I can't reproduce your results.
Your Lora is trained by 300 epoch, not 300 steps?
For 5 images of the dog dataset, the one epoch should run 5 steps. but mine 12 steps. Why?
My train results are not workable. tell me your suggestions.
@spencer21 Hi. The example was trained with 300 steps ( = 50 epoch. )
The high number of steps per epoch is due to the num_repeats parameter from the diffusion-pipe.
The dataset will act like it is duplicated as much as num_repeats.
The default value in the notebook was set to 5, assuming the dataset size is small.
It should be user-control parameter, but I noticed it wasn't displayed in the Colab UI.
So I fixed in #31 to make it visible.
Thanks for bringing this up.
My train results are not workable.
Did the LoRA not work after training with 300 steps?