Honggu Liu

Yes, We do! We have trained MesoNet on FaceForensics++, and we will upload the pretrained model soon.

你好，根据你的问题，我认为就是过拟合了，因为sample dataset数据集中样本实在是太少了，很容易过拟合到400个训练视频，可以考虑两种解决办法。 1. 增加额外的数据集，比如选取DFDC full的一部分。 2. 使用数据增强的手段，对训练集数据多样化。 Honggu Sent from Mail for Windows 10 From: HKCityUmian Sent: 2021年4月23日 10:52 To: HongguLiu/Deepfake-Detection Cc: Subscribed Subject: [HongguLiu/Deepfake-Detection] 关于利用Xception训练时数据集大小的疑问？ (#19) honggu，您好。我最近在利用Xception训练deefake，其中我遇到了一些问题：我的训练精度非常高，但是validation和test的acc却很低或者不变。起初我以为是我Dataloader部分的代码写错了，但是我将train dataset作为validation，却能够在每个epoch下acc能够提升。先声明一下，我采用的不是FF++的数据集和Kaggle上DFDC的full数据集（太大了），而是用的Kaggle上给的sample dataset（大概400个训练视频，400个测试视频），并且在提取人脸后也做了样本平衡的操作。所以，我想问一下经验丰富的您，是否是我采用的数据集太小而导致的问题，是否必须采用full...

sorry, it will come down the problem about copyright. So I can not provid the daatset for you directly.

To train a model with multiple gpus, we use `model = nn.DataParallel(model)` . If you have trained a model with multiple gpus, you must test model with `if isinstance(model, torch.nn.DataParallel):...

We usually train our model with multiple gpu. And this code is support of training with multiple gpu.

We used the open source toolkit dlib, you can refer to https://pypi.org/project/dlib/

Sorry about late to reply. In this project, the main purpose is providing the training process. So we would like you to train your own model. And we will upload...

1. We usually train our model in 50-100 epochs. 2. Of course, you can use the pretraiend model and finetune for your task.

We use 720 videos and 140 videos to train and validate, and we randomly extract 50 frames for each video.

How about the details of 6000 + 6000 videos? Are there 6000 fake videos and 6000 real videos?