使用arcfaceloss无法收敛 about insightface_tf HOT 13 CLOSED

Based on my training, it is hard to converge loss to a small number if train with m=0.5. My solution is that train with m =0 at the beginning and then increase m to 0.1, 0.2, 0.4, 0.5. It works well and the loss goes from 6 to 0.3 in my database.

from insightface_tf.

可以先使用softmax进行预训练，之后使用arcface进行fine tune
在原作者的issue中有过相关讨论
https://github.com/deepinsight/insightface/issues/214

from insightface_tf.

@ltcs11 为什么使用arcface训练时一直显示Accuracy为0，

谢谢

from insightface_tf.

@Joll123
你是直接使用softmax预训练之后再接arcface fine tune吗？
如果是，可以在继续训练一段时间，或增大batch_size

from insightface_tf.

不用softmax预训练应该也能收敛，只是效果不好。我那个不收敛是因为minibatch太小只有128，学习率又太大导致。减小学习率就能收敛了

from insightface_tf.

@dl8207531 你取的学习率多大

from insightface_tf.

@dl8207531 老兄用哪个数据集训练的, margin多少?训练的模型在lfw的准确率能达到多少呢?

from insightface_tf.

@Joll123 用softmax预训练可以从0.01开始 如果想测试直接mobilefacenet训练可以试下也从0.01开始，如果不行就0.001肯定能收敛

from insightface_tf.

@jetsmith 就InsightFace提供的那个数据集 margin 0.5 mobilefacenet 准确率99.367% 达不到作者baseline 我猜原因应该是minibatch太小

from insightface_tf.

@dl8207531 那你用softmax预训练训到什么程度转arcface训练呢, 你arcface学习率的schedule是如何设置的呢?

from insightface_tf.

@jetsmith softmax从0.01开始训练到不再收敛保持该学习率改用arc，arc同样训练到不再收敛学习率乘0.1继续训练

from insightface_tf.

@dl8207531 您好，请问 你后面改用arc时是全部参数训练，还是冻结了卷积层呢？

from insightface_tf.

and then increase m to 0.1, 0.2, 0.4, 0.5.

How about the frequent of increasing m?

• Epoch [0-10] m is 0, [10-20] m is 0.1, ... epoch [40-50] m is 0.5. If I do this, my losses will increase after the adjustment.
• or epoch 1 m is 0, epoch 2 m is 0.01, ..., epoch 50 m is 0.5 ?

And lr gradually decay or remain constant at a small value?

from insightface_tf.