zimyang / maskrelation Goto Github PK
View Code? Open in Web Editor NEWCode of masked relation learning for deepfake detection
License: Apache License 2.0
Code of masked relation learning for deepfake detection
License: Apache License 2.0
您好,在跑训练代码训练模型时,我遇到了一些问题,希望能得到您的帮助,非常感谢!
在训练过程中,我发现训练集的ACC比较稳定的提升,但是验证集的Acc一直在50左右波动,AUC则是波动剧烈,在57-90之间波动,
部分训练结果如下:
epoch16: * Acc:52.410 Auc:84.277 Pre:52.157 Rec:100.000 F1:68.557
epoch17:* Acc:52.259 Auc:64.254 Pre:52.259 Rec:100.000 F1:68.645
epoch18: * Acc:48.720 Auc:57.736 Pre:48.681 Rec:100.000 F1:65.484
epoch19: * Acc:49.699 Auc:79.181 Pre:49.047 Rec:100.000 F1:65.814
epoch20: * Acc:50.452 Auc:83.204 Pre:50.152 Rec:100.000 F1:66.801
考虑到硬件限制,我使用的显卡都是24G内存的。在尝试使用config.py文件中为r3d设定的原始batchsize值(40)时,我遇到了cuda内存超出的问题。
因此,我将batchsize值调整为8,同时学习率也相应地调整为原来的五分之一,即:
self.lr = 4e-5 # 原始值是2e-4
self.warmup_lr = 2e-7 # 原始值是1e-6
在训练时,我使用了总共4块显卡,并且其他参数均保持默认设置。预训练模型是从https://download.pytorch.org/models/mc3_18-a90a0ba3.pth下载的,而训练与验证数据集都是完整的faceforensics++训练与验证集。
想请问一下可能导致上述训练时验证集ACC上不去,AUC剧烈波动的原因,非常感谢!
此外,我注意到原始代码中的参数与Masked Relation Learning for DeepFake Detection文章中所提及的参数存在差异。
原始代码中,初始学习速率为2e-4,batchsize为40,epoch为100,采用warm up以及cosine策略。
文章中提到的初始学习速率为1e-4,batchsize为30,epoch为30,每5个epoch学习率变为原来的2分之一。
因此,我还有一个问题:如果我想在faceforensics++数据集上复现文章中的结果,我是否应该将这些参数调整为与文章一致?
同时,文章中的batchsize值是指单GPU的batchsize还是多个GPU的batchsize总和?如果是单GPU的batchsize为30,那么单GPU的内存需求似乎会非常高,接近50G。
期待您的回复,非常感谢!
Thank you for your great work. Fig. 1. in the paper shows facial regions segmentation map and treats them as nodes. How is the seg map obtained, and is it implicit in the code to obtain the facial region?
请问可以提供具体的json文件嘛!特别需要,感谢感谢,跪求!
ffpp_raw_train = "/data2/ziming.yang/datasets/ffpp/ffpp_raw_train.json"
ffpp_raw_val = "/data2/ziming.yang/datasets/ffpp/ffpp_raw_val.json"
ffpp_raw_test = "/data2/ziming.yang/datasets/ffpp/ffpp_raw_test.json"
ffpp_c23_train = "/data2/ziming.yang/datasets/ffpp/ffpp_c23_train.json"
ffpp_c23_val = "/data2/ziming.yang/datasets/ffpp/ffpp_c23_val.json"
ffpp_c23_test = "/data2/ziming.yang/datasets/ffpp/ffpp_c23_test.json"
ffpp_c40_train = "/data2/ziming.yang/datasets/ffpp/ffpp_c40_train.json"
ffpp_c40_val = "/data2/ziming.yang/datasets/ffpp/ffpp_c40_val.json"
ffpp_c40_test = "/data2/ziming.yang/datasets/ffpp/ffpp_c40_test.json"
Has anyone successfully run this paper? I ran the pytorch version and it performed very well on the training set, but it performed very poorly on the validation set, can anyone help me?
您好,我想请问一下有数据预处理的代码吗,或者可以麻烦提供一下预处理后数据的存储的结构吗,非常感谢!
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.