Pytorch中tensor与numpy的数据转换

import numpy as np
import torch

# cpu tensor 转 numpy
arr = torch.zeros(3, device='cpu')  # 等同于 arr = torch.zeros(3)
print(arr)  # tensor([0., 0., 0.])
arr2 = arr.numpy()
print(arr2)  # [0. 0. 0.]

# numpy 转 cpu tensor
num = np.zeros(3)
print(num)  # [0. 0. 0.]
num2 = torch.from_numpy(num)
print(num2)  # tensor([0., 0., 0.], dtype=torch.float64)

# GPU tensor 转 numpy
arr3 = torch.zeros(3, device='cuda')
print(arr3)  # tensor([0., 0., 0.], device='cuda:0')
arr4 = arr3.detach().cpu().numpy()
print(arr4)  # [0. 0. 0.]

  二值交叉损失函数 binary_cross_entropy

import numpy as np
import torch
import torch.nn.functional as F

if __name__ == "__main__":
    print(1)
    inputs = torch.randn(3, 2)
    target = torch.rand(3, 2)
    loss = F.binary_cross_entropy(torch.sigmoid(inputs), target)
    print(inputs.shape)  # torch.Size([3, 2])
    print(target.shape)  # torch.Size([3, 2])
    print(loss)  # tensor(0.7674)

    data = []
    label = []
    for i in range(4):
        data.append(inputs.numpy())
        label.append(target.numpy())
    data_tenser = torch.from_numpy(np.array(data))
    label_tenser = torch.from_numpy(np.array(label))
    print(data_tenser.shape)  # torch.Size([4, 3, 2])
    print(label_tenser.shape)  # torch.Size([4, 3, 2])
    error = F.binary_cross_entropy(torch.sigmoid(data_tenser), label_tenser)
    print(error)  # tensor(0.7674)