Scan barcode
A review by _walter_
Machine Learning Design Patterns: Solutions to Common Challenges in Data Preparation, Model Building, and Mlops by Sara Robinson, Michael Munn, Valliappa Lakshmanan
3.0
I feel a little conflicted giving this book a lower rating, and though I'm sure I'll reach for it again someday as a reference, I think it's still a pretty fair assessment of its strengths and weaknesses.
The good:
The subjects covered encompass the entire ML lifecycle, from data ingestion and cleaning, to model tuning and deployment. If you have some experience in this area, a lot of these techniques will be familiar to you, but even then, the way the authors frame each of them as a "pattern" is exquisitely done.
The book is broken down into main themes, and then further into specific patterns. For each of these they present the problem, solution, why it works, and tradeoffs/alternatives. This organizational style makes it very valuable as a reference, and as a refresher for why we do things sometimes.
The not so good:
The book should have been named "Machine Learning Design Patterns...WITH GOOGLE CLOUD SERVICES, LOL".
Seriously, while the authors attempt to provide the reader with some variety regarding the tools one might use to enforce these aforementioned patterns, the examples invariably end up coming back to the Google Cloud Stack. This means you should expect lots of TensorFlow examples, BigQuery ML, etc. Which is fine, but then don't say you'll present "options" when 99% are just one. Anyways...
They really can't help themselves, and sometimes hilarity ensues. Case-and-point: there's a chapter where they attempt to implement one of their patterns using XGBoost, saying something along the lines of "the last few examples have been done using TensorFlow, so will take a look at other tools..." They make the smallest effort to get it going but then they construct this weird hypothetical case so they can get back to using what they know and like, and it goes like this, "but now suppose someone comes in and wants to use TENSORFLOW mkay? Let's do it..."
So there you have it. I think it is still worth to keep as a reference and to fill in the gaps and blind spots you'll ultimately develop as an ML practitioner, but if you are expecting something more hands-on with non-Google stack tools, look elsewhere.
Recommended with caveats as above.
The good:
The subjects covered encompass the entire ML lifecycle, from data ingestion and cleaning, to model tuning and deployment. If you have some experience in this area, a lot of these techniques will be familiar to you, but even then, the way the authors frame each of them as a "pattern" is exquisitely done.
The book is broken down into main themes, and then further into specific patterns. For each of these they present the problem, solution, why it works, and tradeoffs/alternatives. This organizational style makes it very valuable as a reference, and as a refresher for why we do things sometimes.
The not so good:
The book should have been named "Machine Learning Design Patterns...WITH GOOGLE CLOUD SERVICES, LOL".
Seriously, while the authors attempt to provide the reader with some variety regarding the tools one might use to enforce these aforementioned patterns, the examples invariably end up coming back to the Google Cloud Stack. This means you should expect lots of TensorFlow examples, BigQuery ML, etc. Which is fine, but then don't say you'll present "options" when 99% are just one. Anyways...
They really can't help themselves, and sometimes hilarity ensues. Case-and-point: there's a chapter where they attempt to implement one of their patterns using XGBoost, saying something along the lines of "the last few examples have been done using TensorFlow, so will take a look at other tools..." They make the smallest effort to get it going but then they construct this weird hypothetical case so they can get back to using what they know and like, and it goes like this, "but now suppose someone comes in and wants to use TENSORFLOW mkay? Let's do it..."
So there you have it. I think it is still worth to keep as a reference and to fill in the gaps and blind spots you'll ultimately develop as an ML practitioner, but if you are expecting something more hands-on with non-Google stack tools, look elsewhere.
Recommended with caveats as above.