• 模型压缩（二）yolov5剪枝

    一、yolov5s

    在yolov5s.ymal文件中，

    depth_multiple: 0.33  # model depth multiple
    width_multiple: 0.50  # layer channel multiple。

    通道深度（残差数）及宽度（通道数）相对标准的比例。

    标准的backbone中的C3的number分别为：3、6、9、3

    yolov5s的backbone中的C3的number为：1,、2、3、1  （depth_multiple*number）

    同理网络宽度width_multiple*args[0]。

    head类似。

    1. -------------------------------------0-P1/2----------------------------------------------
    2. model.0.conv.weight --------- torch.Size([32, 3, 6, 6])
    3. model.0.bn.weight --------- torch.Size([32])
    4. model.0.bn.bias --------- torch.Size([32])
    5. -------------------------------------1-P2/4----------------------------------------------
    6. model.1.conv.weight --------- torch.Size([64, 32, 3, 3])
    7. model.1.bn.weight --------- torch.Size([64])
    8. model.1.bn.bias --------- torch.Size([64])
    9.  
    10. -------------------------------------C3----------------------------------------------
    11. **cv1**
    12. model.2.cv1.conv.weight --------- torch.Size([32, 64, 1, 1])
    13. model.2.cv1.bn.weight --------- torch.Size([32])  ***
    14. model.2.cv1.bn.bias --------- torch.Size([32])    ***
    15. **cv2**
    16. model.2.cv2.conv.weight --------- torch.Size([32, 64, 1, 1])
    17. model.2.cv2.bn.weight --------- torch.Size([32])
    18. model.2.cv2.bn.bias --------- torch.Size([32])
    19. **cv3**
    20. model.2.cv3.conv.weight --------- torch.Size([64, 64, 1, 1])
    21. model.2.cv3.bn.weight --------- torch.Size([64])
    22. model.2.cv3.bn.bias --------- torch.Size([64])
    23.  
    24. bneck:*1
    25. model.2.m.0.cv1.conv.weight --------- torch.Size([32, 32, 1, 1])
    26. model.2.m.0.cv1.bn.weight --------- torch.Size([32])        ***
    27. model.2.m.0.cv1.bn.bias --------- torch.Size([32])          ***
    28. model.2.m.0.cv2.conv.weight --------- torch.Size([32, 32, 3, 3])
    29. model.2.m.0.cv2.bn.weight --------- torch.Size([32])      ***
    30. model.2.m.0.cv2.bn.bias --------- torch.Size([32])        ***
    31.  
    32. -------------------------------------3-P3/8----------------------------------------------
    33. model.3.conv.weight --------- torch.Size([128, 64, 3, 3])
    34. model.3.bn.weight --------- torch.Size([128])
    35. model.3.bn.bias --------- torch.Size([128])
    36.  
    37. -------------------------------------C3----------------------------------------------
    38. **cv1**
    39. model.4.cv1.conv.weight --------- torch.Size([64, 128, 1, 1])
    40. model.4.cv1.bn.weight --------- torch.Size([64])        ***
    41. model.4.cv1.bn.bias --------- torch.Size([64])        ***
    42. **cv2**
    43. model.4.cv2.conv.weight --------- torch.Size([64, 128, 1, 1])
    44. model.4.cv2.bn.weight --------- torch.Size([64])
    45. model.4.cv2.bn.bias --------- torch.Size([64])
    46. **cv3**
    47. model.4.cv3.conv.weight --------- torch.Size([128, 128, 1, 1])
    48. model.4.cv3.bn.weight --------- torch.Size([128])
    49. model.4.cv3.bn.bias --------- torch.Size([128])
    50. **bneck1**
    51. model.4.m.0.cv1.conv.weight --------- torch.Size([64, 64, 1, 1])
    52. model.4.m.0.cv1.bn.weight --------- torch.Size([64])
    53. model.4.m.0.cv1.bn.bias --------- torch.Size([64])
    54. model.4.m.0.cv2.conv.weight --------- torch.Size([64, 64, 3, 3])
    55. model.4.m.0.cv2.bn.weight --------- torch.Size([64])
    56. model.4.m.0.cv2.bn.bias --------- torch.Size([64])
    57. **bneck2**
    58. model.4.m.1.cv1.conv.weight --------- torch.Size([64, 64, 1, 1])
    59. model.4.m.1.cv1.bn.weight --------- torch.Size([64])
    60. model.4.m.1.cv1.bn.bias --------- torch.Size([64])
    61. model.4.m.1.cv2.conv.weight --------- torch.Size([64, 64, 3, 3])
    62. model.4.m.1.cv2.bn.weight --------- torch.Size([64])
    63. model.4.m.1.cv2.bn.bias --------- torch.Size([64])
    64.  
    65. -------------------------------------5-P4/16----------------------------------------------
    66. model.5.conv.weight --------- torch.Size([256, 128, 3, 3])
    67. model.5.bn.weight --------- torch.Size([256])
    68. model.5.bn.bias --------- torch.Size([256])
    69.  
    70.  
    71. 。。。。。。
    72.

    二、C3模块

            本文选择yolov5s进行通道剪枝，同样根据BN层稀疏化达到剪枝效果。在yolov5s结构中存在shortcut与cat，主路与支路合并操作。其中shortcut是将前层与后层特征相加，cat是通道连接，而shortcut必须保证前后层的通道数一致才可相加。如果shortcut的前后层参与剪枝，就无法保证前后层的通道数一致，所以剪枝过程中必须剔除参与shortcut操作的卷积层，而cat操作则不影响。

    yolov5s的C3模块的Bottleneck结构中存在shortcut操作。为了避免BN层稀疏后，通道数不匹配，所以所有的残差结构都不剪枝。

    C3：

    1. class Bottleneck(nn.Module):
    2.     # Standard bottleneck
    3.     def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
    4.         super().__init__()
    5.         c_ = int(c2 * e)  # hidden channels
    6.         self.cv1 = Conv(c1, c_, 1, 1)
    7.         self.cv2 = Conv(c_, c2, 3, 1, g=g)
    8.         self.add = shortcut and c1 == c2  #通道相同直接相加。
    9.  
    10.     def forward(self, x):
    11.         return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
    12.  
    13.  
    14. class C3(nn.Module):
    15.     # CSP Bottleneck with 3 convolutions
    16.     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
    17.         super().__init__()
    18.         c_ = int(c2 * e)  # hidden channels
    19.         self.cv1 = Conv(c1, c_, 1, 1)
    20.         self.cv2 = Conv(c1, c_, 1, 1)#支路
    21.         self.cv3 = Conv(2 * c_, c2, 1)  # act=FReLU(c2)
    22.         self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
    23.         # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])
    24.  
    25.     def forward(self, x):
    26.         return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))

     C3结构：

    所以C3结构中cv1、cv2参与剪枝。

    三、剪枝操作

    1、稀疏训练

    剔除C3结构中不参与剪枝的卷积层 。

    1.  #-------------------------------parse---------------------------
    2.             srtmp=opt.sr*(1-0.9*epoch/epochs)
    3.             if opt.st:
    4.                 ignore_bn_list=[]
    5.                 #记录bottleneck中所有bn层
    6.                 #C3结构中第一个卷积层与bneck中conv层不剪枝
    7.                 #即参与add操作有三层conv
    8.                 for k,m in model.named_modules():
    9.                     if isinstance(m,Bottleneck):
    10.                         if m.add:
    11.                             ignore_bn_list.append(k.split('.',2)[0]+'.cv1.bn')
    12.                             ignore_bn_list.append(k+ '.cv1.bn')
    13.                             ignore_bn_list.append(k + '.cv2.bn')
    14.                     if isinstance(k,nn.BatchNorm2d) and (k  not in ignore_bn_list):
    15.                         m.weight.grad.data.add_(srtmp*torch.sign(m.weight.data))
    16.                         m.bias.grad.data.add_(opt.sr*10 * torch.sign(m.weight.bias))
    17.                 print(ignore_bn_list)

    2、剪枝操作

    规整剪枝与正常剪枝。

    正常剪枝

    需剪枝的bn层

    1.    bn_layers= {}
    2.     ignore_bn_layers=[]
    3.     for layer_name,layer_model in model.named_modules():
    4.         if isinstance(layer_model,Bottleneck):
    5.             if layer_model.add:
    6.                 ignore_bn_layers.append(layer_name.rsplit('.',2)[0]+'.cv1.bn')#C3中第一个conv
    7.                 ignore_bn_layers.append(layer_name+'.cv1.bn')#bottleneck中第一个conv
    8.                 ignore_bn_layers.append(layer_name+'.cv2.bn')#bottleneck中第一个conv
    9.         if isinstance(layer_model,nn.BatchNorm2d) and (layer_name not in ignore_bn_layers):
    10.                 # print(ignore_bn_layers,layer_name)
    11.                 #未剔除全，主要是每次遍历进入C3中时，cv1没剔除，直到bneck中才开始。
    12.                 bn_layers[layer_name]=layer_model
    13.  
    14.     # print(ignore_bn_layers,)
    15.     # print(len(ignore_bn_layers))
    16.     # print(bn_layers)
    17.     # print(len(bn_layers))
    18.     # exit()
    19.     #再次过滤4个C3中的第一个cv层
    20.     bn_layers= {k:v for k,v in bn_layers.items() if k not in ignore_bn_layers}
    21.     # print(bn_names)
    22.     # print(len(bn_names))
    23.     # exit()

    统计所有BN层通道数量及各通道的权重值，对权重进行排序，并计算得到索引阈值。

    1.   bn_size=[da.weight.data.shape[0] for da in bn_layers.values()]
    2.     total_size=sum(bn_size)
    3.     print(total_size)
    4.     bn_weights=torch.zeros(total_size)
    5.     start=0
    6.     for i,w in enumerate(bn_layers.values()):
    7.         size=w.weight.data.shape[0]
    8.         bn_weights[start:(start+size)] = w.weight.data.abs().clone()
    9.         start+=bn_size[i]
    10.     print(bn_weights,bn_weights.shape)
    11.  
    12.     bn_data,id=torch.sort(bn_weights)
    13.  
    14.     thresh_index=int(percent*total_size)
    15.     thresh_weight=bn_data[thresh_index]
    16.     print(thresh_index,thresh_weight)
    17.     print(f'Gamma value that less than {thresh_weight:.4f} are set to zero!')
    18.     print("=" * 94)
    19.     print(f"|\t{'layer name':<25}{'|':<10}{'origin channels':<20}{'|':<10}{'remaining channels':<20}|")

    存在问题：

    根据阈值来分隔，可能存在某一BN层所有通道均小于阈值，如果将其过滤掉，会造成层层之间的断开，此时需要做判断进行限制，使得每层最少有一个通道得以保留。

    解决方法：获取每个bn层的权重的最大值，然后在这些最大值中取最小值与设定的阈值进行对比，如果小于阈值，则提示修改。
     

    1.  # 避免剪掉所有channel的最高阈值(每个BN层的gamma的最大值的最小值即为阈值上限)
    2.     highest_thre = []
    3.     for bnlayer in bn_layers.values():
    4.         highest_thre.append(bnlayer.weight.data.abs().max().item())
    5.     # print("highest_thre:",highest_thre)
    6.     highest_thre = min(highest_thre)
    7.     # 找到highest_thre对应的下标对应的百分比
    8.     percent_limit = (bn_data == highest_thre).nonzero()[0, 0].item() / len(bn_weights)
    9.  
    10.     print(f'Suggested Gamma threshold should be less than {highest_thre:.4f}.')
    11.     print(f'The corresponding prune ratio is {percent_limit:.3f}, but you can set higher.')
    12.

    重新设置模型文件

    1.   pruned_num=0
    2.     pruned_yaml = {}
    3.     nc = model.model[-1].nc
    4.     with open(cfg, encoding='ascii', errors='ignore') as f:
    5.         model_yamls = yaml.safe_load(f)  # model dict
    6.     # # Define model
    7.     pruned_yaml["nc"] = model.model[-1].nc
    8.     pruned_yaml["depth_multiple"] = model_yamls["depth_multiple"]
    9.     pruned_yaml["width_multiple"] = model_yamls["width_multiple"]
    10.     pruned_yaml["anchors"] = model_yamls["anchors"]
    11.     anchors = model_yamls["anchors"]
    12.     pruned_yaml["backbone"] = [
    13.         [-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2
    14.         [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
    15.         [-1, 3, C3Pruned, [128]],
    16.         [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
    17.         [-1, 6, C3Pruned, [256]],
    18.         [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
    19.         [-1, 9, C3Pruned, [512]],
    20.         [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
    21.         [-1, 3, C3Pruned, [1024]],
    22.         [-1, 1, SPPFPruned, [1024, 5]],  # 9
    23.     ]
    24.     pruned_yaml["head"] = [
    25.         [-1, 1, Conv, [512, 1, 1]],
    26.         [-1, 1, nn.Upsample, [None, 2, 'nearest']],
    27.         [[-1, 6], 1, Concat, [1]],  # cat backbone P4
    28.         [-1, 3, C3Pruned, [512, False]],  # 13
    29.  
    30.         [-1, 1, Conv, [256, 1, 1]],
    31.         [-1, 1, nn.Upsample, [None, 2, 'nearest']],
    32.         [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    33.         [-1, 3, C3Pruned, [256, False]],  # 17 (P3/8-small)
    34.  
    35.         [-1, 1, Conv, [256, 3, 2]],
    36.         [[-1, 14], 1, Concat, [1]],  # cat head P4
    37.         [-1, 3, C3Pruned, [512, False]],  # 20 (P4/16-medium)
    38.  
    39.         [-1, 1, Conv, [512, 3, 2]],
    40.         [[-1, 10], 1, Concat, [1]],  # cat head P5
    41.         [-1, 3, C3Pruned, [1024, False]],  # 23 (P5/32-large)
    42.  
    43.         [[17, 20, 23], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
    44.     ]

    模型重构：

    1.    maskbndict={}
    2.     remain_num=0
    3.     for name,layer in model.named_modules():
    4.         if isinstance(layer,nn.BatchNorm2d):
    5.             bn_model=layer
    6.             mask=obtain_bn_mask(bn_model,thresh_weight)
    7.             # print(mask)
    8.             if name in ignore_bn_layers:
    9.                 # print('-----')
    10.                 mask=torch.ones(layer.weight.data.size()).cuda()
    11.             maskbndict[name]=mask
    12.             # print(mask)
    13.             remain_num+=int(mask.sum())
    14.             bn_model.weight.data.mul_(mask)
    15.             bn_model.bias.data.mul_(mask)
    16.             print(f"|\t{name:<25}{'|':<10}{bn_model.weight.data.size()[0]:<20}{'|':<10}{int(mask.sum()):<20}|")
    17.             assert int(
    18.                 mask.sum()) > 0, "Current remaining channel must greater than 0!!! please set prune percent to lower thesh, or you can retrain a more sparse model..."
    19.  
    20.     print("=" * 94)
    21.  
    22.  
    23.     pruned_model=ModelPruned(maskbndict=maskbndict,cfg=pruned_yaml,ch=3).cuda()
    24.     for m in pruned_model.modules():
    25.         if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model]:
    26.             m.inplace = True  # pytorch 1.7.0 compatibility
    27.         elif type(m) is Conv:
    28.             m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
    29.     from_to_map=pruned_model.from_to_map
    30.     pruned_model_state=pruned_model.state_dict()

    参数拷贝：

    1.   #-----------------------------参数拷贝----------------------------
    2.     modelstate = model.state_dict()
    3.     changed_state=[]
    4.     for((layername,layermodel),(pruned_layername,pruned_layermodel)) in zip(model.named_modules(),pruned_model.named_modules()):
    5.         if isinstance(layermodel,nn.Conv2d) and not layername.startswith("model.24"):
    6.             convname=layername[:-4]+"bn"
    7.             if convname in from_to_map.keys():
    8.                 former=from_to_map[convname]
    9.                 if isinstance(former,str):
    10.                     out_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[layername[:-4] + "bn"].cpu().numpy())))
    11.                     in_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[former].cpu().numpy())))
    12.                     w = layermodel.weight.data[:, in_idx, :, :].clone()
    13.  
    14.                     if len(w.shape) == 3:  # remain only 1 channel.
    15.                         w = w.unsqueeze(1)
    16.                     w = w[out_idx, :, :, :].clone()
    17.  
    18.                     pruned_layermodel.weight.data = w.clone()
    19.                     changed_state.append(layername + ".weight")
    20.                 if isinstance(former, list):
    21.                     orignin = [modelstate[i + ".weight"].shape[0] for i in former]
    22.                     formerin = []
    23.                     for it in range(len(former)):
    24.                         name = former[it]
    25.                         tmp = [i for i in range(maskbndict[name].shape[0]) if maskbndict[name][i] == 1]
    26.                         if it > 0:
    27.                             tmp = [k + sum(orignin[:it]) for k in tmp]
    28.                         formerin.extend(tmp)
    29.                     out_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[layername[:-4] + "bn"].cpu().numpy())))
    30.                     w = layermodel.weight.data[out_idx, :, :, :].clone()
    31.                     pruned_layermodel.weight.data = w[:, formerin, :, :].clone()
    32.                     changed_state.append(layername + ".weight")
    33.                 else:
    34.                     out_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[layername[:-4] + "bn"].cpu().numpy())))
    35.                     w = layermodel.weight.data[out_idx, :, :, :].clone()
    36.                     assert len(w.shape) == 4
    37.                     pruned_layermodel.weight.data = w.clone()
    38.                     changed_state.append(layername + ".weight")
    39.  
    40.                 if isinstance(layermodel, nn.BatchNorm2d):
    41.                     out_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[layername].cpu().numpy())))
    42.                     pruned_layermodel.weight.data = layermodel.weight.data[out_idx].clone()
    43.                     pruned_layermodel.bias.data = layermodel.bias.data[out_idx].clone()
    44.                     pruned_layermodel.running_mean = layermodel.running_mean[out_idx].clone()
    45.                     pruned_layermodel.running_var = layermodel.running_var[out_idx].clone()
    46.                     changed_state.append(layername + ".weight")
    47.                     changed_state.append(layername + ".bias")
    48.                     changed_state.append(layername + ".running_mean")
    49.                     changed_state.append(layername + ".running_var")
    50.                     changed_state.append(layername + ".num_batches_tracked")
    51.  
    52.                 if isinstance(layermodel, nn.Conv2d) and layername.startswith("model.24"):
    53.                     former = from_to_map[layername]
    54.                     in_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[former].cpu().numpy())))
    55.                     pruned_layermodel.weight.data = layermodel.weight.data[:, in_idx, :, :]
    56.                     pruned_layermodel.bias.data = layermodel.bias.data
    57.                     changed_state.append(layername + ".weight")
    58.                     changed_state.append(layername + ".bias")
    59.  
    60.     missing = [i for i in pruned_model_state.keys() if i not in changed_state]
    61.  
    62.     pruned_model.eval()
    63.     pruned_model.names = model.names
    64.     # =============================================================================================== #
    65.     torch.save({"model": model}, "weights/pruned_model/orign_model.pt")
    66.     model = pruned_model
    67.     torch.save({"model": model}, "weights/pruned_model/pruned_model.pt")
    68.     model.cuda().eval()

    参考：

    YOLOv5模型剪枝压缩(2)-YOLOv5模型简介和剪枝层选择_MidasKing的博客-CSDN博客_yolov5剪枝

    yolov5模型压缩之模型剪枝_小小小绿叶的博客-CSDN博客_yolov5模型裁剪

    GitHub - midasklr/yolov5prune

  • 相关阅读:
    mysql分组concat() concat_ws() group_concat()
    【刷题】代码随想录算法训练营第二十三天|669、修剪二叉搜索树，108、将有序数组转换为二叉搜索树，538、把二叉搜索树转换为累加树
    【虹科分享】不要使用Windows解决方案来保护Linux服务器
    java计算机毕业设计内容校园热点新闻推送网站源码+mysql数据库+系统+lw文档+部署
    TwinCAT3安装教程
    Linux云服务器如何快速安装宝塔面板？
    十 计算机网络
    驱动 私有数据传参点灯
    MySQL的视图,用户管理,C语言连接
    怎样做一个简易而温馨的原木风居室空间
  • 原文地址：https://blog.csdn.net/m0_37264397/article/details/126292621