zhongyy / face-transformer Goto Github PK

https://github.com/deepinsight/insightface/tree/master/recognition/arcface_torch

Datasets backbone IJBC(1e-05) IJBC(1e-04) agedb30 cfp_fp lfw
MS1MV3-Arcface r100-fp16 95.31 96.81 98.48 99.06 99.85

Hi @zhongyy , All,

This is a bit urgent. Would appreciate anyone with thoughts on this. In the repo the link to MS1M-RetinaFace dataset is given, and further, a FaceDatset loader is given to load the same. I am surprised that the usual pre-processing such as normalization/standardization is not done. Is it not needed for this repo? Could someone please get back on this asap?

Thanks,

感谢作者的工作！ 看了下源码好像只是简单的randomflip了。
1.请问作者在数据增强部分放弃了transformer比较常用的mixup和cutmix是吗？
2.请问是没有进行归一化吗？是发现不进行归一化效果更好吗？
如果能抽空回答，感激不尽。

Maybe the data is converted to bin files for faster dataloader, but I am confused about how to visualize the data such as the figure in your paper.

Hi Dear,

I am outside of China and I am unable to access datasets because they require a mobile number. Please upload the testing datasets on google drive or dropbox.

Hi @zhongyy

In the abstract, you listed that we evaluated on the IJB-C database but this dataset is not present in your repo. If possible, please put the link to the IJB-C dataset.

Regards,
Khawar

because of some thing...cuda 'no run.
so.i want use CPU run,But it misstated that my gpu id wasn‘t specified.

Hi!

I am trying to install requirements but always I obtain the following problem loading data:

Exception ignored in: <function MXRecordIO.del at 0x7f742061ff70>
Traceback (most recent call last):
File "/home/dparres/miniconda3/envs/fr1/lib/python3.8/site-packages/mxnet/recordio.py", line 88, in del
File "/home/dparres/miniconda3/envs/fr1/lib/python3.8/site-packages/mxnet/recordio.py", line 262, in close
TypeError: super() argument 1 must be type, not None

What I can do? I am using python 3.8 and the command "pip install -r requirements.txt" does not work for me.

Hi,

I have some questions about your input normalization. Specifically, I checked your code and notice that you leave out normalization for input images which means the input value is ranged [0, 255]. I tried to normalize the input to [-1, 1], and the results dropped a little, so is this a training trick or there is some insights behind it?

Hi, I have been trying to re-produce your network's result on IJB-C using insightface's evaluation code but have been unable to do so. Would it be possible that you upload your evaluation code?

Thanks a lot.

Please can you help me with the training since i need to train the model with VGG-Face or LFW
Thanks

Thank you @zhongyy for your pre-trained models. I verified all ViTP12S8 results in the paper and all results are the same as the paper. Thank you @zhongyy. After seeing your model name, I think you stop your model in Epoch_2_Batch_12000. Am i right?

I would like to ask that if I will not use pre-trained models and try to reproduce results. Is that possible?

I am running
CUDA_VISIBLE_DEVICES='0,1,2,3' python3 -u train.py -b 480 -w 0,1,2,3 -d retina -n VITs -head CosFace --outdir ./results/ViT-P12S8_ms1m_cosface_s1 --warmup-epochs 1 --lr 3e-4

Still, the ACC is 50%

highest_acc: [0.5375, 0.538, 0.5048333333333334, 0.517, 0.5651428571428572, 0.5091666666666667]
Epoch 1 Batch 15110	Speed: 6.00 samples/s	Training Loss 33.7626 (33.8095)	Training Prec@1 0.000 (0.000)
Epoch 1 Batch 15120	Speed: 208.26 samples/s	Training Loss 33.7692 (33.7805)	Training Prec@1 0.000 (0.000)
Learning rate 0.000001
Perform Evaluation on ['lfw', 'talfw', 'calfw', 'cplfw', 'cfp_fp', 'agedb_30'] , and Save Checkpoints...
(12000, 512)

Dear @zhongyy,

If you look into the testing datasets, we have a .bin file of each dataset for testing.
Do you have any idea or help me how we can create a .bin file of other datasets FGNET for testing?
InsightFace does not contain FGNET and MORPH datasets
I will be much appreciated it,

Thank you @zhongyy

Hi, Can you please tell which python version you are using? as I am using python 3.7.5 and having issues with packages while running train.py

I would like to ask one question about the table I in your Face Transformer paper.
How you calculate the speed of Img/Sec in Table I?

I am using the below library but I have no idea how to calculate inference speed?
https://github.com/sovrasov/flops-counter.pytorch

Hi
Can this model be used only for feature extraction leaving out classification? Will it outperform Feature pyramid networks(FPNs) and it's variants??

有些工作指出Transformer在图像分类上，就算扣掉很多像素，也能有很好的精度，远超CNN。
为什么你们的实验结果表明，Transformer在人脸识别任务上遮挡鲁棒性不如CNN呢？
可以解释一下吗？

您好！
想请问下如果想复现您的实验，但是requirements文件中的mxnet~=1.8.0.post0 与sklearn~=0.0 依赖无法通过pip安装（且官网中找不到klearn=0.0 的兼容版本，其中可以找到的最早版本为0.4版本；此外我的虚拟环境中可安装的mxnet版本最高似乎只到1.7.0.post2），有没有什么好的解决办法呢？希望我没有理解错您requirements文件中的版本要求信息，十分期待与感谢您的回复！

It is nice to have latency and memory usage comparison between different models.

Hi @zhongyy

I have noted that while training all GPUs are working but when performance evaluation is started only 1 GPU is used. Would it be possible to use four GPUs during performance evaluation? It will make the process faster.

Regards,
Khawar

为什么不直接在insightface项目上用Vit/Vits模型替换resnet模型，开始训练？
读取数据，加载模型，验证等代码和insightface一样，而loss代码不同呢？

您好，请问在lr=1e-3时训20个epoch后我这里的acc只有60不到？

Hii, thank you for your excellent work! I used your shared code to train my personal dataset, but I am curious about the significance of the number "93431" in the "property" file for the "ms1m_retinaface" dataset.

The content of the file is as follows: "93431,112,112", where the last two numbers represent the image size after pre-processing.
Could you please explain the meaning of the first number? Thank you in advance, and I look forward to your response!

Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cuda:1! (when checking argument for argument index in method wrapper_CUDA_scatter__value)

Hi!
Can you share your training log?

请问一下作者的代码使用GPU跑的么，为什么train.py里面parser那里的worker_ids是CPU

一般人脸识别不是都-127.5，然后除128，这里好像没有做？直接用的0-255做的输入吗？

I am doing training as same as your GPU Tesla V100. How many days are required for one epoch?

Hi,

Have you ever tried to train the model with patch size 16? And can you please share the performance on CFP-FP dataset?

I tried to train, but the best result is about 92%+ on CFP-FP(not from your repo). So I want to check if it is the problem with my implementation.

Thanks~

Thank you!

Hi,
where can I download property for DATA_ROOT in test.py?