如何有效地检索 Torch 张量中最大值的索引?

Posted

技术标签:

【中文标题】如何有效地检索 Torch 张量中最大值的索引?【英文标题】:How to efficiently retrieve the indices of maximum values in a Torch tensor? 【发布时间】:2019-04-12 05:49:54 【问题描述】:

假设有一个火炬张量,例如以下形状:

x = torch.rand(20, 1, 120, 120)

我现在想要的是获取每个 120x120 矩阵的最大值的索引。为了简化问题,我首先将x.squeeze() 使用形状[20, 120, 120]。然后我想得到火炬张量,它是一个形状为[20, 2]的索引列表。

我怎样才能快速做到这一点?

【问题讨论】:

为什么需要[20, 2] 矩阵。您是否想要每个 120 * 120 矩阵的行最大值和列最大值? 是的,或者换句话说:对于 20 个120 * 120 矩阵中的每一个,我想要具有最大值的单元格的[x, y] 坐标 如果你想知道***k元素的索引,使用torch.topk()。 这能回答你的问题吗? Extracting the top-k value-indices from a 1-D Tensor 【参考方案1】:

torch.topk() 是您正在寻找的。从文档中,

torch.topk(input, k, dim=None, largest=True, sorted=True, out=None) -> (Tensor, LongTensor)

返回给定input张量的k最大元素 给定的维度。

如果没有给出dim,则选择输入的最后一个维度。

如果largestFalse,则返回k 个最小的元素。

返回一个 (values, indices) 的命名元组,其中索引是原始输入张量中元素的索引。

布尔选项sorted如果True,将确保返回的k个元素本身是排序的

【讨论】:

有用的功能知道,但它没有回答原来的问题。 OP 希望获得该矩阵中最大元素的 20 个 120x120 矩阵中的每一个的索引。也就是说,她想要 20 个 2D 坐标,每个矩阵一个。 topk 仅返回最大化维度中最大元素的索引。 请注意,topk 的文档在返回索引的含义方面存在混淆。它给人的印象是该函数为原始张量提供索引,而实际上它仅返回最大化维度的索引。请参阅试图澄清它的 pytorch 问题github.com/pytorch/pytorch/issues/50331#issue-782748956。【参考方案2】:

如果我理解正确,您不需要值,而是需要索引。不幸的是,没有开箱即用的解决方案。有一个 argmax() 函数,但我看不出如何让它完全按照你的意愿去做。

所以这里有一个小变通方法,因为我们只是划分张量,所以效率应该还可以:

n = torch.tensor(4)
d = torch.tensor(4)
x = torch.rand(n, 1, d, d)
m = x.view(n, -1).argmax(1)
# since argmax() does only return the index of the flattened
# matrix block we have to calculate the indices by ourself 
# by using / and % (// would also work, but as we are dealing with
# type torch.long / works as well
indices = torch.cat(((m / d).view(-1, 1), (m % d).view(-1, 1)), dim=1)
print(x)
print(indices)

n 代表您的第一个维度,d 代表最后两个维度。我在这里取较小的数字来显示结果。但当然这也适用于n=20d=120

n = torch.tensor(20)
d = torch.tensor(120)
x = torch.rand(n, 1, d, d)
m = x.view(n, -1).argmax(1)
indices = torch.cat(((m / d).view(-1, 1), (m % d).view(-1, 1)), dim=1)
#print(x)
print(indices)

这是n=4d=4 的输出:

tensor([[[[0.3699, 0.3584, 0.4940, 0.8618],
          [0.6767, 0.7439, 0.5984, 0.5499],
          [0.8465, 0.7276, 0.3078, 0.3882],
          [0.1001, 0.0705, 0.2007, 0.4051]]],


        [[[0.7520, 0.4528, 0.0525, 0.9253],
          [0.6946, 0.0318, 0.5650, 0.7385],
          [0.0671, 0.6493, 0.3243, 0.2383],
          [0.6119, 0.7762, 0.9687, 0.0896]]],


        [[[0.3504, 0.7431, 0.8336, 0.0336],
          [0.8208, 0.9051, 0.1681, 0.8722],
          [0.5751, 0.7903, 0.0046, 0.1471],
          [0.4875, 0.1592, 0.2783, 0.6338]]],


        [[[0.9398, 0.7589, 0.6645, 0.8017],
          [0.9469, 0.2822, 0.9042, 0.2516],
          [0.2576, 0.3852, 0.7349, 0.2806],
          [0.7062, 0.1214, 0.0922, 0.1385]]]])
tensor([[0, 3],
        [3, 2],
        [1, 1],
        [1, 0]])

我希望这是你想要的! :)

编辑:

这是一个稍微修改过的,可能会更快(我猜不是很多:),但它更简单更漂亮:

而不是像以前那样:

m = x.view(n, -1).argmax(1)
indices = torch.cat(((m // d).view(-1, 1), (m % d).view(-1, 1)), dim=1)

已经对argmax 值进行了必要的整形:

m = x.view(n, -1).argmax(1).view(-1, 1)
indices = torch.cat((m // d, m % d), dim=1)

但正如 cmets 中所述。我认为不可能从中得到更多。

你可以做的一件事,如果真的对你来说获得最后一点可能的性能改进很重要,那就是将上面的函数实现为低级扩展(如C++) 用于 pytorch。

这只会给你一个你可以调用它的函数,并且会避免缓慢的 python 代码。

https://pytorch.org/tutorials/advanced/cpp_extension.html

【讨论】:

是的,这就是我想要的输出。我修改它以将m 转换为.float(),然后使用// 除以d。你提出的是一个解开,类似于numpy.unravel_indices()。如果你能想到一个更快的方法当然会更好。 @Chris 我刚刚做了一个短时间的测试。实际上我认为它非常有效,我想目前没有更快的方法:调用argmax() 本身需要大约 10 倍,只要计算下一行中的索引 - 在 CPU 上,我也可以稍后在 GPU 上检查。但是操作真的很简单而且很严格,所以即使这是一种解决方法,从理论的角度来看它也应该是相当有效的。 不,它一点也不慢,我在 Telsa Volta 上需要大约 5.5 毫秒。我只需要最大化它,但我同意,argmax 是一个线性运算,因为张量是无序的。可能这是最慢的组件,无法加速。 @Chris 最后我做了一个小编辑,版本稍微好一点。但我不希望在性能方面真正有所作为,可能大致相同 - 领先半纳秒。如果充分利用它真的很重要,您可能希望使用 C++ 中的自定义扩展。但考虑到代码的小sn-p,收益可能不会那么多。 谢谢,效果很好。我在评估中也犯了一个错误,似乎只是 0.5 毫秒而不是 5 毫秒。【参考方案3】:

这是torch 中的unravel_index 实现:

def unravel_index(
    indices: torch.LongTensor,
    shape: Tuple[int, ...],
) -> torch.LongTensor:
    r"""Converts flat indices into unraveled coordinates in a target shape.

    This is a `torch` implementation of `numpy.unravel_index`.

    Args:
        indices: A tensor of (flat) indices, (*, N).
        shape: The targeted shape, (D,).

    Returns:
        The unraveled coordinates, (*, N, D).
    """

    coord = []

    for dim in reversed(shape):
        coord.append(indices % dim)
        indices = indices // dim

    coord = torch.stack(coord[::-1], dim=-1)

    return coord

然后,您可以使用torch.argmax 函数来获取“扁平化”张量的索引。

y = x.view(20, -1)
indices = torch.argmax(y)
indices.shape  # (20,)

并使用 unravel_index 函数解开索引。

indices = unravel_index(indices, x.shape[-2:])
indices.shape  # (20, 2)

【讨论】:

这是最接近真实、通用的答案!要更直接地回答原始问题,即询问如何获取最大值的索引,您可能需要编辑以显示如何首先使用 argmax 获取索引,然后解开它们。 我最终不得不对与 argmax 的连接进行编码,所以请检查我的答案。随意将我所做的纳入您的。【参考方案4】:

接受的答案仅适用于给定的示例。

tejasvi88 的回答很有趣,但无助于回答最初的问题(正如我在那里的评论中所解释的那样)。

我相信弗朗索瓦的答案是最接近的,因为它处理的是更通用的情况(任意数量的维度)。但是,它没有与argmax 连接,并且显示的示例没有说明该函数处理批处理的能力。

因此,我将在此处以弗朗索瓦的回答为基础,并添加代码以连接到 argmax。我编写了一个新函数batch_argmax,它返回批处理中最大值的索引。批次可以按多个维度进行组织。我还包括一些测试用例来说明:

def batch_argmax(tensor, batch_dim=1):
    """
    Assumes that dimensions of tensor up to batch_dim are "batch dimensions"
    and returns the indices of the max element of each "batch row".
    More precisely, returns tensor `a` such that, for each index v of tensor.shape[:batch_dim], a[v] is
    the indices of the max element of tensor[v].
    """
    if batch_dim >= len(tensor.shape):
        raise NoArgMaxIndices()
    batch_shape = tensor.shape[:batch_dim]
    non_batch_shape = tensor.shape[batch_dim:]
    flat_non_batch_size = prod(non_batch_shape)
    tensor_with_flat_non_batch_portion = tensor.reshape(*batch_shape, flat_non_batch_size)

    dimension_of_indices = len(non_batch_shape)

    # We now have each batch row flattened in the last dimension of tensor_with_flat_non_batch_portion,
    # so we can invoke its argmax(dim=-1) method. However, that method throws an exception if the tensor
    # is empty. We cover that case first.
    if tensor_with_flat_non_batch_portion.numel() == 0:
        # If empty, either the batch dimensions or the non-batch dimensions are empty
        batch_size = prod(batch_shape)
        if batch_size == 0:  # if batch dimensions are empty
            # return empty tensor of appropriate shape
            batch_of_unraveled_indices = torch.ones(*batch_shape, dimension_of_indices).long()  # 'ones' is irrelevant as it will be empty
        else:  # non-batch dimensions are empty, so argmax indices are undefined
            raise NoArgMaxIndices()
    else:   # We actually have elements to maximize, so we search for them
        indices_of_non_batch_portion = tensor_with_flat_non_batch_portion.argmax(dim=-1)
        batch_of_unraveled_indices = unravel_indices(indices_of_non_batch_portion, non_batch_shape)

    if dimension_of_indices == 1:
        # above function makes each unraveled index of a n-D tensor a n-long tensor
        # however indices of 1D tensors are typically represented by scalars, so we squeeze them in this case.
        batch_of_unraveled_indices = batch_of_unraveled_indices.squeeze(dim=-1)
    return batch_of_unraveled_indices


class NoArgMaxIndices(BaseException):

    def __init__(self):
        super(NoArgMaxIndices, self).__init__(
            "no argmax indices: batch_argmax requires non-batch shape to be non-empty")

这里是测试:

def test_basic():
    # a simple array
    tensor = torch.tensor([0, 1, 2, 3, 4])
    batch_dim = 0
    expected = torch.tensor(4)
    run_test(tensor, batch_dim, expected)

    # making batch_dim = 1 renders the non-batch portion empty and argmax indices undefined
    tensor = torch.tensor([0, 1, 2, 3, 4])
    batch_dim = 1
    check_that_exception_is_thrown(lambda: batch_argmax(tensor, batch_dim), NoArgMaxIndices)

    # now a batch of arrays
    tensor = torch.tensor([[1, 2, 3], [6, 5, 4]])
    batch_dim = 1
    expected = torch.tensor([2, 0])
    run_test(tensor, batch_dim, expected)

    # Now we have an empty batch with non-batch 3-dim arrays' shape (the arrays are actually non-existent)
    tensor = torch.ones(0, 3)  # 'ones' is irrelevant since this is empty
    batch_dim = 1
    # empty batch of the right shape: just the batch dimension 0,since indices of arrays are scalar (0D)
    expected = torch.ones(0)
    run_test(tensor, batch_dim, expected)

    # Now we have an empty batch with non-batch matrices' shape (the matrices are actually non-existent)
    tensor = torch.ones(0, 3, 2)  # 'ones' is irrelevant since this is empty
    batch_dim = 1
    # empty batch of the right shape: the batch and two dimension for the indices since we have 2D matrices
    expected = torch.ones(0, 2)
    run_test(tensor, batch_dim, expected)

    # a batch of 2D matrices:
    tensor = torch.tensor([[[1, 2, 3], [6, 5, 4]], [[2, 3, 1], [4, 5, 6]]])
    batch_dim = 1
    expected = torch.tensor([[1, 0], [1, 2]])  # coordinates of two 6's, one in each 2D matrix
    run_test(tensor, batch_dim, expected)

    # same as before, but testing that batch_dim supports negative values
    tensor = torch.tensor([[[1, 2, 3], [6, 5, 4]], [[2, 3, 1], [4, 5, 6]]])
    batch_dim = -2
    expected = torch.tensor([[1, 0], [1, 2]])
    run_test(tensor, batch_dim, expected)

    # Same data, but a 2-dimensional batch of 1D arrays!
    tensor = torch.tensor([[[1, 2, 3], [6, 5, 4]], [[2, 3, 1], [4, 5, 6]]])
    batch_dim = 2
    expected = torch.tensor([[2, 0], [1, 2]])  # coordinates of 3, 6, 3, and 6
    run_test(tensor, batch_dim, expected)

    # same as before, but testing that batch_dim supports negative values
    tensor = torch.tensor([[[1, 2, 3], [6, 5, 4]], [[2, 3, 1], [4, 5, 6]]])
    batch_dim = -1
    expected = torch.tensor([[2, 0], [1, 2]])
    run_test(tensor, batch_dim, expected)


def run_test(tensor, batch_dim, expected):
    actual = batch_argmax(tensor, batch_dim)
    print(f"batch_argmax of tensor with batch_dim batch_dim is\nactual\nExpected:\nexpected")
    assert actual.shape == expected.shape
    assert actual.eq(expected).all()

def check_that_exception_is_thrown(thunk, exception_type):
    if isinstance(exception_type, BaseException):
        raise Exception(f"check_that_exception_is_thrown received an exception instance rather than an exception type: "
                        f"exception_type")
    try:
        thunk()
        raise AssertionError(f"Should have thrown exception_type")
    except exception_type:
        pass
    except Exception as e:
        raise AssertionError(f"Should have thrown exception_type but instead threw e")

【讨论】:

【参考方案5】:
ps=ps.numpy()
ps=ps.tolist()

mx=[max(l) for l in ps]
mx=max(mx)
for i in range(len(ps[0])):
  if mx==ps[0][i]:
    print("The digit is "+str(i))
    break

这对我很有效

【讨论】:

以上是关于如何有效地检索 Torch 张量中最大值的索引?的主要内容,如果未能解决你的问题,请参考以下文章

如何有效地乘以重复行的火炬张量而不将所有行存储在内存中或迭代?

PyTorch: 张量的拼接切分索引

pytorch - torch.gather 的倒数

colab pytorch张量操作

如何将张量列表转换为 Torch::Tensor?

PyTorch 1.0 中文文档:torch.sparse