deepspeedai / deepspeedai/DeepSpeedExamples

DeepSpeed-Chat Step-3 tensorboard loss figures with multiple training epochs

Open
#811 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
6.8k
Forks
1.1k
Avg merge
2d 16h
Merged PRs (30d)
1

Description

Hi, when training RLHF step-3, I set parameters related to epochs as:

  • ppo_epochs = 1
  • num_train_epochs = 30

and I found that the numbers of lines in "actor_loss", "actor_loss_sum", "critic_loss", "critic_loss_sum" and "reward" (saved in the folder "step3_tensorboard_logs") are the same as the parameter "num_train_epochs", like :

Screenshot 2023-11-16 at 10 03 43
Screenshot 2023-11-16 at 10 04 05

The question is:

How can I make them display as the figure for "train_loss" or "lr" (saved in the folder "ds_tensorboard_logs") like the below one, where the line for the whole 30 epochs as a whole ( a single line):

Screenshot 2023-11-16 at 09 57 49
Screenshot 2023-11-16 at 09 58 20

Thanks!

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by comparing how logging is performed for the step-3_tensorboard_logs and ds_tensorboard_logs folders, using the ppo_epochs and num_train_epochs settings described in the issue. Determine whether the loss and reward values are logged per epoch or aggregated across all 30 epochs; done means the figures use the requested whole-training aggregation.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning, observability
Issue type
Bug
Difficulty
3/5
Estimated time
1-2 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.