search

YiyanXu / DiFashion

Diffusion Models for Generative Outfit Recommendation
10 stars 0 forks source link

Questions about the training detail #2

Closed zhengrchan closed 3 months ago

zhengrchan commented 3 months ago

Congrats for the acceptance by SIGIR! About training details, how many steps and batch size did your model train on the ifashion/polyvore dataset? Which dataset was the “checkpoint-15000” trained on?

YiyanXu commented 3 months ago

Thanks for your interest in DiFashion! During training, the batch size is set to 10 for both datasets. We optimize DiFashion by selecting optimal checkpoints and hyper-parameters based on the average improvement ratio across four key metrics (FID, CIS, Compatibility, and Personalization) relative to the pre-trained SD-v2. Specifically, for the iFashion dataset, DiFashion undergoes training for 15k steps. For the Polyvore dataset, we finetune the model on top of the model trained on iFashion, which involves an additional 9k training steps. The checkpoint “checkpoint-15000” released here is trained solely on iFashion.