How to improve accuracy of some model?

1

From this turorial accuracy of trained network is only 54%
https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html

Is the way to improve accuracy of this network?

Sorry if this is a bit basic of a question, but for some reason I could not find much online to guide me on this. I have googled a lot, read different articles but nothing helps me.

Thanks

2

Hi,

Well this is a very general question indeed. You have many ways to improve such a score. Here are a few possibilities:

  1. Try more complex architectures such as the state of the art model for ImageNet (basically GO DEEPER and at some point you can also make use of “smart modules” such as inception module for instance).
  2. Investigate over fitting of your training by measuring as well the accuracy on the training set and, in case you find a huge gap, look for ways to generalize better your training (dropout, regularization penalties etc…)
  3. The accuracy variance between classes is quite large so it can be due to many different facts (some classes might be underrepresented in the data set or just harder to detect etc…) so you could try to improve the accuracy on classes like frog or cat with some tricks (sur-sampling for instance).
  4. Additional data would also certainly help but this is generaly not what people means by improve the accuracy of a model as adding data almost always improve accuracy.
    You might have many other possibilities but here are a few that should definitely improve your performances.

Hope this help,

1 Like
3

Hi

Thank you fro reply.

Please maybe you can provide some links which explain how to make network deeper.

4

Hi,

Go deeper basically means add more layers. But in order to do it in a smart way you can have a look at this article:

Where are listed the state of the art CNN architectures for ImageNet over the years. Maybe you can learn from that evolution over the years and design something adapted to your problem later. But anyway increasing the amount of layers or the amount of filters per layer increase the amount of parameters of your model so, at first, should improve the performances of your classifier.

2 Likes
5

Thank you! This helps!

6

how can we add layers , please help

7

how did you add more layers can you help me please. I am getting error