JSGD: SGD for large-scale classification

---

This a Stochastic Gradient Descent algorithm used to train linear multiclass classifiers. It is biased towards large classification problems (many classes, many examples, high dimensional data).

The SGD implementation corresponds to the articles:

• Good Practice in Large-Scale Learning for Image Classification , Zeynep Akata,Florent Perronnin, Zaid Harchaoui and Cordelia Schmid, TPAMI 2013
• Towards Good Practice in Large-Scale Learning for Image Classification , Florent Perronnin, Zeynep Akata, Zaid Harchaoui and Cordelia Schmid, CVPR 2012.

---

Source

The latest source is available here: jsgd-61.tgz

Previous (stable) version: jsgd-56.tgz

---

Features

The Package includes:

• A readable Matlab implementation of the SGD method based on the paper;
• A C implementation of SGD, intended to be fast;
• Matlab and Python interfaces to the C package;
• Test scripts;
• An optimization method that finds the best parameters for the SGD by cross-validation.

Documentation is in the README file of the package.

Data files

Here are a few associated datafiles. They all contain descriptor in .fvecs format, computed on Imagenet subsets.

• groupFungus_k64_nclass10_nex10.tgz (3 MB): 4096D Fisher descriptors for 10 classes of the Fungus group, 10 examples per class (+ 10 for test).
• groupFungus_k64_nclass134_nex50 (200 MB): 4096D Fisher descriptors for all classes of the Fungus group, 50 examples per class (+ 50 for test).
• groupUngulate_BOW4096.tgz (430 MB): 4096D BOW descriptors for all images of the Ungulate group.
• k1024_nclass50_nex50.tgz (2.6 GB): 128kD Fisher descriptors for the 50 first classes of the ILSVRC2011 benchmark, 50 examples per class (+ 50 for test).
• fvecs_read.m: script to load .fvecs files into matlab.

---

Contact

Feedback is welcome to

• mattis (dot) paulin (at) inria (dot) fr

---

Last modified: June 19, 2013.