out-of-domain-library 0.2

A library for out-of-domain tasks.

Out of Domain Library

This library provides tools for feature extraction and similarity checking using various pre-trained models. It can be used to determine whether an image is in-domain or out-of-domain based on cosine similarity with a set of gallery features.

Installation

To use this library, simply clone the repository and install the required packages.

bash

git clone https://github.com/suraj385/out_of_domain_library.git

cd out_of_domain_library

pip install -r requirements.txt

or

pip install out-of-domain-library

Note

The library achieved 99.77% accuracy with a threshold of 0.85 with resnet50

Available Models

The following pre-trained models are available for feature extraction:

resnet18,

resnet34,

resnet50,

resnet101,

resnet152,

vgg16,

vgg19,

inception_v3,

densenet121,

densenet169,

efficientnet_b0,

efficientnet_b7,

Usage

Extract features using ResNet-50 and save to a file

import torch

from torchvision import models, transforms

import pandas as pd

import numpy as np

from out_of_domain.feature_extraction import save_gallery_features

t_image_folder = "test_images" #path to your folder

model_name = "resnet50"

output_file = f"{model_name}.npy"

save_gallery_features(t_image_folder, model_name)

print("Test passed: save_gallery_features function works as expected.")

Check if the test image and folder is in-domain using ResNet-50, create a csv , Evaluate the Model on Multiple Folders and Test Accuracy !

import torch

import numpy as np

import pandas as pd

from out_of_domain.similarity_check import is_in_domain, evaluate_model, eval_image_folder

def test_similarity_check():

test_image_folder = "test_images"

model_name = "resnet50"

threshold = 0.85

output_file = f"{model_name}.npy"



gallery_features = np.load(output_file)


print("Checking if a single image is in-domain...")

test_image_path = "/Users/surajgautam/out_of_domain_library/test_images/image_0.jpg"

result = is_in_domain(test_image_path, gallery_features, model_name, threshold)

print(f"Test image is {'in-domain' if result else 'out-of-domain'}")

# Evaluate the model on multiple folders

test_folder_in_domain = "test_folder_in_domain"

test_folder_out_of_domain = "test_folder_out_of_domain"


test_folders = [
    (test_folder_in_domain, True), #true for indomain
    (test_folder_out_of_domain, False) #false for out-of-domain
]

# Evaluate the model

print("Evaluating the model on multiple folders...")

results = evaluate_model(gallery_features, test_folders, model_name, threshold)

for folder, (correct, total, accuracy) in results.items():
    
    print(f"Folder: {folder}, Correct: {correct}, Total: {total}, Accuracy: {accuracy:.4f}")

# Evaluate the image folder

print("Evaluating the image folder...")

eval_image_folder(test_image_folder, gallery_features, model_name, threshold)

# Load and print the results from the CSV file

results_df = pd.read_csv("image_domain_evaluation.csv")

print("Results from image_domain_evaluation.csv:")

print(results_df)

Run the test function

if name == "main":

test_similarity_check()