版权声明：本文为博主原创文章，遵循 CC 4.0 BY 版权协议，转载请附上原文出处链接和本声明。
在深度学习中进行张量操作的时候经常会遇到这样的切片形式，一开始总会感到头疼，所以现在总结一下用法，欢迎大家进行补充。
data[x:y:z]
data[::-1]
data[:-1]
data[-3:]
data[1:4, 2:4]
data[…, 1:2]
data[:, None]
这是最近在看的代码，进行loss的计算，用到了很多切片的知识，所以进行总结，希望看完这篇博客以后大家也能看懂类似的代码
def compute_loss(p, targets, model): # predictions, targets, model
ft = torch.cuda.FloatTensor if p[0].is_cuda else torch.Tensor
lxy, lwh, lcls, lconf = ft([0]), ft([0]), ft([0]), ft([0])
# build_targets对targets向量进行处理
txy, twh, tcls, indices = build_targets(model, targets)
# Define criteria
MSE = nn.MSELoss()
CE = nn.CrossEntropyLoss() # (weight=model.class_weights)
BCE = nn.BCEWithLogitsLoss()
# Compute losses
h = model.hyp # hyperparameters
bs = p[0].shape[0] # batch size
k = bs # loss gain
for i, pi0 in enumerate(p): # layer i predictions, i
b, a, gj, gi = indices[i] # image, anchor, gridy, gridx
tconf = torch.zeros_like(pi0[..., 0]) # conf
# Compute losses
if len(b): # number of targets
pi = pi0[b, a, gj, gi] # predictions closest to anchors
tconf[b, a, gj, gi] = 1 # conf
# pi[..., 2:4] = torch.sigmoid(pi[..., 2:4]) # wh power loss (uncomment)
lxy += (k * h['xy']) * MSE(torch.sigmoid(pi[..., 0:2]), txy[i]) # xy loss
lwh += (k * h['wh']) * MSE(pi[..., 2:4], twh[i]) # wh yolo loss
lcls += (k * h['cls']) * CE(pi[..., 5:], tcls[i]) # class_conf loss
# pos_weight = ft([gp[i] / min(gp) * 4.])
# BCE = nn.BCEWithLogitsLoss(pos_weight=pos_weight)
lconf += (k * h['conf']) * BCE(pi0[..., 4], tconf) # obj_conf loss
loss = lxy + lwh + lconf + lcls
return loss, torch.cat((lxy, lwh, lconf, lcls, loss)).detach()
1. 单列表切片
对于一个普遍形式：data[x:y:z] 讲一下他的意义：
The syntax
[x:y:z]means "take everyzth element of a list from indexxto indexy". Whenzis negative, it indicates going backwards. Whenxisn’t specified, it defaults to the first element of the list in the direction you are traversing the list. Whenyisn’t specified, it defaults to the last element of the list. So if we want to take every 2th element of a list, we use[::2].
– from github:python-is-cool
翻译过来就是：x是起始坐标，y是结束坐标（不包含），z是代表每隔z个数取一个数,如果z是负数代表方向是从后往前。
看一个例子：
>>> data = list(range(10))
>>> data
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
>>> data[1:5] # 代表[1,5)
[1, 2, 3, 4]
>>> data[-3:] # 代表从倒数第3个到最后， 方向默认还是从前往后
[7, 8, 9]
>>> data[:-3] # 代表从第1个到倒数第4个，方向还是默认从前向后
[0, 1, 2, 3, 4, 5, 6]
>>> data[1:8:2] # 从[1,2,3,4,5,6,7]中每两个选一个，得到[1,3,5,7]
[1, 3, 5, 7]
>>> data[1:8:-2] # 注意如果z是负数，那么前边也需要是负数，方向需要正确
[]
>>> data[::-2] # 两个冒号代表取全体
[9, 7, 5, 3, 1]
>>> data[-1:-8:-2]
# 注意data[-1:-8:-1]=[9,8,7,6,5,4,3]意思是倒数第1个到倒数第7个数
# 这里的x,y依然意义是起始和结尾
[9, 7, 5, 3]
>>>
总结：
-
注意起始和结尾的[x,y] 实际的数学意义是[x,y), 前取到，后取不到
-
注意z是负数的情况下，整个数列方向是从后往前数的，对应的[x,y]也应该是负数，从后往前，这里的x,y意义依然是起始和结尾。
-
双冒号
::代表取全部 -
注意整个数列方向问题，默认都是从前往后，只有z的正负性能控制方向。
-
只有一个冒号的情况是两个冒号的特殊情况，省略了最后一个冒号，如：
>>> data[3:] [3, 4, 5, 6, 7, 8, 9] >>> data[3::1] # 默认方向从前往后，且z的值为1 [3, 4, 5, 6, 7, 8, 9] >>>
2. 多维列表
data[..., 1:2] data[1:2, 2:4, 5:9] data[:,None] data[1:2, 2:4] == data [1:2][2:4] ? data[:]
干货：
- 对于三个点的情况, 代表前边所有维度
- 对于逗号分隔的情况，每个逗号将以上一个单独的分割开来
- None代表增加一个维度
- data[1:2, 2:4] 与 data [1:2] [2:4] 并不相等
- data[:]代表取全体数据
举例子：
- data[…, 1:2] 举例
>>> import numpy as np
>>> data1 = np.arange(27).reshape(3,3,3)
>>> data1
array([[[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8]],
[[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]],
[[18, 19, 20],
[21, 22, 23],
[24, 25, 26]]])
>>> data1[...,1:3]
array([[[ 1, 2],
[ 4, 5],
[ 7, 8]],
[[10, 11],
[13, 14],
[16, 17]],
[[19, 20],
[22, 23],
[25, 26]]])
>>> data1[:,:,1:3]
array([[[ 1, 2],
[ 4, 5],
[ 7, 8]],
[[10, 11],
[13, 14],
[16, 17]],
[[19, 20],
[22, 23],
[25, 26]]])
# 可以看出data1[:,:,1:3]与data1[...,1:3]是等价的，三个点的存在简化了前边维度，
# 不管多少维度存在都可以使用，取代前边所有维度。
- data[1:2, 2:4, 5:9]
>>> import numpy as np
>>> data1 = np.arange(27).reshape(3,3,3)
>>> data1[0:2]
array([[[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8]],
[[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]]])
>>> data1[0:2,0:2]
array([[[ 0, 1, 2],
[ 3, 4, 5]],
[[ 9, 10, 11],
[12, 13, 14]]])
>>> data1[0:2,0:2,0:2]
array([[[ 0, 1],
[ 3, 4]],
[[ 9, 10],
[12, 13]]])
# 一步一步的切片可以看出，每一步都是单独的一个单列表切片，满足单列表切片原则
- data[:,None]
>>> data1.shape
(3, 3, 3)
>>> data1[:,None].shape
(3, 1, 3, 3)
>>> data1[...,None].shape
(3, 3, 3, 1)
>>> data1[None,...].shape
(1, 3, 3, 3)
# None代表增加一个维度,通常使用在数据预处理，给图片添加batch列
- data[1:2, 2:4] == data[1：2] [2:4]
>>> import numpy as np
>>> data1 = np.arange(27)
>>> data1
array([ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,17, 18, 19, 20, 21, 22, 23, 24, 25, 26])
>>> data1[1:7][3:6] # 两个可以理解为切片的切片
array([4, 5, 6])
>>> data1[1:7,3:6] # 只有一维，所以无法运行
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
IndexError: too many indices for array
>>> data1 = np.arange(27).reshape(3,3,3)
>>> data1
array([[[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8]],
[[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]],
[[18, 19, 20],
[21, 22, 23],
[24, 25, 26]]])
>>> data1[0:2]
array([[[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8]],
[[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]]])
>>> data1[0:2][1:3]
# 切片的切片， 且过一次以后就只有0,1两个维度了，所以只能取第一维
array([[[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]]])
>>> data1[0:2,1:3] # 所有满足第一维为0,1 第二维为1,2的
array([[[ 3, 4, 5],
[ 6, 7, 8]],
[[12, 13, 14],
[15, 16, 17]]])
>>>
- data[:] 表示取全体
>>> data1[:]
array([[[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8]],
[[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]],
[[18, 19, 20],
[21, 22, 23],
[24, 25, 26]]])
>>> data1[:,:,:]
array([[[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8]],
[[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]],
[[18, 19, 20],
[21, 22, 23],
[24, 25, 26]]])
>>>
后记：这篇博客总结了目前常见到的问题，在深度学习中经常用到，tensor的切片常常让人有点微微头疼，所以特意总结一下，大家遇见新的切片方式可以在评论中补充，有问题的也可以提出。之后还有一些张量操作会单独出一篇博客。