marcellodebernardi / loss-landscapes Goto Github PK
View Code? Open in Web Editor NEWApproximating neural network loss landscapes in low-dimensional parameter subspaces for PyTorch
License: MIT License
Approximating neural network loss landscapes in low-dimensional parameter subspaces for PyTorch
License: MIT License
Hi,
I think that random noise is gaussian.
rand_u_like(start_point)โ rand_n_like(start_point)
https://github.com/marcellodebernardi/loss-landscapes/blob/master/loss_landscapes/main.py#L133
original code
https://github.com/tomgoldstein/loss-landscape/blob/master/net_plotter.py#L82
Hi,
It there any way to run loss_landscapes.random_plane on GPU? I tried to run it on Google's Colab with GPU but got an error "
RuntimeError: expected backend CUDA and dtype Float but got backend CPU and dtype Float"
Thank you for the beautiful implementation!
For the bias term in filter normalize, normalize is applied to filter wise multiple times in your implementation.ใ
I believe the following implementation is correct. What do you think?
def filter_normalize_(self, ref_point: 'ModelParameters', order=2):
"""
In-place filter-wise normalization of the tensor.
:param ref_point: use this model's filter norms, if given
:param order: norm order, e.g. 2 for L2 norm
:return: none
"""
for l in range(len(self.parameters)):
# normalize one-dimensional bias vectors
if len(self.parameters[l].size()) == 1:
self.parameters[l] *= (ref_point.parameters[l].norm(order) / self.parameters[l].norm(order))
else:
# normalize two-dimensional weight vectors
for f in range(len(self.parameters[l])):
self.parameters[l][f] *= ref_point.filter_norm((l, f), order) / (self.filter_norm((l, f), order))
For this commit:
8d34610
could you please create a new release here? https://pypi.org/project/loss-landscapes/
Thank you!
Hi
The paper "[Visualizing the loss landscape of neural nets]" also uses PCA Directions to plot the optimizer path. Which I think is not implemented here. It would be great if its implemented.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. ๐๐๐
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.