accuracy =0

[Mebigger]

您好，请问在训练到iteration=100时，accuracy=0是正常现象吗，loss在39到40之间
用的函数是train（resume=None）
感谢您的回复

[CaptainEven]

@Mebigger 训练过程需要初始化权重，通过ImageNet训练好的vgg16模型权重初始化，使用我刚上传的InitRepNet.py训练就没问题了。

[Mebigger]

@CaptainEven
非常感谢您的回复，
您的意思是把train函数中的net=RepNet 改成 net=InitRepNet 吗，如果是的话，我改完之后accuary还是0（我已经导入了vgg16网络和预训练的权重）。

[CaptainEven]

@Mebigger 检查数据和标签是否正确，是否是按照ProcessVehicleID.py处理的数据？

[Mebigger]

@CaptainEven 对的，标签先经过了ProcessVehicleID.py的处理，print了vgg.state_dict(),确认权重模型没问题

[CaptainEven]

@Mebigger 贴训练的中间输出看看

[Mebigger]

@CaptainEven
C:\Softwares\python.exe C:/WYX-Py_projects/Veichle_ReID/Program/RepNet-VehicleReID-master/RepNet.py
=> out_ids: 10086, out_attribs: 257
=> in dim: 1000, out dim: 10086
=> in dim: 1024, out dim: 10086
=> Mix difference network:
InitRepNet(
(conv1_1): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv1_2): ReLU(inplace)
(conv1_3): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv1_4): ReLU(inplace)
(conv1_5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(conv1): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv2_1): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv2_2): ReLU(inplace)
(conv2_3): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv2_4): ReLU(inplace)
(conv2_5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(conv2): Sequential(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv3_1): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv3_2): ReLU(inplace)
(conv3_3): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv3_4): ReLU(inplace)
(conv3_5): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv3_6): ReLU(inplace)
(conv3_7): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(conv3): Sequential(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv4_1_1): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv4_1_2): ReLU(inplace)
(conv4_1_3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv4_1_4): ReLU(inplace)
(conv4_1_5): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv4_1_6): ReLU(inplace)
(conv4_1_7): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(conv4_1): Sequential(
(0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv4_2_1): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv4_2_2): ReLU(inplace)
(conv4_2_3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv4_2_4): ReLU(inplace)
(conv4_2_5): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv4_2_6): ReLU(inplace)
(conv4_2_7): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(conv4_2): Sequential(
(0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv5_1_1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv5_1_2): ReLU(inplace)
(conv5_1_3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv5_1_4): ReLU(inplace)
(conv5_1_5): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv5_1_6): ReLU(inplace)
(conv5_1_7): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(conv5_1): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv5_2_1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv5_2_2): ReLU(inplace)
(conv5_2_3): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv5_2_4): ReLU(inplace)
(conv5_2_5): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(conv5_2_6): ReLU(inplace)
(conv5_2_7): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(conv5_2): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(FC6_1_1): Linear(in_features=25088, out_features=4096, bias=True)
(FC6_1_2): ReLU(inplace)
(FC6_1_3): Dropout(p=0.5)
(FC6_1_4): Linear(in_features=4096, out_features=4096, bias=True)
(FC6_1_5): ReLU(inplace)
(FC6_1_6): Dropout(p=0.5)
(FC6_1): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace)
(2): Dropout(p=0.5)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace)
(5): Dropout(p=0.5)
)
(FC6_2_1): Linear(in_features=25088, out_features=4096, bias=True)
(FC6_2_2): ReLU(inplace)
(FC6_2_3): Dropout(p=0.5)
(FC6_2_4): Linear(in_features=4096, out_features=4096, bias=True)
(FC6_2_5): ReLU(inplace)
(FC6_2_6): Dropout(p=0.5)
(FC6_2): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace)
(2): Dropout(p=0.5)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace)
(5): Dropout(p=0.5)
)
(FC7_1): Linear(in_features=4096, out_features=1000, bias=True)
(FC7_2): Linear(in_features=4096, out_features=1000, bias=True)
(FC_8): Linear(in_features=2000, out_features=1024, bias=True)
(attrib_classifier): Linear(in_features=1000, out_features=257, bias=True)
(arc_fc_br2): ArcFC()
(arc_fc_br3): ArcFC()
(shared_layers): Sequential(
(0): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(1): Sequential(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(2): Sequential(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
(branch_1_feats): Sequential(
(0): Sequential(
(0): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(1): Sequential(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(2): Sequential(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
(1): Sequential(
(0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(2): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
(branch_1_fc): Sequential(
(0): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace)
(2): Dropout(p=0.5)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace)
(5): Dropout(p=0.5)
)
(1): Linear(in_features=4096, out_features=1000, bias=True)
)
(branch_1): Sequential(
(0): Sequential(
(0): Sequential(
(0): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(1): Sequential(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(2): Sequential(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
(1): Sequential(
(0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(2): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
(1): Sequential(
(0): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace)
(2): Dropout(p=0.5)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace)
(5): Dropout(p=0.5)
)
(1): Linear(in_features=4096, out_features=1000, bias=True)
)
)
(branch_2_feats): Sequential(
(0): Sequential(
(0): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(1): Sequential(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(2): Sequential(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
(1): Sequential(
(0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(2): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
(branch_2_fc): Sequential(
(0): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace)
(2): Dropout(p=0.5)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace)
(5): Dropout(p=0.5)
)
(1): Linear(in_features=4096, out_features=1000, bias=True)
)
(branch_2): Sequential(
(0): Sequential(
(0): Sequential(
(0): Sequential(
(0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(1): Sequential(
(0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(2): Sequential(
(0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
(1): Sequential(
(0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(2): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU(inplace)
(2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(3): ReLU(inplace)
(4): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(5): ReLU(inplace)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
(1): Sequential(
(0): Sequential(
(0): Linear(in_features=25088, out_features=4096, bias=True)
(1): ReLU(inplace)
(2): Dropout(p=0.5)
(3): Linear(in_features=4096, out_features=4096, bias=True)
(4): ReLU(inplace)
(5): Dropout(p=0.5)
)
(1): Linear(in_features=4096, out_features=1000, bias=True)
)
)
)
=> total 75063 samples loaded in train mode
=> total 3894 samples loaded in test mode
=> optimize all layers.

Training...
=> Epoch Train loss Train acc Test acc
=> epoch 1 iter 0/4692, total_iter 0 | loss 40.965 | accuracy 0.000%

[CaptainEven]

@Mebigger 时间有点久，你先训练几个epoch看看，iter100,accuracy为0很正常，我自己这边也检查一下

[Mebigger]

@CaptainEven
好的，再次感谢
tips：我需要把InitRepNet.py中的
feats=vgg_oirg.model.features._modules
classifier=vgg_oirg.model.classifier._modules中的.model删掉才可以训练

[CaptainEven]

@Mebigger 刚刚重新上传了RepNet.py和InitRepNet.py，现在应该没有errrors了

[Mebigger]

@CaptainEven 已经全部改成您的文件了，目前，accuracy还是0，我继续训练试试，再次感谢您的回复

[CaptainEven]

多训练几个epoch应该就慢慢收敛了

[CaptainEven]

@Mebigger from scratch训练我自己没尝试过，我估计也能慢慢收敛

[Mebigger]

@CaptainEven
好的，
您的机器太快了，我的1080ti，不知道要训练到什么时候，
但是我发现gpu的利用率并不是很高，cpu的利用率反而非常高

[CaptainEven]

@Mebigger 细粒度分类是比较难训练的，比如人脸识别....只能让及其慢慢跑了，不行的话，尝试把LR调高点...

[CaptainEven]

@Mebigger 不知道什么原因，我的gpu利用率是100%...

[CaptainEven]

@Mebigger 训练几个epoch之后开始慢慢往收敛走了...

=> epoch 4 iter 1293/4689, total_iter 15360 | loss 19.908 | accuracy 25.000%
=> epoch 4 iter 1303/4689, total_iter 15370 | loss 20.616 | accuracy 12.500%
=> epoch 4 iter 1313/4689, total_iter 15380 | loss 20.296 | accuracy 18.750%
=> epoch 4 iter 1323/4689, total_iter 15390 | loss 25.566 | accuracy 0.000%
=> epoch 4 iter 1333/4689, total_iter 15400 | loss 21.102 | accuracy 6.250%
=> epoch 4 iter 1343/4689, total_iter 15410 | loss 23.225 | accuracy 6.250%
=> epoch 4 iter 1353/4689, total_iter 15420 | loss 23.098 | accuracy 6.250%
=> epoch 4 iter 1363/4689, total_iter 15430 | loss 20.185 | accuracy 12.500%
=> epoch 4 iter 1373/4689, total_iter 15440 | loss 21.110 | accuracy 12.500%
=> epoch 4 iter 1383/4689, total_iter 15450 | loss 21.762 | accuracy 6.250%
=> epoch 4 iter 1393/4689, total_iter 15460 | loss 22.846 | accuracy 6.250%
=> epoch 4 iter 1403/4689, total_iter 15470 | loss 20.441 | accuracy 6.250%
=> epoch 4 iter 1413/4689, total_iter 15480 | loss 19.164 | accuracy 12.500%
=> epoch 4 iter 1423/4689, total_iter 15490 | loss 19.489 | accuracy 6.250%
=> epoch 4 iter 1433/4689, total_iter 15500 | loss 22.107 | accuracy 6.250%
=> epoch 4 iter 1443/4689, total_iter 15510 | loss 21.369 | accuracy 18.750%

[Mebigger]

@CaptainEven
终于跑起来了，我的得改成batchsize=8，内存不够用
之前GPU利用率为0是并没有用gpu，最后找到原因是----os.environ['CUDA_VISIBLE_DEVICES'] = '1'这个地方的“1”，我需要改成“0”，因为我只有一个gpu
下边是网友的解释：
os.environ[“CUDA_DEVICE_ORDER”] = “PCI_BUS_ID” # 按照PCI_BUS_ID顺序从0开始排列GPU设备 
os.environ[“CUDA_VISIBLE_DEVICES”] = “0” #设置当前使用的GPU设备仅为0号设备  设备名称为'/gpu:0'
我觉得和device = torch.device('cuda: 0' if torch.cuda.is_available() else 'cpu')这句话也有关。

非常感谢您的耐心指导以及回复

[CaptainEven]

@Mebigger 对的，因为我这里有8块GPU，通常0号GPU专门用来测试，并不用来训练，所以不是0

[Mebigger]

@CaptainEven
您好，现在训练的有点过拟合，train达到99.99%，test只有49左右，请问您有什么小技巧吗，我正尝试该参数呢，您的CarReID_data是自己创建的数据集吗，对应的pair_set_car.txt内容是什么格式的呢？

[CaptainEven]

@Mebigger 过拟合是因为相对于模型容量大小，训练数据量偏少，但是不影响车辆比对，在进行比对时只需要提取特征向量，计算夹角余弦，跟设定阈值比对即可：

，pair_set_car.txt是VehicleID中留作test的数据(可以参考我刚上传的dataset.py中process_car和gen_test_pairs函数的处理)，格式：每一行一条数据，包括，两张图片的路径和标签（0 or 1,1表示是同一个对象）

[Mebigger]

这个是我用train训练的结果：

下边这是用train_mc训练的结果：


为什么加上FocalLoss之后效果反而到不好了呢？

[CaptainEven]

@Mebigger Foclaloss主要针对正负样本不均衡的情况，特别像object detection中负样本远远大于正样本的情况，至于在分类中这种样本不均衡的情况要小得多，所以FocalLoss是否变现更好是不一定的，跟数据集本身的特点有关。

[Mebigger]

@CaptainEven
您好，您说的arc loss 是arcFace Loss吗？

为什么在train函数中计算veicle_id的时候使用Foclaloss，而不是arc loss呢？

[CaptainEven]

@Mebigger 无论Arc loss还是Focal loss都是基于Softmax loss(或CrossEntropy loss)的改进，本质上仍是Softmax loss，从上面的公式也可看出来，无非是对输入到最后Softmax的向量元素权重有不同。所以，Arc loss和Focal loss可同时使用。当然你可以在最后一层经过Arc loss计算之后直接送入CrossEntropy loss.

[Mebigger]

@CaptainEven
再次感谢您快速的回答；
对的，您说的非常好，又对两个loss有了进一步的认识，显然我没有理解到两个loss的本质。
另外，branch==5时，应该是RepNet的结构吧，为什么只在test的时候用呢，因为我发现上边的图片ID预测的并不是很准确。

[CaptainEven]

@Mebigger branch==5时提取最终的1024维特征向量，test准确率较低的根本原因在于训练数据的不足，导致过拟合。如果要test准确率达到train差不多的水平，估计需要至少10W数量级的训练数据。

[ly0303521]

@CaptainEven 你好，请问你的的训练集有多少张图片，我现在的数据集现在有30万张，有20多万张是我自己的数据，这批数据只有车辆ID的标签，但是因该对模型训练应该有帮助，现在已经训练了14个epoch，现在训练准确率93％，测试准确率只有33％（我的图片尺寸被改成了160，最后的特征向量也改成了256）

[tianmaxingkong666]

@CaptainEven @Mebigger 你好，请教一下，train和train_mc的主要区别是什么呢？

[tkone2018]

@Mebigger 你好，请问你那个结果展示图是博主给的测试脚本吗？我也是想显示以下测试图片的品牌，谢谢