torch_npu.npu_ifmr

API接口

torch_npu.npu_ifmr(Tensor data, Tensor data_min, Tensor data_max, Tensor cumsum, float min_percentile, float max_percentile, float search_start, float search_end, float search_step, bool with_offset) -> (Tensor, Tensor)

功能描述

使用“begin,end,strides”数组对ifmr结果进行计数。

参数说明

• data (Tensor) - 特征图张量。
• data_min (Tensor) - 特征图最小值的张量。
• data_max (Tensor) - 特征图最大值的张量。
• cumsum (Tensor) - cumsum bin数据张量。
• min_percentile (Float) - 最小初始化百分位数。
• max_percentile (Float) - 最大初始化百分位数。
• search_start (Float) - 搜索起点。
• search_end (Float) - 搜索终点。
• search_step (Float) - 搜索步长。
• with_offset (Bool) - 是否使用offset。

输出说明

• scale (Tensor) - 最优尺度。
• offset (Tensor) - 最优offset。

示例

>>> input = torch.rand((2,2,3,4),dtype=torch.float32).npu()
>>> input
tensor([[[[0.4508, 0.6513, 0.4734, 0.1924],
          [0.0402, 0.5502, 0.0694, 0.9032],
          [0.4844, 0.5361, 0.9369, 0.7874]],
        [[0.5157, 0.1863, 0.4574, 0.8033],
          [0.5986, 0.8090, 0.7605, 0.8252],
          [0.4264, 0.8952, 0.2279, 0.9746]]],
        [[[0.0803, 0.7114, 0.8773, 0.2341], 
         [0.6497, 0.0423, 0.8407, 0.9515], 
         [0.1821, 0.5931, 0.7160, 0.4968]],
          [[0.7977, 0.0899, 0.9572, 0.0146],
          [0.2804, 0.8569, 0.2292, 0.1118],
          [0.5747, 0.4064, 0.8370, 0.1611]]]], device='npu:0')  
>>> min_value = torch.min(input)  
>>> min_value  
tensor(0.0146, device='npu:0')
>>> max_value = torch.max(input)  
>>> max_value  
tensor(0.9746, device='npu:0')  
>>> hist = torch.histc(input.to('cpu'),
                         bins=128,
                         min=min_value.to('cpu'),
                         max=max_value.to('cpu'))  
>>> hist  
tensor([1., 0., 0., 2., 0., 0., 0., 1., 1., 0., 1., 0., 1., 0., 0., 0., 0., 0.,
          0., 1., 0., 0., 2., 1., 0., 0., 0., 0., 2., 1., 0., 0., 0., 0., 0., 1.,
          0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0.,
          1., 0., 0., 0., 1., 1., 0., 1., 1., 0., 1., 0., 1., 0., 0., 1., 0., 1.,
          0., 0., 1., 0., 0., 2., 0., 0., 0., 0., 0., 0., 2., 0., 0., 0., 0., 0., 
         0., 0., 1., 1., 0., 0., 0., 0., 0., 1., 0., 0., 0., 1., 1., 2., 0., 0.,
          1., 1., 1., 0., 1., 0., 0., 1., 0., 1., 1., 0., 0., 0., 1., 0., 1., 1.,
          0., 1.])  >>> cdf = torch.cumsum(hist,dim=0).int().npu()  
>>> cdf  
tensor([ 1,  1,  1,  3,  3,  3,  3,  4,  5,  5,  6,  6,  7,  7,  7,  7,  7,  7,
          7,  8,  8,  8, 10, 11, 11, 11, 11, 11, 13, 14, 14, 14, 14, 14, 14, 15,
          15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 16, 16,
          17, 17, 17, 17, 18, 19, 19, 20, 21, 21, 22, 22, 23, 23, 23, 24, 24, 25, 
         25, 25, 26, 26, 26, 28, 28, 28, 28, 28, 28, 28, 30, 30, 30, 30, 30, 30, 
         30, 30, 31, 32, 32, 32, 32, 32, 32, 33, 33, 33, 33, 34, 35, 37, 37, 37,
          38, 39, 40, 40, 41, 41, 41, 42, 42, 43, 44, 44, 44, 44, 45, 45, 46, 47,
          47, 48], device='npu:0', dtype=torch.int32)  
>>> scale, offset = torch_npu.npu_ifmr(input,
                                     min_value, 
                                    max_value,  
                                   cdf, 
                                    min_percentile=0.999999, 
                                    max_percentile=0.999999,
                                     search_start=0.7, 
                                    search_end=1.3, 
                                    search_step=0.01,
                                     with_offset=False)  
>>> scale  tensor(0.0080, device='npu:0')  
>>> offset  tensor(0., device='npu:0')