search

microsoft / ADBench

Benchmarking various AD tools.
MIT License
107 stars 40 forks source link

BA and Hand graph X-axis correcting #137

Closed mikhailnikolaev closed 5 years ago

mikhailnikolaev commented 5 years ago

Related to issue #126.

X-axes for BA and Hand graphs are corrected. Now for BA X-axis shows all scalar variable count (11n + 3m + 2p) and for Hand X-axis shows count of correspondencies. This info is extracted from data test file names, so, such files are renamed with adding respected postfixes.

Graph examples

Static

BA Jacobian - Release Graph HAND (Complicated, Small) Jacobian - Release Graph HAND (Simple, Big) Jacobian - Release Graph HAND (Simple, Small) Jacobian - Release Graph HAND (Complicated, Big) Jacobian - Release Graph

Plotly

image image

msdmkats commented 5 years ago

If x-axis for hand shows the number of correspondences, then, maybe we should make it say so? Because the number of independent variables is 26 for simple cases and 26 + 2 * n_correspondences for complicated ones.

mikhailnikolaev commented 5 years ago

If x-axis for hand shows the number of correspondences, then, maybe we should make it say so? Because the number of independent variables is 26 for simple cases and 26 + 2 * n_correspondences for complicated ones.

It looks reasonable. I've done this

tomjaguarpaw commented 5 years ago

I'm not familiar with HAND, but does Manual really not get slower for bigger problem sizes?

mikhailnikolaev commented 5 years ago

I'm not familiar with HAND, but does Manual really not get slower for bigger problem sizes?

It does, but in relation to others tools slowdown doesn't look very significant. I mean, it's very hard to notice deceleration in such a scale. You can look at plots from PR #105, for example. You will see the same situation, but that PR was closed about a month ago, so, there is not possibility that the current PR could break graph plotting for Manual HAND.