video_backbone剖析

sparse_head_simulation/report.md

@@ -0,0 +1,108 @@
+# Sparse Head Execution Simulation
+
+This report uses zero-order hold over logged stepwise action/state outputs.
+For a sparse scheme, the final output at step 49 is approximated by the most recent checkpoint output.
+
+## Experiment 1.1
+
+### dense
+
+- Head exec count: 50
+- Head compute saving ratio: 0.0000
+- Final action cosine vs dense: mean=1.0000, median=1.0000
+- Final state cosine vs dense: mean=1.0000, median=1.0000
+
+### sparse_10
+
+- Head exec count: 10
+- Head compute saving ratio: 0.8000
+- Final action cosine vs dense: mean=0.9341, median=0.9357
+- Final state cosine vs dense: mean=0.9130, median=0.9207
+
+### sparse_4
+
+- Head exec count: 4
+- Head compute saving ratio: 0.9200
+- Final action cosine vs dense: mean=1.0000, median=1.0000
+- Final state cosine vs dense: mean=1.0000, median=1.0000
+
+### sparse_5
+
+- Head exec count: 5
+- Head compute saving ratio: 0.9000
+- Final action cosine vs dense: mean=0.7747, median=0.7703
+- Final state cosine vs dense: mean=0.7463, median=0.7481
+
+### sparse_8
+
+- Head exec count: 8
+- Head compute saving ratio: 0.8400
+- Final action cosine vs dense: mean=1.0000, median=1.0000
+- Final state cosine vs dense: mean=1.0000, median=1.0000
+
+### tail_heavy_10
+
+- Head exec count: 10
+- Head compute saving ratio: 0.8000
+- Final action cosine vs dense: mean=1.0000, median=1.0000
+- Final state cosine vs dense: mean=1.0000, median=1.0000
+
+### tail_heavy_6
+
+- Head exec count: 6
+- Head compute saving ratio: 0.8800
+- Final action cosine vs dense: mean=1.0000, median=1.0000
+- Final state cosine vs dense: mean=1.0000, median=1.0000
+
+### tail_heavy_8
+
+- Head exec count: 8
+- Head compute saving ratio: 0.8400
+- Final action cosine vs dense: mean=1.0000, median=1.0000
+- Final state cosine vs dense: mean=1.0000, median=1.0000
+
+### tail_only_4
+
+- Head exec count: 4
+- Head compute saving ratio: 0.9200
+- Final action cosine vs dense: mean=1.0000, median=1.0000
+- Final state cosine vs dense: mean=1.0000, median=1.0000
+
+### tail_only_6
+
+- Head exec count: 6
+- Head compute saving ratio: 0.8800
+- Final action cosine vs dense: mean=1.0000, median=1.0000
+- Final state cosine vs dense: mean=1.0000, median=1.0000
+
+## Experiment 1.2
+
+- Target threshold: 0.95
+- Action earliest checkpoint mean: 46.0179
+- State earliest checkpoint mean: 46.5357
+- Unique min head steps needed for action: [1.0]
+- Unique min head steps needed for state: [1.0]
+
+## Experiment 1.3
+
+### budget_10_uniform_vs_tail
+
+- Tail better action share: 1.0000
+- Tail better state share: 1.0000
+- Mean tail-minus-uniform action cosine: 0.0659
+- Mean tail-minus-uniform state cosine: 0.0870
+
+### budget_4_uniform_vs_tail
+
+- Tail better action share: 0.0000
+- Tail better state share: 0.0000
+- Mean tail-minus-uniform action cosine: 0.0000
+- Mean tail-minus-uniform state cosine: 0.0000
+
+### budget_8_uniform_vs_tail
+
+- Tail better action share: 0.0000
+- Tail better state share: 0.0000
+- Mean tail-minus-uniform action cosine: 0.0000
+- Mean tail-minus-uniform state cosine: 0.0000
+