search

bytedance / R2Former

Official repository for R2Former: Unified Retrieval and Reranking Transformer for Place Recognition
Apache License 2.0
83 stars 6 forks source link

Accuracy #11

Closed LKELN closed 1 year ago

LKELN commented 1 year ago

Thanks for your work! But I still can't get the same accuracy as you on train reranking , even if the environment and train_reranking.sh parameters I use are the same as you provided, can you tell me why? info.log

Jeff-Zilence commented 1 year ago

Since you are using 2GPUs, the learning needs to be tuned. Also you need to do the finetuning to match our final results. See the second command in train_reranking.sh.

LKELN commented 1 year ago

Even though I'm using two GPUs, all my training parameters are the same as yours, and as you said, it should be around 88.4 without fine-tuning.And When I train the end-to-end model, the training accuracy is as you said. This accuracy fluctuates too much, can you fix this problem?

szhu-bytedance commented 1 year ago

The batch size per GPU will have impact on the performance for all distributed training pipeline. I am not able to debug for you given the limited information.