Git Product home page Git Product logo

Comments (29)

shicai avatar shicai commented on May 23, 2024 7

finetuning很简单,你在开头加data层,结尾加loss/acc层,改变fc6的名字和类别。
注意用param { lr_mult: 0 decay_mult: 0 }固定batchnorm层的参数

from densenet-caffe.

limin1130 avatar limin1130 commented on May 23, 2024 1

我训练的时候,accuracy一直为0,这个是怎么回事

from densenet-caffe.

yefanhust avatar yefanhust commented on May 23, 2024 1

我是用python重写了一遍网络结构再生成的prototxt 没用这里的

from densenet-caffe.

shicai avatar shicai commented on May 23, 2024

All these models are converted from https://github.com/liuzhuang13/DenseNet
So you can ask Zhuang Liu for more details.

from densenet-caffe.

limin1130 avatar limin1130 commented on May 23, 2024

我是想基于你发布的caffe模型进行微调,因为看到你发布的prototxt的layer名与Zhuang Liu发布的配置文件不同,所以想问问你进行训练时用的配置文件

from densenet-caffe.

gittigxuy avatar gittigxuy commented on May 23, 2024

那请问一下,data层的crop_size必须指定为ImageNet里面的224*224的吗?你这个应该是deploy.prototxt吧?100多MB的caffemodel是准确率最高的caffemodel吗?然后我能否利用你已经上传的DenseNet的caffemodel来进行图片的性别分类呢?还是需要重新利用你给的网络模型结构来进行训练呢?

from densenet-caffe.

shicai avatar shicai commented on May 23, 2024

data层crop_size在训练的时候是224,你可以不指定到这个大小。
这是deploy文件,准确率已经写在表格里。
可以,你可以利用这个预训练模型,也可以自己从头训练,你自己看着办都行。

from densenet-caffe.

limin1130 avatar limin1130 commented on May 23, 2024

使用161 fine-tuning 的时候,使用的图片是224*224,
train batchsize=64,出现error == cudaSuccess (2 vs. 0) out of memory
改成batchsize=2,还是报错,检查显卡使用了11G多,显卡用的是K40
改batchsize=1,运行train成功,显卡使用2000M的样子,奇怪,怎么改成batchsize=1,会降低那么多,是哪里有问题吗?
另外,finetuning的时候发现loss震荡较大。

from densenet-caffe.

OPPOA113 avatar OPPOA113 commented on May 23, 2024

@shicai
你好 网络有实验过检测任务吗 效果会怎样?

from densenet-caffe.

shicai avatar shicai commented on May 23, 2024

@limin1130 关于显存,作者最近提出了一个内存优化的方案,见论文《Memory-Efficient Implementation of DenseNets

@OPPOA113 我自己没用DenseNet做过检测,不过作者透露将很快发布一篇论文,使用DenseNet做检测也能获得state of the art的结果。所以,DenseNet做检测应该没什么问题。

from densenet-caffe.

gittigxuy avatar gittigxuy commented on May 23, 2024

@OPPOA113
用DenseNet做检测任务已经做好了吗?

from densenet-caffe.

OPPOA113 avatar OPPOA113 commented on May 23, 2024

@gittigxuy
NO 做过自己的数据的分类实验,参数没调试好,效果不理想。你那边可有检测?

from densenet-caffe.

gittigxuy avatar gittigxuy commented on May 23, 2024

@OPPOA113 ,我还没做检测,我做图片2分类,但是效果不太理想,可能是初学的原因吧,再略加修改一下,你那的图片分类的项目代码可否给我看一下?如果可以的话,麻烦发邮件到[email protected],谢谢了

from densenet-caffe.

OPPOA113 avatar OPPOA113 commented on May 23, 2024

@gittigxuy
就是caffe的代码训练的cifar10 没有什么其他的代码

from densenet-caffe.

gittigxuy avatar gittigxuy commented on May 23, 2024

@shicai ,求大佬给个用到DenseNet的项目案例,我train - loss图波动很大,找不到原因

from densenet-caffe.

yefanhust avatar yefanhust commented on May 23, 2024

Hi, 我从零开始训练DenseNet121,将fc6换成InnerProduct layer,然后加上Accuracy和SoftmaxWithLoss,loss基本不变,accuracy一直为0。是什么原因呢?

from densenet-caffe.

shicai avatar shicai commented on May 23, 2024

@yefanhust 只凭你的只言片语,我也不确定原因是什么。即使是随机的初始化,accuracy也会接近0.1%,不会是0的。

from densenet-caffe.

yefanhust avatar yefanhust commented on May 23, 2024

DenseNet121.txt
@shicai 这里附上我修改过的DenseNet121.prototxt neng 能麻烦帮我看下吗?训练中loss和accuracy一直停留在最初的水平。

from densenet-caffe.

yefanhust avatar yefanhust commented on May 23, 2024

@shicai 再附上solver的config,完全是按照原始paper中的数据设置的。
solver.txt

from densenet-caffe.

shicai avatar shicai commented on May 23, 2024

@yefanhust (1) 确定你的batch size是256,我看文件中是8;(2) 如果从头训练,需要给Convolution和Scale层加上param参数,控制lr和wd,有几个参数就加几个param,比如conv不带bias就加一个,scale层带bias就加两个:param { lr_mult: 1 decay_mult: 0 }

from densenet-caffe.

yefanhust avatar yefanhust commented on May 23, 2024

@shicai (1) 在服务器上训练时是256,8是在单机上测试用的 (2) 好的,我会根据你的建议加上相应的参数
谢谢

from densenet-caffe.

nerddd avatar nerddd commented on May 23, 2024

@yefanhust 请问你训练的时候256,为什么我训练的时候batch_size为10,都out of memory,我的显卡是GeForce GTX TITAN X

from densenet-caffe.

yefanhust avatar yefanhust commented on May 23, 2024

@nerddd 我是在DGX-1上训练的,用的8块P100

from densenet-caffe.

xiaocmxiao avatar xiaocmxiao commented on May 23, 2024

@shicai
你好。我用你提供的deploy改写的train_val进行训练,loss一直是2.5左右,不下降。我改train_val只改了开头的数据层和在结尾加了accuracy和softmaxwithloss,请问还有哪里需要修改?

from densenet-caffe.

jackchowtop1 avatar jackchowtop1 commented on May 23, 2024

@yefanhust 想问一下你问题解决了吗?解决了的话,可不可以给看一下修改后的prototxt文件?

from densenet-caffe.

RuWang15 avatar RuWang15 commented on May 23, 2024

I modified the prototxt according to the comments under this issue. The only difference is that my batchsize is 10, and the loss does not converge and it explodes. I wonder must the batchsize be 256? And is there a correct version of train_val.prototxt for densenet?

from densenet-caffe.

JaosonMa avatar JaosonMa commented on May 23, 2024

@yefanhust 你可以把你的生成densenet的python脚本分享一下吗?或者将densenet-caffe的train_test.prototxt和solver.prototxt分享一下,我的邮箱 [email protected]
谢谢

from densenet-caffe.

Joeyeee avatar Joeyeee commented on May 23, 2024

@yefanhust 四块P100表示hold不住batchsize256 多显卡跑caffe 你的cudnn禁用了么

from densenet-caffe.

XUYUNYUN666 avatar XUYUNYUN666 commented on May 23, 2024

@shicai
你好。我用你提供的deploy改写的train_val进行训练,loss一直是2.5左右,不下降。我改train_val只改了开头的数据层和在结尾加了accuracy和softmaxwithloss,请问还有哪里需要修改?

请问能加个联系方式吗?我也准备finetuning ,也只是改了开头和加了结尾,我的QQ1443563995

from densenet-caffe.

Related Issues (20)

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.