Training Hyperparameters of t2i model training

[leileqiTHU]

Hi there! Thanks for your great work on the T2I model. I'm currently studying its training process and have some questions about the Stage 1 training (on LAION-50M dataset). Could you kindly share the related details? It would be a huge help for my learning/reproduction work.

My questions are as follows:

1. How many epochs were used for Stage 1 training on LAION-50M?
2. Was multinode training adopted for Stage 1?
   • If yes: How many nodes were used, and what was the total training time?
3. Could you also list other key training hyperparameters for Stage 1 (e.g., batch size per GPU, learning rate, optimizer type, image resolution, etc.)?

Thanks again for your time and help!

[zwxu064]

@leileqiTHU Hi, Have you figured out what the .jsonl file is in autoregressive/train/extract_codes_t2i.py? Thanks.