fixed rendering during eval

[m2kulkarni]

No description provided.

[greptile-apps]

Greptile Summary

Switched from Transformer to Recurrent policy architecture and enhanced rendering system for adaptive driving agents during evaluation. Added support for multi-scenario rendering with scenario counters, dual-policy population play (ego vs co-player), centered camera positioning for full map views, and new human replay video rendering functionality.

Key changes:

• Configuration updates to use Recurrent architecture instead of Transformer
• Added k_scenarios support for rendering multiple consecutive scenarios per evaluation
• Implemented population play infrastructure (ego agents + co-player policy with conditioning)
• Enhanced visualization with scenario counter overlays and improved camera centering
• Added render_human_replay_videos function for evaluation video generation
• Unified RNN/Transformer loading logic to use single rnn_name parameter

Issue found:

• Critical logic bug in pufferlib/utils.py:500 where the rendering loop doesn't iterate over different maps

Confidence Score: 2/5

• Contains a critical logic bug that will cause incorrect behavior during human replay rendering
• While most changes are well-structured (config updates, architecture refactoring, visualization enhancements), there's a critical bug in render_human_replay_videos where the loop renders the same map multiple times instead of iterating through different maps. This defeats the purpose of the num_maps parameter and will produce duplicate videos.
• pufferlib/utils.py requires immediate attention - the render_human_replay_videos function has a logic error in the rendering loop

Important Files Changed

Filename	Overview
pufferlib/config/ocean/adaptive.ini	Config changes to switch from Transformer to Recurrent architecture, adjust agent counts, and add rendering parameters
pufferlib/ocean/drive/drive.c	Fixed typo in demo comment and hardcoded demo values to avoid config dependency
pufferlib/ocean/env_config.h	Added k_scenarios and co-player population play configuration parsing
pufferlib/pufferl.py	Unified RNN/Transformer loading via rnn_name, added human replay video rendering during eval
pufferlib/ocean/drive/visualize.c	Added multi-scenario rendering, dual-policy population play support, centered camera for full map view, and scenario counter overlay
pufferlib/utils.py	Added render_human_replay_videos function and updated render_videos to use correct INI file and pass rendering parameters

_{Last reviewed commit: 6d8d963}

[greptile-apps]

_{6 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Loop renders same map repeatedly instead of different maps. The cmd doesn't change between iterations - no --map-name specified. Either remove the loop (set num_maps=1) or add map selection logic like in render_videos at line 380-383.