search

m-bain / frozen-in-time

Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval [ICCV'21]
https://arxiv.org/abs/2104.00650
MIT License
342 stars 43 forks source link

How to select checkpoint when fine-tuning on MSRVTT #49

Closed chaochen99 closed 2 years ago

chaochen99 commented 2 years ago

Dear Sir,

During fine-tuning of MSRVTT, we found no validation set in the code. How to choose a checkpoint?

Looking forward to your reply.

Thanks！

m-bain commented 2 years ago

Hi, correct, unfortunately for msr-vtt people use validation set and test set the same. You can just train for 50 epochs or so which is when train/val converges I believe. Best SOTA numbers by other works likely use val set performance to pick checkpoint.