Build personalized machine learning models for Tinder based on your historical preference using Python.
There are three parts to this:
- A function to build a database which records everything about the profiles you've liked and disliked.
- A function to train a model to your database.
- A function to use the trained model to automatically like and dislike new profiles.
The last layer of a CNN trained for facial classification can be used as a feature set which describes an individual's face. It just so happens that this feature set is related to facial attractiveness.
tindetheus let's you build a database based on the profiles that you like and dislike. You can then train a classification model to your database. The model training first uses a MTCNN to detect and box the faces in your database. Then a facenet model is run on the faces to extract the embeddings (last layer of the CNN). A logistic regression model is then fit to the embeddings. The logistic regression model is saved, and this processes is repeated in automation to automatically like and dislike profiles based on your historical preference.
This blog post has a short description of how tindetheus works.
For a more detailed description of how and why this works see [1] https://arxiv.org/abs/1803.04347
tindetheus browsebuild a database by liking and disliking profiles on Tinder. The database contains all the profile information as a numpy array, while the profile images are saved in a different folder.
tindetheus browse --distance=20by default tindetheus starts with a 5 mile radius, but you can specify a search distance by specifying --distance. The above example is to start with a 20 mile search radius. It is important to note that when you run out of nearby users, tindethesus will ask you if you'd like to increase the search distance by 5 miles.
tindetheus trainUse machine learning to build a personalized model of who you like and dislike based on your database. The more profiles you've browsed, the better your model will be.
tindetheus likeUse your personalized model to automatically like and dislike profiles. The profiles which you have automatically liked and disliked are stored in al_database. By default this will start with a 5 mile search radius, which increases by 5 miles until you've used 100 likes. You can change the default search radius by using
tindetheus like --distance=20which would start with a 20 mile search radius.
If you use Windows you may want to read this guide on how to install tindetheus on Windows.
- Install pynder from source (pynder on pip has not been updated)
git clone https://github.com/charliewolf/pynder.git
[sudo] pip install ./pynder- Install tindetheus
[sudo] pip install tindetheus- After you have installed tindetheus. Create a new folder that will be your Tinder database.
mkdir my_tinder_data
cd my_tinder_data- You need your facebook auth token. There are many discussions on this on the internet to find this. You can find your facebook auth token by using a man in the middle (MIM) attack to sniff out the requests. You are looking for access_token=. The MIM attack can be conducted by creating a proxy with ssl certificate. If you are still lost, perhaps check out this or this.
- Create a config.txt file that contains the following two lines exactly
facebook_token = YYYY
model_dir = 20170512-110547
where YYYY is replaced with your facebook token in order to login using pynder.
-
Download a pretrained facenet model. I recommend using this model 20170512-110547 mirror. You must download 20170512-110547.zip and extract the contents in your my_tinder_data folder. The contents will be a folder named 20170512-110547. You should specify the pretrained model that you use in the second line of the config.txt tile. You can use other pretrained facenet models as long as you include the model directory in your folder and change the config.txt accordingly.
-
You need to initialize git in your my_tinder_data folder which is used to track revision history. Run the following commands to initialize git.
git init
git add .
git commit -m "first commit"- Start building your database. Manually reviewing 20-40 profiles will be a good starting point, but you can do it with less. Before you start training a model you have to be sure that you've liked and disliked at leach one profile.
tindetheus browse- After browsing profiles you can train your personalized classification model at any time. Just run
tindetheus trainto build your personalized model. With more profiles you can build a more accurate model, so feel free to browse more profiles at any time and build to your database. Newly browsed profiles aren't automatically added to the model, so you must manually run tindetheus train to update your model.
- You can automatically like and dislike profiles based on your trained model. To do this simply run
tindetheus likewhich will use your latest trained model to automatically like and dislike profiles. The application will start with a 5 mile search radius, and automatically like and dislike the people in this radius. After running out of people, the search radius is increased by 5 miles and the processes repeats. This goes on until you've used 100 likes, at which point the application stops.
-
This is all in the early stages, so after each session I highly recommend you backup your my_tinder_data folder by creating an archive of the folder.
-
If you want to manually browse your database, check out this example file.
You can now store all default optional parameters in the config.txt! This means you can set your starting distance, number of likes, and image_batch size without manually specifying the options each time. This is an example config.txt file:
facebook_token = XXXXXXX # your facebook token hash
# alternatively you can use the XAuthToken
XAuthToken = xxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxx
model_dir = 20170512-110547 # the location of your facenet model directory
# see https://github.com/davidsandberg/facenet#pre-trained-models for other
# pretrained facenet models
image_batch = 1000 # number of images to load in a batch during train
# the larger the image_batch size, the faster the training process, at the
# cost of additional memory. A 4GB machine may struggle with 1000 images.
distance = 5 # Set the starting distance in miles
likes = 100 # set the number of likes you want to use
# note that free Tinder users only get 100 likes in 24 hours
As of Version 0.4.0, tindetheus now includes a validate function. This validate functions applies your personally trained tinder model on an external set of images. If there is a face in the image, the model will predict whether you will like or dislike this face. The results are saved in validation.csv.
First you'll need to get a validation data set. I've created a small subset of the hot or not database for testing purposes. You can download the validation.zip here which is a a subset of the female images in [2], and extract it to your tinder database directory.
Then execute
tindetheus validate
to run the pretrained tindetheus model on your validation image set. You could run the tindetheus trained model on the entire hot or not database to give you an idea of how your model reacts in the wild. Note that validate will attempt to rate each face in your image database, while tindetheus only considers the images with just one face.
The validate function only looks at images within folders in the validation folder. All images directly within the validation folder will be ignored. The following directory structure considers the images in the validation/females and validation/movie_stars directories.
my_tinder_project
│ config.txt
| validation.csv
│
└───validation
| | this_image_ignored.jpg
│ │
│ └───females
│ │ │ image00.jpg
│ │ │ image01.jpg
│ │ │ ...
│ └───movie_stars
│ │ image00.jpg
│ │ image01.jpg
│ │ ...
- 2019/05/05 Version 0.4.3. Add option to log in using XAuthToken thanks to charlesduponpon. Add like_folder command line option to create al/like and al/dislike folders based on the historically liked and disliked profiles. Allows quick access to asses model quality.
- 2019/04/29 Version 0.4.1. Fix issue where line endings that were causing authentication failure. Fix handling of config.txt.
- 2018/12/02 Version 0.4.0. New validate function to apply your tindetheus model to a new dataset. See README on how to use this function. Fix issues with lossy integer conversions. Some other small bug fixes.
- 2018/11/25 Version 0.3.3. Update how facenet TensorFlow model is based into object. Fixes session recursion limit.
- 2018/11/04 Version 0.3.1. Fix bug related to Windows and calc_avg_emb(), which wouldn't find the unique classes. Version 0.3.2, tindetheus will now exit gracefully if you have used all of your free likes while running tindetheus like.
- 2018/11/03 Version 0.3.0. Major refresh. Bug fix related to calling a tindetheus.export_embeddings function. Added version tracking and parser with --version. New optional parameters: likes (set how many likes you have remaining default=100), and image_batch (set the number of images to load into facenet when training default=1000). Now all optional settings can be saved in config.txt. Saving the same filename in your database no longer bombs out on Windows. Code should now follow pep8.
- 2018/05/11 Added support for latest facenet models. The different facenet models don't appear to really impact the accuracy according to this post. You can now specify which facenet model to use in the config.txt file. Updated facenet clone implementation. Now requires minimum tensorflow version of 1.7.0. Added example script for inspecting your database manually.
tindetheus uses the following open source libraries:
Tindetheus is a combination of Tinder (the popular online dating application) and the Greek Titans: Prometheus and Epimetheus. Prometheus signifies "forethought," while his brother Epimetheus denotes "afterthought". In synergy they serve to improve your Tinder experience.
Epimetheus creates a database from all of the profiles you review on Tinder.
Prometheus learns from your historical preferences to automatically like new Tinder profiles.
[1] Jekel, C. F., & Haftka, R. T. (2018). Classifying Online Dating Profiles on Tinder using FaceNet Facial Embeddings. arXiv preprint arXiv:1803.04347.
[2] Donahue, J., & Grauman, K. (2011). Annotator rationales for visual recognition. http://vision.cs.utexas.edu/projects/rationales/