How to 'troubleshoot' model accuracy issues (dog breeds example)

Part 1 (2018) Beginner (2018)
1

Hi.
I’m testing my model for dog breeds identification.

  1. i get results with ~90-91% of accuracy, score on kaggle ~0.25 . This result seems to be far from what Jeremy gets in video, although i reproduce same steps on same data. Does anyone have same problem ? Can this depend on environment i run it on ? I use kaggle data, resnext50, make same steps as in video …

  2. although 91% is much less, than possible, it still seems to be decent. But when i run predictions on individual images from google, i got 7 mistakes from 7 tries. Not even close guess to the breed on the image, e.g.:

    image
    image.jpg732×733 235 KB

I resize images to something around 300-600 px, dogs are in the middle… but my model’s predictions are disappointing.
Any ideas on how to ‘troubleshoot’ quality issues with the model ?

1 Like
2

Same here,
Did you solve the problem?
I am desperate :confused:

3

No, I’m stuck :frowning:

1 Like
4

Chiming in here with the same issue.

1 Like
5

can you tell what the value of “sz” in your notebook?