Hi, thank you for this awesome repo. I have tried to generate embs on KDEF dataset. This dataset contains samples such as: I have used SVM for classifying the generated embs but got a very low acc of 50%. Please note that I have removed images of faces which are depicting the profile. Based on my explanation, what can be the problem? Should I remove the images which are partially straight also (images like the second image)?

Hello, I'm new to PyTorch.

I have a problem on training.

First, I write the "ms1m_train.list" for few datasets and train the model by run "./run.sh" below is the ms1m_train.list:

then I run command "./eval.sh /Volumes/jc_sandisk/MS1Mv2_dataset/mymagfacemodel/test/00010.pth official 100" for eval, but I got the error "=> Not all weights loaded, model params: 925, loaded params: 0".

Maybe it's a simple problem, but it stuck me.

By the way, "MODEL_ARC" in "run.sh" is "iresnet100".

Hi All, A bit urgent. Hence would really be thankful if someone with a quick answer to take the time to reply to me.

A look at the following class https://github.com/IrvingMeng/MagFace/blob/dbd0c5b587d62b3ee51933e41b68f70e0e6db8bf/inference/gen_feat.py#L52 shows that there is no pre-processing done, apart from just loading the image using cv2. In contrast, Insightface (which also uses the same meta files) performs face alignment https://github.com/deepinsight/insightface/blob/d4d4531a702e22cc7666cda2de6db53f4dc2e4db/recognition/evaluation/ijb/ijb_evals.py#L238 . While using the same meta, I am wondering why there is a difference. Am I to understand that magface does not use any pre-processing? (is there no crop+align using landmark here?)

Thanks,

Hello, first of all thank you for providing the code, models and preprocessed images!

I tried to reproduce the results on LFW and with the preprocessed images you provided I also get the values you reported. But when I preprocess the images from LFW by myself using the preprocessing from the ArcFace repository (https://github.com/deepinsight/insightface), the performance is a lot worse (~0.96) than with your images (~0.998).

I don't know if the authors of Arcface have changed the preprocessing code over time, but would it be possible to share the preprocessing script you used to create the images? Or if not, could you tell me what you changed to create the images?

Thanks in advance and thanks again for providing the code and models! Best, Marco

Hello! According to the paper, I trained on MS1MV2, extract the feature directly by iResNet50 and use the code in ArcFace when validating on LFW, CFP-FF, CFP-FP and AgeDB-30. However, I evaluate the model on every epoch and only get best accuracy of 0.98283 on LFW, 0.97814 on CFP-FF, 0.77314 on CFP-FP, and 0.89300 on AgeDB-30. Here is the training log on MS1MV2 output.log. I changed the learning rate and batch size to 1.25 times the original while training.

Is there anything wrong?

Hi, Expert Nice work to embedded magnitude into latent space of features. Then I train on masked and non-masked MS1M, and test on LFW, CFP, Megaface But I wonder the magnitude without any discrimination on testing dataset May you have any suggestion about this situation?

magnitude.jpg

Hi, thank you for this repo and open-sourcing it. I have generated embedding of approximately 7k perfectly-matched-to-face images (such as the following) and then tried to classify them using SVM (sklearn --> linear and RBF kernels) but I have got around 85% accuracy on the task. On the other hand when I used FACENET-generated embedding on the exact same SVM I got around 94%. Please tell me if I am doing sth wrong because the results are not at all SOTA.

Thank you for your great research and the open source code. We want to evaluate the improvement for the cluster task but the code in "eval/eval_cluster" is empty. Do you have the plan to public the code of this part? Do you have any suggestions for cluster testing on IJBB? Thank you!

Hello, I tried to adapt ArcFace evaluation script for MagFace, you can see it here https://github.com/kantengri/MagFace/tree/try-arcface-eval

Data & model used: LFW test dataset (lfw-deepfunneled), resized to 112x112. list file: lfw_test_pair.txt (from ArcFace) model magface_epoch_00025.pth (Trained by original codes, https://drive.google.com/file/d/1Bd87admxOZvbIOAyTkGEntsEz3fyMt7H/view?usp=sharing)

test.py (modified) output: (7696, 512) total time is 96.21539497375488, average time is 0.20003200618244257 truncating identity_list from 7701 to 7696 skipped 5 lfw face verification accuracy: 0.6071726438698916 threshold: 0.48544353

How the accuracy can be improved ? It's much below declared performance (99.83%, see MagFace paper).

https://github.com/IrvingMeng/MagFace/blob/dbd0c5b587d62b3ee51933e41b68f70e0e6db8bf/eval/eval_quality/eval_quality.py#L19

嘿，对于数据集 LFW，代码里这个 pair_list 是怎么得到的？ 是直接用 LFW 提供的pairs.txt 吗？ 但是，我使用pairs.txt 并没有复现成功 paper 画的 EVRC。

https://github.com/IrvingMeng/MagFace/blob/dbd0c5b587d62b3ee51933e41b68f70e0e6db8bf/eval/eval_quality/eval_quality.py#L134

嘿，为什么这样计算 fnmr ？ 我查阅其他的资料，我的理解应该是 fnmr = 1.0 * num_acc / num_query 。 能不能解释一下，谢谢！

Hi. Thank you for this awesome repo. In the inference mode I don't have any problem. But when it comes to training the architecture of the model has the following problem. RuntimeError: mat1 and mat2 shapes cannot be multiplied (2x100352 and 25088x512) This error happens in File "/home/marefat/projects/MagFace/run/../models/iresnet.py", line 145, in forward x = self.fc(x)

I tried to change the following: self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_classes) to fix this problem and match the dimensions. But after this I face core dump problem. Any recommendation will be appreciated.

Hi, when I train the model with the script run_dist.sh, the checkpoint can not be saved when the first epoch finishes. However, the gpu is still occupied and there is no more logs. Any advise ?

Hi, Can anyone give me any suggestions on how can I generate pair.list file for lfw like dataset used in this project? I was able to evaluate the existing data that is provided in the repo. Thanks in advance.

Hey, I have for some hours now tried to use MagFace with no luck. I want to take a lot of images, check the image quality regarding biometric features and print that data. However, I get the RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.

Is there a tutorial/documentation for how to use MagFace with all prerequisites?

Thank you