How to improve the training accuracy as much as possible?

[tms2003]

I found that the current instructions have multiple different weights, including coco, coco+365, and 365 pre-trained. There is also a default weight (automatically downloaded during training). I want to train on my own dataset, so what training parameters or tuning parameters should I use? Should I start with the 365 pre-trained weights or the coco+365 pre-trained weights?and how many epoch is better?

[Peterande]

Thank you for your question!

• Model Tuning:
  The best approach is to use the pretrained model with the suffix obj365 for tuning.

• Class Mapping:
  To further retain the pretrained capabilities, you can follow the steps in Usage → Custom Dataset → Modify Class Mappings to map the Objects365 categories to your dataset.

  • If there are no duplicate categories, or you don't mind convergence speed differences, you can skip this step.

• Epochs and Hyperparameters:
  You need to test the epochs and hyperparameters based on your dataset, as fine-tuning a pre-trained model on a small dataset may lead to overfitting.

  • My suggestion: Set a relatively large number of epochs and observe the convergence.
  • Adjust the number of epochs for further training based on the results.

• Config Update:
  I have already added configurations for Tuning on Custom Dataset and updated the README with the steps in Usage → Custom Dataset → Tuning on Custom Dataset.

Thank you for your support of D-FINE!