backbones/renet.py

[1303463822]

in resnet.py, the ResNet101 uses basicblock with depthes = [3, 4, 23, 3] rather than bottleneck. Are there any errors?

[leondgarse]

Not error, just following the usage in official insightface/iresnet.py, and it's accuracy is bit higher. You can also find some training log using typical ResNet50V2/ResNet100V2 in Resnet50V2 / Resnet101V2 swish using SGD + L2 regularizer + cosine lr decay training on MS1MV3 dataset. Or just use keras.applications.ResNet101 by using "resnet101" rather than "r100" when initializing basic_model, built by models.py#L39-L45.

import models
basic_model = models.buildin_models("resnet101", dropout=0, emb_shape=512, output_layer='E', bn_momentum=0.9, bn_epsilon=1e-5)
...