Skip to content

Pytorch

numpy 转到 torch

numpy_tensor = np.random.randn(10, 20)
pytorch_tensor1 = torch.Tensor(numpy_tensor)
pytorch_tensor2 = torch.from_numpy(numpy_tensor)

参考：L1aoXingyu/code-of-learn-deep-learning-with-pytorch: This is code of book "Learn Deep Learning with PyTorch"

运算类

矩阵相乘

torch.mm()

矩阵相乘用torch.mm(a,b)。

AB = A.mm(B) # computes A.B (matrix multiplication)
# or
AB = torch.mm(A, B)
# or
AB = torch.matmul(A, B)
# or, even simpler
AB = A @ B # Python 3.5+

不要用torch.dot(a,b)，因为dot()会将矩阵转为一维再操作，向量长不相同时会报错。

torch.dot()与numpy.dot()功能有差异，具体讨论参见：torch dot function consistent with numpy · Issue #138 · pytorch/pytorch

torch.mul()

矩阵上对应点（对应元素）相乘。

weights = torch.FloatTensor([1.0, -1.0, 2.0])
inputs = torch.FloatTensor([2, 3, 4])
inputs.mul(weights)

torch.matmul()

高维矩阵运算，支持广播。

>>> tensor1 = torch.randn(3, 4)
>>> tensor2 = torch.randn(4)
>>> torch.matmul(tensor1, tensor2).size()

torch — PyTorch master documentation

详见：torch.Tensor 的 4 种乘法 - da_kao_la 的博客 - CSDN 博客

学习资料 📚