About Training
#1
by
chenxluo
- opened
Excellent work—the model is impressively accurate. I'm curious about how you trained it. I'd love to fine-tune it further to enhance its performance in specific areas, such as precise recognition of compositional perspectives. Would you be willing to share your training code as a reference?Thank you.
The training and inference code is available at: https://github.com/celll1/tagutl
I recommend that you train from the original model. It has been confirmed that performing additional training on an already fine-tuned model will decrease performance.
The simplest, most reliable, and also the most tedious, way to improve the model's accuracy is to go through the data manually, piece by piece, and correct any missing tags.