Skip to content

Commit 9b2a2a1

Browse files
houman1359houman1359
committed
cleanup/optimzed experiments
1 parent 8b1b091 commit 9b2a2a1

54 files changed

Lines changed: 353 additions & 353 deletions

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

‎README.md‎

Lines changed: 10 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -85,18 +85,18 @@ Cross-layer halo analysis tracks downstream dependencies to predict cascade effe
8585
```
8686
alignment/
8787
├── configs/
88-
│ ├── cluster_analysis/ # Cluster-based analysis configs
89-
│ ├── paper/ # Paper experiment configs
90-
│ └── examples/ # Example configs
88+
| ├── cluster_analysis/ # Cluster-based analysis configs
89+
| ├── paper/ # Paper experiment configs
90+
| └── examples/ # Example configs
9191
├── scripts/
92-
│ ├── run_experiment.py # Main entry point
93-
│ └── run_analysis.py # Post-hoc analysis
92+
| ├── run_experiment.py # Main entry point
93+
| └── run_analysis.py # Post-hoc analysis
9494
├── src/alignment/
95-
│ ├── analysis/ # Visualization, clustering, cascade analysis
96-
│ ├── experiments/ # Experiment classes
97-
│ ├── metrics/ # Importance metrics
98-
│ ├── models/ # Model wrappers
99-
│ └── pruning/ # Pruning strategies
95+
| ├── analysis/ # Visualization, clustering, cascade analysis
96+
| ├── experiments/ # Experiment classes
97+
| ├── metrics/ # Importance metrics
98+
| ├── models/ # Model wrappers
99+
| └── pruning/ # Pruning strategies
100100
├── tests/ # Unit tests
101101
└── docs/ # Documentation
102102
```

‎configs/README.md‎

Lines changed: 10 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -7,17 +7,17 @@ configs/
77
├── template.yaml # Complete template with all options
88
├── unified_template.yaml # Unified format template
99
├── vision_prune/ # Vision model pruning configs
10-
│ ├── resnet18_cifar10_full.yaml
11-
│ ├── resnet18_cifar10_unified.yaml # Unified format version
12-
│ ├── resnet50_imagenet100.yaml
13-
│ ├── vgg16_cifar10_full.yaml
14-
│ └── mobilenetv2_cifar10_full.yaml
10+
| ├── resnet18_cifar10_full.yaml
11+
| ├── resnet18_cifar10_unified.yaml # Unified format version
12+
| ├── resnet50_imagenet100.yaml
13+
| ├── vgg16_cifar10_full.yaml
14+
| └── mobilenetv2_cifar10_full.yaml
1515
├── prune_llm/ # LLM pruning configs
16-
│ ├── llama3_8b_full.yaml
17-
│ ├── llama3_8b_unified.yaml # Unified format version
18-
│ ├── llama2_7b_full.yaml
19-
│ ├── mistral_7b_full.yaml
20-
│ └── qwen2_7b_full.yaml
16+
| ├── llama3_8b_full.yaml
17+
| ├── llama3_8b_unified.yaml # Unified format version
18+
| ├── llama2_7b_full.yaml
19+
| ├── mistral_7b_full.yaml
20+
| └── qwen2_7b_full.yaml
2121
└── examples/ # Example configs
2222
├── mnist_basic.yaml
2323
├── resnet_pruning.yaml

‎configs/examples/cnn2p2_pruning_comprehensive.yaml‎

Lines changed: 7 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -87,13 +87,13 @@ cnn:
8787
# Higher score = more important (keep) when selection_mode="low"
8888
#
8989
# INTERPRETATION GUIDE:
90-
# - rayleigh_quotient: High = aligned with data → keep (prune low)
91-
# - conditional_rayleigh_quotient: High = aligned with class-specific data → keep (prune low)
92-
# - mi_about_class: High = informative about class → keep (prune low)
93-
# - average_redundancy: High = MORE redundant → prune (use selection_mode="high" to prune high scorers)
94-
# - activation_l2_norm: High = active neuron → keep (prune low)
95-
# - composite_importance: Combines RQ + class_MI - redundancy → prune low
96-
# - alignment_minus_redundancy: RQ - R → prune low
90+
# - rayleigh_quotient: High = aligned with data -> keep (prune low)
91+
# - conditional_rayleigh_quotient: High = aligned with class-specific data -> keep (prune low)
92+
# - mi_about_class: High = informative about class -> keep (prune low)
93+
# - average_redundancy: High = MORE redundant -> prune (use selection_mode="high" to prune high scorers)
94+
# - activation_l2_norm: High = active neuron -> keep (prune low)
95+
# - composite_importance: Combines RQ + class_MI - redundancy -> prune low
96+
# - alignment_minus_redundancy: RQ - R -> prune low
9797

9898
pruning:
9999
enabled: true

‎configs/examples/llama2_7b_pruning.yaml‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -12,10 +12,10 @@
1212
#
1313
# From SCAR paper Table 5 (Cross-Model Generalization at 50% sparsity):
1414
# ┌─────────────────┬──────────────┬──────────────────┐
15-
# │ Model │ Method │ PPL↓ │ Acc.↑ │
15+
# | Model | Method | PPLdown | Acc.up |
1616
# ├─────────────────┼──────────────┼───────┼─────────┤
17-
# │ Llama-2-7B │ Wanda │ 19.4 │ 62.3% │
18-
# │ │ SCAR │ 13.1 │ 66.8% │
17+
# | Llama-2-7B | Wanda | 19.4 | 62.3% |
18+
# | | SCAR | 13.1 | 66.8% |
1919
# └─────────────────┴──────────────┴───────┴─────────┘
2020
#
2121
# EXPECTED RUNTIME: ~6-10 hours on H100

‎configs/examples/llama3_comprehensive_pruning.yaml‎

Lines changed: 41 additions & 41 deletions
Original file line numberDiff line numberDiff line change
@@ -7,26 +7,26 @@
77
# This config runs a comprehensive comparison of:
88
#
99
# ┌─────────────────────────────────────────────────────────────────────────┐
10-
# │ CATEGORY │ METHOD │ REFERENCE │
10+
# | CATEGORY | METHOD | REFERENCE |
1111
# ├────────────────────┼───────────────────────────────┼────────────────────┤
12-
# │ ALIGNMENT-BASED │ rayleigh_quotient (RQ) │ Our method │
13-
# │ │ gaussian_mi_analytic (MI) │ Related to RQ │
14-
# │ │ average_redundancy │ Info-theoretic │
12+
# | ALIGNMENT-BASED | rayleigh_quotient (RQ) | Our method |
13+
# | | gaussian_mi_analytic (MI) | Related to RQ |
14+
# | | average_redundancy | Info-theoretic |
1515
# ├────────────────────┼───────────────────────────────┼────────────────────┤
16-
# │ SCAR METRICS │ scar_loss_proxy │ Activation + Grad │
17-
# │ (activation+grad) │ scar_activation_power │ Raw activation │
18-
# │ │ scar_taylor │ First-order term │
19-
# │ │ scar_curvature │ Second-order term │
16+
# | SCAR METRICS | scar_loss_proxy | Activation + Grad |
17+
# | (activation+grad) | scar_activation_power | Raw activation |
18+
# | | scar_taylor | First-order term |
19+
# | | scar_curvature | Second-order term |
2020
# ├────────────────────┼───────────────────────────────┼────────────────────┤
21-
# │ SUPERNODE-AWARE │ supernode_protection_score │ Protects important │
22-
# │ │ supernode_connectivity_score │ Connectivity-based │
21+
# | SUPERNODE-AWARE | supernode_protection_score | Protects important |
22+
# | | supernode_connectivity_score | Connectivity-based |
2323
# ├────────────────────┼───────────────────────────────┼────────────────────┤
24-
# │ GENERALIZED │ generalized_importance │ No outlier needed │
24+
# | GENERALIZED | generalized_importance | No outlier needed |
2525
# ├────────────────────┼───────────────────────────────┼────────────────────┤
26-
# │ MAGNITUDE-BASED │ activation_l2_norm │ Common baseline │
26+
# | MAGNITUDE-BASED | activation_l2_norm | Common baseline |
2727
# ├────────────────────┼───────────────────────────────┼────────────────────┤
28-
# │ SOTA BASELINES │ wanda │ Sun et al. 2023 │
29-
# │ │ sparsegpt │ Frantar+Alistarh'23│
28+
# | SOTA BASELINES | wanda | Sun et al. 2023 |
29+
# | | sparsegpt | Frantar+Alistarh'23|
3030
# └─────────────────────────────────────────────────────────────────────────┘
3131
#
3232
# EVALUATION BENCHMARKS:
@@ -364,32 +364,32 @@ visualization:
364364
# ├── results_YYYYMMDD_HHMMSS.json # All metrics and scores
365365
# ├── experiment.log
366366
# ├── plots/
367-
# │ ├── pruning/
368-
# │ │ ├── pruning_comparison.png # All methods, perplexity
369-
# │ │ ├── pruning_comparison_loss.png # All methods, loss
370-
# │ │ ├── pruning_comparison_accuracy_mmlu.png # All methods, MMLU
371-
# │ │ ├── pruning_comparison_accuracy_*.png # Other benchmarks
372-
# │ │ └── pruning_<method>_comparison_*.png # Per-method plots
373-
# │ ├── histograms/
374-
# │ │ ├── histogram_rayleigh_quotient.png
375-
# │ │ ├── histogram_scar_loss_proxy.png
376-
# │ │ └── histogram_*.png
377-
# │ ├── scatter/
378-
# │ │ ├── scatter_activation_l2_norm_vs_rayleigh_quotient.png
379-
# │ │ └── scatter_*.png
380-
# │ ├── supernode/
381-
# │ │ ├── supernode_comparison_*.png
382-
# │ │ └── supernode_score_dist_*.png
383-
# │ ├── supernode_robustness/
384-
# │ │ ├── jaccard_heatmap_*.png # Metric overlap heatmaps
385-
# │ │ ├── spearman_heatmap_*.png # Score correlation heatmaps
386-
# │ │ ├── bootstrap_stability_*.png # Per-neuron stability
387-
# │ │ ├── consistency_bars_*.png # Cross-metric consistency
388-
# │ │ └── score_scatter_matrix_*.png # Metric pair correlations
389-
# │ ├── redundancy/
390-
# │ │ └── redundancy_heatmap_*.png
391-
# │ └── scar/
392-
# │ ├── scar_loss_proxy_layers.png
393-
# │ └── scar_metrics_heatmap.png
367+
# | ├── pruning/
368+
# | | ├── pruning_comparison.png # All methods, perplexity
369+
# | | ├── pruning_comparison_loss.png # All methods, loss
370+
# | | ├── pruning_comparison_accuracy_mmlu.png # All methods, MMLU
371+
# | | ├── pruning_comparison_accuracy_*.png # Other benchmarks
372+
# | | └── pruning_<method>_comparison_*.png # Per-method plots
373+
# | ├── histograms/
374+
# | | ├── histogram_rayleigh_quotient.png
375+
# | | ├── histogram_scar_loss_proxy.png
376+
# | | └── histogram_*.png
377+
# | ├── scatter/
378+
# | | ├── scatter_activation_l2_norm_vs_rayleigh_quotient.png
379+
# | | └── scatter_*.png
380+
# | ├── supernode/
381+
# | | ├── supernode_comparison_*.png
382+
# | | └── supernode_score_dist_*.png
383+
# | ├── supernode_robustness/
384+
# | | ├── jaccard_heatmap_*.png # Metric overlap heatmaps
385+
# | | ├── spearman_heatmap_*.png # Score correlation heatmaps
386+
# | | ├── bootstrap_stability_*.png # Per-neuron stability
387+
# | | ├── consistency_bars_*.png # Cross-metric consistency
388+
# | | └── score_scatter_matrix_*.png # Metric pair correlations
389+
# | ├── redundancy/
390+
# | | └── redundancy_heatmap_*.png
391+
# | └── scar/
392+
# | ├── scar_loss_proxy_layers.png
393+
# | └── scar_metrics_heatmap.png
394394
# └── checkpoints/
395395

‎configs/examples/llama3_extended_analysis.yaml‎

Lines changed: 13 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -192,20 +192,20 @@ visualization:
192192
# Expected outputs:
193193
# - results/llama3_extended_analysis_YYYYMMDD_HHMMSS/
194194
# ├── metrics/
195-
# │ ├── layer_metrics.json
196-
# │ ├── halo_redundancy_results.json
197-
# │ ├── multi_supernode_results.json
198-
# │ └── cross_layer_results.json
195+
# | ├── layer_metrics.json
196+
# | ├── halo_redundancy_results.json
197+
# | ├── multi_supernode_results.json
198+
# | └── cross_layer_results.json
199199
# ├── plots/
200-
# │ ├── halo/
201-
# │ │ ├── halo_redundancy_by_depth.png
202-
# │ │ ├── halo_redundancy_comprehensive.png
203-
# │ │ └── halo_redundancy_heatmap.png
204-
# │ ├── multi_supernode/
205-
# │ │ └── cluster_analysis_*.png
206-
# │ └── cross_layer/
207-
# │ ├── cross_layer_redundancy.png
208-
# │ └── layer_efficiency.png
200+
# | ├── halo/
201+
# | | ├── halo_redundancy_by_depth.png
202+
# | | ├── halo_redundancy_comprehensive.png
203+
# | | └── halo_redundancy_heatmap.png
204+
# | ├── multi_supernode/
205+
# | | └── cluster_analysis_*.png
206+
# | └── cross_layer/
207+
# | ├── cross_layer_redundancy.png
208+
# | └── layer_efficiency.png
209209
# └── evaluation/
210210
# └── benchmark_results.json
211211
# ============================================================================

0 commit comments

Comments
 (0)