fei-wang / insightface Goto Github PK

View Code? Open in Web Editor NEW

104.0 104.0 30.0 4.11 MB

implementation of insightface by using Tensorflow

License: MIT License

Python 100.00%

arcface insightface retinaface tensorflow

insightface's Introduction

Hi 👋, I'm Fei Wang

In the code I weave, machine learning thrives; together we achieve, as data comes alive.

insightface's People

Contributors

Stargazers

Watchers

insightface's Issues

可以提供一些retinaface这个模型在大型公开数据集的测试效果吗？

About the recognition(resnet) pretrained model

hi, could you share the pretrained model? thanks~

RetinaFace pretrained model

Thanks for porting this to tensorflow. Is there a retinaface r50 pretrained model for tensorflow, like there is for mxnet?

Bad Implementation of ArcFace Model. [Not Recommended to clone.]

ｖalid acc always is 0.5

when I train this recognition model , the valid result always is 0.5 ,and the total loss change from 40 to 8.
can you explain the results?

File "/opt/zhongls/PersonFace/insightface-master/retinaface/losses/loss.py", line 54, in _decode_label
idxes = self._match_gt_anchor(label, pred[loc[0], loc[1], loc[2]])
File "/opt/anaconda3/envs/pyzls/lib/python3.6/site-packages/tensorflow/python/ops/array_ops.py", line 680, in _slice_helper
name=name)
File "/opt/anaconda3/envs/pyzls/lib/python3.6/site-packages/tensorflow/python/ops/array_ops.py", line 846, in strided_slice
shrink_axis_mask=shrink_axis_mask)
File "/opt/anaconda3/envs/pyzls/lib/python3.6/site-packages/tensorflow/python/ops/gen_array_ops.py", line 9967, in strided_slice
_six.raise_from(_core._status_to_exception(e.code, message), None)
File "", line 3, in raise_from
tensorflow.python.framework.errors_impl.InvalidArgumentError: slice index 162 of dimension 1 out of bounds. [Op:StridedSlice] name: strided_slice/

Which dataset you used for training and what is performance of this implementation

Dear,

insightface used multiple dataset like faces_emore or MS1M_Celab etc to training and validation accuracy. Can you explicitly mentioned that which dataset you tried and what are final loss and accuracy on training and different validation sets.
If you can update the links and information of dataset on your readme page, that would be extremely helpful to start working on this code if someone who never work on face recognition problem.

Best

Could you share the annotation(label) file of wider face if you use it for RetinaFace training

I got some annotation from offical insightface but it does not include image size w.h in the label file.

Anchor match?

we find that you do the anchor matching in feature level in _match_gt_anchor(gt, anchor)

Why this? Is there any mystery in training Retinaface?

Can you explain config file param means?

loss_type: 'logit'   # logit or triplet
logits_scale: 64.0
logits_margin1: 1.0   # m1: sphereface should >= 1
logits_margin2: 0.2   # m2: cosineface should >= 0
logits_margin3: 0.3   # m3: arcface    should >= 0

center_loss_factor: 0.0  # center loss
center_alpha: 0.9   # center update rate

alpha: 0.2     # triplet margin

# run params
thresh: 0.2
below_fpr: 0.001        # fpr should below this

Sphere loss definition

Thanks for your great work!

Could you please tell me where did you define the sphere loss function in your script?

I am also reimplementing it.

Thanks a lot!

ids = list(os.listdir(image_dir)) 24 ids.sort() 25 cat_num = len(ids) FileNotFoundError: [Errno 2] No such file or directory: '/root/data/insightface/recognition/train'

Fully-connected layer in Arcface

Hi Wei,

Thanks for you awesome work, its really nice to see someone start coding in TF2.0!

I have one question about the last fully connected layer in arcface, which is the NormDense layer defined in recognition/models/models.py. Have you tried training with very large number of identities? The model would be ridiculously big with just ~10k identities. In the ArcFace paper, the author even reported results on some dataset with 94k identities.

Do you find it sufficient to train on just 1000 identities and model would generalise well to other unseen classes? Or it is generally required to fine-tune the model using triplet loss afterwards to make learned feature transferable?

Thanks for your help!

about base anchors

config 中的base anchors是怎么计算出来的，如何与stride 匹配？

多谢

fei-wang / insightface Goto Github PK

insightface's Introduction

Hi 👋, I'm Fei Wang

In the code I weave, machine learning thrives; together we achieve, as data comes alive.

insightface's People

Contributors

Stargazers

Watchers

Forkers

insightface's Issues

Recommend Projects

Recommend Topics

Recommend Org