update cluster experiments, dataloaders and pruning strategies for vision nets

.github/workflows/test.yml

@@ -45,8 +45,12 @@ jobs:
 
     - name: Run tests
       run: |
-        pytest tests/ -v --cov=src/alignment --cov-report=xml --cov-report=html --tb=short -ra
-
+        pytest tests/ -v --cov=src/alignment --cov-report=xml --cov-report=html --cov-report=term-missing --tb=short -ra
+
+    - name: Check coverage threshold
+      run: |
+        coverage report --fail-under=25
+
     - name: Upload coverage to Codecov
       uses: codecov/codecov-action@v3
       with:

.pre-commit-config.yaml

@@ -46,3 +46,12 @@ repos:
     hooks:
       - id: detect-secrets
         args: ['--baseline', '.secrets.baseline']
+
+  - repo: local
+    hooks:
+      - id: fast-tests
+        name: Fast unit tests
+        entry: python -m pytest tests/unit/ -x -q -m "not slow and not integration and not gpu"
+        language: system
+        pass_filenames: false
+        always_run: true

README.md

@@ -85,18 +85,18 @@ Cross-layer halo analysis tracks downstream dependencies to predict cascade effe
 ```
 alignment/
 ├── configs/
-│   ├── cluster_analysis/   # Cluster-based analysis configs
-│   ├── paper/              # Paper experiment configs
-│   └── examples/           # Example configs
+|   ├── cluster_analysis/   # Cluster-based analysis configs
+|   ├── paper/              # Paper experiment configs
+|   └── examples/           # Example configs
 ├── scripts/
-│   ├── run_experiment.py   # Main entry point
-│   └── run_analysis.py     # Post-hoc analysis
+|   ├── run_experiment.py   # Main entry point
+|   └── run_analysis.py     # Post-hoc analysis
 ├── src/alignment/
-│   ├── analysis/           # Visualization, clustering, cascade analysis
-│   ├── experiments/        # Experiment classes
-│   ├── metrics/            # Importance metrics
-│   ├── models/             # Model wrappers
-│   └── pruning/            # Pruning strategies
+|   ├── analysis/           # Visualization, clustering, cascade analysis
+|   ├── experiments/        # Experiment classes
+|   ├── metrics/            # Importance metrics
+|   ├── models/             # Model wrappers
+|   └── pruning/            # Pruning strategies
 ├── tests/                  # Unit tests
 └── docs/                   # Documentation
 ```

configs/README.md

@@ -7,17 +7,17 @@ configs/
 ├── template.yaml              # Complete template with all options
 ├── unified_template.yaml      # Unified format template
 ├── vision_prune/              # Vision model pruning configs
-│   ├── resnet18_cifar10_full.yaml
-│   ├── resnet18_cifar10_unified.yaml  # Unified format version
-│   ├── resnet50_imagenet100.yaml
-│   ├── vgg16_cifar10_full.yaml
-│   └── mobilenetv2_cifar10_full.yaml
+|   ├── resnet18_cifar10_full.yaml
+|   ├── resnet18_cifar10_unified.yaml  # Unified format version
+|   ├── resnet50_imagenet100.yaml
+|   ├── vgg16_cifar10_full.yaml
+|   └── mobilenetv2_cifar10_full.yaml
 ├── prune_llm/                 # LLM pruning configs
-│   ├── llama3_8b_full.yaml
-│   ├── llama3_8b_unified.yaml  # Unified format version
-│   ├── llama2_7b_full.yaml
-│   ├── mistral_7b_full.yaml
-│   └── qwen2_7b_full.yaml
+|   ├── llama3_8b_full.yaml
+|   ├── llama3_8b_unified.yaml  # Unified format version
+|   ├── llama2_7b_full.yaml
+|   ├── mistral_7b_full.yaml
+|   └── qwen2_7b_full.yaml
 └── examples/                  # Example configs
     ├── mnist_basic.yaml
     ├── resnet_pruning.yaml

configs/examples/cnn2p2_pruning_comprehensive.yaml

@@ -87,13 +87,13 @@ cnn:
 # Higher score = more important (keep) when selection_mode="low"
 #
 # INTERPRETATION GUIDE:
-# - rayleigh_quotient: High = aligned with data → keep (prune low)
-# - conditional_rayleigh_quotient: High = aligned with class-specific data → keep (prune low)
-# - mi_about_class: High = informative about class → keep (prune low)
-# - average_redundancy: High = MORE redundant → prune (use selection_mode="high" to prune high scorers)
-# - activation_l2_norm: High = active neuron → keep (prune low)
-# - composite_importance: Combines RQ + class_MI - redundancy → prune low
-# - alignment_minus_redundancy: RQ - R → prune low
+# - rayleigh_quotient: High = aligned with data -> keep (prune low)
+# - conditional_rayleigh_quotient: High = aligned with class-specific data -> keep (prune low)
+# - mi_about_class: High = informative about class -> keep (prune low)
+# - average_redundancy: High = MORE redundant -> prune (use selection_mode="high" to prune high scorers)
+# - activation_l2_norm: High = active neuron -> keep (prune low)
+# - composite_importance: Combines RQ + class_MI - redundancy -> prune low
+# - alignment_minus_redundancy: RQ - R -> prune low
 
 pruning:
   enabled: true

configs/examples/llama2_7b_pruning.yaml

@@ -12,10 +12,10 @@
 #
 # From SCAR paper Table 5 (Cross-Model Generalization at 50% sparsity):
 # ┌─────────────────┬──────────────┬──────────────────┐
-# │ Model           │ Method       │ PPL↓  │ Acc.↑   │
+# | Model           | Method       | PPLdown  | Acc.up   |
 # ├─────────────────┼──────────────┼───────┼─────────┤
-# │ Llama-2-7B      │ Wanda        │ 19.4  │ 62.3%   │
-# │                 │ SCAR         │ 13.1  │ 66.8%   │
+# | Llama-2-7B      | Wanda        | 19.4  | 62.3%   |
+# |                 | SCAR         | 13.1  | 66.8%   |
 # └─────────────────┴──────────────┴───────┴─────────┘
 #
 # EXPECTED RUNTIME: ~6-10 hours on H100
@@ -139,7 +139,7 @@ supernode_summary:
 pruning:
   enabled: true
 
-  sparsity_levels: [0.25, 0.5, 0.75]
+  sparsity_levels: [0, 0.25, 0.5, 0.75]
 
   selection_modes: ["low", "high"]
 

configs/examples/llama3_comprehensive_pruning.yaml

@@ -7,26 +7,26 @@
 # This config runs a comprehensive comparison of:
 # 
 # ┌─────────────────────────────────────────────────────────────────────────┐
-# │ CATEGORY           │ METHOD                        │ REFERENCE          │
+# | CATEGORY           | METHOD                        | REFERENCE          |
 # ├────────────────────┼───────────────────────────────┼────────────────────┤
-# │ ALIGNMENT-BASED    │ rayleigh_quotient (RQ)        │ Our method         │
-# │                    │ gaussian_mi_analytic (MI)     │ Related to RQ      │
-# │                    │ average_redundancy            │ Info-theoretic     │
+# | ALIGNMENT-BASED    | rayleigh_quotient (RQ)        | Our method         |
+# |                    | gaussian_mi_analytic (MI)     | Related to RQ      |
+# |                    | average_redundancy            | Info-theoretic     |
 # ├────────────────────┼───────────────────────────────┼────────────────────┤
-# │ SCAR METRICS       │ scar_loss_proxy               │ Activation + Grad  │
-# │ (activation+grad)  │ scar_activation_power         │ Raw activation     │
-# │                    │ scar_taylor                   │ First-order term   │
-# │                    │ scar_curvature                │ Second-order term  │
+# | SCAR METRICS       | scar_loss_proxy               | Activation + Grad  |
+# | (activation+grad)  | scar_activation_power         | Raw activation     |
+# |                    | scar_taylor                   | First-order term   |
+# |                    | scar_curvature                | Second-order term  |
 # ├────────────────────┼───────────────────────────────┼────────────────────┤
-# │ SUPERNODE-AWARE    │ supernode_protection_score    │ Protects important │
-# │                    │ supernode_connectivity_score  │ Connectivity-based │
+# | SUPERNODE-AWARE    | supernode_protection_score    | Protects important |
+# |                    | supernode_connectivity_score  | Connectivity-based |
 # ├────────────────────┼───────────────────────────────┼────────────────────┤
-# │ GENERALIZED        │ generalized_importance        │ No outlier needed  │
+# | GENERALIZED        | generalized_importance        | No outlier needed  |
 # ├────────────────────┼───────────────────────────────┼────────────────────┤
-# │ MAGNITUDE-BASED    │ activation_l2_norm            │ Common baseline    │
+# | MAGNITUDE-BASED    | activation_l2_norm            | Common baseline    |
 # ├────────────────────┼───────────────────────────────┼────────────────────┤
-# │ SOTA BASELINES     │ wanda                         │ Sun et al. 2023    │
-# │                    │ sparsegpt                     │ Frantar+Alistarh'23│
+# | SOTA BASELINES     | wanda                         | Sun et al. 2023    |
+# |                    | sparsegpt                     | Frantar+Alistarh'23|
 # └─────────────────────────────────────────────────────────────────────────┘
 #
 # EVALUATION BENCHMARKS:
@@ -364,32 +364,32 @@ visualization:
 # ├── results_YYYYMMDD_HHMMSS.json    # All metrics and scores
 # ├── experiment.log
 # ├── plots/
-# │   ├── pruning/
-# │   │   ├── pruning_comparison.png                    # All methods, perplexity
-# │   │   ├── pruning_comparison_loss.png              # All methods, loss
-# │   │   ├── pruning_comparison_accuracy_mmlu.png     # All methods, MMLU
-# │   │   ├── pruning_comparison_accuracy_*.png        # Other benchmarks
-# │   │   └── pruning_<method>_comparison_*.png        # Per-method plots
-# │   ├── histograms/
-# │   │   ├── histogram_rayleigh_quotient.png
-# │   │   ├── histogram_scar_loss_proxy.png
-# │   │   └── histogram_*.png
-# │   ├── scatter/
-# │   │   ├── scatter_activation_l2_norm_vs_rayleigh_quotient.png
-# │   │   └── scatter_*.png
-# │   ├── supernode/
-# │   │   ├── supernode_comparison_*.png
-# │   │   └── supernode_score_dist_*.png
-# │   ├── supernode_robustness/
-# │   │   ├── jaccard_heatmap_*.png          # Metric overlap heatmaps
-# │   │   ├── spearman_heatmap_*.png         # Score correlation heatmaps
-# │   │   ├── bootstrap_stability_*.png      # Per-neuron stability
-# │   │   ├── consistency_bars_*.png         # Cross-metric consistency
-# │   │   └── score_scatter_matrix_*.png     # Metric pair correlations
-# │   ├── redundancy/
-# │   │   └── redundancy_heatmap_*.png
-# │   └── scar/
-# │       ├── scar_loss_proxy_layers.png
-# │       └── scar_metrics_heatmap.png
+# |   ├── pruning/
+# |   |   ├── pruning_comparison.png                    # All methods, perplexity
+# |   |   ├── pruning_comparison_loss.png              # All methods, loss
+# |   |   ├── pruning_comparison_accuracy_mmlu.png     # All methods, MMLU
+# |   |   ├── pruning_comparison_accuracy_*.png        # Other benchmarks
+# |   |   └── pruning_<method>_comparison_*.png        # Per-method plots
+# |   ├── histograms/
+# |   |   ├── histogram_rayleigh_quotient.png
+# |   |   ├── histogram_scar_loss_proxy.png
+# |   |   └── histogram_*.png
+# |   ├── scatter/
+# |   |   ├── scatter_activation_l2_norm_vs_rayleigh_quotient.png
+# |   |   └── scatter_*.png
+# |   ├── supernode/
+# |   |   ├── supernode_comparison_*.png
+# |   |   └── supernode_score_dist_*.png
+# |   ├── supernode_robustness/
+# |   |   ├── jaccard_heatmap_*.png          # Metric overlap heatmaps
+# |   |   ├── spearman_heatmap_*.png         # Score correlation heatmaps
+# |   |   ├── bootstrap_stability_*.png      # Per-neuron stability
+# |   |   ├── consistency_bars_*.png         # Cross-metric consistency
+# |   |   └── score_scatter_matrix_*.png     # Metric pair correlations
+# |   ├── redundancy/
+# |   |   └── redundancy_heatmap_*.png
+# |   └── scar/
+# |       ├── scar_loss_proxy_layers.png
+# |       └── scar_metrics_heatmap.png
 # └── checkpoints/
 

configs/examples/llama3_extended_analysis.yaml

@@ -192,20 +192,20 @@ visualization:
 # Expected outputs:
 # - results/llama3_extended_analysis_YYYYMMDD_HHMMSS/
 #   ├── metrics/
-#   │   ├── layer_metrics.json
-#   │   ├── halo_redundancy_results.json
-#   │   ├── multi_supernode_results.json
-#   │   └── cross_layer_results.json
+#   |   ├── layer_metrics.json
+#   |   ├── halo_redundancy_results.json
+#   |   ├── multi_supernode_results.json
+#   |   └── cross_layer_results.json
 #   ├── plots/
-#   │   ├── halo/
-#   │   │   ├── halo_redundancy_by_depth.png
-#   │   │   ├── halo_redundancy_comprehensive.png
-#   │   │   └── halo_redundancy_heatmap.png
-#   │   ├── multi_supernode/
-#   │   │   └── cluster_analysis_*.png
-#   │   └── cross_layer/
-#   │       ├── cross_layer_redundancy.png
-#   │       └── layer_efficiency.png
+#   |   ├── halo/
+#   |   |   ├── halo_redundancy_by_depth.png
+#   |   |   ├── halo_redundancy_comprehensive.png
+#   |   |   └── halo_redundancy_heatmap.png
+#   |   ├── multi_supernode/
+#   |   |   └── cluster_analysis_*.png
+#   |   └── cross_layer/
+#   |       ├── cross_layer_redundancy.png
+#   |       └── layer_efficiency.png
 #   └── evaluation/
 #       └── benchmark_results.json
 # ============================================================================