pppoe / tensorflow-sphereface-asoftmax Goto Github PK

cos_sign = tf.sign(cos_th)

res = 2*tf.multiply(tf.sign(cos_th), tf.square(cos_th)) - 1

what's the mean of the sign function in calculate the cos(m*th)

Since m=3 hasn't been implemented, will you please provide the formula for that.

what is the meaning of 'f' and 'ff' ? 0.5???

Hi author.
There is a new published paper proposed additive margin softmax(AM-softmax), which seems easier to train than sphereface.
https://arxiv.org/abs/1801.05599
I implemented main part of this paper here by tensorflow:
https://github.com/Joker316701882/Additive-Margin-Softmax
With exact hyper-parameters with author, I can also only achieve 98.x% accu. So I'm wondering is it the problem of tensorflow low-level implementation (like optimizer) different from caffe so that with same parameters, it's hard to reach exact performance. Have you ever tried loss in paper like AM-softmax(cosface), arcface etc.

Will be glad to know your idea about this.

hi, I change the Network from TFlearn code to TF slim code, the tf-slim code as follow
`def prelu(_x, scope=None):
"""parametric ReLU activation"""
with tf.variable_scope(name_or_scope=scope, default_name="prelu"):
_alpha = tf.get_variable("prelu", shape=_x.get_shape()[-1],
dtype=_x.dtype, initializer=tf.constant_initializer(0.1))
return tf.maximum(0.0, _x) + _alpha * tf.minimum(0.0, _x)

def Network(data_input, training = True):
x = slim.conv2d(data_input, 32, [3,3], stride = 1, activation_fn=prelu)
x = slim.conv2d(x, 32, [3, 3], stride = 2, activation_fn=prelu)
x = slim.conv2d(x, 64, [3, 3], activation_fn=prelu)
x = slim.conv2d(x, 64, [3, 3], stride = 2, activation_fn=prelu)
x = slim.conv2d(x, 128, [3, 3], activation_fn=prelu)
x = slim.conv2d(x, 128, [3, 3], stride = 2, activation_fn=prelu)
x = slim.flatten(x)
feat = slim.fully_connected(x, 2)
return feat
`
when I train the network, the loss does not decline, how can i solve this problem?Thanks

0 2.3063145
100 2.3025854
200 2.3025854
300 2.3025854
400 2.3025854
500 2.3025854
600 2.3025854

sloved! just change "feat = slim.fully_connected(x, 2)" to "feat = slim.fully_connected(x, 2,activation_fn=prelu)"

I tested your code for 20-layer setting of Sphereface paper with the CNN and residual units but the loss value for m=4 became weirdly huge.

I tried to reproduce sphereface via tensorflow on CASIA-WebFace, but I found the loss function is very difficult to convey, and loss function sometimes be very large (10e6). Are there any solutions?