4 ms·

On the topic of exploring the innards of machine learning models, here are some visualizations I created of the hidden internal representation layers within the

• ShamelessC 3y ago
  All of your posts are fantastic. I’m amazed you have the patience for the data cleanup required for the plants finetune. I remember wanting to do a Pokemon finetune (before such things were commonplace) and the dataset cleanup was the worst part. Stability issues and lack of decent multi-GPU setup being a close second.
  • nuclearsugar 3y ago
    Thanks! Indeed the cleanup required for the 'Nature Artificial' dataset was intense and forced me to consider other techniques. It was around that time that I realized that I could use Stable Diffusion to output 10,000 to 50,000 images, which has proved to be a useful dataset creation tool for training StyleGAN2 or StyleGAN3. Yet nailing down a SD text prompt that will continuously output images in a specific style and context can be challenging. StyleGAN2/3 is quite sensitive to fine-tuning models with a dataset that contains lots of variation, such as the Pokemon dataset. But it will converge much better with a dataset of just Pikachu images in tons of different poses. It seemingly wants to find the most common pattern in the dataset and interpolate within that space. From there nailing down the ideal value for the gamma attribute is difficult, but luckily can be reduced every 1000 to 2000kimg until the model no longer trains favorably. Currently my typical gamma strategy is: 80 > 10 > 5 > 2
• sdwr 3y ago
  Incredible! You should make music videos