tensor是类似于array和matrix的结构，可以在GPU或CPU上运行，在PyTorch中，模型的输入、输出和参数都是tensor格式的。

导入相关的库

  1. import torch
  2. import numpy as np

一、初始化一个tensor

1.1 直接从数据转换

  1. data = [[1, 2], [3, 4]]
  2. x_data = torch.tensor(data)

1.2 从numpy.array转换

  1. np_array = np.array(data)
  2. x_np = torch.from_numpy(np_array)

从tensor转向一个numpy数组

  1. t_array = torch.ones(5)
  2. n_array = t.numpy()

将GPU上的tensor转到CPU上的numpy

  1. cpu_numpy = gpu_tensor.cpu().detatch().numpy()

将tensor数组转至CPU的方法

  1. cpu_tensor = gpu_tensor.cpu()
  2. cpu_tensor = gpu_tensor.to('cpu')

1.3 其他数组传递数组大小

  1. import numpy as np 
  2. import torch
  4. list_data = [[1, 2], [3, 4], [5, 6]]
  5. x_data = torch.tensor(list_data)
  6. x_ones = torch.ones_like(x_data)
  7. x_rand = torch.rand_like(x_data, dtype=torch.float)
  8. print(x_ones, x_rand)

1.4 指定大小创建数组

  1. shape = (2,3,)
  2. rand_tensor = torch.rand(shape)
  3. ones_tensor = torch.ones(shape)
  4. zeros_tensor = torch.zeros(shape)

二、输出tensor的属性

  1. import numpy as np 
  2. import torch
  4. tensor = torch.rand(3,4)
  6. print(f"Shape of tensor: {tensor.shape}")
  7. print(f"Datatype of tensor: {tensor.dtype}")
  8. print(f"Device tensor is stored on: {tensor.device}")

三、tensor转移到GPU上

  1. # We move our tensor to the GPU if available
  2. if torch.cuda.is_available():
  3.     tensor = tensor.to('cuda')

四、tensor的索引

  1. import numpy as np 
  2. import torch
  4. data = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
  5. tensor = torch.tensor(data)
  6. print('First row: ',tensor[0])
  7. print('First column: ', tensor[:, 0])
  8. print('Last column:', tensor[..., -1])
  9. tensor[:,1] = 0        # 第二列全为0
  10. tensor[1] = 0        # 第二行全为0
  11. print(tensor)

五、tensor的拼接

  1. import numpy as np 
  2. import torch
  4. # data = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
  5. data = [[1, 2], [3, 4]]
  6. tensor = torch.tensor(data)
  8. t1 = torch.cat([tensor, tensor, tensor], dim=0)        # 竖着连接3个tensor
  9. t2 = torch.cat([tensor, tensor, tensor], dim=1)        # 横着连接3个tensor

另外有一个函数为torch.stack，但是它和torch.cat有区别，它会增加每次拼接的数组的维度，如下所示：

  1. import numpy as np 
  2. import torch
  4. # data = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
  5. data = [[1, 2], [3, 4]]
  6. tensor = torch.tensor(data)
  8. t1 = torch.stack([tensor, tensor, tensor], dim=0)        # 竖着连接3个tensor
  9. t2 = torch.cat([tensor, tensor, tensor], dim=0)        # 横着连接3个tensor
  10. print(t1)
  11. print(t2)

tensor([[[1, 2],
[3, 4]], [[1, 2],
[3, 4]], [[1, 2],
[3, 4]]]) tensor([[1, 2],
[3, 4],
[1, 2],
[3, 4],
[1, 2],
[3, 4]])

六、计算tensor的相乘和点积

  1. import numpy as np 
  2. import torch
  4. # data = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
  5. data = [[1, 2], [3, 4]]
  6. tensor = torch.tensor(data, dtype=torch.float)
  8. # This computes the matrix multiplication between two tensors. y1, y2, y3 will have the same value
  9. y1 = tensor @ tensor.T
  10. y2 = tensor.matmul(tensor.T)
  12. y3 = torch.rand_like(tensor)
  13. torch.matmul(tensor, tensor.T, out=y3)
  14. print(y1)
  19. # This computes the element-wise product. z1, z2, z3 will have the same value
  20. z1 = tensor * tensor
  21. z2 = tensor.mul(tensor)
  23. z3 = torch.rand_like(tensor)
  24. torch.mul(tensor, tensor, out=z3)
  25. print(z1)

七、tensor的item属性

  1. import numpy as np 
  2. import torch
  4. # data = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
  5. data = [[1, 2], [3, 4]]
  6. tensor = torch.tensor(data,dtype=torch.float)
  8. agg = tensor.sum()
  9. agg_item = agg.item()
  10. print(agg_item, type(agg_item))

八、tensor的in place操作

这个操作我不知道怎么翻译，大致意思就是可以直接对这个tensor数组进行一些操作，

8.1 tensor添加一个数

如这个tensor数组整体加上一个数字：

  1. import numpy as np 
  2. import torch
  4. # data = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
  5. data = [[1, 2], [3, 4]]
  6. tensor = torch.tensor(data,dtype=torch.float)
  8. print(tensor, "\n")
  9. tensor.add_(5)            # 数组中的每个数都加上5
  10. print(tensor)

8.2 tensor的复制

  1. import numpy as np 
  2. import torch
  4. # data = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
  5. data = [[1, 2], [3, 4]]
  6. tensor = torch.tensor(data,dtype=torch.float)
  8. print(tensor, "\n")
  9. tensor_copy = tensor.copy_(tensor)            # 复制这个tensor
  10. tensor_copy.add_(5)
  11. print(tensor_copy)

8.3 tensor的转置

  1. import numpy as np 
  2. import torch
  4. # data = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
  5. data = [[1, 2], [3, 4]]
  6. tensor = torch.tensor(data,dtype=torch.float)
  8. print(tensor, "\n")
  9. tensor_t = tensor.t_()                # 转置数组
  10. print(tensor_t)