2. OpBuilders

如下展示各个OpBuilder的python接口用法.

算子的详细语义可以参考TopsGraph C++ API参考/Operation Semantics.

2.1. Abs

def abs(x)

总结: 计算输入tensor的绝对值

参数:

  • x - 需要计算绝对值的tensor

返回值:

  • x 的绝对值

示例代码:

data = np.array([-1, 0, 1], dtype=np.float32)
output = tops.jit(tgb.abs)(data)
# expect: output = [1, 0, 1]

2.2. Acos

Computes the arccosine (inverse of cosine) of a tensor element-wise.

(2.2.1)\[acos(x)=cos^{-1}(x)\]
Inputs:

  • \(x\) is an op whose result shape is \((*)\), where \(*\) means any number of dimensions.

Outputs:

  • \(result\) with the same shape and data type as the input.

Examples:
>>> data = np.array([-0.4, -0.2, 0.1, 0.3], dtype=np.float32)
>>> output = tops.jit(tgb.acos)(data)
>>> # expect: [1.98231317, 1.77215425, 1.47062891, 1.26610367]

2.3. Add

Add two tensors element-wise with Numpy-style broadcasting support.

(2.3.1)\[result=x+y\]
Inputs:

  • \(x\) is an op whose result shape is \((*)\), where \(*\) means any number of dimensions.

  • \(y\) is an op whose data type is same as \(x\).

Parameters:

  • \(broadcast\_dimensions\) specify the dimensions in the target shape each dimension of the input shape corresponds to. If inputs are static shape and the user want to follow numpy-style broadcasting, just use the default value of empty list. Default: [].

Outputs:

  • \(result\) with the same data type as the inputs.

Examples:
>>> lhs = np.array([1, -2, 3, -4], dtype=np.float32)
>>> rhs = np.array([0, 2, -1, 8],  dtype=np.float32)
>>> output = tops.jit(tgb.add)(lhs, rhs)
>>> # expect: [1., 0., 2., 4.]