YijinHuang / SSiT

SSiT: Saliency-guided Self-supervised Image Transformer for Diabetic Retinopathy Grading
22 stars 4 forks source link

save-interval- checkpoints #10

Closed Wadha-Almattar closed 10 months ago

Wadha-Almattar commented 11 months ago

Hi, On of the parameters at the model training is save-interval where its default value is 20. So, during the training a checkpoint will be saved after each 20 epochs. In the downstream task with fine-tuning, which checkpoint should I utilized in this case. Now, I fine-tune with the last checkpoint which has 300 number. Does it make difference in SSL if using the last one to fine-tune or should I try them all.

YijinHuang commented 10 months ago

All experiments in the paper were trained using the last checkpoint. While I didn’t test other checkpoints, previous works typically utilize the final checkpoint for downstream tasks.