How to break down large SVM classification model?

Question

I have a classification problem with large number of classes: feature set is 512 Dimension, number of classes are around 3000. This is a face identification problem. (identify among 3000 celebrities, whose face it is. The feature is extracted using FaceNet.)

The issue for training such a SVM model is too slow:

I used sklearn SVC, it came with below result:

RAM usage: > 100GB // I eventually ended up using virtual memory of 100GB
Training time: > 30 hours
Classification time: > 1 hour per face
Other issue: Single CPU usage, no parallelization

To summarize, it is not practical to use above naive way of training for such a SVM model.

My Question: What is a practical way of optimizing SVM training / usage of this scale input data? (~ 3000 classes, feature_size = 512)

Answer 1

I think that using SVMs for your model is the main problem.

SVMs (linear or otherwise) are originally intended for binary classification. There are various procedures for extending them to a multi-class problem. The most common methods involve transforming the problem into a set of binary classification problems. The most common strategy is using "One-Versus-All" classifiers (OVA classification), and choosing the class which classifies the test data with the greatest margin. Another strategy is to build a set of One-Versus-One classifiers, and to choose the class that is selected by the most classifiers. Both these strategies are computationally expensive both in training and in testing. I know there are other, more complicated methods, but I don't think that any of them is intended for such a high number of classes.

The most trivial way would be to use a neural-network model for such a problem. If you want to use one of the more classic algorithms, perhaps using decision trees will work (or even k-nearest neighbors).