softwaregift / feathernets_face-anti-spoofing-attack-detection-challenge-cvpr2019 Goto Github PK

Hi, thank you for your sharing, i have download the data from baidu, but it look like need a password to unzip, could you please tell me?

您在论文里说直接输出1024维的特征向量，不使用全连接层，这样在计算focal loss时是不是只使用了向量的前两个元素。这样的话是不是会丢失很多信息？

你好，我对训练过程有两点疑问。
1·比赛公布数据集有RGB，IR和Depth数据，看你的代码是仅仅只用Depth数据训练吗？为什么选择不使用三种类型的数据一起训练分类器呢，混合训练是效果差吗？
2·样本增强方面。在data/our_filelist/2depth_train.txt中，sample-eye-nose_out是对原始深度图直接提取保留眼睛和鼻子部分然后其他区域都置为0吗？如果是这样处理，请问对于网络学习泛化能力提高明显吗？
谢谢分享！

您好，在 RGB-D-NI數據集中，深度圖與近紅外圖的分辨率是一樣的，但是可見光的分辨率與近紅外和深度圖的分辨率都不一樣，這3種圖像該怎麼融合呢？您這邊是分別預測的嗎

根据您给的百度云盘链接下载的mobilenet_v2.pth.tar无法解压；之前的问题中说不解压也可以使用该模型进行训练，但是仍然是无法使用的，在加载模型时，无法读取该模型的epoch等参数，请问是否是该模型无法使用，能重新上传新的模型吗？
测试命令如下：
python main.py --config="cfgs/mobilenetv2.yaml" --resume ./checkpoints/pre-trainedModels/mobilenet_v2.pth.tar --phase-test True --val True --val-save True
显示如下：
Input image size: 224, test size: 224

• Number of FLOPs: 306.17M
  The network has 2226434 params.
  total_params 2226434
  D:\FeatherNets_Face-Anti-spoofing-Attack
  => loading checkpoint './checkpoints/pre-trainedModels/mobilenet_v2.pth.tar'
  Traceback (most recent call last):
  File "main.py", line 406, in
  main()
  File "main.py", line 151, in main
  args.start_epoch = checkpoint['epoch']
  KeyError: 'epoch'

Could you please attach the code of the Data Preprocessing, I want to test the model with my own pictures ...

https://pan.baidu.com/share/init?surl=_caiA9r9SPjRI09fNKsscg 提取码：9ooz
我在这里下载了一份CASIA-SURF数据集，请问如何和
/data/2depth_train.txt
/data/2label_train.txt
/data/depth_val.txt
/data/label_val.txt
/data/depth_test.txt
/data/label_test.txt
对照起来。

run data/fileList.py

Traceback (most recent call last): File "fileList.py", line 78, in <module> for line in fileinput.input("test_private_list.txt"): File "D:\ProgramData\Anaconda3\lib\fileinput.py", line 252, in __next__ line = self._readline() File "D:\ProgramData\Anaconda3\lib\fileinput.py", line 364, in _readline self._file = open(self._filename, self._mode) FileNotFoundError: [Errno 2] No such file or directory: 'test_private_list.txt'
@SoftwareGift

main.py 的validate函数出现这个问题，我分别试过roc-0.3.2和roc-0.2.3都会出现这个问题，求怎么解决？

为什么我把batch_size改到512，训练结果反而变得很差？普通情况是大的batchsize相对来说结果不会比小的差。

在运行代码时遇到这个问题，请问该怎么解决呢？其中：

import torch
print(torch.cuda.is_available())
False

显卡驱动不匹配的原因吗

您好，这个百度云盘链接（提交代码的链接）失效了，您能重新提供一个吗？

作者你好，我看很多作者和我一样，都卡在数据集上了，请问什么时候您能发开您论文中提到的数据集和标签数据？

it seems the MobileNetV2 Model is invalid, could you upload again?

作者好，能否提前把论文挂出来学习一下？谢谢。

请问能描述一下怎么样从FeatherNet的输出向量（长度1024的）计算出score吗？

When I run "python main.py --config="cfgs/fishnet150-32.yaml" --b 32 --lr 0.01 --every-decay 30 --fl-gamma 2 >> fishnet150-train.log". The following error happen:

D:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\upsampling.py:129: UserWarning: nn.Upsample is deprecated. Use nn.functional.interpolate instead. warnings.warn("nn.{} is deprecated. Use nn.functional.interpolate instead.".format(self.name)) Traceback (most recent call last): File "main.py", line 402, in <module> main() File "main.py", line 198, in main train(train_loader, model, criterion, optimizer, epoch) File "main.py", line 230, in train for i, (input, target) in enumerate(train_loader): File "D:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 637, in __next__ return self._process_next_batch(batch) File "D:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 658, in _process_next_batch raise batch.exc_type(batch.exc_msg) FileNotFoundError: Traceback (most recent call last): File "D:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 138, in _worker_loop samples = collate_fn([dataset[i] for i in batch_indices]) File "D:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 138, in <listcomp> samples = collate_fn([dataset[i] for i in batch_indices]) File "F:\CVPR2019\read_data.py", line 94, in __getitem__ depth = Image.open(depth_dir[idx]) File "D:\ProgramData\Anaconda3\lib\site-packages\PIL\Image.py", line 2609, in open fp = builtins.open(filename, "rb") FileNotFoundError: [Errno 2] No such file or directory: 'F:\\CVPR2019\\data/Training/fake_part/CLKJ_BS1010/06_enm_b.rssdk/depth/151.jpg'

@SoftwareGift

请问视频数据集可以下载吗？

我使用命令行
python main.py --config="cfgs/FeatherNetB-32.yaml" --b 32 --lr 0.01 --every-decay 60 --fl-gamma 3 >> MobileLiteNetB-bs32--train.log
但是得到如下报错
Traceback (most recent call last):
File "main.py", line 406, in
main()
File "main.py", line 113, in main
model = models.dictargs.arch
File "/home/plato/FeatherNets_Face-Anti-spoofing-Attack-Detection-Challenge-CVPR2019-regression/models/FeatherNet.py", line 161, in FeatherNetB
model = FeatherNet(se = True,avgdown=True)
File "/home/plato/FeatherNets_Face-Anti-spoofing-Attack-Detection-Challenge-CVPR2019-regression/models/FeatherNet.py", line 87, in init
super(MobileLiteNet, self).init()
NameError: name 'MobileLiteNet' is not defined

请问应该怎么修改？

您好，从百度云下载的数据集没有Val文件夹，请问怎么能获得验证集吗，谢谢。

I got it now.

where can i get the dataset? thanks

你好，我在训练完一个epoch后进行validata后出错，经查是在
_,predicted = torch.max(soft_output.data, 1)
predicted = predicted.to('cpu').detach().numpy()

这里的predicted不是包含0,1的tensor，是一个这样的tensor
Tensor: tensor([920, 744, 920, 818, 416, 920, 818, 920, 919, 920, 818, 818, 920, 974,
920, 855, 919, 744, 503, 818, 744, 920, 818, 684, 818, 503, 920, 919,
920, 818, 919, 974], device='cuda:0')

导致在后面计算confusion_matrix的时候无法正确计算tn, fp, fn, tp，
请问是哪里有问题吗？ 我是从train_list里随机取的2000个样本做val_list的，其他地方没有改动

请问大佬们，模型训练出来有用到系统中的吗？准确率怎么样？

请问有没有测试数据的标签呀?

我们基于Feathernet训练了一个pytorch模型，希望能在嵌入式设备使用。
首先将py模型转成了onnx模型，用python验证onnx模型效果和py模型一致，然后基于ncnn提供的工具onnx2ncnn进行了ncnn模型转换，转换成功，并用C++可以加载运行，但预测效果全无，想请教一下这块问题如何处理

in ColorAugmentation(), in data_aug.py,
How to set self.eig_val and self.eig_vec for single channel near-infrared image?

作者你好，请问ensemble classifier中的多个model，是指多个不同的神经网络结构吗？

Hello, I use your code to train featherNetB with depth images, and get 0.9080% on ACER on val set, which is a good result. But I get 2.1473% on ACER on test set, which is a apparent decline. Could you provide some advice to improve the results except ensembling? Because we want to have a simple and fast model.

您好，请问有单独的test代码吗？能放出来吗？

请问大佬，论文中关于三种特征图融合是在哪一部分呢。我在您的代码中只看到了关于depth的部分。期待大佬回复哈 ;-)

I use the MobileNetV2 and I want to use your model on mobile devices by NCNN.
I convert the model to onnx.When I want to convert the MobileNetV2.onnx by NCNN.The layer ReduceMea n not supported.But NCNN's contributor tell me if I can replace "nn.AvgPool2d" to "nn.AdaptiveAvgPool2d" before the last conv.I will get a usful model.
https://github.com/Tencent/ncnn/issues/1092

But I can't find the "nn.AvgPool2d"?Where I can replace it?Thx.

你好，streaming module将6444特征图拉伸成1024维向量，在做二分类时只用到第一二个值。这样做DWConv是否有意义，是我哪里理解错了？谢谢

大神，您好~参考您提供read_data.py，发现只读取的depth数据集作为mobilenetv2训练样本，ir和rgb不需要吗？

##accorcoding to casia-surf val to commpute
normalize = transforms.Normalize(mean=[0.14300402, 0.1434545, 0.14277956],
std=[0.10050353, 0.100842826, 0.10034215])
how to commpute normalize? do you have any references?
how about training effect using normalize?

Hi @SoftwareGift ,thanks for your great project! I train your model about python3 main.py --config="cfgs/mobilenetv2.yaml" --b 8 --lr 0.001 --every-decay 40 --fl-gamma 2 >> mobilenetv2-bs32-train.log and I obtain more than 95% accuracy. That a small problem, I should mkdir logs file first. But I want to train your best result model, when I use your 6th train method. I use python3 main.py --config="cfgs/FeatherNetB-32.yaml" --b 32 --lr 0.01 --every-decay 60 --fl-gamma 3 >> MobileLiteNetB-bs32--train.log error appears.The feedback is

main.py:91: YAMLLoadWarning: calling yaml.load() without Loader=... is deprecated, as the default Loader is unsafe. Please read https://msg.pyyaml.org/load for full details.
config = yaml.load(f)
Traceback (most recent call last):
File "main.py", line 406, in
main()
File "main.py", line 113, in main
model = models.dictargs.arch
KeyError: 'FeatherNetB'

I don't know how to fix the FeatherNetB. Please help how to train the FeatherNetB correctly.Thank you.

如何根据下面的代码的最后部分获取预测结果啊？ 比如最终的预测分数

# 加载模型
 model = models.fishnet150()

 # 加载模型文件的权重数据到model模型
 checkpoint = torch.load('checkpoints/pre-trainedModels/fishnet150_ckpt.tar')  # 加载指定的模型文件
 model.load_state_dict(checkpoint['state_dict'])  # 导入模型中的权重数据

 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")  # 设置计算硬件，CPU 或 GPU
 torch.cuda.set_device(device)
 # torch.backends.cudnn.benchmark = True
 model = model.module.to(device)
 model.eval()

 # 准备transform
 img_size = 224
 ratio = 224.0 / float(img_size)
 normalize = transforms.Normalize(mean=[0.14300402, 0.1434545, 0.14277956],
                                  std=[0.10050353, 0.100842826, 0.10034215])
 transform = transforms.Compose([transforms.Resize(int(224 * ratio)), transforms.CenterCrop(img_size),
                                 transforms.ToTensor(), normalize])

 # 读取单张人脸RGB图像进行预测
 img = Image.open("face1.jpg")
 img = img.convert('RGB')
 img = img.resize((224, 224))
 img = transform(img)

 img = np.array(img)  # 形状为(3, 224, 224)
 img = np.expand_dims(img, 0)  # 此时形状为(1, 3, 224, 224)

 input = torch.tensor(img, dtype=torch.float32, device=device)  
 output = model(input)  # output形状为[1, 1000]

 soft_output = torch.softmax(output, dim=-1)  # 对output中最后一个维度上的元素进行softmax处理，得到的形状依然是[1, 1000]
 preds = soft_output.to('cpu').detach().numpy()  # 转换为numpy数组，形状为(1, 1000)

会提供tensorflow的代码吗

have pip-installed dependencies in your environment file, but you do not list pip itself

求大神指教 感激不尽

i want get the Available dataset~ please~