okyksl / com-516-project

COM-516 Mini Project

We provide two seperate implementations. One implementation over Jupyter Notebook which can be found under src/mcmc_stub.ipynb, and another full Python implementation is accessible through command-line via runner.py. Note that beta scheduling described in section 3.2.2. is only accessible through the notebook implementation. Equivalence of two implementation is verified by running over the the same random seeds.

Below, explains the command-line usage for Python source including experimentation support and parameters available. For evaluating our solution quickly and cleanly or running extensive experiments with multiple parameters Python source might be a better option. For interactive purposes (editing the code, visualizing results), notebook implementation might be a better option.

Installation

Install dependencies listed in requirements.txt.

pip install -r requirements.txt

Running

Simply run the solver with:

python runner.py

You can use predefined datasets g1 and g2 and even enter customize number of cities n and seed (for reproducibility) as follows:

python runner.py dataset=g1 dataset.n=10 dataset.seed=5

To run the solver over various lambdas:

python runner.py --multirun lmbd=0,0.25,0.5,0.75,1.0

The default options are as follows. Use Naive solver and g2 dataset with different lambda as follows:

dataset: g1
solver: mcmc
lmbd: 0.1

python runner.py solver=naive dataset=g2 lmbd=0.7

MCMC Options

MCMC solver provides the following options with given default values:

beta: 1 # initial beta value
step: 2000 # number of steps
start: empty # 'empty set' or 'binomial'
seed: null # random seed for reproducibility
scheduler: # beta scheduling
  checkpoints: [500, 1000, 1500] # steps to change beta values
  betas: [5, 25, 125] # beta values used after checkpoints
use_best: true # if True, uses the best value over all steps instead of final state
num_trials: 10 # if > 1, runs the chain with different seeds and reports best results found
visualize: False # visualizes the run of chain (only the last trial)

Here is a more complicated example running MCMC solver over 5 instances of G2 dataset with various lambda values:

python runner.py --multirun lmbd=0.6,0.8,1.0,1.2,1.4 dataset=g2 dataset.seed=0,1,2,3,4 solver=mcmc solver.num_trials=5 solver.use_best=true solver.seed=0

Logging

Results our automatically stored under the outputs folder. Depending on the task, hydra will store the results under either outputs/run or outputs/multirun folders. Each run is encapsuled with another file named after current date and time (e.g. 2020-12-14_15-53-02) which contains result logs and experiment parameters.

Learn more about using hydra from here.

Contributing

1. Create a new Solver class under src/solvers.py by implementing _solve function.
2. Define a new hydra solver configuration under conf/solver similar to conf/solver/naive.yaml
3. Run hydra with the new solver solver=newsolver

com-516-project's Issues

Calculate Most Distant Points via Convex Hull