C4 W2 A1 - Why ResNet50 works poorly on my own image?

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

Hello, am wondering why ResNet50 works so poorly on my own set of images and how to improve such situation? The Hint says it’s because of the different distributions between sample data and my own images. Maybe to retrain the model with training data containing my own set of images?
Anyone can come up with some solutions for this? Thanks!

2

Hi @Chixing_Wei,

Yes, that’s a good potential solution.

Best,
Mubsi

3

OK, thanks for your reply!

4

Hi @Chixing_Wei

May you can making batch normalization or Standardize your Data to make all data in specific range so it make all data from one distribution .I hope that help you

Please feel free to ask any additional question,
Thanks,
Abdelrahman

5

I upload my photograph, and once it take me by a dog with 100% probability, and the following time it took me by a cat with the same certainty :thinking:

6

Hello, thank you for your reply. I think batch normalization is already implemented in the resnet50 model in stage 1. And there is also a line of code where data of my own image x is divided by 255.0, which in my opinion is the standardization step. I don’t know whether I understand you correctly, if you mean something else please let me know.

7

Hello,Yes I mean that they reduce distributions between samples…so if that is implemented it will affect on it

Thanks

8

OK, but what I mean is even though these steps are implemented already by default, the ResNet50 still works poorly on my own image. This is what I am asking for help with.

Thank you!