Awesome

Optimal Transport With Order Constraints

The otoc directory contains the following:

1. otoc: optimal transport with order constraints (first-order method).
2. search_otoc_candidates: obtain set of alternative transport plan candidates given a base transport plan.
   • The base transport plan should be a regular OT transport plan (without optimal constriants).
   • Can use the POT package to compute one, see the code examples below.

The above two techniques are described in the below publication; please consider citing if you find this repository useful.

Citation

Fabian Lim, Laura Wynter, Shiau Hong Lim, "Order Constraints in Optimal Transport", 2022. https://arxiv.org/abs/2110.07275.

Usage

To install the python wheel:

pip install git+https://github.com/IBM/otoc@main#egg=otoc

Code Examples

otoc solves the OT with OC formulation with a first-order method.

from otoc import otoc
import numpy as np

C = np.array([
    [1, 5, 3],
    [0, 1, 4],
])
a = np.array([.5, .5])
b = np.array([1./3, 1./3, 1./3])

# solve the optimal transport with order constraints
# meta give some statistics
X, meta = otoc(
    [(0,2)], # X[0,2] >= all others
    a, b, C, 
)
# X
# array([[0.16658407, 0.        , 0.3333329 ],
#        [0.16674927, 0.3333329 , 0.        ]])

X, meta = otoc(
    [(0,2), (1,0)], # X[1,0] >= X[0,2] >= others
    a, b, C, 
)
# X
# array([[0.        , 0.16669247, 0.333301  ],
#        [0.333301  , 0.16664086, 0.        ]])

Alternatively, calling with implementation='iterator' will return an Iterator object, that iterates once per first-order iteration.

# ... intialize a, b, C as above

# returns an iterator object
algo = otoc(
    [(0,2)], 
    a, b, C, 
    implementation='iterator'
)

# this iterate through the first-order method until stopping
for _ in algo:
    pass

# prints solution and cost
print(algo.solution)
# (array([[0.16658407, 0.        , 0.3333329 ],
#         [0.16674927, 0.3333329 , 0.        ]]),
#  1.4999156499311608)

# prints statistics
print(algo.projection_statistics)
# Statistics(res_primal=8.260397888704896e-05, res_dual=2.4625149137936733e-05, norm_dual=3.570708156906683)

# prints iteration count
print(algo.iteration_cnt)
# 41

otoc_candidates learns multiple candidates given a base transport plan by searching OC's in a tree-search manner. This is an Iterator class that iterates as the tree nodes are searched.

# install the python optimal transport package
# https://pythonot.github.io/
# pip install pot==0.7.0

# --- generate a, b, C as above

# compute the OT base plan using the python optimal
# transport package
import ot # imports POT
base_cost, base_plan = ot.lp.emd2(
    a, b, C,
    return_matrix=True,
)

# search candidates on top of base plan
from otoc import search_otoc_candidates
algo_search = search_otoc_candidates(
    a, b, C, 
    base=base_plan['G'],
    thresholds=(1.,1.), # small example, so just relax thresholds
    k1=10,
    k2=5,
    k3=1,
)

# iterate each node (which is an Iterator object of otoc)
# each node corresponds to an otoc with some learnt order constraints
for node in algo_search:

    # run the first order iterations of otoc
    # until stopping
    for _ in node:
        pass

# third best candidate
algo_search.best_solution(n=2)
# array([[0.16676627, 0.        , 0.33333243],
#        [0.16656706, 0.33333243, 0.        ]])

# history of nodes explored
algo_search._history
# each gives the index, bound_value, and the otoc iterator (algo)
# [History(index=[(0, 2)], bound_value=1.3333333333333333, algo=<otoc.iterative.dual_method.DualMethod object at 0x7fb6242dc9d0>, skip=False),
#  History(index=[(1, 1)], bound_value=1.3333333333333333, algo=<otoc.iterative.dual_method.DualMethod object at 0x7fb6242dc1d0>, skip=False),
#  History(index=[(1, 0)], bound_value=1.3333333333333333, algo=<otoc.iterative.dual_method.DualMethod object at 0x7fb6242dc590>, skip=False),
#  History(index=[(0, 0)], bound_value=1.6666666666666665, algo=<otoc.iterative.dual_method.DualMethod object at 0x7fb6242dce10>, skip=False),
#  History(index=[(0, 1)], bound_value=2.3333333533333334, algo=<otoc.iterative.dual_method.DualMethod object at 0x7fb6242dd210>, skip=True),
#  History(index=[(1, 2)], bound_value=2.0, algo=<otoc.iterative.dual_method.DualMethod object at 0x7fb6242dd490>, skip=True)]

# check which explored node was the second best candidate
algo_search.best_history_index(n=2)
# 1

# confirm
algo_search._history[1].algo.solution
# (array([[0.16676627, 0.        , 0.33333243],
#         [0.16656706, 0.33333243, 0.        ]]),
#  1.5000959909228264)

License

The code in the otoc and tests directories are released under the MIT License. The otoc has about 1000 lines of code (ignoring space and comments) and only one package dependency (numpy).