Technical details:

• Base model: google/vit-base-patch16-224
• Dataset: [Recycling Classification (12 classes)] (https://www.kaggle.com/datasets/mostafaabla/garbage-classification)
• Learning rate: 0.0001
• Effective training batch size: 16 (2 items per TPU core over 8 cores.)

Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC)