About EMA

[amazon-science]

About EMA

[aaberdam]

Hi,

Thank you for your interest in our work.

To enable EMA, set ema: True in the configuration file found at this link:

semimtr-text-recognition/configs/semimtr_finetune.yaml

Line 86 in d4921a3

ema: False,

This will cause the program to use the script model_fusion_teacher_student_ema.py.

By default, the teacher gradients are stopped, but if you wish to change this, you can do so also in this configuration.

If you have any more questions, please let me know.
Aviad

[YuNie24]

Thank you for reply :)

but still have some questions.

Would you explain more on how "model_fusion_teacher_student_ema.py" is called?

According to my test of your SemiMTR finetune scripts, "ConsistencyRegularizationFusionModel" model has about 36M trainable parameter and students and teacher use the same model parameters.

However, when using EMA, it seems that the parameters should be doubled because structure have to have a separate ABINetIterModel on student and teacher.

I set ema: True in the configuration file(semimtr_finetune.yaml), but trainable parameter has still 36M and looks not using TeacherStudentFusionEMA.

[YuNie24]

Ah, I just checked another config line.

semimtr-text-recognition/configs/semimtr_finetune.yaml

Line 56 in d4921a3

name: 'semimtr.modules.model_fusion_consistency_regularization.ConsistencyRegularizationFusionModel'

and it looks like it should be name: semimtr.modules.model_fusion_teacher_student_ema.TeacherStudentFusionEMA
instead of semimtr.modules.model_fusion_consistency_regularization.ConsistencyRegularizationFusionModel if we apply EMA.

then TeacherStudentFusionEMA and ConsistencyRegularizationFusionModel cannot be executed simultaneously.
is it right ?