When training the initial colour model, I found that the validation loss was exponentially increasing after only 2/3 epochs. This generally occurs due to overfitting. Overfitting takes place when the model fits to the training data too well; find an example in figure 1 below. Consequently, the model cannot identify general patterns and instead only recognises very specific features. To reduce the effects of overfitting, and to improve the model’s accuracy, I utilised a variety of techniques. The techniques used and their explanations can be found by visiting their documentation:
Figure 1: Visual representation of overfitting and underfitting to data.
Source: Medium