These are three images that I pulled from the complete 10000 image dataset. I made sure to only choose images that were not seen by the model in the training set (images inside the training set between 251.jpeg and 9000.jpeg) since I only trained the model on the first 250 images.
The model returned 0.04274386 for 4827.jpeg compared to the real value 68.8593444824219.
The model returned 0.04743501 for 6212.jpeg compared to the real value 46.7590255737305.
The model returned 0.04669612 for 8197.jpeg compared to the real value 110.827850341797.
I believe that the model was trained on too little data, as I can only manage about 250 images. This is what I believe is giving my model such bad predictions.