将列表转换为张量更准确

Question

我想return在Dataloader中列出一个列表。但是对于 return 它，它需要是张量吧？所以我对它进行了转换，但是在这个过程中信息丢失了，还有其他方法吗？

pt_tensor_from_list = torch.tensor(pose_transform)
pt_tensor_from_list = torch.FloatTensor(pose_transform)

我期望输出：

([[-0.0003000000142492354, -0.0008999999845400453, 0.00039999998989515007, 0], [0.0010000000474974513, -0.00019999999494757503, 0.0003000000142492354, 0], [0.00019999999494757503, -0.0005000000237487257, -0.0008999999845400453, 0], [5.484399795532227, -24.28619956970215, 117.5000991821289, 1])

但它是：

([[ -0.0003, -0.0009, 0.0004, 0.0000], [ 0.0010, -0.0002, 0.0003, 0.0000], [ 0.0002, -0.0005, -0.0009, 0.0000], [ 5.4844, -24.2862, 117.5001, 1.0000]])

Answer 1

在此类转换过程中，您不会丢失任何信息。它看起来更紧凑的原因是，当您打印张量时，它会调用 __str__() 或 __repr__() 方法，这会使您的张量看起来更漂亮。正如您可以找到 here torch.Tensor uses a kind of internal tensor formatter called _tensor_str. If you look inside the code link 一样，您会发现默认情况下参数 precision 设置为 4:

precision: Number of digits of precision for floating point output (default = 4).

这就是为什么在打印张量时只有 4 位数字的张量值。但实际上，存储在张量中的值与原始列表中的值相同。

这里是一个小例子来获得一个想法：

代码：

import torch

test_list = ([[-0.0003000000142492354, -0.0008999999845400453, 0.00039999998989515007, 0],
              [0.0010000000474974513, -0.00019999999494757503, 0.0003000000142492354, 0],
              [0.00019999999494757503, -0.0005000000237487257, -0.0008999999845400453, 0],
              [5.484399795532227, -24.28619956970215, 117.5000991821289, 1]])

print('Original values:')
for i in test_list:
    for j in i:
        print(j)

pt_tensor_from_list = torch.FloatTensor(test_list)

print('When printing FloatTensor:')
print(pt_tensor_from_list.dtype, pt_tensor_from_list, sep='\n')

print('When printing each value separately:')
for i in pt_tensor_from_list:
    for j in i:
        print(j.item())

输出：

Original values:
-0.0003000000142492354
-0.0008999999845400453
0.00039999998989515007
0
0.0010000000474974513
-0.00019999999494757503
0.0003000000142492354
0
0.00019999999494757503
-0.0005000000237487257
-0.0008999999845400453
0
5.484399795532227
-24.28619956970215
117.5000991821289
1
When printing FloatTensor:
torch.float32
tensor([[-3.0000e-04, -9.0000e-04,  4.0000e-04,  0.0000e+00],
        [ 1.0000e-03, -2.0000e-04,  3.0000e-04,  0.0000e+00],
        [ 2.0000e-04, -5.0000e-04, -9.0000e-04,  0.0000e+00],
        [ 5.4844e+00, -2.4286e+01,  1.1750e+02,  1.0000e+00]])
When printing each value separately:
-0.0003000000142492354
-0.0008999999845400453
0.00039999998989515007
0.0
0.0010000000474974513
-0.00019999999494757503
0.0003000000142492354
0.0
0.00019999999494757503
-0.0005000000237487257
-0.0008999999845400453
0.0
5.484399795532227
-24.28619956970215
117.5000991821289
1.0

如您所见，我们在分别打印每个值时得到相同的值。

但是如果您选择了错误的张量类型，您可能会丢失一些信息，例如 HalfTensor 而不是 FloatTensor。这是一个例子：

代码：

pt_tensor_from_list = torch.HalfTensor(test_list)

print('When printing HalfTensor:')
print(pt_tensor_from_list.dtype, pt_tensor_from_list, sep='\n')

print('When printing each value separately:')
for i in pt_tensor_from_list:
    for j in i:
        print(j.item())

输出：

When printing HalfTensor:
torch.float16
tensor([[-2.9993e-04, -8.9979e-04,  4.0007e-04,  0.0000e+00],
        [ 1.0004e-03, -2.0003e-04,  2.9993e-04,  0.0000e+00],
        [ 2.0003e-04, -5.0020e-04, -8.9979e-04,  0.0000e+00],
        [ 5.4844e+00, -2.4281e+01,  1.1750e+02,  1.0000e+00]],
       dtype=torch.float16)
When printing each value separately:
-0.0002999305725097656
-0.0008997917175292969
0.0004000663757324219
0.0
0.0010004043579101562
-0.00020003318786621094
0.0002999305725097656
0.0
0.00020003318786621094
-0.0005002021789550781
-0.0008997917175292969
0.0
5.484375
-24.28125
117.5
1.0

您现在可以注意到，这些值（有一点）不同。访问 pytorch tensor docs 了解更多关于不同类型的 torch.tensor。

将列表转换为张量更准确

Transform List to Tensor more accurat

transform

list

pytorch

tensor