search_experiment.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Tue May 10 17:28:03 2022

@author: cyrilvallez
"""


import numpy as np

from helpers import utils
from fast_search import search

# Force the use of a user input at run-time to specify the path 
# so that we do not mistakenly reuse the path from previous experiments
save_folder = utils.parse_input()

# algorithm = 'SimCLR v2 ResNet50 2x'
# algorithm = 'Dhash 64 bits'
algorithm = 'SimCLR finetuned'
main_dataset = 'BSDS500_original'
# main_dataset = 'Kaggle_templates'
distractor_dataset = 'Flickr500K'
query_dataset = 'BSDS500_attacks'
# query_dataset = 'Kaggle_memes'

# factory_str = 'Flat'
metrics = ['L2', 'L1', 'cosine']


nlist = int(10*np.sqrt(500000))
factory_str = f'IVF{nlist},Flat'
# nprobes = [1, 5, 10, 20, 30, 40, 50, 60, 70, 80]
nprobes = [1, 5, 10, 20, 50, 100, 200]

ks = [1, 2, 3, 4, 5, 10, 20]

filename = save_folder + 'results.json'
search.compare_k_Flat(ks, algorithm, main_dataset, query_dataset,
                        distractor_dataset, filename)
# search.compare_nprobe_IVF(nlist, nprobes, algorithm, main_dataset, query_dataset,
                        # distractor_dataset, filename, k=10)