minivision-ai/silent-face-anti-spoofing

静默活体检测（Silent-Face-Anti-Spoofing）

silent-face-anti-spoofing's Issues

關於train的一些問題

由於我以雜誌上的人去做驗證，發現效果並不佳，所以我想train來提昇這部份
我在train時碰到些問題想詢問一下：

1. 我需要在datasets下依目錄格式去去建立資料夾並放入我要的訓練的資料？
2. patch_info參數應該選org_1_80x60 / 1_80x80 / 2.7_80x80 / 4_80x80，也就是說它只會訓練我選的資料？
3. org_1_80x60，下有0、1、2的資料夾，應該為標籤，請問真臉、假臉2d、假臉3d分別對應標籤0、1、2？
4. 要訓練的資料需要先自行調整這些格式（org_1_80x60 / 1_80x80 / 2.7_80x80 / 4_80x80）？
5. 可以發布每個標籤的訓練資料各一比嗎？希望有個範本來建立自己的資料集

convert caffe model not work well

使用PytorchToCaffe转换出来的模型(转换的prototxt与官方的一样)，输入图片推理得到的3个值一样大小，用官方提供的推理没有问题, 请问是哪里出了问题 ？是否可以提供转换caffe模型的工具

caffe模型和项目中resources文件中模型的区别

请问7月30日开源的caffe模型和项目中resources文件中的模型精度有区别吗

What are the output classes?

@minivision-ailab Many thanks for the great software.

What are the 3 classes of the classification models? Real, Fake and..?

Thanks again

Do you have a validation dataset for training?

Hi you, thanks for your work. I have read your code and seem as you don't use validation in training process.
Whether Am I wrong? Can you clarify it for me. I am very appreciated it. Thank you!!

pc端运行速度

如题，服务器上cpu和gpu运行速度大概都是多少

na

na

数据集分3类，是哪些类，可以写下数据集的格式嘛

为什么是3分类？是哪3个分类鸭？

手机屏幕图片效果误判率高

 你好，我这边调用windows笔记本调用摄像头，然后用手机屏幕打开自拍照，测试效果不是很理想，阈值很高，达到0.99

Why no pading to square before resizing

你好，请问为什么预处理的时候不先把人脸框padding 成square然后再做resize呢，这样不会另到人脸框的aspect ratio 变化造成畸变吗

conversion to caffe Error?

Really nice.
I was trying to convert the given models in caffe using https://github.com/xxradon/PytorchToCaffe
The first model was successfully converted i.e 2.7_80x80_MiniFASNetV2.pth but for model 4_0_0_80x80_MiniFASNetV1SE.pth
error occurs:
conv1.conv
conv: blob1
conv1 was added to layers
139799641391824:conv_blob1 was added to blobs
conv1.bn
batch_norm1 was added to layers
139799641391176:batch_norm_blob1 was added to blobs
bn_scale1 was added to layers
conv1.prelu
prelu1 was added to layers
139799641391680:prelu_blob1 was added to blobs
conv2_dw.conv
conv: prelu_blob1
conv2 was added to layers
139799641392040:conv_blob2 was added to blobs
conv2_dw.bn
batch_norm2 was added to layers
139799641391752:batch_norm_blob2 was added to blobs
bn_scale2 was added to layers
conv2_dw.prelu
prelu2 was added to layers
139799641392256:prelu_blob2 was added to blobs
conv_23.conv.conv
conv: prelu_blob2
conv3 was added to layers
139799641392616:conv_blob3 was added to blobs
conv_23.conv.bn
batch_norm3 was added to layers
139799641392976:batch_norm_blob3 was added to blobs
bn_scale3 was added to layers
conv_23.conv.prelu
prelu3 was added to layers
139799641392184:prelu_blob3 was added to blobs
conv_23.conv_dw.conv
conv: prelu_blob3
conv4 was added to layers
139799641392400:conv_blob4 was added to blobs
conv_23.conv_dw.bn
batch_norm4 was added to layers
139799641391536:batch_norm_blob4 was added to blobs
bn_scale4 was added to layers
conv_23.conv_dw.prelu
prelu4 was added to layers
139799641391320:prelu_blob4 was added to blobs
conv_23.project.conv
conv: prelu_blob4
conv5 was added to layers
139799641391608:conv_blob5 was added to blobs
conv_23.project.bn
batch_norm5 was added to layers
139799641392328:batch_norm_blob5 was added to blobs
bn_scale5 was added to layers
conv_3.model.0.conv.conv
conv: batch_norm_blob5
conv6 was added to layers
139799641391464:conv_blob6 was added to blobs
conv_3.model.0.conv.bn
batch_norm6 was added to layers
139799641392760:batch_norm_blob6 was added to blobs
bn_scale6 was added to layers
conv_3.model.0.conv.prelu
prelu5 was added to layers
139799641392904:prelu_blob5 was added to blobs
conv_3.model.0.conv_dw.conv
conv: prelu_blob5
conv7 was added to layers
139799641392112:conv_blob7 was added to blobs
conv_3.model.0.conv_dw.bn
batch_norm7 was added to layers
139799641391896:batch_norm_blob7 was added to blobs
bn_scale7 was added to layers
conv_3.model.0.conv_dw.prelu
prelu6 was added to layers
139799641392472:prelu_blob6 was added to blobs
conv_3.model.0.project.conv
conv: prelu_blob6
conv8 was added to layers
139799641394992:conv_blob8 was added to blobs
conv_3.model.0.project.bn
batch_norm8 was added to layers
139799641394920:batch_norm_blob8 was added to blobs
bn_scale8 was added to layers
add1 was added to layers
139799641395064:add_blob1 was added to blobs
conv_3.model.1.conv.conv
conv: add_blob1
conv9 was added to layers
139799641394704:conv_blob9 was added to blobs
conv_3.model.1.conv.bn
batch_norm9 was added to layers
139799641394632:batch_norm_blob9 was added to blobs
bn_scale9 was added to layers
conv_3.model.1.conv.prelu
prelu7 was added to layers
139799641394776:prelu_blob7 was added to blobs
conv_3.model.1.conv_dw.conv
conv: prelu_blob7
conv10 was added to layers
139799641394416:conv_blob10 was added to blobs
conv_3.model.1.conv_dw.bn
batch_norm10 was added to layers
139799641394344:batch_norm_blob10 was added to blobs
bn_scale10 was added to layers
conv_3.model.1.conv_dw.prelu
prelu8 was added to layers
139799641394488:prelu_blob8 was added to blobs
conv_3.model.1.project.conv
conv: prelu_blob8
conv11 was added to layers
139799641394128:conv_blob11 was added to blobs
conv_3.model.1.project.bn
batch_norm11 was added to layers
139799641394056:batch_norm_blob11 was added to blobs
bn_scale11 was added to layers
add2 was added to layers
139799641394272:add_blob2 was added to blobs
conv_3.model.2.conv.conv
conv: add_blob2
conv12 was added to layers
139799641393840:conv_blob12 was added to blobs
conv_3.model.2.conv.bn
batch_norm12 was added to layers
139799641393768:batch_norm_blob12 was added to blobs
bn_scale12 was added to layers
conv_3.model.2.conv.prelu
prelu9 was added to layers
139799641393984:prelu_blob9 was added to blobs
conv_3.model.2.conv_dw.conv
conv: prelu_blob9
conv13 was added to layers
139799641393552:conv_blob13 was added to blobs
conv_3.model.2.conv_dw.bn
batch_norm13 was added to layers
139799641393480:batch_norm_blob13 was added to blobs
bn_scale13 was added to layers
conv_3.model.2.conv_dw.prelu
prelu10 was added to layers
139799641393696:prelu_blob10 was added to blobs
conv_3.model.2.project.conv
conv: prelu_blob10
conv14 was added to layers
139799641393264:conv_blob14 was added to blobs
conv_3.model.2.project.bn
batch_norm14 was added to layers
139799640977344:batch_norm_blob14 was added to blobs
bn_scale14 was added to layers
add3 was added to layers
139799641393408:add_blob3 was added to blobs
conv_3.model.3.conv.conv
conv: add_blob3
conv15 was added to layers
139799640976912:conv_blob15 was added to blobs
conv_3.model.3.conv.bn
batch_norm15 was added to layers
139799640976768:batch_norm_blob15 was added to blobs
bn_scale15 was added to layers
conv_3.model.3.conv.prelu
prelu11 was added to layers
139799640977200:prelu_blob11 was added to blobs
conv_3.model.3.conv_dw.conv
conv: prelu_blob11
conv16 was added to layers
139799640976336:conv_blob16 was added to blobs
conv_3.model.3.conv_dw.bn
batch_norm16 was added to layers
139799640976192:batch_norm_blob16 was added to blobs
bn_scale16 was added to layers
conv_3.model.3.conv_dw.prelu
prelu12 was added to layers
139799640976624:prelu_blob12 was added to blobs
conv_3.model.3.project.conv
conv: prelu_blob12
conv17 was added to layers
139799640975760:conv_blob17 was added to blobs
conv_3.model.3.project.bn
batch_norm17 was added to layers
139799640975616:batch_norm_blob17 was added to blobs
bn_scale17 was added to layers
conv_3.model.3.se_module.avg_pool
ave_pool1 was added to layers
139799640975184:ave_pool_blob1 was added to blobs
conv_3.model.3.se_module.fc1
conv: ave_pool_blob1
conv18 was added to layers
139799640976048:conv_blob18 was added to blobs
conv_3.model.3.se_module.bn1
batch_norm18 was added to layers
139799640975040:batch_norm_blob18 was added to blobs
bn_scale18 was added to layers
conv_3.model.3.se_module.relu
relu1 was added to layers
139799640975472:relu_blob1 was added to blobs
conv_3.model.3.se_module.fc2
conv: relu_blob1
conv19 was added to layers
139799640974608:conv_blob19 was added to blobs
conv_3.model.3.se_module.bn2
batch_norm19 was added to layers
139799640974464:batch_norm_blob19 was added to blobs
bn_scale19 was added to layers
mul1 was added to layers
139799640974032:mul_blob1 was added to blobs
WARNING: CANNOT FOUND blob 139799640974896
Traceback (most recent call last):
File "silent_spoof_convert_to_caffe.py", line 36, in
pytorch_to_caffe.trans_net(net,input,name)
File "./pytorch_to_caffe.py", line 786, in trans_net
out = net.forward(input_var)
File "/home/boson/WORKDONE/FaceSpoofing/Silent-Face-Anti-Spoofing/src/model_lib/MiniFASNet.py", line 222, in forward
out = self.conv_3(out)
File "/home/boson/Pytorch-cu9.2/venv/lib/python3.6/site-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, **kwargs)
File "/home/boson/WORKDONE/FaceSpoofing/Silent-Face-Anti-Spoofing/src/model_lib/MiniFASNet.py", line 134, in forward
return self.model(x)
File "/home/boson/Pytorch-cu9.2/venv/lib/python3.6/site-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, **kwargs)
File "/home/boson/Pytorch-cu9.2/venv/lib/python3.6/site-packages/torch/nn/modules/container.py", line 92, in forward
input = module(input)
File "/home/boson/Pytorch-cu9.2/venv/lib/python3.6/site-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, **kwargs)
File "/home/boson/WORKDONE/FaceSpoofing/Silent-Face-Anti-Spoofing/src/model_lib/MiniFASNet.py", line 156, in forward
x = self.se_module(x)
File "/home/boson/Pytorch-cu9.2/venv/lib/python3.6/site-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, **kwargs)
File "/home/boson/WORKDONE/FaceSpoofing/Silent-Face-Anti-Spoofing/src/model_lib/MiniFASNet.py", line 113, in forward
return module_input * x
File "./pytorch_to_caffe.py", line 583, in _mul
bottom=[log.blobs(input), log.blobs(args[0])], top=top_blobs)
File "./Caffe/layer_param.py", line 33, in init
self.bottom.extend(bottom)
TypeError: None has type NoneType, but expected one of: bytes, unicode

Can you help?

根本一点都不准确，太差了，对准图片，手不要抖动拍照就会骗过算法；训练部分用到光流，识别部分也用上吧！！

更本一点都不准确，太差了，对准图片，手不要抖动拍照就会骗过算法；训练部分用到光流，识别部分也用上吧！

Why do we add the model predictions?

In test.py the softmaxed model probabilities are summed.

prediction += model_test.predict(img, os.path.join(model_dir, model_name))

This will result in values > 1 which makes no sense for probabilities.

Is there a better way to proceed?

Thanks!

變為即時影像偵測

我現在在嘗試使用電腦相機去做即時的偵測與顯示，我是以cpu去跑，但其FPS不盡理想，不能像demo上的動圖一樣理想，我是以更改test.py去完成，請問更改test.py需要注意些甚麼?請問影響FPS最大部分code是?

论文链接

您好！论文可以分享一下吗，为什么不是live,spoof二分类，三分类是什么，谢谢！

org_1_heightxwidth是80x60吗？

原图(org_1_heightxwidth),直接将原图resize到固定尺寸(width, height),如图1所示;

原图(org_1_heightxwidth)是把数据集中宽高比为width/height=3/4的图片resize到80*60 后得到的吗？

计算花费时间部分单位有误

time.time()获取的是秒数.毫秒数，直接相减的单位应该是秒数，而不是ms，例如我自己的电脑显示0.11ms 而实际上是0.11s

Dataset Overview?

Hello, i really want to understand what type of dataset you used. I know it has three class: 2D-Fake, 3D-Fake and Real but my question is can you provide a sample images of these three classes.
Also have you used any opensource dataset like CelebA or FFHQ for Real face samples. And How you collected these dataset. Just an overview will help in training.
Also can you describe the procedure how you collected/make the dataset, and size of the dataset (number of images for each class)

Thanks

Campare the single model and bagging model？

Nice job. I read the test.py and find that inference process uses two models to predict the labels prob. Does it get batter result than single model?

Question about model_type, num_classes,image size ?

Hi,
First, thanks for your excellent product.
I am currently testing your ant-spoofing program.
There are some question.

1. What is the role of "model_type"(MiniFASNetV1/MiniFASNetV2/MiniFASNetV1SE/MiniFASNetV2SE) ?
   There is only 4 items or can be added ?

2. What is the role of "num_classes" ?
   There is 3 or 4 in your source ...
   the 3 or 4 means "0" , "1", "2" ?
   ├── datasets
   └── RGB_Images
   ├── org_1_80x60
   ├── 0
   ├── 1
   └── 2
   What is the meaning 0/1/2 respectively ?

   If I want to 2 kind of classes(Real Face , Fake Face) , I can change 3 or 4 to 2 in your source ?

3. I want to train the 2 kind images(Real Face / Fake Face)
   3.1) I can compose dataset like below ?
   It is reasonable ?
   ├── datasets
   └── RGB_Images
   ├── 1_230x230
   ├── Real
   ├── RealFace1.jpg
   ├── ....
   └── Fake
   ├── FakeFace1.jpg
   ├── FakeFace2.jpg
   ├── ...

        In your source the is only org_1x80x60 , 1_80x80 , 2.7_80x80 , 4_80x80.
        I have image (203x203)
        I should make image size as yoy did ?
          for example,  org_1x203x203 , 1_80x80 , 2.7_80x80 , 4_80x80. ...
   

   3.2) after finishing training, how do I name the output => snapshot .pth and copy it to "Silent-Face-Anti-Spoofing-master\resources\anti_spoof_models"
   Do I name it as "1_80x80_MiniFASNetV2.pth" or another name or "1_230x230_MiniFASNetV2.pth" or new what ?
   Of course , I anlayzed your source but it is a little bit confusing .

Thanks in Advance

视频里讲到只用了crop数据增强，不能用像素值扰动，为啥这里还有亮度之类的数据扰动？

Silent-Face-Anti-Spoofing/src/data_io/dataset_loader.py

Lines 18 to 19 in b6d5f04

	trans.ColorJitter(brightness=0.4,
	contrast=0.4, saturation=0.4, hue=0.1),

Can't open deploy.prototxt in function cv::dnn::ReadProtoFromTextFile

self.detector = cv2.dnn.readNetFromCaffe(deploy, caffemodel)
cv2.error: OpenCV(4.2.0) C:\projects\opencv-python\opencv\modules\dnn\src\caffe\caffe_io.cpp:1121: error: (-2:Unspecified error) FAILED: fs.is_open(). Can't open "./resources/detection_model/deploy.prototxt" in function 'cv::dnn::ReadProtoFromTextFile'

我遇到了這樣的問題，這是我opencv的版本 opencv-python==4.2.0.34
嘗試過更改為絕對路徑但並沒有效果
環境:
python==3.8.4
torch==1.6.0+cpu
torchvision==0.7.0+cpu
自己有找過問題解法，可是沒有進展
希望有人能幫助我

训练集分为3类是哪三类？

作者你好，将训练集分为3类是哪三类啊？我理解不是应该只有两类吗？真脸和假脸？

PReLU转化问题

您好，这项工作非常好，但是在实际使用中我想将开源的caffemodel转化为TensorRT，但是TensorRT不支持PReLU层，我想请问一下，这个PReLU层该怎么转化为TensorRT支持的Relu层呢？非常感谢。

門檻值問題

詢問門檻值的調整是否於下列參數？
anti_spoof_predict.py
self.detector_confidence = 0.6

我在這個值所作的調整，並沒有反應到我的預測結果中
0.8,0.7,0.5,1.0接會出現相同的預測結果

想請問決定門檻值的正確參數是？

FAILED: ReadProtoFromTextFile(param_file, param). Failed to parse NetParameter file: ./resources/detection_model/deploy.prototxt in function 'cv::dnn::ReadNetParamsFromTextFileOrDie'

我在windows10跑这个代码，发现加载不了模型，请问需要改什么吗？
File "D:/project/alive_detect/Silent-Face-Anti-Spoofing-master/test.py", line 109, in
test(args.image_name, args.model_dir, args.device_id)
File "D:/project/alive_detect/Silent-Face-Anti-Spoofing-master/test.py", line 35, in test
model_test = AntiSpoofPredict(device_id)
File "D:\project\alive_detect\Silent-Face-Anti-Spoofing-master\src\anti_spoof_predict.py", line 55, in init
super(AntiSpoofPredict, self).init()
File "D:\project\alive_detect\Silent-Face-Anti-Spoofing-master\src\anti_spoof_predict.py", line 32, in init
self.detector = cv2.dnn.readNetFromCaffe(deploy, caffemodel)
cv2.error: OpenCV(4.1.0) C:\projects\opencv-python\opencv\modules\dnn\src\caffe\caffe_io.cpp:1151: error: (-2:Unspecified error) FAILED: ReadProtoFromTextFile(param_file, param). Failed to parse NetParameter file: ./resources/detection_model/deploy.prototxt in function 'cv::dnn::ReadNetParamsFromTextFileOrDie'

how should the dataset be?

Is the dataset you are using an open source dataset or was prepared by you?
I read the data preprocessing section but it was not clear for me.
What should be the dataset structure for training?

Opencv_Loader BGR

opencv cv2.imread is Default BGR，but generate_FT is cv2.COLOR_RGB2GRAY，it's correct?

傅里叶分支问题

你好，傅里叶分支的target为输入图像经过傅立叶变换后和resize后的结果，而看代码，他的pred是神经网络某一层的特征图，这两者似乎不能做对应。直接求MSELoss貌似也不合原理。你那边能做一下解释吗。

发布的apk不防3D模型？

发布的apk不防3D模型， 体验了一下，3D模型防不住

If without FT

Thanks for your excellenct work and share!
I am wondering whether have you done an ablation experiment ,that's how about the model's performance if we remove the FT generator? Because I canot be sure the amazing performance comes from your original CNN model or from the the gain brought by FT.
Looking forward for your reply,thks!

Model Architecture for Scale1_80x80 Training

Hello,

Is it possible to provide just the model architecture file for Scale 1 80x80
Then i can train on my own database and share with you.

Does Scale 1 model needs Squeeze Excitation module for training?

Thanks,

您好 ，您发布的体验版性能大概是多少啊，我拿照片测试时，发现卷一下边可以轻松通过

win/ubuntu推断结果不一致

image_F1.jpg在win下0.73FakeFace，但是在ubuntu18.04下是0.55RealFace，同样的电脑双系统，同样的环境包版本，为什么会推断出不一致的结果？

多尺度模型

resource中anti_spoof_models中只有2.7和4尺度的模型，没有原始尺度的模型（scale=1）,能否提供下

训练完的模型怎么使用

训练完的模型使用时报错，运行test.py在win10报错
Traceback (most recent call last):
File "D:\anaconda3\envs\silence\lib\tarfile.py", line 189, in nti
n = int(s.strip() or "0", 8)
ValueError: invalid literal for int() with base 8: '_rebuil'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "D:\anaconda3\envs\silence\lib\tarfile.py", line 2299, in next
tarinfo = self.tarinfo.fromtarfile(self)
File "D:\anaconda3\envs\silence\lib\tarfile.py", line 1093, in fromtarfile
obj = cls.frombuf(buf, tarfile.encoding, tarfile.errors)
File "D:\anaconda3\envs\silence\lib\tarfile.py", line 1035, in frombuf
chksum = nti(buf[148:156])
File "D:\anaconda3\envs\silence\lib\tarfile.py", line 191, in nti
raise InvalidHeaderError("invalid header")
tarfile.InvalidHeaderError: invalid header

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "D:\anaconda3\envs\silence\lib\site-packages\torch\serialization.py", line 555, in _load
return legacy_load(f)
File "D:\anaconda3\envs\silence\lib\site-packages\torch\serialization.py", line 466, in legacy_load
with closing(tarfile.open(fileobj=f, mode='r:', format=tarfile.PAX_FORMAT)) as tar,
File "D:\anaconda3\envs\silence\lib\tarfile.py", line 1591, in open
return func(name, filemode, fileobj, **kwargs)
File "D:\anaconda3\envs\silence\lib\tarfile.py", line 1621, in taropen
return cls(name, mode, fileobj, **kwargs)
File "D:\anaconda3\envs\silence\lib\tarfile.py", line 1484, in init
self.firstmember = self.next()
File "D:\anaconda3\envs\silence\lib\tarfile.py", line 2311, in next
raise ReadError(str(e))
tarfile.ReadError: invalid header

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "D:/2020project/Silent-Face-Anti-Spoofing-master/test.py", line 110, in
test(args.image_name, args.model_dir, args.device_id)
File "D:/2020project/Silent-Face-Anti-Spoofing-master/test.py", line 60, in test
prediction += model_test.predict(img, os.path.join(model_dir, model_name))
File "D:\2020project\Silent-Face-Anti-Spoofing-master\src\anti_spoof_predict.py", line 87, in predict
self._load_model(model_path)
File "D:\2020project\Silent-Face-Anti-Spoofing-master\src\anti_spoof_predict.py", line 67, in _load_model
state_dict = torch.load(model_path, map_location=self.device)
File "D:\anaconda3\envs\silence\lib\site-packages\torch\serialization.py", line 386, in load
return _load(f, map_location, pickle_module, **pickle_load_args)
File "D:\anaconda3\envs\silence\lib\site-packages\torch\serialization.py", line 559, in _load
raise RuntimeError("{} is a zip archive (did you mean to use torch.jit.load()?)".format(f.name))
RuntimeError: ./resources/anti_spoof_models\1_80x80_MiniFASNetV1.pth is a zip archive (did you mean to use torch.jit.load()?)

在linux报错
python test.py
Traceback (most recent call last):
File "test.py", line 109, in
test(args.image_name, args.model_dir, args.device_id)
File "test.py", line 59, in test
prediction += model_test.predict(img, os.path.join(model_dir, model_name))
File "/ywcx/guanxiao/Silent-Face-Anti-Spoofing/src/anti_spoof_predict.py", line 87, in predict
self._load_model(model_path)
File "/ywcx/guanxiao/Silent-Face-Anti-Spoofing/src/anti_spoof_predict.py", line 64, in _load_model
self.model = MODEL_MAPPINGmodel_type.to(self.device)
KeyError: 'model'

请问这个要怎么解决，在linux我改了模型的后缀，MODEL_MAPPING = {
'MiniFASNetV1': MiniFASNetV1,
'MiniFASNetV2': MiniFASNetV2,
'MiniFASNetV1SE':MiniFASNetV1SE,
'MiniFASNetV2SE':MiniFASNetV2SE
}
这四个都试了，都说模型不对，默认训练的是什么模型啊

1_80x80模型可以提供吗？

confidence大于1

感谢分享，我用你的模型重新做预测，图片resize到80*80，发现每个模型输出的confidence都有大于1的情况。请问模型输出的范围不是0--1吗？还是我做错了什么呢？

train problem

MultiFTNet.py修改： self.model = MiniFASNetV2SE(...) 改为self.model = MiniFASNetV1SE(...) , 图像处理为1_80x80, 训练出来的模型大小是2690，原生resources下的4_0_0_80x80_MiniFASNetV1SE.pth 大小为1813，请问这是怎么回事 ？？

anti_spoof_predict代码有可以优化的地方

把_load_model放到__init__里面，否则用摄像头实时测试的时候，每次都要重新加载一遍模型浪费时间

Model 1_80x80 ?

In your Data Processing README Part, you mentioned about three scaling factors 1_80x80, 2.7_80x80 and 4_80x80 and you have provided 2.7 and 4.0 pretrained models but no 1.0 model? Can you provide 1_80x80 model also.
Thank You

Can't detect photo attack

Almost all face photos are classified as live face.

效果极差

app试了，python代码也试了，感觉都不行，阈值调不调一个样，只要图片贴的足够近，判断real face分数都在0.9以上
还有也没搞懂为啥用两个模型叠加使用，感觉两个模型给出的结果都一样啊
开始还以为是输入图片的尺寸问题，后来图片输入都跟demo一样是480*640，也不行

嗨，您好，将训练集分为3类是什么意思呢，不是应该2类吗？0和1 F和T

Using provided pretrained models for finetuning

Firstly, I would like to thank the minivision group for sharing the source code and trained weights with the open source community. The documentation on the README is great too. You guys even put in the effort to translate them into english. Thank you very much.

I am currently trying to use the pretrained weights provide to fine tune them to my dataset (only 2 class, real or spoof) but i am facing some problems. Is there any tips to doing so?

I note that the state_dict you provided has been pruned before and looking that the way you loaded the models, i see you did the following:

state_dict = torch.load(model_path, map_location=self.device)
       keys = iter(state_dict)
       first_layer_name = keys.__next__()
       if first_layer_name.find('module.') >= 0:
           from collections import OrderedDict
           new_state_dict = OrderedDict()
           for key, value in state_dict.items():
               name_key = key[7:]
               new_state_dict[name_key] = value
           self.model.load_state_dict(new_state_dict)
       else:
           self.model.load_state_dict(state_dict)

where you only used sliced state_dict.

训练数据

@zhuyingSeu 您好！后面会公开训练数据吗

ONNX model !!

Hello, I tried to convert your given model to onnx and converted them successfully, but the output of V1SE model is not consistent with the given pytorch model. I used onnxruntime for inference of the two models.
V2 model is fine but V1SE is not. Can you help?

训练自己的数据

您好，我测试了你们的模型，在我这边的场景下效果并不是很好。我的场景是红外摄像头灰度图，请问一下用你们的工程训练效果会好使吗？或者您那边有一些好的建议吗？

Model performance

Thank you so much for such amazing code - and for making your weights available!

Do you have ROC curves for assessing performance?

Have you run the classifier over any available benchmark data sets?

Thanks again!