Deep_Object_Pose VS trl

I've been doing some research in this area and there are a few deep learning solutions to this problem. For example, NVIDIA's Deep Object Pose Estimation will estimate the 6DOF pose of a known object. But you'll have to train the network if you want to detect a new object. PoseCNN, which someone else mentioned, does a similar thing. CenterPose is more interesting, as it can estimate then pose of an object from a known category; e.g. sneakers, or laptops, rather that one specific object (as DOPE and PoseCNN do).

For our last event of ArchE Week, the Ohio State Underwater Robotics Team (Website, Instagram) is hosting a workshop tonight on machine learning! The workshop is an interactive walkthrough of using machine learning solutions to make predictions. Some example problems we could be trying to solve are predicting a grade, predicting the weather, and the classic recognize a digit problem. Our team personally uses machine learning to do real-time object detection with YOLO and NVidia DOPE, so we may touch on that as well!

Implementation: ORPO has been integrated into popular fine-tuning libraries like TRL, Axolotl, and LLaMA-Factory.

I am using TRL for RLHF fine-tuning the Llama-2-7B model and getting an OOM error (even with batch_size=1). If anyone used TRL for RLHF can please tell me what I am doing wrong? Code details can be found in the GitHub issue.

SFTtrainer from huggingface

For lora - PEFT seems to work. I don't have patience to wait 5 hours, but modifying this example seems to work. You don't even need to modify that much, as their model just as neo-x uses query_key_value name for self-attention.

They have examples of making GPT output more positive (code) by using a sentiment model as reward. There are other examples about reducing toxicity, summarization here: https://github.com/lvwerra/trl/tree/main/examples . Should be fairly simple to modify the sentiment example and try the calculator reward you mentioned above.

You can use this -> https://github.com/lvwerra/trl/blob/main/examples/sentiment/scripts/gpt-neox-20b_peft/merge_peft_adapter.py

Just the hh directly. From the results it seems like it might possibly be enough but I might also try instruction tuning then running the whole process from that base. I will also be running the reinforcement learning by using a Lora using this as an example https://github.com/lvwerra/trl/tree/main/examples/sentiment/scripts/gpt-neox-20b_peft

This package is pretty simple to use! https://github.com/lvwerra/trl