March 2020
Beginner to intermediate
342 pages
8h 38m
English
In Batches Large and Small, we found out that stochastic GD doesn’t work well for our specific problem and neural network configuration. On the other hand, mini-batch GD with a batch size of 32 seems to do okay.
What’s the smallest batch size that gives us better early feedback than plain old batch GD? Find out for yourself by changing the batch sizes in compare_batch_sizes.py, from the 13_batching directory in the book’s source code.
It is also possible that stochastic GD would work better on our problem if we used different hyperparameters. For example, try it with a smaller learning rate. Do you get a smoother loss?
Read now
Unlock full access