Liangyu Chen

Demo of I2T Retrival

1

Hi! Thank you for the great demo in the Colab notebook and hugging face. I am still confused about how to evaluate Image-to-Text retrieval in a quick manner. Do you...

Slow Downloading Speed

6

Hi! I am in Singapore. The download speed of models and datasets is less than 100KB/s, from aliyun servers I think. It is too slow for downloading large datasets and...

Negative LR

1

Hi! Thank you for your great work. When I was distilling with my own dataset, there was very large loss (iter = 0490) and negative learning rate. Could you help...

Hi developers, thank you for completing this wonderful re-implementation. As I am checking the differences between this repo and the original one, I noticed that the original repo also implemented...

[FEATURE]: Data Toolkit

4

### Describe the feature I appreciate your great work of releasing [llama 2 model](https://github.com/hpcaitech/ColossalAI/tree/785802e809ccf26b3864ae811dc908ecdf601a70/applications/Colossal-LLaMA-2). When will Data Processing Toolkit be released?

Does ChatGPT work better than text003

2

Great work! Seems you are using `gpt-3.5-turbo` by default. imo, to use ChatGPT can be more tricky to engineer, and might be empirically worse than text003. My implementation is like...

There are various clustering methods from K-means to K-medoids. Is there any rationale of applying K-medoids algo? Thank you!

I have sufficient memory while not many CPU cores on my server, therefore IO can be the bottleneck of the training trials. I noticed that you have set `pin_memory=False` in...

Load checkpoint from default_local_dir & Save hdfs checkpoints

2

I think there might be a bug, if the resume_mode is auto and default_local_dir is set and default_hdfs is also set. should load checkpoint_folder from self.config.trainer.default_local_dir, we simply dont rusume...