Skip to content

cc-hpc-itwm/GradVis

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

GradVis

A DNN Training Visualization Toolbox

Original Paper: arxiv.org/abs/1909.12108

Usage

Import using

import Visualization as vis
import nn_model
import trajectory_plots as tplot

Main function is visualize:

vis.visualize(nnmodel,filenames,N,save_as,random_dir=False,proz=0.5,v_vec=[],w_vec=[],verbose=False)

Inputs:

- nnmodel: instance of nn_model class
- filename: array of filename of parameters one wishes to plot
- N: number of values in x,y direction. Defines the resolution of the loss landscape
- save_as: filename of file to be saved
- random_dir: If one wants to only plot the loss surface at arbitrary directions around one point use `True`
- proz: Plot additional fraction of the plot (have border around path when plotting)
- v_vec,w_vec: For plots in direction of those vectors

Outputs:

- Numpy .npz file which stores the following data:

If random_dir=False and no vectors specified:
    (X,Y,Z),(Paths_x,y,z),(variation of PCA diections, v vector, w vector)
If random_dir=False and v_vec,w_vec specified:
    (X,Y,Z),(Paths_x,y,z)
If random_dir=True:
    (X,Y,Z),(v vector, w vector)

The nn_model instance is created as follows:

nn_model.PyTorch_NNModel(model,train,converged_params)

Inputs:

- model: PyTorch neural network model
- train: function returning loss value of model (basically training function without doing backpropagation or updating the gradients)
- converged_params: filename of the converged parameters

For plotting there are three functions:

2D contour plots using plot_loss_2D:

tplot.plot_loss_2D(filename,save_as,is_log=False)

Plot contour of loss function

Inputs:

- filename: path to file that was created using `visualize`
- save_as: filename of plot to be saved
- is_log: plot loss landscape in log plot

3D contour plots using plot_loss_3D:

tplot.plot_loss_3D(filename,save_as,is_log=False)

Create 3D plot of loss landscape with optional path

Inputs:

- filename: path to file that was created using `visualize`
- save_as: filename of plot to be saved
- is_log: plot loss landscape in log plot

Interactive 3D plots using plot3D:

tplot.plot3D(X, Y, Z, height=600, xlabel = "X", ylabel = "Y", zlabel = "Z", initialCamera = None)

Make an interactive 3D plot (mainly helpful for jupyterlab)

Inputs: X,Y,Z values of loss landscape

Examples

Two examples are provided showing different use-cases:

  • Multiple_minima.py: Allows multiple minima plotting.

  • Resnet_vis.py: Plots loss landscape of a Resnet20 in 2D and 3D.

Roadmap

  • Distributed computation on multiple nodes

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •  

Languages