visualize_model_performance.py

source modelling visualization analytics

File Path: src/modelling/visualize_model_performance.py

Purpose: Generates advanced visual diagnostics and classification reports for trained models.

Overview

The visualize_metrics function evaluates a model’s performance on a test dataset and produces a comprehensive 3-part visual dashboard for error analysis.

Visual Dashboard Components

1. Global Confusion Matrix (Heatmap):

   • Visualizes overall diagonal sharpness across all classes.
   • Axes are unlabeled (indices only) to keep the view clean for high class counts (e.g., 502).

2. Per-Class Performance (Scatter Plot):

   • Plots the F1-Score for every class ID.
   • Highlights underperforming classes (F1 < 0.5) in red for quick identification.

3. Top Misclassifications (Bar Chart):

   • Specifically identifies the top 20 most frequent “True → Predicted” error pairs.
   • Helps identify semantic similarities causing model confusion.

Outputs

• Visual Dashboard: Saved as {checkpoint_name}-Model_Diagnostics.jpg in the checkpoint directory.
• Classification Report: Saved as {checkpoint_name} - classification_report.csv containing precision, recall, and f1-score for all classes.

Usage

# Can be called via the shared CLI (compatible with onnx_benchmark)
python src/modelling/visualize_model_performance.py --checkpoint_path checkpoints/best.pth

Related Documentation

• train.py - Automated trigger after training completion.
• utils.py - Metadata extraction from checkpoint paths.