- Removing the pooling layer between the two CNN layers
- Reduced beginning loss from 14 to 2
- Removing the standardization of the data
- No change
- Changing the learning rate from .001 to .01
- No change
- Getting rid of the image data generator
- Huge improvement in speed and accuracy
- Dropout layers
- Better validation accuracy
- Dropout layer between the dense layers
- Works if rate < .5
- Dropout layer after CNNs
- Really effective if < .5
- Increasing epochs:
- Worse. Overfits quickly.
- Increasing batch size
- Works up to a point (128) then performance begins to decrease
- Standardizing data by dividing by max value
- Worse performance. Maxes out ~.98
- Supplementing data with augmentations
- Overfit very quickly
- Worth noting that I doubled the test set by randomly rotating the digits between 1 and 90 degrees
- Using a range of 1-15 dramatically improved validation accuration over 1-90.
- 1-10 even better
- 1-5 better still
- Testing accuracy was not any better