Hi! Could you please explain to me this block?
Hi, the main idea of this slide is to explain the importance of having different generators for validation and test data.
First, remember that we can split our data into training, validation, and test sets. Each of these has a specific purpose in the process. Training data helps your model learn and generalize about its performing task. Validation data helps your model know how well it is learning during the training time. Finally, test data is data your model has never seen, so it is the final proof of the quality of your model.
Since each of your sets is different, it makes sense to use different generators. Using batch normalization in training helps the performance and generalization during this stage. Still, it won’t make sense to do it during validation and testing since these two scenarios try to emulate a real-life situation where new unknown incoming data comes one by one.
Additionally, using the test data statistics would be like cheating since you give your model more information about the data, which is no longer entirely unknown. That is why we only use the statistics of the known data, the training data, so we are not “cheating” by giving the model more information than it should have from the new incoming data.
I hope this helps. Please let me know if you have any more questions or if anything is unclear.