yolov2-----源码阅读【reorg层】

Bertha 。 2021-11-22 14:32 366阅读 0赞

官方yolov2 结构如图  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl8zODYzMjI0Ng_size_16_color_FFFFFF_t_70]  
reorg层的作用是将上一层卷积后的特征图38X38X64 拉伸成19X19X256，具体看代码：

elif module_type == "reorg":
     	 stride = int(module["stride"]) #stride 为2
     	 B,C,H,W = x.size()  #B为bach大小暂设为32，c为通道数64 H，W都为38
    	 x = x.view(B, C, int(H / stride), stride, int(W / stride), stride).transpose(3, 4).contiguous()
    	   #将特征图#32 64 38 38 -----------> #调整并转置为#32 64 19 2 19 2
    	 x = x.view(B, C, int(H / stride * W / stride), int(stride * stride)).transpose(2, 3).contiguous()
    	 #32 64 19 2 19 2 --------------> #32 64 361 4
     	 x = x.view(B, C, int(stride * stride), int(H / stride), int(W / stride)).transpose(1, 2).contiguous()
     	 # 32 64 361 4 --------------> #32 64 4 19 19
         x = x.view(B, int(stride * stride * C), int(H / stride), int(W / stride))
          #32 64 4 19 19 -------------> #32 256 19 19

##### 先说np的transpose函数 #####

#生成一个多维数组A
    A = np.array([[[ 0,  1,  2,  3],    #（0,0,0,），（0,0,1）（0,0,2），（0,0,3） 0,1,2,3分别代表
                [ 4,  5,  6,  7]],       # （0,1,0），（0,1,1）（0,1,2），（0,1,3）
    
               [[ 8,  9, 10, 11],     #（1,0,0,），（1,0,1）（1,0,2），（1,0,3）
                [12, 13, 14, 15]]])   #（1,1,0,），（1,1,1）（1,1,2），（1,1,3）
                #大小2*2*4
    #转置一下并打印
    b=A.transpose((1,0,2))  #此处的1，0,2代表将原来的0坐标轴和1坐标轴交换，即将对应坐标的数字重新排一下
    print(b)
    [[[ 0  1  2  3]
      [ 8  9 10 11]]
     [[ 4  5  6  7]
      [12 13 14 15]]]

#### .view（）是pytoch中对张量的维度大小进行转换 ####

import numpy as np
    #rand.seed
    import torch
    b=torch.Tensor(np.random.random((1,4,6,6)))
    print(b)
    tensor([[[[0.1436, 0.1968, 0.2878, 0.3360, 0.7482, 0.9375],
              [0.3952, 0.1303, 0.8594, 0.3781, 0.3427, 0.8003],
              [0.1780, 0.8446, 0.3642, 0.5780, 0.3739, 0.6575],
              [0.2092, 0.3296, 0.7916, 0.5591, 0.2719, 0.4667],
              [0.3777, 0.4153, 0.3515, 0.8864, 0.9704, 0.1552],
              [0.2988, 0.6856, 0.8054, 0.4950, 0.1066, 0.8497]],
    
             [[0.0231, 0.3548, 0.5212, 0.9130, 0.6010, 0.4696],
              [0.6379, 0.1427, 0.3928, 0.6991, 0.4271, 0.3882],
              [0.8887, 0.5123, 0.6829, 0.0290, 0.5958, 0.5498],
              [0.0443, 0.9919, 0.0910, 0.8394, 0.1548, 0.8674],
              [0.9087, 0.6444, 0.3820, 0.8761, 0.0926, 0.2949],
              [0.4235, 0.4609, 0.2493, 0.1540, 0.9204, 0.3543]],
    
             [[0.8178, 0.1309, 0.7330, 0.4258, 0.6299, 0.0394],
              [0.6773, 0.9003, 0.2470, 0.5658, 0.8564, 0.6463],
              [0.3301, 0.7628, 0.9331, 0.6643, 0.7884, 0.6011],
              [0.5981, 0.5726, 0.7588, 0.1942, 0.9418, 0.5945],
              [0.7860, 0.6807, 0.8768, 0.4645, 0.7157, 0.7627],
              [0.0471, 0.0966, 0.0766, 0.2205, 0.5230, 0.3839]],
    
             [[0.7261, 0.7219, 0.1146, 0.2839, 0.4567, 0.6208],
              [0.2844, 0.4079, 0.8501, 0.7609, 0.8347, 0.3463],
              [0.1635, 0.1130, 0.5009, 0.5895, 0.8188, 0.2572],
              [0.1263, 0.0446, 0.4758, 0.4998, 0.0663, 0.2657],
              [0.5530, 0.9403, 0.6419, 0.2198, 0.3392, 0.3970],
              [0.4066, 0.9568, 0.7766, 0.7742, 0.2008, 0.3791]]]])
    
    x=b
    stride=2
    x = x.view(B, C, int(H / stride), stride, int(W / stride), stride).transpose(3, 4).contiguous()
    
    print(x)
    tensor([[[[[[0.1436, 0.1968],
                [0.3952, 0.1303]],
    
               [[0.2878, 0.3360],
                [0.8594, 0.3781]],
    
               [[0.7482, 0.9375],
                [0.3427, 0.8003]]],
    
    
              [[[0.1780, 0.8446],
                [0.2092, 0.3296]],
    
               [[0.3642, 0.5780],
                [0.7916, 0.5591]],
    
               [[0.3739, 0.6575],
                [0.2719, 0.4667]]],
    
    
              [[[0.3777, 0.4153],
                [0.2988, 0.6856]],
    
               [[0.3515, 0.8864],
                [0.8054, 0.4950]],
    
               [[0.9704, 0.1552],
                [0.1066, 0.8497]]]],
    
    
    
             [[[[0.0231, 0.3548],
                [0.6379, 0.1427]],
    
               [[0.5212, 0.9130],
                [0.3928, 0.6991]],
    
               [[0.6010, 0.4696],
                [0.4271, 0.3882]]],
    
    
              [[[0.8887, 0.5123],
                [0.0443, 0.9919]],
    
               [[0.6829, 0.0290],
                [0.0910, 0.8394]],
    
               [[0.5958, 0.5498],
                [0.1548, 0.8674]]],
    
    
              [[[0.9087, 0.6444],
                [0.4235, 0.4609]],
    
               [[0.3820, 0.8761],
                [0.2493, 0.1540]],
    
               [[0.0926, 0.2949],
                [0.9204, 0.3543]]]],
    
    
    
             [[[[0.8178, 0.1309],
                [0.6773, 0.9003]],
    
               [[0.7330, 0.4258],
                [0.2470, 0.5658]],
    
               [[0.6299, 0.0394],
                [0.8564, 0.6463]]],
    
    
              [[[0.3301, 0.7628],
                [0.5981, 0.5726]],
    
               [[0.9331, 0.6643],
                [0.7588, 0.1942]],
    
               [[0.7884, 0.6011],
                [0.9418, 0.5945]]],
    
    
              [[[0.7860, 0.6807],
                [0.0471, 0.0966]],
    
               [[0.8768, 0.4645],
                [0.0766, 0.2205]],
    
               [[0.7157, 0.7627],
                [0.5230, 0.3839]]]],
    
    
    
             [[[[0.7261, 0.7219],
                [0.2844, 0.4079]],
    
               [[0.1146, 0.2839],
                [0.8501, 0.7609]],
    
               [[0.4567, 0.6208],
                [0.8347, 0.3463]]],
    
    
              [[[0.1635, 0.1130],
                [0.1263, 0.0446]],
    
               [[0.5009, 0.5895],
                [0.4758, 0.4998]],
    
               [[0.8188, 0.2572],
                [0.0663, 0.2657]]],
    
    
              [[[0.5530, 0.9403],
                [0.4066, 0.9568]],
    
               [[0.6419, 0.2198],
                [0.7766, 0.7742]],
    
               [[0.3392, 0.3970],
                [0.2008, 0.3791]]]]]])
    
    
    x = x.view(B, C, int(H / stride * W / stride), int(stride * stride)).transpose(2, 3).contiguous()
    print(x)
    
    tensor([[[[0.1436, 0.2878, 0.7482, 0.1780],
              [0.3642, 0.3739, 0.3777, 0.3515],
              [0.9704, 0.1968, 0.3360, 0.9375],
              [0.8446, 0.5780, 0.6575, 0.4153],
              [0.8864, 0.1552, 0.3952, 0.8594],
              [0.3427, 0.2092, 0.7916, 0.2719],
              [0.2988, 0.8054, 0.1066, 0.1303],
              [0.3781, 0.8003, 0.3296, 0.5591],
              [0.4667, 0.6856, 0.4950, 0.8497]],
    
             [[0.0231, 0.5212, 0.6010, 0.8887],
              [0.6829, 0.5958, 0.9087, 0.3820],
              [0.0926, 0.3548, 0.9130, 0.4696],
              [0.5123, 0.0290, 0.5498, 0.6444],
              [0.8761, 0.2949, 0.6379, 0.3928],
              [0.4271, 0.0443, 0.0910, 0.1548],
              [0.4235, 0.2493, 0.9204, 0.1427],
              [0.6991, 0.3882, 0.9919, 0.8394],
              [0.8674, 0.4609, 0.1540, 0.3543]],
    
             [[0.8178, 0.7330, 0.6299, 0.3301],
              [0.9331, 0.7884, 0.7860, 0.8768],
              [0.7157, 0.1309, 0.4258, 0.0394],
              [0.7628, 0.6643, 0.6011, 0.6807],
              [0.4645, 0.7627, 0.6773, 0.2470],
              [0.8564, 0.5981, 0.7588, 0.9418],
              [0.0471, 0.0766, 0.5230, 0.9003],
              [0.5658, 0.6463, 0.5726, 0.1942],
              [0.5945, 0.0966, 0.2205, 0.3839]],
    
             [[0.7261, 0.1146, 0.4567, 0.1635],
              [0.5009, 0.8188, 0.5530, 0.6419],
              [0.3392, 0.7219, 0.2839, 0.6208],
              [0.1130, 0.5895, 0.2572, 0.9403],
              [0.2198, 0.3970, 0.2844, 0.8501],
              [0.8347, 0.1263, 0.4758, 0.0663],
              [0.4066, 0.7766, 0.2008, 0.4079],
              [0.7609, 0.3463, 0.0446, 0.4998],
              [0.2657, 0.9568, 0.7742, 0.3791]]]])

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl8zODYzMjI0Ng_size_16_color_FFFFFF_t_70]: /images/20211122/3e0489c1528d478b8f9adf6f32c15227.png