torch_npu.npu_normalize_batch

API接口

torch_npu.npu_normalize_batch(self, seq_len, normalize_type=0) -> Tensor

功能描述

执行批量归一化。

参数说明

• self (Tensor) - 支持float32数据类型，shape为(n, c, d)。
• seq_len (Tensor) - 支持Int32数据类型，shape为(n, )， 表示每批次标准化数据量 。
• normalize_type (Int，默认值为0) - 支持 "per_feature"或"all_features"。值为0表示 "per_feature"，值为1表示"all_features"。

示例

>>> a=np.random.uniform(1,10,(2,3,6)).astype(np.float32)
>>> b=np.random.uniform(3,6,(2)).astype(np.int32)
>>> x=torch.from_numpy(a).to("npu")
>>> seqlen=torch.from_numpy(b).to("npu")
>>> out = torch_npu.npu_normalize_batch(x, seqlen, 0)
>>> out
tensor([[[ 1.1496, -0.6685, -0.4812,  1.7611, -0.5187,  0.7571],
        [ 1.1445, -0.4393, -0.7051,  1.0474, -0.2646, -0.1582],
        [ 0.1477,  0.9179, -1.0656, -6.8692, -6.7437,  2.8621]],

        [[-0.6880,  0.1337,  1.3623, -0.8081, -1.2291, -0.9410],
        [ 0.3070,  0.5489, -1.4858,  0.6300,  0.6428,  0.0433],
        [-0.5387,  0.8204, -1.1401,  0.8584, -0.3686,  0.8444]]],
      device='npu:0')