Virtual epochs for PyTorch

A common problem when training neural networks is the size of the data1. There are several strategies for storing and querying large amounts of data, or for increasing model throughput to speed up training when there are large amounts of data, but scale causes problems in much more mundane …

Superconvergence in PyTorch

In Super-Convergence: Very fast training of neural networks using large learning rates1, Smith and Tobin present evidence for a learning rate parametrization scheme that can result in a 10x decrease in training time, while maintaining similar accuracy. Specifically, they propose the use of a cyclical learning rate, which starts …

A faster way to generate thin plate splines

In Evading real-time person detectors by adversarial t-shirt1, Xu and coauthors show that the adversarial patch attack described by Thys, Van Ranst, and Goedemé2 is less successful when applied to flexible media like fabric, due to the warping and folding that occurs.

low success rate with adversarial patch from AUTHORS

They propose to remedy this failure …

How to combine variable length sequences in PyTorch DataLoaders

If you're getting started with PyTorch for text, you've probably encountered an error that looks something like:

Sizes of tensors must match except in dimension 0.

The short explanation for this error is that sequences are often different lengths, but tensors are required to be rectangular. The fix for this …

Adding data augmentation to torchtext datasets

It is universally acknowledged that artificially augmented datasets lead to models which are both more accurate and more generalizable. They do this by introducing variability which is likely to be encountered in ecologically valid settings but is not present in the training data; and, by providing negative examples of spurious …

A multi-file torchtext data loader

To help models generalize, it's common to use some form of data augmentation. This is where the original training data are modified in some way that preserves their semantics while changing their values. The model is fitted to the original training data, plus one or more augmented versions of it …

A faster way to generate lagged values

At Novi Labs, we spend a lot of time working with timeseries data. Generically speaking, these data are formatted something like this:

id  time  value
a     0      1
a     1      2
a     2      3
b     0      4
b     1      5
c     0      6

where we have individual sensors represented by …