整理 Q1-Q3 实验代码与结果

2026-09-24 16:25:15 +08:00
parent 0261ecdfba
commit 8f5c2c3be6
247 changed files with 69828 additions and 19 deletions
@@ -270,6 +270,7 @@ stay fixed. Emotion labels are not used.
 cd Q1
 uv run python -m q1.tsfa_experiment --device cuda --seed 42
 uv run python -m q1.tsfa_alignment_only_eval --device cuda --seed 42
+uv run python -m q1.tsfa_emotion_probe --device cuda --seed 42
 ```

 The first command fits one semantic branch and equal-seed correspondence probes
@@ -282,5 +283,101 @@ paired video-group confidence intervals, monotonicity/span/entropy diagnostics,
 plots, manifests, and checkpoints are written to `outputs/tsfa/`; the small
 `report_bundle/` contains shareable
 summaries and figures. See [RESULTS.md](RESULTS.md) for the measured tradeoff
-and interpretation limits.
+and interpretation limits. The third command evaluates the same frozen
+representations with the grouped, five-segment logistic-regression and Ridge
+probes; emotion labels are used only by these downstream probes.
+
+With the math OOF predictions and fold assignment CSV available at their
+workspace paths, compare TSFA against B0–B4 on the identical held-out samples
+and bootstrap by source video:
+
+```bash
+uv run python -m q1.compare_emotion_probes --bootstrap-repeats 2000 --seed 42
+```
+
+The local Audio–Vision edge and modality-private residual experiment runs a
+fixed 2×2 ablation with the same five grouped folds, temporal windows,
+self-supervised objective, and emotion probes:
+
+```bash
+uv run python -m q1.tsfa_av_private_ablation --device cuda --seed 42
+uv run python -m q1.compare_emotion_probes --tsfa-predictions outputs/tsfa_av_private_ablation/predictions.csv --candidate-method TSFA-T+Private --output-dir outputs/tsfa_av_private_ablation --output-stem private_vs_math --bootstrap-repeats 2000 --seed 42
+```
+
+The ablation files and figure are in `outputs/tsfa_av_private_ablation/`.
+
+## TSFA-SPR shared/private factorization
+
+Run the first grouped five-fold experiment with M4 temporal checkpoints frozen:
+
+```bash
+cd Q1
+uv run python -m q1.tsfa_shared_private --device cuda --seed 42
+```
+
+The experiment pools each fold-standardized original text, audio, and vision
+source stream with the matching frozen `M4_sourceTime` attention weights. A
+shared MLP and three independent private MLPs learn from same-slot cross-modal
+InfoNCE, within-modality orthogonality, and per-modality reconstruction. No
+emotion labels enter representation training. The run also fits a train-fold
+PCA control on raw private slots to match the 256-dimensional-per-slot SPR
+control, then evaluates all representations with the existing grouped
+LogisticRegression/Ridge probes. Outputs include the fold checkpoints,
+training history, OOF predictions, grouped bootstrap contrasts, source and
+reconstruction diagnostics, and figures under
+`outputs/tsfa_shared_private/`.
+
+If you only need to regenerate the derived source contrasts, assessment, and
+plots from an existing run without retraining, use:
+
+```bash
+uv run python -m q1.tsfa_shared_private --finalize-existing
+```
+
+For the first seed's interpretation, diagnostic caveats, and next-step
+decision, see [RESULTS.md](RESULTS.md#tsfa-spr时序共享私有分解) and
+[`outputs/tsfa_shared_private/README.md`](outputs/tsfa_shared_private/README.md).
+
+## Shared information definition comparison
+
+Compare three concrete definitions of shared information using the same
+fold-standardized original source features pooled by the frozen M4 temporal
+checkpoints: Similarity (reuse the existing SPR weights), Correlation
+(regularized MAXVAR/GCCA), and Predictability (cross-modal Ridge prediction
+and its private residual). The Similarity network is loaded for inference and
+is not retrained. Emotion labels are used only by the identical outer-fold
+emotion probes.
+
+```bash
+cd Q1
+uv run python -m q1.shared_definition_comparison --device cuda --seed 42
+```
+
+The run reuses the fixed five-fold `video_id` splits, compares same-slot,
+shifted, and within-video-shuffled predictive controls, and reports shared
+variance, residual predictability, dimension-matched 1,280D emotion probes,
+private-source ablations, and paired video-group bootstrap contrasts against
+RawPrivate-PCA, Similarity-SPR, TSFA+RawPrivate, and math B0/B4. The math
+artifacts are read-only inputs. Tables, figures, fold assignments, and the run
+manifest are written to `outputs/shared_definition_comparison/`; measured
+results and limitations are in [RESULTS.md](RESULTS.md#shared-information-definition-similarity-vs-correlation-vs-predictability).
+
+## Shared time coordinate with modality-private content
+
+This follow-up uses the corrected definition: M4 supplies a frozen 50-slot
+temporal coordinate, while BERT text, eGeMAPS audio, and DeiT vision retain
+modality-specific content. Optional same-slot product/difference features are
+evaluated only as task-specific interactions; they do not assert shared
+semantics. Feature standardization, PCA, and emotion probes are fitted inside
+each training fold. The math outputs are read-only historical controls.
+
+```bash
+cd Q1
+uv run python -m q1.shared_time_private_content --device cuda --seed 42 --bootstrap-draws 2000
+```
+
+The run writes 19 new representation OOF probes, prior-result controls,
+video-group bootstrap contrasts, temporal monitoring, fold/PCA audits, and nine
+figures to `outputs/shared_time_private_content/`. See the dedicated
+[results and seven-question interpretation](RESULTS.md#shared-time-private-content修正版验证).