Week 2 Model Misclassifies Cats

My model with augmented data has over 80% accuracy for both training and validation. But testing it on the images I pulled of the internet for the ungraded assignment (simple model 70% accuracy) I have many more misclassifications.

(I added a cell with code for manually uploading images from week 1)

All the dogs are classified as dogs. But of the six images of cats only two were identified as cats. The model from week 1 identified five of the six as cats.

What could the reason be?

Here are some reasons:

  1. Model performance is ~80% on training data. It’s still possible for the model to make mistakes even if images on which the model was trained on was used. Build a better model to address this issue.
  2. Cat images used in testing the model aren’t the kind of images the model was trained on. For instance, if you train your model with cat images in good lighting and expose the model to images with snow background, bad results can be expected. Include this data as part of model training / perform data augmentation to address this. See drift