bubbliiiing / arcface-pytorch Goto Github PK

博主，Resnet的预训练权重网站上找不到。

ValueError: Cannot have number of splits n_splits=10 greater than the number of samples: n_samples=0.

Hi,

How can I convert those pth models to onnx models? I tried several torch2onnx.py (https://github.com/deepinsight/insightface/blob/master/recognition/arcface_torch/torch2onnx.py) from different repositories but fails.

Many thanks!

-Scott

use this weights to train. but unfortunately, it raised an error as below:

arcface-pytorch/train.py", line 144, in
pretrained_dict = {k: v for k, v in pretrained_dict.items() if np.shape(model_dict[k]) == np.shape(v)}
KeyError: 'stage1.0.0.weight'

你好，看到你提供的预训练模型在4M左右，为什么我训练出来的在9M左右呢？会是什么原因呢？

为什么我下载了预训练好的权重arcface_mobilefacenet.pth以后，设置backbone为mobilefacenet，运行eval_LFW.py结果accuracy只有百分之60多呢

每次都是在 5888/6000 这个地方出现的, 请问可能是什么问题呢

Test Epoch: [5888/6000 (96%)]: : 24it [00:12,  1.86it/s]
Traceback (most recent call last):
  File "eval_LFW.py", line 66, in <module>
    test(test_loader, model, png_save_path, log_interval, batch_size, cuda)
  File "/home/zk/project/arcface-pytorch/utils/utils_metrics.py", line 150, in test
    tpr, fpr, accuracy, val, val_std, far, best_thresholds = evaluate(
  File "/home/zk/project/arcface-pytorch/utils/utils_metrics.py", line 14, in evaluate
    val, val_std, far = calculate_val(thresholds, distances,
  File "/home/zk/project/arcface-pytorch/utils/utils_metrics.py", line 81, in calculate_val
    f = interpolate.interp1d(far_train, thresholds, kind='slinear')
  File "/home/zk/.conda/envs/dl/lib/python3.8/site-packages/scipy/interpolate/_interpolate.py", line 571, in __init__
    self._spline = make_interp_spline(xx, yy, k=order,
  File "/home/zk/.conda/envs/dl/lib/python3.8/site-packages/scipy/interpolate/_bsplines.py", line 1252, in make_interp_spline
    raise ValueError("Expect x to not have duplicates")
ValueError: Expect x to not have duplicates

每次arcface.pth模型读取人脸png之后生成的参数不一样，不可以复用，想问一下怎么样才能固定？

博主 什么时候能出一个arcface与retinaface结合的程序呀 例如您之前发布的retinaface+facenet一样

hello,up, I just want to thank you and learn from you

use my own datasets,however,the acc of train.py always equals 0.At the same time,the loss function does not converge.Why? i need adjust hyper-parameter(s or m)?

请问，如果我想使用MS_Celeb_1M数据集训练，是否只要把net.arcface.py里的embedding_size和 num_classes更改就可以了？还有你的arcface_training.py的作用是什么呢~~

开始进行LFW数据集的验证。
Traceback (most recent call last):
File "train.py", line 337, in
fit_one_epoch(model_train, model, loss_history, optimizer, epoch, epoch_step, epoch_step_val, gen, gen_val, Epoch, Cuda, LFW_loader, lfw_eval_flag, fp16, scaler, save_period, save_dir, local_rank)
File "/root/autodl-tmp/arcface-pytorch-main/utils/utils_fit.py", line 114, in fit_one_epoch
_, _, accuracy, _, _, _, _ = evaluate(distances,labels)
File "/root/autodl-tmp/arcface-pytorch-main/utils/utils_metrics.py", line 13, in evaluate
val, val_std, far = calculate_val(thresholds, distances,
File "/root/autodl-tmp/arcface-pytorch-main/utils/utils_metrics.py", line 72, in calculate_val
f = interpolate.interp1d(far_train, thresholds, kind='slinear')
File "/root/miniconda3/lib/python3.8/site-packages/scipy/interpolate/_interpolate.py", line 571, in init
self._spline = make_interp_spline(xx, yy, k=order,
File "/root/miniconda3/lib/python3.8/site-packages/scipy/interpolate/_bsplines.py", line 1252, in make_interp_spline
raise ValueError("Expect x to not have duplicates")
ValueError: Expect x to not have duplicates

请问进行评估的时候，为什么Best_thresholds会超过1呀

您好，我想要学习您的arcface代码并尝试使用在项目中，但是项目需要模型输出标签得分。那就需要为模型添加fc层，请问应该如何添加呢。初入领域，有些一筹莫展

非常感谢大佬无私地分享！:)

请问一下，lfw_pair.txt里的数字都表示什么含义呀？

Up有试过将arcface_mobilefacenet.pt转ncnn吗？
我试了下转onnx再转ncnn，ncnn输出的结果全是1或-1， 但是pt和onnx结果都是正常且一致。

百度网络LFW数据集下载失败

请问下在predict里只能预测两张人脸的相似度吗，不能预测一个图片是不是人脸吗？

我下载下来CASIA-WebFaces数据集是96×112的，不是112×112。是我的问题，还是数据集就是96×112的呢？

作者只使用了LFW测试数据集，如果想要增加AgeDB30和CFP-FP测试数据集，是否独立添加？callback如何修改呢？

我在租的服务器上跑的时候遇到了“RuntimeError: expected device cuda:0 and dtype Float but got device cuda:0 and dtype Long”这个错误，
torch是1.2.0版本的，网上说是因为版本太低，也可以通过更改数据类型解决，
我是把onehot替换为onehot.float()之后就解决了。

我用你提供的lfw数据和模型，得到如下结果。 与期望的accuracy 99.1，有些差距。

Accuracy: 0.98800+-0.00722
Best_thresholds: 1.17000
Validation rate: 0.92400+-0.02086 @ FAR=0.00100

请问在该实现中训练的时候image_shape必须是112x112x3的吗，是否可以使用不同长宽和长宽比的图片？

您好，不好意思打扰您了，想请问一下为什么加载了您提供的预训练的model以后，在CASIA-WebFace数据集训练时，刚开始的准确率还是比较低呢（从0开始）？正常来说您的模型是在CASIA-WebFace数据集训练的，那么一开始的准确率也应该比较高吧。

如何断点重新训练呢

如果我想用Agedb30数据集去替换LFW数据来验证模型，怎么修改程序？谢谢大神

感谢博主开源！从读研开始一直关注博主，谢谢博主贡献代码。
下边是我运行时候遇到的两个问题，如果有遇到同样问题可以参考。
在运行时候会遇到两个问题，在win系统上，会报错ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (6000,) + inhomogeneous part（pytorch版本为2.0.1）.linux系统上会报错RuntimeError: CuDNN error: CUDNN_STATUS-NOT_INITIALIZED（torch==1.2.0）。
查了下一个是数据长度不匹配，一个是cuda加速问题，在linux上cuda=false，用cpu就没这个问题。其他方法也懒得在尝试了
第二个是predict.py文件，如果显示匹配不到qt类似问题，打不开图片，环境装个pyqt5的包即可解决。

i have one problem during traing CASIA-WebFace ,the accuracy is droping dwon i used same model parameters

请问为什么我采用这这个模型，我将模型最后一层的输出改为嵌入向量维度不就可以吗，为什么准确率一直是0呢，请问问题出在哪里呢，我看着B导mobileNetV1的模型也是这么做的啊，为啥这么做就不可以呢

class MobileNetV4(nn.Module):
def init(self, model,embedding_size=128,dropout_keep_prob=0.5, pretrained=False):
# MobileNetV4ConvSmall MobileNetV4ConvMedium MobileNetV4ConvLarge
# MobileNetV4HybridMedium MobileNetV4HybridLarge
"""Params to initiate MobilenNetV4
Args:
model : support 5 types of models as indicated in
"https://github.com/tensorflow/models/blob/master/official/vision/modeling/backbones/mobilenet.py"
"""
super().init()
assert model in MODEL_SPECS.keys()
self.model = model
self.spec = MODEL_SPECS[self.model]

    # conv0
    self.conv0 = build_blocks(self.spec['conv0'])
    # layer1
    self.layer1 = build_blocks(self.spec['layer1'])
    # layer2
    self.layer2 = build_blocks(self.spec['layer2'])
    # layer3
    self.layer3 = build_blocks(self.spec['layer3'])
    # layer4
    self.layer4 = build_blocks(self.spec['layer4'])
    # layer5
    self.layer5 = build_blocks(self.spec['layer5'])
    self.layer6 = nn.Conv2d(1280,512,kernel_size=1)
    #embeding
    self.layer7 = nn.Linear(512*49, embedding_size)
    # self.layer6 = nn.Conv2d(1280,embedding_size,1)
    self.features = nn.BatchNorm1d(embedding_size, eps=1e-05)
    self.dropout = nn.Dropout(p=dropout_keep_prob, inplace=True)
def forward(self, x):
    x0 = self.conv0(x)

    x1 = self.layer1(x0)
    x2 = self.layer2(x1)
    x3 = self.layer3(x2)
    x4 = self.layer4(x3)
    x5 = self.layer5(x4)
    x6 = self.layer6(x5)

    x7 = torch.flatten(x6, 1)
    # 或者 x5 = x5.view(x5.size(0), -1)
    x8 = self.layer7(x7)
  #  x6 = self.features(x6)
    print(x8.shape)
    return x8

请问 IResent和Resnet有什么区别？

一般 K 折交叉是训练的时候用在训练集和测试集划分上吧, 为什么在评估 LFW 的时候还要使用 K 折交叉的方式来划分 LFW 呢?

你好！我想更换一个新的backbone，但是这两变量不知道指什么，感谢！

mobilenetv1 前向传播stage3后面的部分是什么意思

Accuracy: 0.90542+-0.00699
Best_thresholds: 1.29000
Validation rate: 0.65617+-0.01883 @ FAR=0.00117
could you help me hoe to increase the validation rate
ands what does FAR MEAN?

您好！
请问下predict模式输入两张图时，asrface.py
output1 = self.net(photo_1).cpu().numpy()
output2 = self.net(photo_2).cpu().numpy()
output1和output2 输出的是模型提取出的图像的特征向量吗？

when I download LFW dataset from original dataset it have shape (250, 250, 3) and eval, it have result 0.77. When I use your dataset LFW in baidu, I check img shape (96, 112, 3) and result like in table. I read your code, it resize shape (112, 112, 3) in data loader. Why it wrong like that. Thanks you