YOLOv8剪枝

前言

手写AI推出的全新模型剪枝与重参课程。记录下个人学习笔记，仅供自己参考。

本次课程主要讲解YOLOv8剪枝。

课程大纲可看下面的思维导图

1.Overview

YOLOV8剪枝的流程如下：

结论：在VOC2007上使用yolov8s模型进行的实验显示，预训练和约束训练在迭代50个epoch后达到了相同的mAP(:0.5)值，约为0.77。剪枝后，微调阶段需要65个epoch才能达到相同的mAP50。修建后的ONNX模型大小从43M减少到36M。

注意：我们需要将网络结构和网络权重区分开来，YOLOv8的网络结构来自yaml文件，如果我们进行剪枝后保存的权重文件的结构其实是和原始的yaml文件不符合的，需要对yaml文件进行修改满足我们的要求。

2.Pretrain(option)

步骤如下：

• git clone https://github.com/ultralytics/ultralytics.git
• use VOC2007, and modify the VOC.yaml(去除VOC2012的相关内容)
• disable amp(禁用amp混合精度)

# FILE: ultralytics/yolo/engine/trainer.py
...
def check_amp(model):# Avoid using mixed precision to affect finetunereturn False # <============== modified(修改部分)device = next(model.parameters()).device  # get model deviceif device.type in ('cpu', 'mps'):return False  # AMP only used on CUDA devicesdef amp_allclose(m, im):# All close FP32 vs AMP results...

3.Constrained Training

约束训练是为了筛选哪些channel比较重要，哪些channel没有那么重要，也就是我们上节课所说的稀疏训练

• prune the BN layer by adding L1 regularizer.

# FILE: ultralytics/yolo/engine/trainer.py
...
# Backward
self.scaler.scale(self.loss).backward()# <============ added(新增)
l1_lambda = 1e-2 * (1 - 0.9 * epoch / self.epochs)
for k, m in self.model.named_modules():if isinstance(m, nn.BatchNorm2d):m.weight.grad.data.add_(l1_lambda * torch.sign(m.weight.data))m.bias.grad.data.add_(1e-2 * torch.sign(m.bias.data))# Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
if ni - last_opt_step >= self.accumulate:self.optimizer_step()last_opt_step = ni
...

注意1：在剪枝时，我们选择加载last.pt而非best.pt，因为由于迁移学习，模型的泛化性比较好，在第一个epoch时mAP值最大，但这并不是真实的，我们需要稳定下来的一个模型进行prune

注意2：我们在对Conv层进行剪枝时，我们只考虑1v1(如BottleNeck,C2f and SPPF)、1vm(如Backbone,Detect)的情形，并不考虑mv1的情形。

思考：Constrained Training的必要性？

约束训练可以使得模型更易于剪枝。在约束训练中，模型会学习到一些通道或者权重系数比较不重要的信息，而这些信息在剪枝过程中得到应用，从而达到模型压缩的效果。而如果直接进行剪枝操作，可能会出现一些问题，比如剪枝后的模型精度大幅下降、剪枝不均匀等。因此，在进行剪枝操作之前，通过稀疏训练的方式，可以更好地准确地确定哪些通道或者权重系数可以被剪掉，从而避免上述问题的发生。

4.Prune

4.1 检查BN层的bias

• 剪枝后，确保BN层的大部分bias足够小(接近于0)，否则重新进行稀疏训练

for name, m in model.named_modules():if isinstance(m, torch.nn.BatchNorm2d):w = m.weight.abs().detach()b = m.bias.abs().detach()ws.append(w)bs.append(b)print(name, w.max().item(), w.min().item(), b.max().item(), b.min().item())

4.2 设置阈值和剪枝率

• threshold：全局或局部
• factor：保持率，裁剪太多不推荐

factor = 0.8
ws = torch.cat(ws)
threshold = torch.sort(ws, descending=True)[0][int(len(ws) * factor)]
print(threshold)

4.3 最小剪枝Conv单元的TopConv

Top-Bottom Conv如下图所示：

TopConv剪枝的示例代码如下：

def prune_conv(conv1: Conv, conv2: Conv):gamma = conv1.bn.weight.data.detach()beta  = conv1.bn.bias.data.detach()keep_idxs = []    local_threshold = thresholdwhile len(keep_idxs) < 8:keep_idxs = torch.where(gamma.abs() >= local_threshold)[0] local_threshold = local_threshold * 0.5n = len(keep_idxs)print(n / len(gamma) * 100)  # 打印我们保留了多少的channel# pruneconv1.bn.weight.data = gamma[keep_idxs]conv1.bn.bias.data   = beta[keep_idxs]conv1.bn.running_mean.data = conv1.bn.running_mean.data[keep_idxs]conv1.bn.running_var.data  = conv1.bn.running_var.data[keep_idxs]conv1.bn.num_features   = nconv1.conv.weight.data  = conv1.conv.weight.data[keep_idxs]conv1.conv.out_channels = nif conv1.conv.bias is not None:conv1.conv.bias.data = conv1.conv.bias.data[keep_idxs]# pattern to prune
# 1. prune all 1 vs 1 TB pattern e.g. bottleneck
for name, m in model.named_modules():if isinstance(m, Bottleneck):prune_conv(m.cv1, m.cv2)

注意：由于NVIDIA的硬件加速的原因，我们保留的channels应该大于等于8，我们可以通过设置local_threshold，尽量小点，让更多的channel保留下来。

4.4 最小剪枝Conv单元的BottomConv

BottomConv剪枝的示例代码如下：

def prune_conv(conv1: Conv, conv2: Conv):...if not isinstance(conv2, list):conv2 = [conv2]for item in conv2:if item is not None:if isinstance(item, Conv):conv = item.convelse:conv = itemconv.in_channels = nconv.weight.data = conv.weight.data[:, keep_idxs]

注意：BottomConv存在两种情形，一种是单个Conv，还有一种是Conv列表。TopConv需要考虑conv2d+bn+act，而BottomConv只需要考虑conv2d

4.5 Seq剪枝

Seq剪枝的示例代码如下：

def prune(m1, m2):if isinstance(m1, C2f):m1 = m1.cv2if not isinstance(m2, list):m2 = [m2]for i, item in enumerate(m2):if isinstance(item, C2f) or isinstance(item, SPPF):m2[i] = item.cv1prune_conv(m1, m2)# 2. prune sequential
seq = model.model
for i in range(3, 9):if i in [6, 4, 9]: continueprune(seq[i], seq[i+1])

注意：我们不考虑1vm的情形，因此在4,6,9的module我们是不进行剪枝的，后续head进行Concat时是对4,6,9的module进行拼接的。考虑到前几个conv的特征提取的重要性，我们也不剪枝它们。(那感觉没剪几个module呀😂)

4.6 Detect-FPN剪枝

Detect-FPN剪枝的示例代码如下：

# 3. prune FPN related block
detect: Detect = seq[-1]last_inputs = [seq[15], seq[18], seq[21]]
colasts     = [seq[16], seq[19], None]for last_input, colast, cv2, cv3 in zip(last_inputs, colasts, detect.cv2, detect.cv3):prune(last_input, [colast, cv2[0], cv3[0]])prune(cv2[0], cv2[1])prune(cv2[1], cv2[2])prune(cv3[0], cv3[1])prune(cv3[1], cv3[2])for name, p in yolo.model.named_parameters():p.requires_grad = True

注意：一定要设置所有参数为需要训练。因为加载后的model会给弄成False，导致报错

4.7 完整示例代码

完整的示例代码如下：

from ultralytics import YOLO
import torch
from ultralytics.nn.modules import Bottleneck, Conv, C2f, SPPF, Detect# Load a model
yolo = YOLO("epoch-50-constrained_weights/last.pt")  # build a new model from scratch
model = yolo.modelws = []
bs = []for name, m in model.named_modules():if isinstance(m, torch.nn.BatchNorm2d):w = m.weight.abs().detach()b = m.bias.abs().detach()ws.append(w)bs.append(b)print(name, w.max().item(), w.min().item(), b.max().item(), b.min().item())
# keep
factor = 0.8
ws = torch.cat(ws)
threshold = torch.sort(ws, descending=True)[0][int(len(ws) * factor)]
print(threshold)def prune_conv(conv1: Conv, conv2: Conv):gamma = conv1.bn.weight.data.detach()beta  = conv1.bn.bias.data.detach()# if gamma.abs().min() > f or beta.abs().min() > 0.1:#     return# idxs = torch.argsort(gamma.abs() * coeff + beta.abs(), descending=True)keep_idxs = []local_threshold = thresholdwhile len(keep_idxs) < 8:keep_idxs = torch.where(gamma.abs() >= local_threshold)[0]local_threshold = local_threshold * 0.5n = len(keep_idxs)# n = max(int(len(idxs) * 0.8), p)print(n / len(gamma) * 100)# scale = len(idxs) / nconv1.bn.weight.data = gamma[keep_idxs]conv1.bn.bias.data   = beta[keep_idxs]conv1.bn.running_var.data = conv1.bn.running_var.data[keep_idxs]conv1.bn.running_mean.data = conv1.bn.running_mean.data[keep_idxs]conv1.bn.num_features = nconv1.conv.weight.data = conv1.conv.weight.data[keep_idxs]conv1.conv.out_channels = nif conv1.conv.bias is not None:conv1.conv.bias.data = conv1.conv.bias.data[keep_idxs]if not isinstance(conv2, list):conv2 = [conv2]for item in conv2:if item is not None:if isinstance(item, Conv):conv = item.convelse:conv = itemconv.in_channels = nconv.weight.data = conv.weight.data[:, keep_idxs]def prune(m1, m2):if isinstance(m1, C2f):      # C2f as a top convm1 = m1.cv2if not isinstance(m2, list): # m2 is just one modulem2 = [m2]for i, item in enumerate(m2):if isinstance(item, C2f) or isinstance(item, SPPF):m2[i] = item.cv1prune_conv(m1, m2)for name, m in model.named_modules():if isinstance(m, Bottleneck):prune_conv(m.cv1, m.cv2)seq = model.model
for i in range(3, 9):if i in [6, 4, 9]: continueprune(seq[i], seq[i+1])detect:Detect = seq[-1]
last_inputs   = [seq[15], seq[18], seq[21]]
colasts       = [seq[16], seq[19], None]
for last_input, colast, cv2, cv3 in zip(last_inputs, colasts, detect.cv2, detect.cv3):prune(last_input, [colast, cv2[0], cv3[0]])prune(cv2[0], cv2[1])prune(cv2[1], cv2[2])prune(cv3[0], cv3[1])prune(cv3[1], cv3[2])# ***step4，一定要设置所有参数为需要训练。因为加载后的model他会给弄成false。导致报错
# pipeline：
# 1. 为模型的BN增加L1约束，lambda用1e-2左右
# 2. 剪枝模型，比如用全局阈值
# 3. finetune，一定要注意，此时需要去掉L1约束。最终final的版本一定是去掉的
for name, p in yolo.model.named_parameters():p.requires_grad = True# 1. 不能剪枝的layer，其实可以不用约束
# 2. 对于低于全局阈值的，可以删掉整个module
# 3. keep channels，对于保留的channels，他应该能整除n才是最合适的，否则硬件加速比较差
#    n怎么选，一般fp16时，n为8
#                int8时，n为16
#    cp.async.cg.shared
#yolo.val()
# yolo.export(format="onnx")
# yolo.train(data="VOC.yaml", epochs=100)
print("done")

5.YOLOv8剪枝总结

关于yolov8剪枝有以下几点值得注意：

Pipeline:

• 1. 为模型的BN增加L1约束，lambda用1e-2左右
• 2. 剪枝模型使用的是全局阈值
• 3. finetune模型时，一定要注意，此时需要去掉L1约束，最终的final的版本一定是去掉的(在ultralytics/yolo/engine/trainer.py中注释)
• 4. 对于yolo.model.named_parameters()循环，需要设置p.requires_grad为True

Future work:

• 1. 不能剪枝的layer，其实可以不用约束
• 2. 对于低于全局阈值的，可以删掉整个module
• 3. keep channels，对于保留的channels，它应该能整除n才是最合适的，否则硬件加速比较差

• n怎么选呢？一般fp16时，n为8；int8时，n为16

总结

本次课程学习了YOLOv8的剪枝，主要是对前面剪枝课程的一个总结和实现吧，大体流程就是稀疏训练后进行剪枝最后微调，看着虽然简单，但实际细节把控还是非常多的，比如说哪些部分好剪，哪些部分不好剪，剪枝的过程中如何通过model获取想要prune的module等等，需要对YOLOv8整体网络结构和对ONNX模型的操作非常熟练，这还只是基础理论，实操部分的坑还没踩呢，在之后好好练习练习吧😄