model: embed_dim: 384 depth: 8 num_heads: 6 mlp_ratio: 4 target_frames: 1024 mel_bins: 64 patch_time: 16 patch_freq: 16 decoder_layers: 6 decoder_groups: 16 decoder_kernel: 3 objective: kind: mae mae_norm_pix: true num_classes: 527 clone_batch: 1 mask_prob: 0.8 mask_length: 5 mask_prob_adjust: 0.07 average_top_k_layers: 0 utterance_weight: 1.0 normalize_utterance_target: false ema_decay: 0.9998 ema_end_decay: 0.99999 ema_anneal_steps: null data: data_dir: /workspace/data/asmel_flat train_split: train val_split: balanced_train revision: '' batch_size: 48 num_workers: 12 prefetch_factor: 4 pin_memory: true persistent_workers: true drop_last: true seed: 0 max_shards: 0 mixture: cluster_index: /workspace/data/asmel_flat/train_cluster_index.npy distributions: null dist_id: null weights: - 0.0 - 0.015146484375 - 0.37251171875 - 0.0 - 0.25891015625 - 0.001552734375 - 0.058642578125 - 0.001099609375 - 0.000388671875 - 0.021359375 - 0.011974609375 - 0.114892578125 - 0.00032421875 - 0.028673828125 - 0.006279296875 - 0.0 - 0.000388671875 - 0.00213671875 - 0.0 - 0.10571875 repetition_cap: 1000.0 distinct_fraction: null optim: lr: 0.0003 betas: - 0.9 - 0.95 eps: 1.0e-08 weight_decay: 0.05 clip_norm: 4.0 warmup_frac: 0.1 min_lr_scale: 0.002 loop: budget_clips: 1920000 schedule_clips: null gradient_accumulation_steps: 1 precision: bfloat16 compile: true compile_mode: default log_every_n_steps: 20 heartbeat_every_n_steps: 50 max_steps: null eval: enabled: true every_n_steps: 1000 num_clips: 2048 mask_seed: 12345 batch_size: 48 checkpoint: save_every_n_steps: 5000 export_every_n_steps: 5000 keep_last_n_exports: 0 tracking: campaign_id: phase5 trial_id: p5-mae-15m-d214 wandb_project: mixture-transfer wandb_entity: null wandb_mode: online wandb_run_id: null hub_repo_id: null hub_private: false hub_upload: false runtime: seed: 0 data_loader_seed: null objective_seed: null deterministic: false allow_tf32: true