the values at the boundary of adjacent pooling cells are used by both cells.
|
||||
When set to False, the values are not reused. Defaults to False.
|
||||
deterministic(bool, optional): When set to True, a fixed pooling region
|
||||
will be used when iterating over a FractionalAvgPool node in the computation graph. Mainly
|
||||
used in unit test to make FractionalAvgPool deterministic. When set to False,
|
||||
fixed pool regions will not be used. Defaults to False.
|
||||
seed(int, optional): If either seed or seed2 are set to be non-zero, the random number generator
|
||||
is seeded by the given seed. Otherwise, it is seeded by a random seed. Defaults to 0.
|
||||
seed2(int, optional): An second seed to avoid seed collision. Defaults to 0.
|
||||
pseudo_random(bool, optional): Generate the pooling sequence either randomly or pseudo-randomly.
|
||||
If the pseudo_random parameter is set to True, the sequence will be generated in a
|
||||
pseudo-random fashion, otherwise it will be generated randomly.
|
||||
Refer to `Fractional Max-Pooling <https://arxiv.org/pdf/1412.6071>`_
|
||||
by Benjamin Graham to understand the distinction between the two.
|
||||
Default: False.
|
||||
overlapping(bool, optional): When set to True, the values at the boundary of adjacent pooling cells
|
||||
will be shared by both cells during pooling process. When set to False, the values are not reused.
|
||||
Default: False.
|
||||
deterministic(bool, optional): If deterministic is set to True, a fixed pooling region will be used
|
||||
in the computation graph, ensuring that the FractionalAvgPool is deterministic.
|
||||
This is often used in unit tests. When set to False, fixed pool regions will not be used.
|
||||
Default: False.
|
||||
seed(int, optional): If either seed or seed2 are set to a non-zero value, the random number
|
||||
generator will be seeded using the specified seed. If neither seed nor seed2 are set,
|
||||
the generator will be seeded by a random seed.
|
||||
Default: 0.
|
||||
seed2(int, optional): The second seed to avoid seed collision.
|
||||
Default: 0.
|
||||
|
||||
Inputs:
|
||||
- **x** (Tensor) -The data type must be one of the following types: float32, float64, int32, int64.
|
||||
|
|
|
|||
Loading…
Reference in New Issue