fix dynamic filtering

[Django-Jiang]

Solve #570

[gemini-code-assist]

Summary of Changes

Hello @Django-Jiang, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses an issue in dynamic filtering by enhancing the robustness of the check_reward_nonzero_std function. The changes improve the precision of reward standard deviation calculations and refine the threshold for determining a non-zero standard deviation, which helps prevent filtering errors due to floating-point arithmetic and ensures more reliable dynamic sampling.

Highlights

• Data Type Precision: Changed the torch.tensor data type from torch.float to torch.float64 within the check_reward_nonzero_std function to improve numerical precision when calculating reward standard deviations.
• Standard Deviation Threshold: Modified the standard deviation check from > 0.0 to > 1e-6 to account for potential floating-point inaccuracies, ensuring that very small, near-zero standard deviations are correctly handled as zero.

Changelog

• miles/rollout/filter_hub/dynamic_sampling_filters.py
  • Updated the check_reward_nonzero_std function to use torch.float64 for reward tensor creation, improving numerical stability.
  • Adjusted the standard deviation comparison threshold from 0.0 to 1e-6 to better handle floating-point precision issues.

Activity

• No specific review activity or comments have been recorded for this pull request yet.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

The pull request addresses dynamic filtering by improving the numerical stability and correctness of the check_reward_nonzero_std function. The change to torch.float64 enhances precision for standard deviation calculations, and the introduction of an epsilon 1e-6 for comparison correctly handles floating-point inaccuracies, preventing samples with effectively zero standard deviation from passing the filter. This is a good fix for the identified issue.

[gemini-code-assist]

The use of a small epsilon 1e-6 for comparing floating-point standard deviation is a good practice to account for numerical inaccuracies. However, this value is a magic number. It would be more maintainable and readable to define it as a named constant, perhaps at the module level. This improves clarity and makes it easier to adjust the threshold if needed in the future.

    ZERO_STD_EPSILON = 1e-6
    keep = torch.tensor(rewards, dtype=torch.float64).std() > ZERO_STD_EPSILON