Upload 4 files

1118_muqmulan/ckpt/full_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c4609e02573014b08e7a73e9a883abdd6615b9713e6c0c959904dc4d777029b
+size 2735874200

1118_muqmulan/config.yaml

@@ -0,0 +1,144 @@
+DEVICES: '7'
+accelerate:
+  mixed_precision: bf16
+basics:
+  random_seed: 42
+  save_dir: /data/yrb/musicarena/Haiwen/offline_data/cmi-arena/experiments/reward_model
+dataset:
+  audio_dropout:
+    apply_to_eval: true
+    apply_to_ref: true
+    enabled: true
+    eval_only_on_training: true
+    max_duration: 1000
+    min_duration: 250
+  cache_dir: /data/yrb/musicarena/Haiwen/offline_data/cmi-arena/tmp
+  db_path: /data/yrb/musicarena/Haiwen/offline_data/cmi-arena/backend/database.db
+  duration: 600.0
+  embedding_dir: /data/yrb/musicarena/Haiwen/offline_data/cmi-arena/supervised_embeddings
+  max_samples: null
+  max_val_samples: null
+  preference_file: /data/yrb/musicarena/Haiwen/offline_data/cmi-arena/CMI-Training/unbiased_qwen/train.json
+  sample_rate: 24000
+  use_preextracted: true
+  val_preference_file: null
+loss:
+  IF_ratio: 0.5
+  filter_ties: true
+  label_smoothing: 0.0
+  reduction: mean
+model:
+  attention_mode: CA
+  attn_dropout: 0.0
+  category_embeddings: null
+  dim: 768
+  dim_head: 64
+  downsample:
+    configs:
+      conv2_4x:
+        factor: 4
+        kernel_size: 5
+        kind: conv*2
+        use_layernorm: true
+      conv_4x:
+        factor: 4
+        kernel_size: 5
+        kind: conv
+        stage: 1
+        use_layernorm: true
+      glu_4x:
+        factor: 4
+        kernel_size: 5
+        kind: gluconv*2+pw
+        use_layernorm: true
+      mean:
+        factor: 2
+        kind: mean
+      mean_4x:
+        dropout: 0.0
+        factor: 30
+        kind: mean+mlp
+        mlp_ratio: 2.0
+      none:
+        factor: 1
+        kind: none
+    eval: mean_4x
+    ref: null
+    text: none
+  ff_dropout: 0.0
+  ff_mult: 4
+  freeze_audio: true
+  freeze_text: true
+  gradient_checkpointing: false
+  heads: 8
+  joint_tf_depth: 1
+  load_config:
+    checkpoint_path: null
+    frozen_from_pretrained: true
+    pretrained_name: OpenMuQ/MuQ-MuLan-large
+    strict: false
+  mlp_dim: 768
+  mode: text_only
+  model_name: OpenMuQ/MuQ-MuLan-large
+  name: reward
+  null_embedding:
+    audio:
+      dropout: 0.5
+      length: 10
+    lyrics:
+      dropout: 0.3
+      length: 10
+    text:
+      dropout: 0.2
+      length: 10
+  output_dim: 2
+  prompt_tf_depth: 1
+  sr: 24000
+  text_encoder:
+    name: muq_mulan
+    tune: null
+  text_lora_config: null
+  train_muq_depth: 0
+  train_muqmulan: false
+  use_layer_idx: -1
+project_root: /data/yrb/musicarena/Haiwen/offline_data/cmi-arena
+run_name: null
+train:
+  batch_size: 24
+  betas:
+  - 0.9
+  - 0.99
+  ema_decay: 0.9999
+  ema_update_every: 1
+  enable_gradient_checkpointing: true
+  force_clear_prev_results: false
+  grad_accum_every: 2
+  log_tensorboard: true
+  lr_schedule:
+    min_lr_ratio: 0.001
+    name: linear_cosine
+    total_steps: 30000
+    warmup_steps: 300
+  max_grad_norm: 100
+  mlp_lr: 0.0002
+  num_train_steps: 30000
+  num_valid_batches: 10
+  num_workers: 8
+  other_lr: null
+  resume: /data/yrb/musicarena/Haiwen/offline_data/cmi-arena/experiments/contrastive/20260118_0146/ckpt/reward_model.best_15760.pt
+  resume_optimizer: false
+  save_model_every: 2000
+  use_checkpoint_config: false
+  use_ema: true
+  use_lion: false
+  valid_batch_size: 20
+  valid_every: 2000
+  valid_frac: 0.1
+  verify_weights_on_load: true
+validate:
+  checkpoint: null
+  dir: null
+  max_failure_cases: 30
+  num_batches: null
+  only: false
+  trust_checkpoint: true