Xuenan Xu

由于exposure bias的存在，用lr scheduler不断减小学习率，loss降到3.6左右就不会往下了，否则validation上效果会变差

你有微信之类的联系方式吗？加你详聊

比赛提交的是多个模型的ensemble，每个都比较大，为了方便，这里就放了一个CNN14+BERTbase的ckpt

Hi, do you have access to ViT-L checkpoint? I am also looking for this.

1. The articles are provided. However, since I have reorganized the codes several times, the detailed step-to-step instruction is not included. For example, [run_condition_adverse.py](https://github.com/wsntxxn/AudioCaption/blob/master/captioning/ignite_runners/run_condition_adverse.py) is for [diversity-controllable captioning](https://ieeexplore.ieee.org/document/9746834) but the...

Sorry for the procrastination, the codes and sample configuration files for reinforcement learning and ensemble have been uploaded.

The audio-text contrastive learning code has also been uploaded to https://github.com/wsntxxn/DCASE2022T6_CLAP. You can refer to it.