模型在mvtec标准数据集上训练，可以问一下为什么效果不好吗？有什么解决方案吗？

joinusers commented 10 months ago

(SimpleNet) bash run.sh #开始train INFO:main:Command line arguments: main.py --gpu 0 --seed 0 --log_group simplenet_mvtec --log_project MVTecAD_Results --results_path results --run_name firstone net -b wideresnet50 -le layer2 -le layer3 --pretrain_embed_dimension 1536 --target_embed_dimension 1536 --patchsize 3 --meta_epochs 40 --embedding_size 256 --gan_epochs 50 --noise_std 0.015 --dsc_hidden 1024 --dsc_layers 2 --dsc_margin .5 --pre_proj 1 dataset --batch_size 2 --resize 329 --imagesize 288 -d toothbrush mvtec DataHere #命令行参数 INFO:main:Dataset: train=60 test=42 #训练集(60good)、测试集(12good，30fefective) INFO:main:Evaluating dataset [mvtec_toothbrush] (1/1)... #训练单类别数据集“牙刷” /home/jovyan/anaconda3/envs/SimpleNet/lib/python3.8/site-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and will be removed in 0.15, please use 'weights' instead. warnings.warn(/home/jovyan/anaconda3/envs/SimpleNet/lib/python3.8/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or None for 'weights' are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing weights=Wide_ResNet50_2_Weights.IMAGENET1K_V1. You can also use weights=Wide_ResNet50_2_Weights.DEFAULT to get the most up-to-date weights. warnings.warn(msg) INFO:main:Training models (1/1) #单类别训练结果 INFO:main:instance_auroc（I_auroc): 0.986 #表征图片good/defective分类效果
INFO:main:full_pixel_auroc(P_auroc): 0.985 #表征图片分割效果（像素级）

INFO:utils:instance_auroc（I_auroc): 0.986 #多类别平均训练效果 INFO:utils:full_pixel_auroc(P_auroc): 0.985

使用相同的命令行参数，选择test模式： INFO:main:instance_auroc（I_auroc): 0.508 INFO:main:full_pixel_auroc(P_auroc): 0.383

Onlydrink commented 10 months ago

你好方便问一下你这个可视化是怎么展示出来的

joinusers commented 10 months ago

使用了utils文件定义的可视化函数

chenxixxa commented 10 months ago

我也遇到了一样的问题，训练的精度很高有99%，测试的就很低 INFO:utils:instance_auroc: 0.414 INFO:utils:full_pixel_auroc: 0.272 INFO:utils:anomaly_pixel_auroc: 0.229 楼主可以一块讨论下哇

joinusers commented 10 months ago

我也遇到了同样的问题，训练的精度有99%，测试的就很低 INFO:utils:instance_auroc: 0.414 INFO:utils:full_pixel_auroc: 0.272 INFO:utils:anomaly_pixel_auroc: 0.229 楼主可以讨论下哇

好哇，怎么联系你呢

chenxixxa commented 10 months ago

我也遇到了同样的问题，训练的精度有99%，测试的就很低 INFO:utils:instance_auroc: 0.414 INFO:utils:full_pixel_auroc: 0.272 INFO:utils:anomaly_pixel_auroc: 0.229 楼主可以讨论下哇

好哇，怎么联系你呢

chenxixxa2023@163.com

456sjd commented 10 months ago

请问下有没有直接训好的模型我想直接测试下我的数据集

Onlydrink commented 10 months ago

请问下有没有直接训好的模型我想直接测试下我的数据集

他给的参数已经训练好较好的效果，你可以直接用你的数据集试一下。这个是他训练的结果

nanbowan6662 commented 10 months ago

你好，我也遇到了这个问题，请问你解决了吗

rikichou commented 10 months ago

我也复现不了，有已经复现了的朋友吗？

Row Names,instance_auroc,full_pixel_auroc mvtec_screw,0.5925394548063128,0.5941589700607617 mvtec_pill,0.4162575013638843,0.6108649797694258 mvtec_capsule,0.3382528919026725,0.5600852752470121 mvtec_carpet,0.5325040128410915,0.6535846697401231 mvtec_grid,0.5964912280701754,0.43871774004861214 mvtec_tile,0.6792929292929294,0.3910630009927796 mvtec_wood,0.8342105263157895,0.6367744026677434 mvtec_zipper,0.8353466386554622,0.5154027106212969 mvtec_cable,0.41547976011994003,0.5281190755351722 mvtec_toothbrush,0.5083333333333333,0.3827801596608477 mvtec_transistor,0.5295833333333334,0.4716959814170349 mvtec_metal_nut,0.5796676441837731,0.6306286149491259 mvtec_bottle,0.673015873015873,0.5059354621788167 mvtec_hazelnut,0.7253571428571429,0.6382359505905498 mvtec_leather,0.36888586956521746,0.2398478236301243 Mean,0.5750145426437954,0.5198596544739619

joinusers commented 10 months ago

我也复现不了，有已经复现了的朋友吗？

Row Names,instance_auroc,full_pixel_auroc mvtec_screw,0.5925394548063128,0.5941589700607617 mvtec_pill,0.4162575013638843,0.6108649797694258 mvtec_capsule,0.3382528919026725,0.5600852752470121 mvtec_carpet,0.5325040128410915,0.6535846697401231 mvtec_grid,0.5964912280701754,0.43871774004861214 mvtec_tile,0.6792929292929294,0.3910630009927796 mvtec_wood,0.8342105263157895,0.6367744026677434 mvtec_zipper,0.8353466386554622,0.5154027106212969 mvtec_cable,0.41547976011994003,0.5281190755351722 mvtec_toothbrush,0.5083333333333333,0.3827801596608477 mvtec_transistor,0.5295833333333334,0.4716959814170349 mvtec_metal_nut,0.5796676441837731,0.6306286149491259 mvtec_bottle,0.673015873015873,0.5059354621788167 mvtec_hazelnut,0.7253571428571429,0.6382359505905498 mvtec_leather,0.36888586956521746,0.2398478236301243 Mean,0.5750145426437954,0.5198596544739619

可以问一下你训练了多少轮次吗

FFFox-abc commented 10 months ago

训练的精度很好99%，测试精度太低了是怎么回事呀

JoyWang1123 commented 10 months ago

请问您解决了吗是因为什么呢为什么训练之后没有权重文件呢(刚入门不是很懂)

jssyzsfzy commented 9 months ago

我这边也是到底啥问题呢

Leon-Lee-WC commented 9 months ago

因为test里面导入模型的时候有问题，那两个键名和模型名字都有问题

PHaiJun commented 9 months ago

因为test里面导入模型的时候有问题，那两个键名和模型名字都有问题

确实如此，train()保存模型和test()加载模型不一致，改一下test()就可以了

 if "pretrained_enc" in state_dicts:
     self.feature_enc.load_state_dict(state_dicts["pretrained_enc"])
 if "pretrained_dec" in state_dicts:
     self.feature_dec.load_state_dict(state_dicts["pretrained_dec"])

改为：

if "discriminator" in state_dicts:
    self.discriminator.load_state_dict(state_dicts["discriminator"])
if "pre_projection" in state_dicts:
    self.pre_projection.load_state_dict(state_dicts["pre_projection"])

模型文件名也要保持一致

jssyzsfzy commented 8 months ago

效果就是不行换方法了 EfficientAD可以试试效果不错

在 2023-12-18 16:59:18，"joinusers" @.***> 写道：

Closed #32 as completed.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

wu33learn commented 8 months ago

因为test里面导入模型的时候有问题，那两个键名和模型名字都有问题

确实如此，train()保存模型和test()加载模型不一致，改一下test()就可以了
 if "pretrained_enc" in state_dicts:
     self.feature_enc.load_state_dict(state_dicts["pretrained_enc"])
 if "pretrained_dec" in state_dicts:
     self.feature_dec.load_state_dict(state_dicts["pretrained_dec"])
改为：
if "discriminator" in state_dicts:
  self.discriminator.load_state_dict(state_dicts["discriminator"])
if "pre_projection" in state_dicts:
  self.pre_projection.load_state_dict(state_dicts["pre_projection"])
模型文件名也要保持一致

因为test里面导入模型的时候有问题，那两个键名和模型名字都有问题

确实如此，train()保存模型和test()加载模型不一致，改一下test()就可以了
 if "pretrained_enc" in state_dicts:
     self.feature_enc.load_state_dict(state_dicts["pretrained_enc"])
 if "pretrained_dec" in state_dicts:
     self.feature_dec.load_state_dict(state_dicts["pretrained_dec"])
改为：
if "discriminator" in state_dicts:
  self.discriminator.load_state_dict(state_dicts["discriminator"])
if "pre_projection" in state_dicts:
  self.pre_projection.load_state_dict(state_dicts["pre_projection"])
模型文件名也要保持一致

我改了，但是test模式下前后结果一样啊，都非常低

wu33learn commented 8 months ago

因为test里面导入模型的时候有问题，那两个键名和模型名字都有问题

请问你如何修改的，效果很好吗

Leon-Lee-WC commented 8 months ago

因为test里面导入模型的时候有问题，那两个键名和模型名字都有问题

请问你如何修改的，效果很好吗

和论文里面差不多

1s3ss commented 7 months ago

@wu33learn 请问你的加载路径是怎么改的？

1s3ss commented 7 months ago

@wu33learn ok，找到了，谢谢

sdracula commented 7 months ago

@wu33learn ok，找到了，谢谢

请问你现在test效果好吗

Cutemoon1 commented 7 months ago

@sdracula 还是不行 INFO:utils:instance_auroc : 0.508 INFO:utils:full_pixel_auroc : 0.383

sdracula commented 7 months ago

@sdracula 还是不行 INFO:utils:instance_auroc : 0.508 INFO:utils:full_pixel_auroc : 0.383

test or train?

sdracula commented 7 months ago

@sdracula 还是不行 INFO:utils:instance_auroc : 0.508 INFO:utils:full_pixel_auroc : 0 我现在能复现出来了，之前有点问题也是只有0.5

Cutemoon1 commented 7 months ago

@sdracula 还是不行 INFO:utils:instance_auroc : 0.508 INFO:utils:full_pixel_auroc : 0 我现在能复现出来了，之前有点问题也是只有0.5

是test，请问你是怎么做出来的？

sdracula commented 7 months ago

@sdracula 还是不行 INFO:utils:instance_auroc : 0.508 INFO:utils:full_pixel_auroc : 0 我现在能复现出来了，之前有点问题也是只有0.5

是test，请问你是怎么做出来的？

可能是测试集的问题，重新解压一下数据集，再试一下

kimsax commented 5 months ago

因为test里面导入模型的时候有问题，那两个键名和模型名字都有问题

请问你如何修改的，效果很好吗

和论文里面差不多

您好可以分享一下您的代码，和模型文件吗，我这边一直复现不出来好的效果，谢谢

kimsax commented 5 months ago

请问下有没有直接训好的模型我想直接测试下我的数据集

他给的参数已经训练好较好的效果，你可以直接用你的数据集试一下。这个是他训练的结果这个库里没有参数文件和模型文件吧，请问你在哪找到的，可以分享一下吗

DonaldRR / SimpleNet

模型在mvtec标准数据集上训练，可以问一下为什么效果不好吗？有什么解决方案吗？ #32