Question about batch size

[fs302]

Hi Shuai,

I find the code in gnn.py takes 1 record per batch, will it suffers the problem of overfitting or takes more time in training and inference? And what is your motivation for this design.

Anyway, I learned a lot from your implementation. Thanks!

[itxiud2015]

May I know the applicable version of tensorflow for the code?

[Calemsy]

Hi Shuai,

I find the code in gnn.py takes 1 record per batch, will it suffers the problem of overfitting or takes more time in training and inference? And what is your motivation for this design.

Anyway, I learned a lot from your implementation. Thanks!

thanks for your good questions.
I can't implement the version of vectorization, so take one example per batch, and really suffers the problem of time consumptions as you said, but how to understand this can lead to overfitting.

[Calemsy]

May I know the applicable version of tensorflow for the code?

This code base on tensorflow 1.12.0.

[fs302]

Hi Shuai,
I find the code in gnn.py takes 1 record per batch, will it suffers the problem of overfitting or takes more time in training and inference? And what is your motivation for this design.
Anyway, I learned a lot from your implementation. Thanks!

thanks for your good questions.
I can't implement the version of vectorization, so take one example per batch, and really suffers the problem of time consumptions as you said, but how to understand this can lead to overfitting.

Well, forget it(overfitting). I rethought this problem, the major challenge is the computing cost, both in training and inference. I will try to extend it and if that works I will let you know.

[Calemsy]

Hi Shuai,
I find the code in gnn.py takes 1 record per batch, will it suffers the problem of overfitting or takes more time in training and inference? And what is your motivation for this design.
Anyway, I learned a lot from your implementation. Thanks!

thanks for your good questions.
I can't implement the version of vectorization, so take one example per batch, and really suffers the problem of time consumptions as you said, but how to understand this can lead to overfitting.

Well, forget it(overfitting). I rethought this problem, the major challenge is the computing cost, both in training and inference. I will try to extend it and if that works I will let you know.

trust you and good luck for you, hope learn from each other.