bleakie / maskinsightface Goto Github PK

楼主您好，在运行gen_datasets.py时，出现prn模型读取错误（已从readme提供的百度网盘链接下载模型，并放入Data/nrt-data里边），请问您知道是什么原因吗？非常感谢！
代码：
self.pos_predictor = PosPrediction(self.resolution_inp, self.resolution_op)
prn_path = os.path.join(prefix, 'Data/net-data/256_256_resfcn256_weight')
if not os.path.isfile(prn_path + '.data-00000-of-00001'):
print("please download PRN trained model first.")
exit()
self.pos_predictor.restore(prn_path)

错误：
DataLossError (see above for traceback): Unable to open table file /home/../MaskInsightface/PRNet_Mask/Data/net-data/256_256_resfcn256_weight: Failed precondition: /../MaskInsightface/PRNet_Mask/Data/net-data/256_256_resfcn256_weight: perhaps your file is in a different file format and you need to use a different restore operator?

File "train.py", line 1031, in
main()
File "train.py", line 1028, in main
train_net(args)
File "train.py", line 744, in train_net
sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
File "train.py", line 220, in get_symbol
print('init VarGFaceNet', args.num_layers)
AttributeError: 'Namespace' object has no attribute 'num_layers'

作者能同时发布128模型吗？

Hello, can you give a pre training model, that Baidu online link expired! @bleakie bleakie

百度网盘链接被取消了

如题，非常感谢大佬的开源项目，学习中想先测试一下，谢谢大佬啦

Hi @bleakie,

Congratulations for the nice work.

I am trying to use SSR-Net pretrained model in ncnn for edge device inference. For that I need to convert this model to onnx. However with this keras2onnx converter i am facing following trouble.

(ghimire36) C:\Users\ghimire\Desktop\keras2onnx>python keras_onnx.py
Traceback (most recent call last):
  File "keras_onnx.py", line 7, in <module>
    model = load_model('ssrnet_3_3_3_112_0.75_1.0.h5')
  File "C:\ProgramData\Anaconda3\envs\ghimire36\lib\site-packages\tensorflow_core\python\keras\saving\save.py", line 143, in load_model
    return hdf5_format.load_model_from_hdf5(filepath, custom_objects, compile)
  File "C:\ProgramData\Anaconda3\envs\ghimire36\lib\site-packages\tensorflow_core\python\keras\saving\hdf5_format.py", line 159, in load_model_from_hdf5
    raise ValueError('No model found in config file.')
ValueError: No model found in config file.

keras_onnx.py

from tensorflow.python.keras.models import load_model
import onnx
import keras2onnx

onnx_model_name = 'ssrnet_3_3_3_112_0.75_1.0.onnx'

model = load_model('ssrnet_3_3_3_112_0.75_1.0.h5')
onnx_model = keras2onnx.convert_keras(model, model.name)
keras2onnx.save_model(onnx_model, onnx_model_name)

Any help will be highly appreciated!

Thanks.

Hi, I could not download "FaceQnet pretrained model .h5 file". The page is broken. Thank you.

请问模型y2-res2-6-10-2-dim256与y2-res4-8-16-4-dim256是同样的数据集训练的吗，数据集是居于glint和私有数据训练，私有数据大概多少？我用单用glint数据训练，效果不太好，CFP_FP的识别率差

按照我的理解，人脸对齐步骤如下
1.统计facial landmark相对检测框的坐标，求取平均脸的landmark坐标。
2.将landmark相对坐标转化为112x112或其他大小检测框的绝对坐标。
3.利用相似性变换计算待对齐人脸与标准脸的变换矩阵。

不知我上述理解是否有问题。这里面对于第二点，landmark相对坐标该如何转化到绝对坐标呢？按照步骤1的统计，绝对坐标是根据检测框大小求来的，但是针对不同姿态的人脸，检测框的宽高比可能就有很大变化，这样就应该没法直接套用112x112或112x96大小的人脸。

请问关于这一点，您有什么建议么？

预训练模型的下载链接没有开放权限啊，下载不了

以0.5为分界线，经常把女的识别成男的

人脸质量检测facequality效果好像不好啊

Hello!

Have you estimated what is the effect of your input image normalization approach (face wrapping + background removing) on the final quality of recognition, compared to the conventional way of normalization (like insightface do, for instance) ?

Thank you!

hello，can you tell me the code mening？
if args.num_classes != arg_params['fc7_weight'].shape[0]:
print ("discarding fc7 weights...")
arg_params.pop('fc7_weight')

y2-res2-6-10-2-dim256 这个是已经训练好的没有用mask的模型吗，我看CFP_FP(%)能到97.18，精度有这么高吗。
@bleakie

i want to know how did you use Trillionpairs Datasets to generate .lst file? In your code, you mentioned 'webface celeb facescrub megaface fgnet ytf clfw' . so is there anyone of Trillionpairs?