
#Computational infrastructure used to run the experiments 

The experiments ran on an Ubuntu 22.04.3 LTS machine with 3.16TB hard disk and an
NVIDIA GeForce RTX 2080 Ti GPU with 11264 MiB RAM. Our source code was implemented in Python 3.9.

#Requirements
- Anaconda GPLv2 (Free software)
- Python 3.9
- scallopy 0.2.2: https://github.com/scallop-lang/scallop (MIT License)
- highspy 1.5.3: https://pypi.org/project/highspy/ (MIT License)
- OR-Tools: https://developers.google.com/optimization/install/python (Apache-2.0 license)
- PySDD 0.2.11: https://pypi.org/project/PySDD/ (Apache-2.0 license)
- networkx 3.2.1: https://pypi.org/project/networkx/ (BSD License)
- tqdm 4.66
- numpy 1.26.2 
- PyTorch 2.1.1 
- PyTorch Vision 0.16.1

File requirements.txt has all the requirements to create our virtual environment via conda.  

#More
- We used code from https://github.com/MediaBrain-SJTU/RECORDS-LTPLL to implemenet RECORDS (MIT License).
- We used code from https://github.com/hbzju/SoLar to implemenet the sliding-window based technique for estimating the priors.

#Datasets
 -MNIST (Creative Commons Attribution-Share Alike 3.0)
 -CIFAR-10: https://www.cs.toronto.edu/~kriz/cifar.html
 -HWF: https://liqing.io/NGS/ 

The first two datasets are available via PyTorch, so the user does not need to download them in advance.  
The last has to be downloaded by the users.

#Folder structure 
- train_scallop.py: Python code to run SL (with/without RECORDS), LP(EMP), LP(ALG1) for MAX-M, SUM-M, HWF-M
- train_smallest_parent.py: script to run SL (with/without RECORDS), LP(EMP), and LP(ALG1) for Smallest Parent
- testing_time: Python code to run LA and CAROT for an input model. Supports 
 -- imbalance-free training and testing, as well as, 
 -- long-tail training and long-tail testing.
- scripts: example linux shell scripts for running our experiments using SL, SL+RECORDS, LP(EMP), LP(ALG1).

#Training arguments
- ```--size_partial_dataset```: number of partial training samples.
- ```--M```: number of instances per training sample, as in our paper.   
- ```--records```: if it is on, then SL runs jointly with RECORDS (see the experiments for further details). 
- ```--top-k```: number of top-k proofs to consider during training (for efficiency reasons, all our experiments on MAX-M, SUM-M, HWF-M ran with k=1).
- ```--ilp-training```: if it is on, then training is carried under LP.
- ```--epsilon_ilp```: the epsilon approximation in Eq. (6).
- ```--ilp_solver```: solver for linear programs. At the moment, pywraplp is supported from https://developers.google.com/optimization/install/python.
- ```--continuous-relaxation```: if set to true, then continuous relaxations are adopted to solve Eq. (6).
- ```--gamma```, ```--est_epochs```, and ```--estimated-label-ratio``` are defined as in SoLar for empirical distribution estimation.
- ```--estimated-label-ratio```: techique to compute the partial ratios. ```gold``` is for the ratios computed using the ground-truth hidden gold labels. ```partial``` is for using Algorithm 1 to compute the ratios, 
```naive``` is for using SoLar's sliding window-based approximate technique.   

Instructions to run the Python code under folder "testing_time" are given inside the files. 

