Update
Outcome
- Completed essential transfer from Minedojo to MineRL.
- Add more wrappers including operation logic of interaction for Minecraft Client, task setting…
- still use Minedojo’s defined tasks, task categories, task allocation
- Add helper functions for env initialisation: inventory, block, player’s slot(item, health, food), partial settings are based on different tasks, but now we could only set fixed env settings
- Delete uncessary/repeated functions and make the project volume smaller.
- Exitsintg problem； lidar aprt(Minedojo). Needs to overwrite Malmo (lowest level: build.gradle)
TODO with problems
- Implement transfer learning plan：
- existing problem: Train existing model with LoRA, using peft?
openai/clip-vit-base-patch16https://radekosmulski.com/how-to-fine-tune-a-tranformer-pt-2/ check: https://huggingface.co/openai/clip-vit-base-patch16 - Prepare additional dataset? if so, should I use additional training dataset？
- existing problem: Train existing model with LoRA, using peft?
About this Post
This post is written by Rui Xu, licensed under CC BY-NC 4.0.