🐉Afrodita-DiT is a text-driven DiT implementation, I hope you like it :b

Train File test with MNIST is TrainDiT2TestMnist.py

Terminal:

python TrainDiT2TestMnist.py  --batch-size 4 --epochs 10

If you have a 24GB graphics card you can train with it with these settings:

python TrainDiT2TestMnist.py  --batch-size 48 --epochs 10

Train File with Flickr30k is TrainDiTFlickr30k.py

Terminal:

python TrainDiTFlickr30k.py  --batch-size 10 --epochs 10

First training test with MNIST

Image generated at epoch 0 and step 2900📸	Real image at epoch 0 and step 2900📸	Image generated at epoch 1 and step 1100📸	Real image at epoch 1 and step 1100📸

Training tests are being done in Colab with the model size "DiT-S/8" and the first checkpoint that was saved has the size of 621MB, then I will do a test with some dataset like MSCOCO or flickr30k

I have based myself on the Meta repo: https://github.com/facebookresearch/DiT.git