Implementations and examples of use cases of loss functions used in contrastive representation learning.
Implemented in TensorFlow, orginally for use in genetic data, here shown for other applications as well.
The required packages are all included in the Apptainer definition file image.def.
Build the Apptainer image:
$ sudo apptainer build image.sif image.def
Run the image, the --nv flag exposes the NVIDIA GPU to the container:
$ apptainer run --nv image.sif
As an alternative, the file requirements.txt contains the needed packages. Creating a python venv, and executing
pip install -r requirements.txt
will install the packages needed to run the code.
The program run_CL.py is called for all implemented examples, at the time of writing MNIST and Cifar10.
To run, the user need to state whether we want to train a model anew or to plot already saved model states, which dataset to run, and where to save the logs.
For example, to train a model on the MNIST dataset, execute the following:
$ python3 -u run_CL.py train --data=mnist --dir=./test_mnist
There is also an optional argument to restart training from a previously saved state, by appending the --load_path=/path_to_saved_model argument.
To plot results for saved model states in a directory and evaluate the KNN-classification accuracy, run
$ python3 -u run_CL.py plot --data=mnist --dir=./test_mnist
Currently, this also runs PCA and t-SNE to compare with. Note that t-SNE may be relatively slow to run.
Within the singularity container running the example of contrastive learning on the MNIST dataset as described above for approximately 400 epochs, generates the following output: 
A similar dataset in size, but with clothing items instead is the Fashion-MNIST dataset.
$ python3 -u run_CL.py train --data=fashion_mnist --dir=./test_fashion_mnist
The code yields the following results:
