Do you know why the accuracy is lower than the original project

[xmuszq]

@auroua E.g., the ACC for LFW in original project could go up to 99.8%. Do you know what's missing here?

[auroua]

The author use 100 layer network. Please compare with the LResNet50E-IR.
Another reason is the author using a large batchsize which I can't use because of the memory limit.

[siahewei]

very strange problem? i cannot train well using this project, my best accuracy 96% on lfw, so i changed my training method which based on amloss ,as follows: 1. use the model which had trained well by others. 2. keep inference logit as set trainable=false 2. only train param 'w' in arcface loss part. but get bad result, the inference loss is alway about 25. And i have trained "Additive-Margin-Softmax", i get accuracy 99.3% very easy. so i doubt the 'arc face loss' really wok?

[ruobop]

@siahewei
I can achieve 99.6% on LFW at 400K iters by simply using this project's original code. I am still waiting for better result with more iters. You may double check your train data and consider increasing batch size a little bit.

[xmuszq]

@ruobop let us know when you get better results. Thanks.

[HsuTzuJen]

@auroua In the TF document, It says that using NCHW is better than using NHWC in training mode with CUDA, and NHWC in inference mode.
Is there any way to achieve this?

[darchonyzx]

I come into the same problem with @siahewei . I use 96x96 images to train a model, but it seems to up to the bottleneck 0.96 in lfw test.