search

syfafterzy / SVDNet-for-Pedestrian-Retrieval

Code for https://arxiv.org/abs/1703.05693
138 stars 47 forks source link

Loss no convergence #9

Open TengliEd opened 6 years ago

TengliEd commented 6 years ago

I downloaded the raw caffemodel from http://dl.caffe.berkeleyvision.org/bvlc_reference_caffenet.caffemodel and modified the data paths and then trained it using the provided config prototxt file (solver_linear.prototxt). But loss did not converge around 6.5. What else modifications are needed?

TengliEd commented 6 years ago

Or someone who has repeated this experiment can specify steps to make it from this code repository? Cheers!

ghost commented 6 years ago

Maybe you can show your log, and you can also check the IDE baseline (https://github.com/zhunzhong07/IDE-baseline-Market-1501).

发送自 Windows 10 版邮件应用

发件人: Teng Li 发送时间: 2018年3月13日 12:00 收件人: syfafterzy/SVDNet-for-Pedestrian-Retrieval 抄送: Subscribed 主题: Re: [syfafterzy/SVDNet-for-Pedestrian-Retrieval] Loss no convergence(#9)

Or someone who has repeated this experiment can specify steps to make it from this code repository? Cheers! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

TengliEd commented 6 years ago

train_linear.log

TengliEd commented 6 years ago

I suppose I made bad lmdb. Could you plz provide files to produce lmdb data?

ghost commented 6 years ago
  1. I check you log, and Finetuning from SVDNet/caffenet/1024d_linear.caffemodel seems strange.

Do you train 1024d_linear.caffemodel by yourself? Step 2: Train a model with a linear fully-connected layer based on caffenet or resnet.

  1. About LMDB:

    see here https://github.com/syfafterzy/SVDNet-for-Pedestrian-Retrieval/issues/1 or http://caffe.berkeleyvision.org/gathered/examples/imagenet.html

TengliEd commented 6 years ago

Many Thanks @Simon4john . I found I used person IDs as labels. Now I have no idea extracting features on test data to evaluate.

ghost commented 6 years ago

@TengliEd To evaluate the trained model, you can see here

Good luck!

TengliEd commented 6 years ago

I evaluated the caffenet trained by RRI but results did not achieve as high level as mentioned in the paper (R1: 72.62%; mAP: 44.85%). I used Euclidean distance to measure similarity. Should I use other measurements to achieve your model performance?

ghost commented 6 years ago

During testing, the image needs to subtract mean. Be sure the mean is same for training and testing.

ghost commented 6 years ago

测试时候，需要减去的均值文件和训练的必须保持一致，你检查一下。

wjzh1 commented 6 years ago

@Simon4john 请问我用的是resnet18网络，在分类层前面加了一个fc层，在用作者matlab代码转换的时候出现索引超出矩阵维度，请问要怎么修改change_resnet_W.m这代码

ghost commented 6 years ago

@wjzh1 检查一下https://github.com/syfafterzy/SVDNet-for-Pedestrian-Retrieval/blob/439ae0e086304cc8e38c8112b2a96e1a8ebdf647/matlab/SVDNet/change_resnet_W.m#L1 deploy renset的时候你得prototext和model是对应的吗？按照给了resnet50的案例照着理一遍，就可以训练出来的。仔细一点。

wjzh1 commented 6 years ago

------------------ 原始邮件 ------------------ 发件人: "John"notifications@github.com; 发送时间: 2018年4月16日(星期一) 下午4:58 收件人: "syfafterzy/SVDNet-for-Pedestrian-Retrieval"SVDNet-for-Pedestrian-Retrieval@noreply.github.com; 抄送: "beyond"1678824369@qq.com; "Mention"mention@noreply.github.com; 主题: Re: [syfafterzy/SVDNet-for-Pedestrian-Retrieval] Loss no convergence(#9)

@wjzh1 检查一下https://github.com/syfafterzy/SVDNet-for-Pedestrian-Retrieval/blob/439ae0e086304cc8e38c8112b2a96e1a8ebdf647/matlab/SVDNet/change_resnet_W.m#L1 deploy renset的时候你得prototext和model是对应的吗？按照给了resnet50的案例照着理一遍，就可以训练出来的。仔细一点。

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.