[WIP] low rank sinkhorn, solve_sample, OTResultLazy + test functions #542
Conversation
|
|
||
| ################################## LR-DYSKTRA ALGORITHM ########################################## | ||
|
|
||
| def LR_Dysktra(eps1, eps2, eps3, p1, p2, alpha, dykstra_p): |
There was a problem hiding this comment.
a bit more documentation with reeference to the paper and to the algorithm number in the paper please
| q3_1, q3_2, v1_, v2_, q1, q2 = dykstra_p | ||
|
|
||
| # POT backend | ||
| eps1, eps2, eps3, p1, p2 = list_to_array(eps1, eps2, eps3, p1, p2) |
There was a problem hiding this comment.
not needs LR_dykstra is used inside POT function and will never be interfaced: i will receive arrays not lists
| eps1, eps2, eps3, p1, p2 = list_to_array(eps1, eps2, eps3, p1, p2) | ||
| q3_1, q3_2, v1_, v2_, q1, q2 = list_to_array(q3_1, q3_2, v1_, v2_, q1, q2) | ||
|
|
||
| nx = get_backend(eps1, eps2, eps3, p1, p2, q3_1, q3_2, v1_, v2_, q1, q2) |
There was a problem hiding this comment.
this should be done only when nx is not given to the function and set to None, else use the new from the calling function (you need to add nx as a parameter to the function)
|
|
||
| ns, nt = X_s.shape[0], X_t.shape[0] | ||
| if a is None: | ||
| a = nx.from_numpy(unif(ns), type_as=X_s) |
There was a problem hiding this comment.
| a = nx.from_numpy(unif(ns), type_as=X_s) | |
| a = unif(ns, type_as=X_s) |
| if a is None: | ||
| a = nx.from_numpy(unif(ns), type_as=X_s) | ||
| if b is None: | ||
| b = nx.from_numpy(unif(nt), type_as=X_s) |
| b = nx.from_numpy(unif(nt), type_as=X_s) | ||
|
|
||
| # Compute cost matrix | ||
| M = dist(X_s,X_t, metric=metric) |
There was a problem hiding this comment.
M should never be computed as an ntimes n marix It shoud be stired as afctorized version (see discussion on the paper with D=AB) and used only as factorized version later (when computing dot ).
There was a problem hiding this comment.
in fcat I'm OK with implementin only squared_euclidena and raise NotImplemented for other metrics until we have afficent function for obtainin low rank factorization of the metrics
| C_t_Q = nx.dot(M.T,Q) | ||
| diag_g = (1/g)[:,None] | ||
|
|
||
| eps1 = nx.exp(-gamma*(nx.dot(CR,diag_g)) - ((gamma*reg)-1)*nx.log(Q)) |
There was a problem hiding this comment.
adda efw more comments with references to equations and alg line nume rin the paper please
| ## Function isn't complete, still work in progress for reg == 0 / reg is None case (and unbalanced cases) | ||
|
|
||
|
|
||
| from .utils import unif, list_to_array, dist, OTResultLazy |
There was a problem hiding this comment.
those should be at the top (move it when nearing merge)
| potentials = None | ||
| lazy_plan = None | ||
|
|
||
| if max_iter is None: |
There was a problem hiding this comment.
move those before calling empirical_sinkhorn function those values are relevant to this function
Hello Rémi,
Here are the modifications we made to the repo:
We also added test functions (and a new test_lowrank.py file).
Don't hesitate to give us feedback.