search

navervision / lincir

Official Pytorch implementation of LinCIR: Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024)
Other
100 stars 5 forks source link

About ViT-G backbone pretrained model #15

Closed lan-lw closed 4 months ago

lan-lw commented 5 months ago

Hi, I tried to pretrained a phi model with ViT-G backbone, but the results are not as good. Can you provide a pretrained model with ViT-G as backbone?

geonm commented 5 months ago

Try reducing the validation steps to 50 or 100 and retrain the model.

Based on experience, in the case of Giga, we need to consider various methods to avoid overfitting.

However, proposing such methods is beyond the scope of this study, so we used the same training parameters for all training.

Using Giga as the backbone and training a phi model with LinCIR can lead to very quick convergence.

lan-lw commented 5 months ago

Try reducing the validation steps to 50 or 100 and retrain the model.

Based on experience, in the case of Giga, we need to consider various methods to avoid overfitting.

However, proposing such methods is beyond the scope of this study, so we used the same training parameters for all training.

Using Giga as the backbone and training a phi model with LinCIR can lead to very quick convergence.

Got it, thanks!