点击上方蓝字关注我们
微信公众号:OpenCV学堂 关注获取更多计算机视觉与深度学习知识
torchvision分类介绍
Torchvision高版本支持各种SOTA的图像分类模型,同时还支持不同数据集分类模型的预训练模型的切换。使用起来十分方便快捷,Pytroch中支持两种迁移学习方式,分别是:
- Finetune模式基于预训练模型,全链路调优参数- 冻结特征层模式这种方式只修改输出层的参数,CNN部分的参数冻结
上述两种迁移方式,分别适合大量数据跟少量数据,前一种方式计算跟训练时间会比第二种方式要长点,但是针对大量自定义分类数据效果会比较好。
自定义分类模型修改与训练
加载模型之后,feature_extracting 为true表示冻结模式,否则为finetune模式,相关的代码如下:
def set_parameter_requires_grad(model, feature_extracting): if feature_extracting: for param in model.parameters(): param.requires_grad = False
以resnet18为例,修改之后的自定义训练代码如下:
代码语言:javascript复制model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
# Here the size of each output sample is set to 5.
# Alternatively, it can be generalized to nn.Linear(num_ftrs, len(class_names)).
model_ft.fc = nn.Linear(num_ftrs, 5)
model_ft = model_ft.to(device)
criterion = nn.CrossEntropyLoss()
# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)
# Decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)
model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,
num_epochs=25)
数据集是flowers-dataset,有五个分类分别是:
daisydandelionrosessunflowerstulips
全链路调优,迁移学习训练CNN部分的权重参数
代码语言:javascript复制Epoch 0/24
----------
train Loss: 1.3993 Acc: 0.5597
valid Loss: 1.8571 Acc: 0.7073
Epoch 1/24
----------
train Loss: 1.0903 Acc: 0.6580
valid Loss: 0.6150 Acc: 0.7805
Epoch 2/24
----------
train Loss: 0.9095 Acc: 0.6991
valid Loss: 0.4386 Acc: 0.8049
Epoch 3/24
----------
train Loss: 0.7628 Acc: 0.7349
valid Loss: 0.9111 Acc: 0.7317
Epoch 4/24
----------
train Loss: 0.7107 Acc: 0.7669
valid Loss: 0.4854 Acc: 0.8049
Epoch 5/24
----------
train Loss: 0.6231 Acc: 0.7793
valid Loss: 0.6822 Acc: 0.8049
Epoch 6/24
----------
train Loss: 0.5768 Acc: 0.8033
valid Loss: 0.2748 Acc: 0.8780
Epoch 7/24
----------
train Loss: 0.5448 Acc: 0.8110
valid Loss: 0.4440 Acc: 0.7561
Epoch 8/24
----------
train Loss: 0.5037 Acc: 0.8170
valid Loss: 0.2900 Acc: 0.9268
Epoch 9/24
----------
train Loss: 0.4836 Acc: 0.8360
valid Loss: 0.7108 Acc: 0.7805
Epoch 10/24
----------
train Loss: 0.4663 Acc: 0.8369
valid Loss: 0.5868 Acc: 0.8049
Epoch 11/24
----------
train Loss: 0.4276 Acc: 0.8504
valid Loss: 0.6998 Acc: 0.8293
Epoch 12/24
----------
train Loss: 0.4299 Acc: 0.8529
valid Loss: 0.6449 Acc: 0.8049
Epoch 13/24
----------
train Loss: 0.4256 Acc: 0.8567
valid Loss: 0.7897 Acc: 0.7805
Epoch 14/24
----------
train Loss: 0.4062 Acc: 0.8559
valid Loss: 0.5855 Acc: 0.8293
Epoch 15/24
----------
train Loss: 0.4030 Acc: 0.8545
valid Loss: 0.7336 Acc: 0.7805
Epoch 16/24
----------
train Loss: 0.3786 Acc: 0.8730
valid Loss: 1.0429 Acc: 0.7561
Epoch 17/24
----------
train Loss: 0.3699 Acc: 0.8763
valid Loss: 0.4549 Acc: 0.8293
Epoch 18/24
----------
train Loss: 0.3394 Acc: 0.8788
valid Loss: 0.2828 Acc: 0.9024
Epoch 19/24
----------
train Loss: 0.3300 Acc: 0.8834
valid Loss: 0.6766 Acc: 0.8537
Epoch 20/24
----------
train Loss: 0.3136 Acc: 0.8906
valid Loss: 0.5893 Acc: 0.8537
Epoch 21/24
----------
train Loss: 0.3110 Acc: 0.8901
valid Loss: 0.4909 Acc: 0.8537
Epoch 22/24
----------
train Loss: 0.3141 Acc: 0.8931
valid Loss: 0.3930 Acc: 0.9024
Epoch 23/24
----------
train Loss: 0.3106 Acc: 0.8887
valid Loss: 0.3079 Acc: 0.9024
Epoch 24/24
----------
train Loss: 0.3143 Acc: 0.8923
valid Loss: 0.5122 Acc: 0.8049
Training complete in 25m 34s
Best val Acc: 0.926829
冻结CNN部分,只训练全连接分类权重
代码语言:javascript复制Params to learn:
fc.weight
fc.bias
Epoch 0/24
----------
train Loss: 1.0217 Acc: 0.6465
valid Loss: 1.5317 Acc: 0.8049
Epoch 1/24
----------
train Loss: 0.9569 Acc: 0.6947
valid Loss: 1.2450 Acc: 0.6829
Epoch 2/24
----------
train Loss: 1.0280 Acc: 0.6999
valid Loss: 1.5677 Acc: 0.7805
Epoch 3/24
----------
train Loss: 0.8344 Acc: 0.7426
valid Loss: 1.1053 Acc: 0.7317
Epoch 4/24
----------
train Loss: 0.9110 Acc: 0.7250
valid Loss: 1.1148 Acc: 0.7561
Epoch 5/24
----------
train Loss: 0.9049 Acc: 0.7346
valid Loss: 1.1541 Acc: 0.6341
Epoch 6/24
----------
train Loss: 0.8538 Acc: 0.7465
valid Loss: 1.4098 Acc: 0.8293
Epoch 7/24
----------
train Loss: 0.9041 Acc: 0.7349
valid Loss: 0.9604 Acc: 0.7561
Epoch 8/24
----------
train Loss: 0.8885 Acc: 0.7468
valid Loss: 1.2603 Acc: 0.7561
Epoch 9/24
----------
train Loss: 0.9257 Acc: 0.7333
valid Loss: 1.0751 Acc: 0.7561
Epoch 10/24
----------
train Loss: 0.8637 Acc: 0.7492
valid Loss: 0.9748 Acc: 0.7317
Epoch 11/24
----------
train Loss: 0.8686 Acc: 0.7517
valid Loss: 1.0194 Acc: 0.8049
Epoch 12/24
----------
train Loss: 0.8492 Acc: 0.7572
valid Loss: 1.0378 Acc: 0.7317
Epoch 13/24
----------
train Loss: 0.8773 Acc: 0.7432
valid Loss: 0.7224 Acc: 0.8049
Epoch 14/24
----------
train Loss: 0.8919 Acc: 0.7473
valid Loss: 1.3564 Acc: 0.7805
Epoch 15/24
----------
train Loss: 0.8634 Acc: 0.7490
valid Loss: 0.7822 Acc: 0.7805
Epoch 16/24
----------
train Loss: 0.8069 Acc: 0.7644
valid Loss: 1.4132 Acc: 0.7561
Epoch 17/24
----------
train Loss: 0.8589 Acc: 0.7492
valid Loss: 0.9812 Acc: 0.8049
Epoch 18/24
----------
train Loss: 0.7677 Acc: 0.7688
valid Loss: 0.7176 Acc: 0.8293
Epoch 19/24
----------
train Loss: 0.8044 Acc: 0.7514
valid Loss: 1.4486 Acc: 0.7561
Epoch 20/24
----------
train Loss: 0.7916 Acc: 0.7564
valid Loss: 1.0575 Acc: 0.8049
Epoch 21/24
----------
train Loss: 0.7922 Acc: 0.7647
valid Loss: 1.0406 Acc: 0.7805
Epoch 22/24
----------
train Loss: 0.8187 Acc: 0.7647
valid Loss: 1.0965 Acc: 0.7561
Epoch 23/24
----------
train Loss: 0.8443 Acc: 0.7503
valid Loss: 1.6163 Acc: 0.7317
Epoch 24/24
----------
train Loss: 0.8165 Acc: 0.7583
valid Loss: 1.1680 Acc: 0.7561
Training complete in 20m 7s
Best val Acc: 0.829268
测试结果:
零代码训练演示
我已经完成torchvision中分类模型自定义数据集迁移学习的代码封装与开发,支持基于收集到的数据集,零代码训练,生成模型。图示如下:
轻松支持十种主流的CNN模型
self.models_combox.addItem("resnet18")self.models_combox.addItem("resnet34")self.models_combox.addItem("resnet50")self.models_combox.addItem("resnet101")self.models_combox.addItem("inception")self.models_combox.addItem("densenet")self.models_combox.addItem("wide_resnet50")self.models_combox.addItem("wide_resnet101")self.models_combox.addItem("resnext50_32x4d")self.models_combox.addItem("resnext101_32x8d")
扫码获取YOLOv5 TensorRT INT8量化脚本与视频教程
扫码查看OpenCV OpenVIO Pytorch系统化学习路线图
推荐阅读
CV全栈开发者说 - 从传统算法到深度学习怎么修炼
2022入坑深度学习,我选择Pytorch框架!
Pytorch轻松实现经典视觉任务
教程推荐 | Pytorch框架CV开发-从入门到实战
OpenCV4 C 学习 必备基础语法知识三
OpenCV4 C 学习 必备基础语法知识二
OpenCV4.5.4 人脸检测 五点landmark新功能测试
OpenCV4.5.4人脸识别详解与代码演示
OpenCV二值图象分析之Blob分析找圆
OpenCV4.5.x DNN YOLOv5 C 推理
OpenCV4.5.4 直接支持YOLOv5 6.1版本模型推理
OpenVINO2021.4 YOLOX目标检测模型部署测试
比YOLOv5还厉害的YOLOX来了,官方支持OpenVINO推理