mindspore.ops.fractional_max_pool3d

mindspore.ops.fractional_max_pool3d(input_x, kernel_size, output_size=None, output_ratio=None, return_indices=False, _random_samples=None)[source]

This operator applies a 3D fractional max pooling over an input signal. The input is composed of several input planes. The max-pooling operation is applied in kD x kH x kW regions by a stochastic step size determined by the target output size.The number of output features is equal to the number of input planes.

Refer to the paper Fractional MaxPooling by Ben Graham for more details.

The input and output data format can be “NCDHW”. N is the batch size, C is the number of channels, D the feature depth, H is the feature height, and W is the feature width.

Parameters

  • input_x (Tensor) – The input of FractionalMaxPool3d, which is a 4D or 5D tensor. Tensor of data type: float16, float32, double, int32, int64. Supported shape \((N, C, D_{in}, H_{in}, W_{in})\).

  • kernel_size (Union[int, tuple[int]]) – The size of kernel used to take the maximum value, is an int number that represents depth, height and width of the kernel, or a tuple of three int numbers that represent depth, height and width respectively. The value must be a positive integer.

  • output_size (Union[int, tuple[int]], optional) – The Shape of the target output_size, is an int number that represents depth, height and width, or a tuple of three int numbers that represent depth, height and width respectively. The value must be a positive integer. Default: None.

  • output_ratio (Union[float, tuple[float]], optional) – The ratio of target output shape to input shape. Specifying the size of the output tensor by using a ratio of the input size. Data type: float16, float32, double, and value is between (0, 1). Default: None.

  • return_indices (bool, optional) – If return_indices is True, the indices of max value would be output. Default: False.

  • _random_samples (Tensor, optional) – The random step of FractionalMaxPool3d, which is a 3D tensor. Tensor of data type: float16, float32, double, and value is between (0, 1). Supported shape \((N, C, 3)\).

Returns

  • y (Tensor) - A tensor, the output of FractionalMaxPool3d. Has the same data type with imput_x. Tensor of shape \((N, C, D, H, W)\) .

  • argmax (Tensor) - The indices along with the outputs, which is a Tensor, with the same shape as the y and int32 data type. It will output only when return_indices is True.

Raises

  • TypeError – If input_x is not a 4D or 5D tensor.

  • TypeError – If _random_samples is not a 3D tensor.

  • TypeError – If data type of imput_x is not float16, float32, double, int32, int64.

  • TypeError – If dtype of _random_samples is not float16, float32, double.

  • TypeError – If dtype of argmax is not int32, int64.

  • ValueError – If output_size is a tuple and if output_size length is not 3.

  • ValueError – If kernel_size is a tuple and if kernel_size length is not 3.

  • ValueError – If numbers in output_size or kernel_size is not positive.

  • ValueError – if output_size and output_ratio are None at the same time.

  • ValueError – If the first dimension size of input_x and _random_samples is not equal.

  • ValueError – If the second dimension size of input_x and _random_samples is not equal.

  • ValueError – If the third dimension size of _random_samples is not 3.

Supported Platforms:

GPU CPU

Examples

>>> x = Tensor(np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16])
...            .reshape([1, 1, 2, 2, 4]), mstype.float32)
>>> _random_samples = Tensor(np.array([0.7, 0.7, 0.7]).reshape([1, 1, 3]), mstype.float32)
>>> output, argmax = ops.fractional_max_pool3d(x, kernel_size=(1.0, 1.0, 1.0), output_size=(1, 1, 3),
...                                            _random_samples=_random_samples, return_indices=True)
>>> print(output)
[[[[[13. 14. 16.]]]]]
>>> print(argmax)
[[[[[12 13 15]]]]]
>>> output, argmax = ops.fractional_max_pool3d(x, kernel_size=(1.0, 1.0, 1.0), output_ratio=(0.5, 0.5, 0.5),
...                                            _random_samples=_random_samples, return_indices=True)
>>> print(output)
[[[[[13. 16.]]]]]
>>> print(argmax)
[[[[[12 15]]]]]