Beans and Eurosat Image Training

Beans

When I trained the beans dataset, I got some interesting outputs.

loss: 0.8523 - accuracy: 0.6336 - val_loss: 0.7224 - val_accuracy: 0.7212

My accuracy was about 63%, and my val_accuracy was around 72%. This is an okay performance for a dataset with only about 1,300 images to use. 1,034 of those images were used for training, 128 for testing, and 133 for validation. I believe that if we had more images in all 3 datasets, the accuracy would have been higher.

Eurosat

These were the outputs I got when I trained the eurosat dataset:

loss: 0.7017 - accuracy: 0.7433 - val_loss: 0.5920 - val_accuracy: 0.7811

Both the accuracy and the val_accuracy were higher in this case. I think that this is because there were 27,000 samples in this dataset. 80% of that was for training (so 21,600), 10% were for testing (2,700), and 10% were for validation (also 2,700). Having more samples means that the model has more of a chance to generalize and predict the testing/validation images better.

Image Augmentation

Beans

My accuracies after image augmentation on the beans dataset are both lower than the original accuracies. I chose the second option of creating a random generator.

loss: 1.0987 - accuracy: 0.3325 - val_loss: 1.0992 - val_accuracy: 0.3077

I was able to run 3 epochs, and it took around 12 minutes.

Eurosat

My accuracies for the eurosat dataset were also significantly lower than the original ones:

loss: 1.8328 - accuracy: 0.2973 - val_loss: 1.6687 - val_accuracy: 0.3489

I also ran 3 epochs, with a total of about 5 minutes.

In conclusion, the image augmentation using the random generator method did not help the model, in fact, it hurt the model.