How to train using different datasets?

hello. In Lab2, there is a code that trains with the dataset.

Currently, I am running Lab2 in a GPU environment and wanted to train it using a different dataset.

The image I used as training data is in the link below. (Pictures of vegetables)

The images were created as npy files and ran for 32 epochs with a batch size of 70.

However, we confirmed that the noise was still not removed as shown below.
How can I solve this?