A Foo walks into a Bar...

Why SSD and why bother with AlexNet?

This choice is a bit strange, especially AlexNet few people use. But it’s not the final architecture. When building a very complex system from scratch, I try to build it one piece at a time, and move on to the next piece when the first piece is proven working. I already have a dataset where I know AlexNet works well.

Why would it be more accurate?

I think part of the issue with the original Smiling Bot’s model is the small training dataset. I only have ~10,000 labeled images of smilinng / not smiling faces. Dataset of this size is nothing to write home about. But I can get way more training data for just faces, without knowing their emotions, and use transfer learning.

Will buffering help?

Even if you add a buffer (via shuffle or prefetch), this buffer will be eventually depleted before the training finishes. Buffering takes time, and the more you want to postpone the buffer depletion by making a larger buffer, the more prefetching the buffer will take, and this equation will never balance the way you intend.

How to save to disk?

So it’s completely sensible to start with a pipeline on the fly, iron out the data until you’re confident that the feature vectors are correct, and then to save the large dataset to disk. Saving the data to disk takes only very few lines of code. I tried it by using tf.Dataset.shard in combination with this guide, but for some reason it works very slow.

What to expect?

Expect a lot of hard work if you’re actually going for the certificate (if you are just taking some courses, then pick your own expectations). The requirements listed on the certificate page are strict, so please review them. 10 hours of work a week + several weekends spent on the final project, and that’s if you take one course at a time.