Closed maerduduqi closed 6 days ago
Hi, If you are training the model from scratch on AliMeeting (meaning that you are not fine-tuning a model previously trained on simulated conversations), then that behavior is expected. A randomly initialized model struggles to learn with such a small training set. You should train on simulated conversations or start from this shared model https://github.com/BUTSpeechFIT/DiaPer/tree/main/models/10attractors/SC_LibriSpeech_2spk_adapted1-10
I use AliMeeting data set, after modifying the corresponding path, train with train _ 2spraker configuration, train 100 rounds, starting from 20 rounds,DER bit is 100发自我的手机-------- 原始邮件 --------发件人: fnlandini @.>日期: 2024年7月25日周四 半夜11:17收件人: BUTSpeechFIT/DiaPer @.>抄送: maerduduqi @.>, Author @.>主 题: Re: [BUTSpeechFIT/DiaPer] Training indicators are abnormal (Issue #4) Hi, If you are training the model from scratch on AliMeeting (meaning that you are not fine-tuning a model previously trained on simulated conversations), then that behavior is expected. A randomly initialized model struggles to learn with such a small training set. You should train on simulated conversations or start from this shared model https://github.com/BUTSpeechFIT/DiaPer/tree/main/models/10attractors/SC_LibriSpeech_2spk_adapted1-10
—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: @.***>
You should do the fine-tuning on AliMeeting from a model already adapted to variable number of speakers. I have not tried fine-tuning from the model trained only with 2-speaker data because most likely it will not work well.
Closing due to inactivity
When training with the AliMeeting dataset, train_DER is 100