[MRG] Laplace regularized OTDA #140
Conversation
examples/plot_otda_laplacian.py
Outdated
| # -*- coding: utf-8 -*- | ||
| """ | ||
| ======================== | ||
| OT for domain adaptation |
There was a problem hiding this comment.
add proper titke since it sill appear in the documentation
| References | ||
| ---------- | ||
|
|
||
| .. [5] N. Courty; R. Flamary; D. Tuia; A. Rakotomamonjy, |
There was a problem hiding this comment.
please add the paper that provide the expression and gradients here and in the readme.
You should also update the readme at the line:
Optimal transport for domain adaptation with group lasso regularization [5]
to add Laplacian regularization and also add yourself as a contributor ;)
There was a problem hiding this comment.
added Laplacian regularization to readme and workshop citation to da.py in both emd_laplace and EMDLaplace and to readme
There was a problem hiding this comment.
and myself to contributors
ot/da.py
Outdated
| + (1 - alpha) * np.trace(np.dot(xs.T, np.dot(G, np.dot(lT, np.dot(G.T, xs))))) | ||
|
|
||
| def df(G): | ||
| return alpha * np.dot(lS + lS.T, np.dot(G, np.dot(xt, xt.T)))\ |
There was a problem hiding this comment.
pre compute (lS + lS.T) and np.dot(xt, xt.T) (lT + lT.T)
| lS = laplacian(sS) | ||
| lT = laplacian(sT) | ||
|
|
||
| def f(G): |
There was a problem hiding this comment.
could you please implement both position regularization and displacement regularization?
It only change the M matrix and provides more generality to teh function?
See equation (11) vs (9)
https://remi.flamary.com/biblio/flamary2014optlaplace.pdf
ot/da.py
Outdated
| in NIPS Workshop on Optimal Transport and Machine Learning OTML, 2014. | ||
| """ | ||
|
|
||
| def __init__(self, reg_type='pos', reg_lap=1., reg_src=1., alpha=0.5, |
There was a problem hiding this comment.
how can you pass nn or sim to the emd_laplace function?
this is necessary, you should use a kwargs and store it for the fit
There was a problem hiding this comment.
added dict kwargs in fit that is passed to emd_laplace
| xs : np.ndarray (ns,d) | ||
| samples in the source domain | ||
| xt : np.ndarray (nt,d) | ||
| samples in the target domain |
There was a problem hiding this comment.
sim not documented. You should add it to the list and define also the parameters for each sim : 'nn' and the gaussian kernel width.
There was a problem hiding this comment.
documented sim and added a common parameter 'similarity_param'
ot/da.py
Outdated
| if sim == 'gauss': | ||
| if 'rbfparam' not in kwargs: | ||
| kwargs['rbfparam'] = 1 / (2 * (np.mean(dist(xs, xs, 'sqeuclidean')) ** 2)) | ||
| sS = kernel(xs, xs, method=kwargs['sim'], sigma=kwargs['rbfparam']) |
There was a problem hiding this comment.
do not use kwargs['sim'], use 'gauss'
There was a problem hiding this comment.
I left it in case one day another kernel wille be added to kernel but can fix it
rflamary
changed the title
Uh oh!
There was an error while loading. Please reload this page.