| Parameter,Value | |
| 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 | |