Running the pizza/steak/sushi dataset individually through the Section 6 model gives 43% accuracy (despite test results being 90%)

[Abu-AM]

I was trying my hand at the section 6 exercises and decided to individually convert each image in the dataset and run it through an already trained section 6 model (using transfer learning) and found the accuracy doesn't match up. I thought it might have to do with the dataset I was using so I went to Daniel's solutions and found the same thing happening. Is this normal or I'm doing something wrong?

https://colab.research.google.com/drive/1zm_iN_4F-Pw0LXjDV8Ravg5vDlN7qQk1?usp=sharing