Autoencoders are neural networks designed to project input data onto a low-dimensional space (the latent space) and then reconstruct it as accurately as possible. This package provides an implementation of an autoencoder with a spherical latent space. This constraint makes it possible to navigate and visualize images projected onto the autoencoder's latent space using the Spheriscope app.
To install the package from the git repository, run:
pip install git+https://github.com/MalloryWittwer/spheriscope.git
To train a model, you need a dataset of images. You can use a subset of the MNIST dataset as an example.
Ensure that:
- The images can be resized to 64x64 pixels without losing significant visual quality. The autoencoder operates at this image size. If the images are not already this size, they will be resized internally by the autoencoder.
- The images are grayscale with pixel values between 0 and 255. If the images are in RGB, they will be converted to grayscale.
Save your images in PNG, JPG, or TIF format in a dataset folder. For example:
images/
├── img1.png
├── img2.png
├── ...
Then, run the training script:
spheriscope train <images_dir> <model_output_dir> --epochs 1000 --batch_size 32This will save a model file encoder.keras in the specified output directory.
To project an image onto the autoencoder's latent space, use the script:
spheriscope predict <image_file> <model_path>This will print the polar coordinates of the image in the latent space, for example: encoded_image={'theta': 2.096, 'phi': 0.549}.
To visualize a dataset of images on the autoencoder's latent space, install and run the Spheriscope app.
Then, upload the images to the backend server using:
spheriscope upload <images_dir> <model_output_dir>After the command completes, reload the page at http://localhost:3000. You should see the projected image dataset 🎉.