- deep convolutional generative adversarial network, dcgan, for creating images resembling those taken by the hubble space telescope.
- the trained generator takes in 100 normally distributed numbers in the range [0, 1] and makes tensors of dim 128, 128, 3 which represent an rgb png resembling the hubble sample data imgs
unprocessed kaggle dataset: https://www.kaggle.com/datasets/redwankarimsony/top-100-hubble-telescope-images?resource=download
training sample: opo0010a.png
the results get worse lol most likely because of overfitting and bad hyperparameters
sm1 save pikachu! 😭 epoch 65:
epoch 70:
epoch 79:
epoch 106:
- make sure you have python and jupyter notebook installed
git clone 'https://github.com/hashirkz/hubble_telescope_gan' <local directory to save to>- run all cells in
./usage.ipynb
- the unprocessed kaggle imgs are in compressed .tif format
- the imgs are rgba 4 color channel imgs represented in numpy as m x n x 4 tensors
- be careful reading the imgs into a np.ndarray as some are very large
- recommend to retrain model to around 60 epochs to get organic images
- currently at 110 epochs model is overfitted / hyperparameters may be off a bit because the images look well pretty bad