Exp10 routing temperature analysis #54
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This comprehensive experiment explores how routing temperature affects MoE training: New Components: - Temperature-aware router with dynamic temperature control - Temperature-aware MoE layer with detailed routing statistics - Custom trainer that tracks routing dynamics over training - Temperature scheduling support (linear, cosine, exponential, step) Experiments (13 total): - Temperature ablation: 8 temperatures from 0.5 to 10.0 - Temperature schedules: 4 different scheduling strategies - Extended training: 1 longer run with best temperature Metrics Tracked: - Performance: loss, accuracy, perplexity - Routing: entropy, selection confidence, expert utilization - Specialization: Gini coefficient, utilization variance - Load balancing: auxiliary loss Visualization Suite: - Temperature comparison plots (loss, accuracy, entropy vs temp) - Routing dynamics analysis (entropy evolution, confidence trends) - Expert utilization patterns (per-expert bars, heatmaps) - Schedule comparison (loss curves, temperature evolution) - Specialization analysis (Gini coefficients, variance) Analysis Tools: - Comprehensive plotting (plot_results.py) - Expert specialization analysis (analyze_specialization.py) - Summary report generation Features: - Uses optimal Muon settings from exp9 - Comprehensive documentation (README, EXPERIMENT_CARD, EXPERIMENT_SUMMARY) - Quick demo script for rapid testing - Modular design for easy extension Expected Insights: - Optimal routing temperature for MoE training - Trade-offs between exploration and exploitation - Expert specialization patterns under different temperatures - Effectiveness of temperature scheduling strategies This experiment will generate significant new knowledge about MoE routing dynamics.
- Changed references from 'final_metrics.loss' to 'final_metrics.val_loss' and 'final_metrics.accuracy' to 'final_metrics.val_accuracy' across multiple scripts for consistency in reporting. - Added 'seaborn' to requirements.txt for enhanced visualization capabilities. - Updated optimizer setup in tracking_trainer.py to support multiple optimizers and their respective learning rate schedulers.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.