| \begin{table}[htbp] | |
| \centering | |
| \footnotesize | |
| \caption{Hyperparameters of the proposed model.} | |
| \label{tab:hyperparameters} | |
| \begin{tabular}{lc} | |
| \toprule | |
| Parameter & Value \\ | |
| \midrule | |
| data.cache\_frames & 24 \\ | |
| data.cache\_seconds & 6.0 \\ | |
| data.face\_crop.box\_scale & 1.8 \\ | |
| data.face\_crop.decode\_max\_side & 640 \\ | |
| data.face\_crop.detect\_frames & 4 \\ | |
| data.face\_crop.detector & yunet \\ | |
| data.face\_crop.enabled & True \\ | |
| data.frame\_size & 224 \\ | |
| data.n\_frames & 16 \\ | |
| data.primary & fakeavceleb \\ | |
| data.sample\_rate & 16000 \\ | |
| data.split\_fractions & [0.65, 0.15, 0.2] \\ | |
| data.split\_protocol & strict\_identity \\ | |
| data.split\_seed & 42 \\ | |
| data.subsample\_seed & 0 \\ | |
| data.window\_seconds & 4.0 \\ | |
| features.audio\_model & microsoft/wavlm-base-plus \\ | |
| features.audio\_tokens & 50 \\ | |
| features.dtype & float16 \\ | |
| features.eval\_view\_offset & 4 \\ | |
| features.extract\_batch\_size & auto \\ | |
| features.qacp\_variants\_per\_clip & 4 \\ | |
| features.shard\_clips & 512 \\ | |
| features.shard\_clips\_audio & 4096 \\ | |
| features.train\_view\_offsets & [0, 4, 8] \\ | |
| features.video\_model & MCG-NJU/videomae-base \\ | |
| features.video\_spatial\_grid & 4 \\ | |
| model.compose\_quadrant & False \\ | |
| model.d\_model & 768 \\ | |
| model.dropout & 0.1 \\ | |
| model.n\_fusion\_layers & 4 \\ | |
| model.n\_heads & 8 \\ | |
| model.use\_disentangle & True \\ | |
| model.use\_sync & True \\ | |
| train.phase\_b.betas & [0.9, 0.999] \\ | |
| train.phase\_b.early\_stopping\_patience & 3 \\ | |
| train.phase\_b.effective\_batch & 16 \\ | |
| train.phase\_b.epochs & 6 \\ | |
| train.phase\_b.gradient\_checkpointing & True \\ | |
| train.phase\_b.loss\_weights.a & 1.0 \\ | |
| train.phase\_b.loss\_weights.disentangle & 0.1 \\ | |
| train.phase\_b.loss\_weights.loc & 0.5 \\ | |
| train.phase\_b.loss\_weights.quad & 0.5 \\ | |
| train.phase\_b.loss\_weights.sync & 0.1 \\ | |
| train.phase\_b.loss\_weights.v & 1.0 \\ | |
| train.phase\_b.lr & 3e-05 \\ | |
| train.phase\_b.lr\_encoder & 1e-05 \\ | |
| train.phase\_b.max\_grad\_norm & 1.0 \\ | |
| train.phase\_b.micro\_batch & auto \\ | |
| train.phase\_b.min\_lr\_ratio & 0.01 \\ | |
| train.phase\_b.modality\_dropout & 0.15 \\ | |
| train.phase\_b.nonfinite\_patience & 20 \\ | |
| train.phase\_b.sampler & sqrt\_balanced \\ | |
| train.phase\_b.sampler\_key & quadrant \\ | |
| train.phase\_b.scheduler & warmup\_cosine \\ | |
| train.phase\_b.selection\_metric & val/mean\_auc \\ | |
| train.phase\_b.temporal\_augment & True \\ | |
| train.phase\_b.unfreeze\_top\_blocks\_audio & 2 \\ | |
| train.phase\_b.unfreeze\_top\_blocks\_video & 2 \\ | |
| train.phase\_b.warmup\_ratio & 0.1 \\ | |
| train.phase\_b.weight\_decay & 0.05 \\ | |
| train.qacp.betas & [0.9, 0.999] \\ | |
| train.qacp.early\_stopping\_patience & 8 \\ | |
| train.qacp.effective\_batch & 256 \\ | |
| train.qacp.epochs & 40 \\ | |
| train.qacp.items\_per\_epoch & 4096 \\ | |
| train.qacp.lr & 0.0003 \\ | |
| train.qacp.max\_grad\_norm & 1.0 \\ | |
| train.qacp.micro\_batch & auto \\ | |
| train.qacp.min\_lr\_ratio & 0.01 \\ | |
| train.qacp.nonfinite\_patience & 20 \\ | |
| train.qacp.pseudo\_classes & ["RVRA", "RVFA", "FVRA", "FVFA", "MISMATCH"] \\ | |
| train.qacp.scheduler & warmup\_cosine \\ | |
| train.qacp.selection\_metric & val/qacp\_loss \\ | |
| train.qacp.temperature & 0.1 \\ | |
| train.qacp.warmup\_ratio & 0.05 \\ | |
| train.qacp.weight\_decay & 0.05 \\ | |
| train.stage1.betas & [0.9, 0.999] \\ | |
| train.stage1.early\_stopping\_patience & 6 \\ | |
| train.stage1.effective\_batch & 64 \\ | |
| train.stage1.epochs & 30 \\ | |
| train.stage1.init\_from\_qacp & True \\ | |
| train.stage1.loss\_weights.a & 1.0 \\ | |
| train.stage1.loss\_weights.disentangle & 0.1 \\ | |
| train.stage1.loss\_weights.loc & 0.5 \\ | |
| train.stage1.loss\_weights.quad & 0.5 \\ | |
| train.stage1.loss\_weights.sync & 0.1 \\ | |
| train.stage1.loss\_weights.v & 1.0 \\ | |
| train.stage1.lr & 0.0001 \\ | |
| train.stage1.max\_grad\_norm & 1.0 \\ | |
| train.stage1.micro\_batch & auto \\ | |
| train.stage1.min\_lr\_ratio & 0.01 \\ | |
| train.stage1.modality\_dropout & 0.15 \\ | |
| train.stage1.nonfinite\_patience & 20 \\ | |
| train.stage1.optimizer & adamw \\ | |
| train.stage1.sampler & sqrt\_balanced \\ | |
| train.stage1.sampler\_key & quadrant \\ | |
| train.stage1.scheduler & warmup\_cosine \\ | |
| train.stage1.selection\_metric & val/mean\_auc \\ | |
| train.stage1.warmup\_ratio & 0.05 \\ | |
| train.stage1.weight\_decay & 0.05 \\ | |
| \bottomrule | |
| \end{tabular} | |
| \end{table} | |