A library for approximating loss landscapes in low-dimensional parameter subspaces
loss-landscapes is a library for approximating the value of neural network loss functions in low-dimensional
subspaces suitable for visualization. The intended purpose of this library is to simplify the production of loss
landscape visualizations such as those seen in Visualizing the Loss Landscape of
What is a Loss Landscape?
L : Parameters -> Real Numbers be a loss function, which maps the vector containing all parameters of a
neural network model to a real number. This is true of loss functions in general, of course, but we focus on
neural networks. Since the network has
n parameters, the loss function
L takes an
n-dimensional input. We
can define the loss landscape as the set of all
(param, L(param)), for all parameter
param. For example, the image below, reproduced from the paper mentioned above by Li et al (2018),
provides a visual representation of what a loss function over a two-dimensional parameter space might look like:
Of course, in reality machine learning models have many more parameters than just two, so the input space of the loss function is virtually never two-dimensional. In fact, the model whose loss landscape is shown in the image above actually has far more than two parameters, too. Because we can't produce visualizations in more than three dimensions, we cannot hope to visualize the "true" shape of the loss landscape. Instead, a number of techniques exist for reducing the parameter space to one or two dimensions, ranging from dimensionality reduction techniques like PCA, to restricting ourselves to a particular subspace of the overall parameter space. For more details, read Li et al's paper above.
This library facilitates the computation of a neural network model's loss landscape. Its core design principles are that the library should not mandate any visualization technique or any particular definition of loss, and it should be agnostic to the numerical computation library in use. As an example, it allows a PyTorch user to examine the topology of a model's loss function around its current parameters with filter-normalized directions by calling
landscape = loss_landscape.random_plane(model, evaluation_function, normalize='filter')
This line would return a 2-dimensional array of loss values, which could in turn be plotted with any software.
Below is a simple contour plot made in
matplotlib that demonstrates what a planar loss landscape could look like.
examples directory for more in-depth examples of what is possible.
Current and Planned Features
Currently, the library has the following features:
- Computing loss functions along a line defined by two points
- Computing loss functions along a random direction, starting from a specified point
- Computing loss functions along a plane defined by three points
- Computing loss functions along a plane defined by a specified point and two random direction vectors
Furthermore, filter normalization can be applied to the random direction vectors. Currently, the only supported numerical computation library is PyTorch.
Future versions will increase the number of loss landscapes that are computable, as well as expand support to TensorFlow, scikit-learn, and more.
The package is available on PyPI. Install using
pip install loss-landscapes.
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size loss_landscapes-0.1.1-py3-none-any.whl (6.4 kB)||File type Wheel||Python version py3||Upload date||Hashes View|
|Filename, size loss_landscapes-0.1.1.tar.gz (6.2 kB)||File type Source||Python version None||Upload date||Hashes View|
Hashes for loss_landscapes-0.1.1-py3-none-any.whl