ArcfaceLoss vs ArcfaceLossSimple

[nadongjin]

Hello
I have another question while looking at your code.
The ArcfaceLoss and ArcfaceLossSimple of your code are actually the same as the current learning criteria, right?
Because ArcfaceLossSimple's cos(y_pred_values+margin) = cos(y_pred_values) * cos(margin) - sin(y_pred_vals) * sin(margin), ArcfaceLoss's theta = tf.cos(tf.acos(y_pred_vals) + self.margin2)

margin==margin2==0.5

class ArcfaceLoss(tf.keras.losses.Loss):

def call(self, y_true, norm_logits):
pick_cond = tf.where(y_true != 0)
y_pred_vals = tf.gather_nd(norm_logits, pick_cond)
if self.margin1 == 1.0 and self.margin2 == 0.0 and self.margin3 == 0.0:
theta = y_pred_vals
elif self.margin1 == 1.0 and self.margin3 == 0.0:
theta = tf.cos(tf.acos(y_pred_vals) + self.margin2)
else:
theta = tf.cos(tf.acos(y_pred_vals) * self.margin1 + self.margin2) - self.margin3

class ArcfaceLossSimple(tf.keras.losses.Loss):

def call(self, y_true, norm_logits):
pick_cond = tf.where(y_true != 0)
y_pred_vals = tf.gather_nd(norm_logits, pick_cond)
theta = y_pred_vals * self.margin_cos - tf.sqrt(1 - tf.pow(y_pred_vals, 2)) * self.margin_sin
theta_valid = tf.where(y_pred_vals > self.threshold, theta, self.theta_min - theta)
arcface_logits = tf.tensor_scatter_nd_update(norm_logits, pick_cond, theta_valid) * self.scale
return tf.keras.losses.categorical_crossentropy(y_true, arcface_logits, from_logits=self.from_logits, label_smoothing=self.label_smoothing)

[leondgarse]

Yes, ArcfaceLossSimple is just a simple version of ArcfaceLoss, that using only margin2. Their theta results are same.

[nadongjin]

Your code works well without any problems.
However, I think face detection and landmark are more important than anything else in actual accuracy that does not provide a separate landmark such as IJBB and IJBC, so I would like to try this.
My ms1mv3 dataset organization work is 75% completed. Please wait a little longer.
I think your learning method is SOTA accuracy with only limited datasets.
Since the category_cross_entropy itself generally uses the loss average, I think there will be not a big problem even in the wrong datasets.
However, if you learn with a good datasets, unnecessary loss will be reduced, so I think it will bring the inside closer and the outside farther away.
In fact, about 2.5% of the ms1mv3 datasets were duplicated or had problems.
The conclusion is that if you are given great data, your model will not be able to keep up with anyone.
Already, Chinese, Korean, or global facial recognition companies are enormous and have excellent data. In fact, some companies can buy data with money.
But you and I can't do that.
For example, your model is generally excellent, but I think it is the cause of a slight decline in performance compared to global companies in terms of accuracy of Asians and blacks.

[leondgarse]

Ya, you probably right.

• By accuracy of Asians and blacks, are you referring MR-ALL validation in insightface model zoo? I haven't run that test myself.
• Using a cleaner dataset may be able to apply a heavier regularizer, like larger weight_decay, larger randaug, or larger margin.
• The dataset glitch may explain why Triplet performing bad when fine-tuning an Arcface trained model.

[nadongjin]

Hello
I have an idea.
When I did the accuracy test, I confirmed that the accuracy was improved by flipping on one image and then averaging it.(Of course, this is included in your code.)
In common, models with nist's good accuracy tend to take up a lot of memory when making inference.
But the model size is not big.
Therefore, I think it would be good to improve performance by applying another random_status differently like the image flip applied during learning. : )
thanks..

[nadongjin]

[example]

[leondgarse]

Do you mean a random process averaging the augmented images during training?

• For augment like random_brightness / random_contrast, it may not help. Take random_brightness alone for instance, averaging it with original image may just result an image more like original image, works like a smaller brightness delta.
• For averaging flipped image with original image, I added some commented code in flip average test. If use it instead of img = tf.image.random_flip_left_right(img), will set a 1/3 probability for either original / flipped / averaged image. I think it may help, but still not sure, as some averaged results went weird.
  
• For averaging random_crop, can be applied similarly like random_flip.