3.1 Training parameters#
Number of images seen during training
Given the following parameters:
batch_size = 16
number_of_steps = 200
number_of_epoches = 50
How many images are seen during training in total? (including duplicates, because images are seen multiple times)
Default number of steps
What is the default value of the number of steps? And why?
Lots of steps or lots of epochs?
In the case of our specific implementation, given a constant number_of_steps * number_of_epochs
. Is it better to train with more steps or with more epochs?
Ratio between training and validation
Some of the patches are used as validation set, while others are used in the training set.
What is the difference between training and validation set? And how much data should you use in the validation set?
Initial learning rate
Why is the learning rate parameter called "initial" here? And why is this important?