pytorch使用笔记一

pytorch 常见数据类型

tensor，variable，numpy

numpy->tensor	tensor->numpy
x_numpy=np.arange(100).reshape(10,10) x_tensor=torch.from_numpy(x_numpy)[默认存放在cpu中] 转换后的tensor与numpy指向同一地址	x_tensor=torch.randn((3, 2)) x_numpy=x_tensor.numpy() 转换后的tensor与numpy指向同一地址
tensor->variable	variable->tensor
from torch.autograd import Variable x_variable = Variable(x_tensor, requires_grad=True)	x_tensor =x_variable.detach()
cpu variable->gpu variable	gpu variable->cpu variable
if torch.cuda.is_available(): gpu_variable= x_variable.cuda()	cpu_variable = gpu_variable.cpu()

pytorch 强制数据类型转换

scalar，tensor中某一个位置的值	tensor 张量
x_scalar = x_tensor[0,0]	x_tensor = torch.randn(3,2)
float tensor -> integer tensor [in place]	float tensor -> integer tensor
x_float32=torch.randn((3,2),dtype=torch.float32)	x_int = x_float32.type(torch.int32) 如果x_float32 为variable变量并带有梯度时会直接剥离原来的梯度

torch.masked_select(input, mask, out=None)

>>> x = torch.randn(3,4)
>>> x
tensor([[-0.6089,  1.0113, -0.5017,  0.0393],
        [-0.8978, -0.0118,  0.3297, -0.4590],
        [ 0.9305, -0.6148, -0.8959,  1.0758]])
>>> mask = x.ge(0.5)
>>> mask
tensor([[False,  True, False, False],
        [False, False, False, False],
        [ True, False, False,  True]])
>>> torch.masked_select(x, mask)
tensor([1.0113, 0.9305, 1.0758])

torch 比较连个tensor的大小

torch.equal(tensor1, tensor2),#返回值为 True or False 为scaler
torch.gt(tensor1, tensor2)，#逐位比较，返回的仍然时bool Tensor
torch.ge(tensor1,tensor2)，#逐位比较，返回的仍然时bool Tensor