Dian

Thanks for the quick response and clarification. For the dataset preparation, I probably had misunderstood what the paper said.  And, how about the training config/code?  > As described...

> Thanks for your reply. It seems that the difference between Chinese and English BERT did cause this problem. I lowered the learning rate and ITM and LM are currently...

> You can modify this config file to pre-train with ViT-L: https://github.com/salesforce/LAVIS/blob/main/lavis/projects/blip2/train/pretrain_stage1.yaml thanks for the insight. also, I downloaded Laion-multi dataset to do my own training, and find the epoch...

> You can modify this config file to pre-train with ViT-L: https://github.com/salesforce/LAVIS/blob/main/lavis/projects/blip2/train/pretrain_stage1.yaml Also, do you know why [pretrain_stage1.yaml](https://github.com/salesforce/LAVIS/blob/main/lavis/projects/blip2/train/pretrain_stage1.yaml) has no validation dataset? only training dataset is used... In this case,...

> @rayvzn119 standford的vicuna主要是全量微调+不开8bit+全长度（2048），他们之前的那个效果一般，不过在最近的V1.1版本效果挺好的，基底是13B模型。我们主要在7B上的模型+lora+8bit上进行训练，由于资源所限，我们目前的目标还是在小资源下如何提高中文能力。目前中文能力确实是不如他们的13B的V1.1版本。 > @fireice009 可以参考这个[issue](https://github.com/Facico/Chinese-Vicuna/issues/48) 期待基于13B做一个效果更好的~

> can you please share more detail here? @chenyzh28 by Chinese bert, do you mean that you change the vocab or something else? what the new learning rate did you...

> Thanks for your awesome work in InstructBLIP. When I want to reproduce the result in Figure 5 in your paper, the result is not ideal. > > ``` >...

> 是的，本项目和minigpt-4都是基于BLIP2的思路，结构类似。具体可以看model文件夹下的代码。 看层数配置起来就是EVA-CLIP。。。暴大的模型 

> I am currently using this tool on a website that uses AWS CloudFront to host all their images. However, if you do too many queries to their urls, you...

Also, the perf looks extrememly slow (considerably slower than local disk)... Is this expected?