coding 记录(二）

torch.expand()和torch.repeat() 区别

torch.expand()

对指定的维度进行数值大小的改变,只能改变维大小为1的维，（单数维）不改变的维可以传入-1或者原来的数值。 返回张量的一个新视图，（并不会重新分配内存）可以将张量的单个维度扩大为更大的尺寸。

torch.repeat()
沿着指定的维度重复tensor, 不同于expand()，本函数复制的是tensor中的数据

x.repeat(3,1) 在轴0（行））上复制3份，在轴1（列）上复制1份

a=torch.Tensor(1,256,18,18)
a.shape
torch.Size([1,256,18,18])
b=a.repeat(2,1,1,1)
b.shape
torch.Size([2,256,18,18])
b=a.repeat(2)
则会报错：
RuntimeError: Number of dimensions of repeat dims can not be smaller than number of dimensions of tensor

区别：当复制的数据发生变化时，repeat后的数据不会变化，而expand后的数据发生变化

torch.tensor()和torch.Tensor()

在Pytorch中，Tensor和tensor都用于生成新的张量,torch.Tensor()是Python类，更明确的说，是默认张量类型torch.FloatTensor()的别名,torch.tensor()是Python函数。

>>> y=torch.Tensor([2,2,2])
>>> y
tensor([2., 2., 2.])
>>> y.shape
torch.Size([3])
>>> y.type()
'torch.FloatTensor'


>>> x=torch.tensor([2,2,2])
>>> x
tensor([2, 2, 2])
>>> x.shape
torch.Size([3])
>>> x.type()
'torch.LongTensor'


>>> x=torch.tensor(1,64,1,1)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: tensor() takes 1 positional argument but 4 were given
>>> x=torch.Tensor(1,64,1,1)
>>> x.shape
torch.Size([1, 64, 1, 1])
>>> x.type()
'torch.FloatTensor'

numpy_to_torch() && torch_to_numpy()

def numpy_to_torch(a:np.ndarray):
   return torch.from_numpy(a).float().permute(2,0,1).unsqueeze(0)
def torch_to_numpy(a:torch.Tensor):
    return a.squeeze(0).permute(1,2,0).numpy()

vot数据集的anno由八个值（Polygon)变为四个值（Rectangle–x,y,w,h）表示

import numpy as np


def convert_vot_anno_to_rect(vot_anno, type):
    if len(vot_anno) == 4:
        return vot_anno

    if type == 'union':
        x1 = min(vot_anno[0::2])
        x2 = max(vot_anno[0::2])
        y1 = min(vot_anno[1::2])
        y2 = max(vot_anno[1::2])
        return [x1, y1, x2 - x1, y2 - y1]
    elif type == 'preserve_area':
        if len(vot_anno) != 8:
            raise ValueError

        vot_anno = np.array(vot_anno)
        cx = np.mean(vot_anno[0::2])
        cy = np.mean(vot_anno[1::2])

        x1 = min(vot_anno[0::2])
        x2 = max(vot_anno[0::2])
        y1 = min(vot_anno[1::2])
        y2 = max(vot_anno[1::2])

        A1 = np.linalg.norm(vot_anno[0:2] - vot_anno[2: 4]) * np.linalg.norm(vot_anno[2: 4] - vot_anno[4:6])
        A2 = (x2 - x1) * (y2 - y1)
        s = np.sqrt(A1 / A2)
        w = s * (x2 - x1) + 1
        h = s * (y2 - y1) + 1

        x = cx - 0.5*w
        y = cy - 0.5*h
        return [x, y, w, h]
    else:
        raise ValueError

argsort（）

>>> import numpy as np
>>> x=np.array([1,2,8,-4,-1,9])
>>> x
array([ 1,  2,  8, -4, -1,  9])
>>> x.argsort()
array([3, 4, 0, 1, 2, 5])
#将x中的元素从小到大排列，提取其对应的index(索引)，然后输出到y
#那如果想从大到小排列呢
>>> x.argsort()[::-1]
array([5, 2, 1, 0, 4, 3])
>>> x.argsort()[0]
3    #找到最小的索引
>>> x.argsort()[-1]
5   #找到最大的索引
>>>