luckycallor

@clhne The currently released pretrained model is based on modified ResNet_v2-50 network, which has lots of weights. In the future we will release pretrained model based on smaller models such...

1 means cosine distance, 0 means euclidean distance. You can use either to measure whether two faces are the same person.

@Wxsang

@x3393467 the finetune code only update the last fully connected layer within the "fixed_epoch_num". I recommend you change the run params.

@Muhammad333 When the distance of two embeddings is small, the two faces are the same person.

@lagidigu When using the dataset opensourced by the official InsightFace project, there is no need to do further cropping.

@BruceDai003 Batch size in configure files are mainly used when training the model. For inference, any batch size is ok.

@YJingyu I am sorry that there is a bug in get_embd.py. Now the bug has been fixed. Thank you for using my code.

@huytq97 @Hamiltonsjtu @SharharZ Did you use the latest get_embd.py? There was a bug in get_embd.py, and I have fixed it two months ago. Please check with the latest code.

For both training and evaluation data, you can see the DataZoo of the official mxnet project: https://github.com/deepinsight/insightface/wiki/Dataset-Zoo