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。

-------------------------------------0-P1/2----------------------------------------------
model.0.conv.weight --------- torch.Size([32, 3, 6, 6])
model.0.bn.weight --------- torch.Size([32])
model.0.bn.bias --------- torch.Size([32])
-------------------------------------1-P2/4----------------------------------------------
model.1.conv.weight --------- torch.Size([64, 32, 3, 3])
model.1.bn.weight --------- torch.Size([64])
model.1.bn.bias --------- torch.Size([64])

-------------------------------------C3----------------------------------------------
**cv1**
model.2.cv1.conv.weight --------- torch.Size([32, 64, 1, 1])
model.2.cv1.bn.weight --------- torch.Size([32])  ***
model.2.cv1.bn.bias --------- torch.Size([32])    ***
**cv2**
model.2.cv2.conv.weight --------- torch.Size([32, 64, 1, 1])
model.2.cv2.bn.weight --------- torch.Size([32])
model.2.cv2.bn.bias --------- torch.Size([32])
**cv3**
model.2.cv3.conv.weight --------- torch.Size([64, 64, 1, 1])
model.2.cv3.bn.weight --------- torch.Size([64])
model.2.cv3.bn.bias --------- torch.Size([64])

bneck:*1
model.2.m.0.cv1.conv.weight --------- torch.Size([32, 32, 1, 1])
model.2.m.0.cv1.bn.weight --------- torch.Size([32])        ***
model.2.m.0.cv1.bn.bias --------- torch.Size([32])          ***
model.2.m.0.cv2.conv.weight --------- torch.Size([32, 32, 3, 3])
model.2.m.0.cv2.bn.weight --------- torch.Size([32])      ***
model.2.m.0.cv2.bn.bias --------- torch.Size([32])        ***

-------------------------------------3-P3/8----------------------------------------------
model.3.conv.weight --------- torch.Size([128, 64, 3, 3])
model.3.bn.weight --------- torch.Size([128])
model.3.bn.bias --------- torch.Size([128])

-------------------------------------C3----------------------------------------------
**cv1**
model.4.cv1.conv.weight --------- torch.Size([64, 128, 1, 1])
model.4.cv1.bn.weight --------- torch.Size([64])        ***
model.4.cv1.bn.bias --------- torch.Size([64])        ***
**cv2**
model.4.cv2.conv.weight --------- torch.Size([64, 128, 1, 1])
model.4.cv2.bn.weight --------- torch.Size([64])
model.4.cv2.bn.bias --------- torch.Size([64])
**cv3**
model.4.cv3.conv.weight --------- torch.Size([128, 128, 1, 1])
model.4.cv3.bn.weight --------- torch.Size([128])
model.4.cv3.bn.bias --------- torch.Size([128])
**bneck1**
model.4.m.0.cv1.conv.weight --------- torch.Size([64, 64, 1, 1])
model.4.m.0.cv1.bn.weight --------- torch.Size([64])
model.4.m.0.cv1.bn.bias --------- torch.Size([64])
model.4.m.0.cv2.conv.weight --------- torch.Size([64, 64, 3, 3])
model.4.m.0.cv2.bn.weight --------- torch.Size([64])
model.4.m.0.cv2.bn.bias --------- torch.Size([64])
**bneck2**
model.4.m.1.cv1.conv.weight --------- torch.Size([64, 64, 1, 1])
model.4.m.1.cv1.bn.weight --------- torch.Size([64])
model.4.m.1.cv1.bn.bias --------- torch.Size([64])
model.4.m.1.cv2.conv.weight --------- torch.Size([64, 64, 3, 3])
model.4.m.1.cv2.bn.weight --------- torch.Size([64])
model.4.m.1.cv2.bn.bias --------- torch.Size([64])

-------------------------------------5-P4/16----------------------------------------------
model.5.conv.weight --------- torch.Size([256, 128, 3, 3])
model.5.bn.weight --------- torch.Size([256])
model.5.bn.bias --------- torch.Size([256])

。。。。。。

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：

class Bottleneck(nn.Module):
    # Standard bottleneck
    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
        super().__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_, c2, 3, 1, g=g)
        self.add = shortcut and c1 == c2  #通道相同直接相加。

    def forward(self, x):
        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))

class C3(nn.Module):
    # CSP Bottleneck with 3 convolutions
    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
        super().__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c1, c_, 1, 1)#支路
        self.cv3 = Conv(2 * c_, c2, 1)  # act=FReLU(c2)
        self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
        # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])

    def forward(self, x):
        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 。

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

   bn_layers= {}
    ignore_bn_layers=[]
    for layer_name,layer_model in model.named_modules():
        if isinstance(layer_model,Bottleneck):
            if layer_model.add:
                ignore_bn_layers.append(layer_name.rsplit('.',2)[0]+'.cv1.bn')#C3中第一个conv
                ignore_bn_layers.append(layer_name+'.cv1.bn')#bottleneck中第一个conv
                ignore_bn_layers.append(layer_name+'.cv2.bn')#bottleneck中第一个conv
        if isinstance(layer_model,nn.BatchNorm2d) and (layer_name not in ignore_bn_layers):
                # print(ignore_bn_layers,layer_name)
                #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。
                bn_layers[layer_name]=layer_model

    # print(ignore_bn_layers,)
    # print(len(ignore_bn_layers))
    # print(bn_layers)
    # print(len(bn_layers))
    # exit()
    #再次过滤4个C3中的第一个cv层
    bn_layers= {k:v for k,v in bn_layers.items() if k not in ignore_bn_layers}
    # print(bn_names)
    # print(len(bn_names))
    # 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。

  bn_size=[da.weight.data.shape[0] for da in bn_layers.values()]
    total_size=sum(bn_size)
    print(total_size)
    bn_weights=torch.zeros(total_size)
    start=0
    for i,w in enumerate(bn_layers.values()):
        size=w.weight.data.shape[0]
        bn_weights[start:(start+size)] = w.weight.data.abs().clone()
        start+=bn_size[i]
    print(bn_weights,bn_weights.shape)

    bn_data,id=torch.sort(bn_weights)

    thresh_index=int(percent*total_size)
    thresh_weight=bn_data[thresh_index]
    print(thresh_index,thresh_weight)
    print(f'Gamma value that less than {thresh_weight:.4f} are set to zero!')
    print("=" * 94)
    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。

 # 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)
    highest_thre = []
    for bnlayer in bn_layers.values():
        highest_thre.append(bnlayer.weight.data.abs().max().item())
    # print("highest_thre:",highest_thre)
    highest_thre = min(highest_thre)
    # 找到highest_thre对应的下标对应的百分比
    percent_limit = (bn_data == highest_thre).nonzero()[0, 0].item() / len(bn_weights)

    print(f'Suggested Gamma threshold should be less than {highest_thre:.4f}.')
    print(f'The corresponding prune ratio is {percent_limit:.3f}, but you can set higher.')

重新设置模型文件

  pruned_num=0
    pruned_yaml = {}
    nc = model.model[-1].nc
    with open(cfg, encoding='ascii', errors='ignore') as f:
        model_yamls = yaml.safe_load(f)  # model dict
    # # Define model
    pruned_yaml["nc"] = model.model[-1].nc
    pruned_yaml["depth_multiple"] = model_yamls["depth_multiple"]
    pruned_yaml["width_multiple"] = model_yamls["width_multiple"]
    pruned_yaml["anchors"] = model_yamls["anchors"]
    anchors = model_yamls["anchors"]
    pruned_yaml["backbone"] = [
        [-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2
        [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
        [-1, 3, C3Pruned, [128]],
        [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
        [-1, 6, C3Pruned, [256]],
        [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
        [-1, 9, C3Pruned, [512]],
        [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
        [-1, 3, C3Pruned, [1024]],
        [-1, 1, SPPFPruned, [1024, 5]],  # 9
    ]
    pruned_yaml["head"] = [
        [-1, 1, Conv, [512, 1, 1]],
        [-1, 1, nn.Upsample, [None, 2, 'nearest']],
        [[-1, 6], 1, Concat, [1]],  # cat backbone P4
        [-1, 3, C3Pruned, [512, False]],  # 13

        [-1, 1, Conv, [256, 1, 1]],
        [-1, 1, nn.Upsample, [None, 2, 'nearest']],
        [[-1, 4], 1, Concat, [1]],  # cat backbone P3
        [-1, 3, C3Pruned, [256, False]],  # 17 (P3/8-small)

        [-1, 1, Conv, [256, 3, 2]],
        [[-1, 14], 1, Concat, [1]],  # cat head P4
        [-1, 3, C3Pruned, [512, False]],  # 20 (P4/16-medium)

        [-1, 1, Conv, [512, 3, 2]],
        [[-1, 10], 1, Concat, [1]],  # cat head P5
        [-1, 3, C3Pruned, [1024, False]],  # 23 (P5/32-large)

        [[17, 20, 23], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
    ]

模型重构：

   maskbndict={}
    remain_num=0
    for name,layer in model.named_modules():
        if isinstance(layer,nn.BatchNorm2d):
            bn_model=layer
            mask=obtain_bn_mask(bn_model,thresh_weight)
            # print(mask)
            if name in ignore_bn_layers:
                # print('-----')
                mask=torch.ones(layer.weight.data.size()).cuda()
            maskbndict[name]=mask
            # print(mask)
            remain_num+=int(mask.sum())
            bn_model.weight.data.mul_(mask)
            bn_model.bias.data.mul_(mask)
            print(f"|\t{name:<25}{'|':<10}{bn_model.weight.data.size()[0]:<20}{'|':<10}{int(mask.sum()):<20}|")
            assert int(
                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..."

    print("=" * 94)

    pruned_model=ModelPruned(maskbndict=maskbndict,cfg=pruned_yaml,ch=3).cuda()
    for m in pruned_model.modules():
        if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model]:
            m.inplace = True  # pytorch 1.7.0 compatibility
        elif type(m) is Conv:
            m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
    from_to_map=pruned_model.from_to_map
    pruned_model_state=pruned_model.state_dict()

参数拷贝：

  #-----------------------------参数拷贝----------------------------
    modelstate = model.state_dict()
    changed_state=[]
    for((layername,layermodel),(pruned_layername,pruned_layermodel)) in zip(model.named_modules(),pruned_model.named_modules()):
        if isinstance(layermodel,nn.Conv2d) and not layername.startswith("model.24"):
            convname=layername[:-4]+"bn"
            if convname in from_to_map.keys():
                former=from_to_map[convname]
                if isinstance(former,str):
                    out_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[layername[:-4] + "bn"].cpu().numpy())))
                    in_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[former].cpu().numpy())))
                    w = layermodel.weight.data[:, in_idx, :, :].clone()

                    if len(w.shape) == 3:  # remain only 1 channel.
                        w = w.unsqueeze(1)
                    w = w[out_idx, :, :, :].clone()

                    pruned_layermodel.weight.data = w.clone()
                    changed_state.append(layername + ".weight")
                if isinstance(former, list):
                    orignin = [modelstate[i + ".weight"].shape[0] for i in former]
                    formerin = []
                    for it in range(len(former)):
                        name = former[it]
                        tmp = [i for i in range(maskbndict[name].shape[0]) if maskbndict[name][i] == 1]
                        if it > 0:
                            tmp = [k + sum(orignin[:it]) for k in tmp]
                        formerin.extend(tmp)
                    out_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[layername[:-4] + "bn"].cpu().numpy())))
                    w = layermodel.weight.data[out_idx, :, :, :].clone()
                    pruned_layermodel.weight.data = w[:, formerin, :, :].clone()
                    changed_state.append(layername + ".weight")
                else:
                    out_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[layername[:-4] + "bn"].cpu().numpy())))
                    w = layermodel.weight.data[out_idx, :, :, :].clone()
                    assert len(w.shape) == 4
                    pruned_layermodel.weight.data = w.clone()
                    changed_state.append(layername + ".weight")

                if isinstance(layermodel, nn.BatchNorm2d):
                    out_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[layername].cpu().numpy())))
                    pruned_layermodel.weight.data = layermodel.weight.data[out_idx].clone()
                    pruned_layermodel.bias.data = layermodel.bias.data[out_idx].clone()
                    pruned_layermodel.running_mean = layermodel.running_mean[out_idx].clone()
                    pruned_layermodel.running_var = layermodel.running_var[out_idx].clone()
                    changed_state.append(layername + ".weight")
                    changed_state.append(layername + ".bias")
                    changed_state.append(layername + ".running_mean")
                    changed_state.append(layername + ".running_var")
                    changed_state.append(layername + ".num_batches_tracked")

                if isinstance(layermodel, nn.Conv2d) and layername.startswith("model.24"):
                    former = from_to_map[layername]
                    in_idx = np.squeeze(np.argwhere(np.asarray(maskbndict[former].cpu().numpy())))
                    pruned_layermodel.weight.data = layermodel.weight.data[:, in_idx, :, :]
                    pruned_layermodel.bias.data = layermodel.bias.data
                    changed_state.append(layername + ".weight")
                    changed_state.append(layername + ".bias")

    missing = [i for i in pruned_model_state.keys() if i not in changed_state]

    pruned_model.eval()
    pruned_model.names = model.names
    # =============================================================================================== #
    torch.save({"model": model}, "weights/pruned_model/orign_model.pt")
    model = pruned_model
    torch.save({"model": model}, "weights/pruned_model/pruned_model.pt")
    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