Hornig’s struggles in convincing President Johnson were chalked up to the advisor being “a subtle man dealing with an un

ars from different eras have always debated about which year has been the worst for humankind and most of them point their fingers at AD 536, not only due to the massive number of deaths but also because of the many events that took place which led to a scaringly high death count and terrible living conditions for most humans. In this article, we will be looking at some of the factors that make this the worst year and why other historians, as well as scientists, beg to differ.

So based on this: if you have large (at least a few thousand samples) real world dataset, the faiss-based version is just plain better. For small toy dataset Scikit-learn is a better choice; however, if you have a GPU, the GPU-accelerated faiss version may turn out faster (I haven’t checked it for fair CPU comparison).

With 25 lines of code, we can get a huge speed and accuracy boost for K-Means clustering for reasonably sized datasets with the faiss library. If you need, you can get even better with GPU, multiple GPUs and more, which is nicely explained in faiss docs.

All of those times have been measured with the time.process_time() function, that measures process time instead of wall clock time, for more accurate results. Results are averages of 100 runs, except for MNIST, where it took too long for Scikit-learn and I had to do 5 runs.

The final quality of clustering is calculated as a sum of in-cluster distances, where for each cluster we calculate a sum of Euclidean distances between points in that cluster and its centroid. This is also called inertia.

As we can see, for K-Means clustering for small datasets (first 4 datasets) faiss-based version is slower for training and has a larger error. For prediction it works universally faster.

For larger MNIST dataset faiss is a clear winner. Training 20.5 times faster is huge, especially since it reduces time from almost 3 minutes to less than 8 seconds! A 1.5 times faster prediction is also nice. The true achievement, however, is a spectacular 27.5 times lower error. This means that for larger real world dataset the faiss-based version is vastly more accurate. And this takes only 25 lines of code!

A great feature of faiss is that it has both installation and build instructions (installation docs) and an excellent documentation with examples (getting started docs). After the installation, we can write the actual clustering. The code is quite simple, since we just mimic the Scikit-learn API.

As we can see from the example, the core of the algorithm is searching for nearest neighbors, specifically nearest centroids, both for training and prediction. And that’s where faiss is orders of magnitude faster than Scikit-learn! It leverages great C implementation, concurrency wherever possible and even GPU, if you want.

In short, what matters most as a beginner in Data Science is that you DO Data Science. So just go with either one of the languages and prioritize getting some projects done while sipping away at your choice of sugary beverage. That’s how you will learn the fastest.

I’ve chosen a few popular datasets available in Scikit-learn for comparison. The train times andpredict times are compared. For easier reading, I’ve explicitly written how many times faster is the faiss-based clustering than Scikit-learn’s. For error comparison I’ve just written how many times lower error the faiss-based clustering achieves (since numbers are large and not very informative).

in prediction we perform kNN search with k = 1, returning indices of nearest centroids from self.cluster_centers_ (index [1], since index.search() returns distances and indices)

In both libraries we have to specify algorithm hyperparameters: number of clusters, number of restarts (each starting with other initial guesses) and maximal number of iterations.

While I may be tempted to just recommend Python straight-away (Python is my main, but I do have some working knowledge of R), I want to present an unbiased evaluation of the effectiveness of the two languages for a beginner. This is mainly because the right choice is most definitely going to depend on your own particular situation.

Catagory :general

Hornig’s struggles in convincing President Johnson were chalked up to the advisor being “a subtle man dealing with an un

© Since 2015 TheWyco