TinyLLaVA / TinyLLaVA_Factory

A Framework of Small-scale Large Multimodal Models
https://arxiv.org/abs/2402.14289
Apache License 2.0
661 stars 69 forks source link

损失函数的值不下降,一直维持在3到4 #118

Open MichealZhangxa opened 2 months ago

MichealZhangxa commented 2 months ago

你好,我在你的代码的基础上训练,前期的损失一直在3到4之间正常吗,训练的也是llava的预训练数据集

shiym2000 commented 1 month ago

您好,请问可以看一下您的实验脚本/配置吗?

shiym2000 commented 1 month ago

Please provide your training script and any modified code (if applicable) so that we can better identify the issues in the code.

shiym2000 commented 1 month ago

You might try using the two encoders separately to train the whole model, ensuring that each individual encoder can properly train the whole model.

shiym2000 commented 1 month ago

And you can adjust the model training length by modifying the num_train_epochs parameter.