Upload folder using huggingface_hub

models/dnr/locoformer-medium.enc-crossattn96dim.dec-crossattn96dim.musical64.learnable-query/checkpoints/model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b878ea0ffe29c67401727d51cdaab8db0394178d8e0eafa9d2ca5e241fd472e2
+size 128207890

models/dnr/locoformer-medium.enc-crossattn96dim.dec-crossattn96dim.musical64.learnable-query/merged_config.yaml

@@ -0,0 +1,167 @@
+seed: 2026
+
+sr: 44100
+n_fft: 2048
+hop_length: 512
+
+n_src: 3
+n_chan: 1
+emb_dim: 128
+
+trainer:
+  max_epochs: 150
+  check_val_every_n_epoch: 1
+  gradient_clip_val: 5.0
+  sync_batchnorm: True
+  deterministic: false
+
+  limit_train_batches: 1354
+  limit_val_batches: 10  # x 8 GPUs
+
+  benchmark: True
+  use_distributed_sampler: False
+  precision: bf16-mixed
+
+  callbacks:
+    - _target_: lightning.pytorch.callbacks.ModelCheckpoint
+      filename: "{epoch:04d}"
+      save_top_k: 5
+      monitor: validation/loss
+      mode: min
+      save_last: True
+    - _target_: lightning.pytorch.callbacks.RichProgressBar
+      refresh_rate: 5
+    - _target_: lightning.pytorch.callbacks.RichModelSummary
+      max_depth: 3
+
+datamodule:
+  _target_: spectral_feature_compression.common.datamodules.hdf5_wav_datamodule.DataModule
+  train_dataset_path: 
+    - ./recipes/dnr/hdf5/tr_speech_seg8sec_sad.hdf5
+    - ./recipes/dnr/hdf5/tr_music_seg8sec_sad.hdf5
+    - ./recipes/dnr/hdf5/tr_sfx_seg8sec_sad.hdf5
+  val_dataset_path: ./recipes/dnr/hdf5/cv_unsegmented.hdf5
+  batch_size: 4
+  val_batch_size: 1
+  sr: ${sr}
+  duration: 6
+  val_duration: null
+  num_workers: 4
+  return_ref: true
+  use_scatter_dataset: false
+  use_dm_dataset: true
+  p_source_dropout: 0.05
+  val_drop_last: false
+
+task:
+  _target_: spectral_feature_compression.core.tasks.sup_task.SupTask
+  n_fft: ${n_fft}
+  hop_length: ${hop_length}
+  css_validation: true
+
+  ema_weight: 0.999
+  ema_update_freq: 1
+
+  model:
+    _target_: spectral_feature_compression.core.model.model_wrapper.ModelWrapper
+    n_fft: ${n_fft}
+    hop_length: ${hop_length}
+    fs: ${sr}
+    scaling: true
+
+    css_segment_size: 12
+    css_shift_size: 6
+
+    model:
+      _target_: spectral_feature_compression.core.model.bslocoformer.BSLocoformer
+      n_src: ${n_src}
+      n_chan: ${n_chan}
+
+      n_layers: 6
+      # general setup
+      emb_dim: ${emb_dim}
+      norm_type: rmsgroupnorm
+      num_groups: 8  # used only in RMSGroupNorm
+      tf_order: ft
+      # self-attention related
+      n_heads: 8
+      flash_attention: true  # available when using mixed precision
+      attention_dim: 128
+      # ffn related
+      ffn_type: 
+        - swiglu_conv1d
+        - swiglu_conv1d
+      ffn_hidden_dim: 
+        - 192
+        - 192
+      conv1d_kernel: 8
+      conv1d_shift: 1
+      dropout: 0.1
+      # others
+      eps: 1.0e-5
+
+      encoder:
+        _target_: spectral_feature_compression.core.model.crossattn_enc_dec.CrossAttnEncoder
+        # general SFC config
+        d_inner: &d_inner 96
+        d_model: ${emb_dim}
+        n_chan: ${n_chan}
+        sample_rate: ${sr}
+        n_fft: ${n_fft}
+        n_bands: &n_bands 64
+        band_config: &band_config musical
+        query_type: &query_type learnable
+        # cross-attention related
+        n_heads: &n_heads 4
+        slope: &slope  [1, 1, 1, 1]
+        learnable_slope: &learnable_slope false
+        learnable_pos_bias: &learnable_pos_bias true
+        mask_outside_bands: &mask_outside_bands false
+        
+
+      decoder:
+        _target_: spectral_feature_compression.core.model.crossattn_enc_dec.CrossAttnDecoder
+        # general SFC config
+        d_inner: *d_inner
+        d_model: ${emb_dim}
+        n_src: ${n_src}
+        n_chan: ${n_chan}
+        sample_rate: ${sr}
+        n_fft: ${n_fft}
+        n_bands: *n_bands
+        band_config: *band_config
+        query_type: *query_type
+        # cross-attention related
+        n_heads: *n_heads
+        slope: *slope
+        learnable_slope: *learnable_slope
+        learnable_pos_bias: *learnable_pos_bias
+        mask_outside_bands: *mask_outside_bands
+ 
+  loss:
+    _target_: spectral_feature_compression.core.loss.snr.ThresSNRLossWithInactiveSource
+    solve_perm: false
+    n_src: ${n_src}
+    zeroref_weight: 0.1
+    only_denominator: false
+
+
+  optimizer_config:
+    _target_: aiaccel.torch.lightning.OptimizerConfig
+    optimizer_generator:
+      _partial_: True
+      _target_: torch.optim.AdamW
+      lr: 1.e-3
+      weight_decay: 1.e-2
+      betas:
+        - 0.9
+        - 0.999
+    scheduler_generator:
+      _partial_: True
+      _target_: spectral_feature_compression.common.lr_schedulers.warmup.WarmUpStepLR
+      warmup_steps: 5000
+      decay_start_step: 101550  # 1354 steps x 75 epochs
+      decay_stop_step: 203100  # 1354 steps x 150 epochs
+      step_size: 1354  # 1 epoch
+      decay: 0.96
+    scheduler_interval: step

models/dnr/locoformer-small.enc-crossattn64dim.dec-crossattn64dim.musical64.learnable-query/checkpoints/model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d648ccdc51ad59dd175a9df2269e9de1924f143757da6c2a6392e4e64a4babd3
+size 46661234

models/dnr/locoformer-small.enc-crossattn64dim.dec-crossattn64dim.musical64.learnable-query/merged_config.yaml

@@ -0,0 +1,167 @@
+seed: 2026
+
+sr: 44100
+n_fft: 2048
+hop_length: 512
+
+n_src: 3
+n_chan: 1
+emb_dim: 96
+
+trainer:
+  max_epochs: 150
+  check_val_every_n_epoch: 1
+  gradient_clip_val: 5.0
+  sync_batchnorm: True
+  deterministic: false
+
+  limit_train_batches: 1354
+  limit_val_batches: 10  # x 8 GPUs
+
+  benchmark: True
+  use_distributed_sampler: False
+  precision: bf16-mixed
+
+  callbacks:
+    - _target_: lightning.pytorch.callbacks.ModelCheckpoint
+      filename: "{epoch:04d}"
+      save_top_k: 5
+      monitor: validation/loss
+      mode: min
+      save_last: True
+    - _target_: lightning.pytorch.callbacks.RichProgressBar
+      refresh_rate: 5
+    - _target_: lightning.pytorch.callbacks.RichModelSummary
+      max_depth: 3
+
+datamodule:
+  _target_: spectral_feature_compression.common.datamodules.hdf5_wav_datamodule.DataModule
+  train_dataset_path: 
+    - ./recipes/dnr/hdf5/tr_speech_seg8sec_sad.hdf5
+    - ./recipes/dnr/hdf5/tr_music_seg8sec_sad.hdf5
+    - ./recipes/dnr/hdf5/tr_sfx_seg8sec_sad.hdf5
+  val_dataset_path: ./recipes/dnr/hdf5/cv_unsegmented.hdf5
+  batch_size: 4
+  val_batch_size: 1
+  sr: ${sr}
+  duration: 6
+  val_duration: null
+  num_workers: 4
+  return_ref: true
+  use_scatter_dataset: false
+  use_dm_dataset: true
+  p_source_dropout: 0.05
+  val_drop_last: false
+
+task:
+  _target_: spectral_feature_compression.core.tasks.sup_task.SupTask
+  n_fft: ${n_fft}
+  hop_length: ${hop_length}
+  css_validation: true
+
+  ema_weight: 0.999
+  ema_update_freq: 1
+
+  model:
+    _target_: spectral_feature_compression.core.model.model_wrapper.ModelWrapper
+    n_fft: ${n_fft}
+    hop_length: ${hop_length}
+    fs: ${sr}
+    scaling: true
+
+    css_segment_size: 12
+    css_shift_size: 6
+
+    model:
+      _target_: spectral_feature_compression.core.model.bslocoformer.BSLocoformer
+      n_src: ${n_src}
+      n_chan: ${n_chan}
+
+      n_layers: 4
+      # general setup
+      emb_dim: ${emb_dim}
+      norm_type: rmsgroupnorm
+      num_groups: 4  # used only in RMSGroupNorm
+      tf_order: ft
+      # self-attention related
+      n_heads: 4
+      flash_attention: true  # available when using mixed precision
+      attention_dim: 96
+      # ffn related
+      ffn_type: 
+        - swiglu_conv1d
+        - swiglu_conv1d
+      ffn_hidden_dim: 
+        - 128
+        - 128
+      conv1d_kernel: 8
+      conv1d_shift: 1
+      dropout: 0.1
+      # others
+      eps: 1.0e-5
+
+      encoder:
+        _target_: spectral_feature_compression.core.model.crossattn_enc_dec.CrossAttnEncoder
+        # general SFC config
+        d_inner: &d_inner 64
+        d_model: ${emb_dim}
+        n_chan: ${n_chan}
+        sample_rate: ${sr}
+        n_fft: ${n_fft}
+        n_bands: &n_bands 64
+        band_config: &band_config musical
+        query_type: &query_type learnable
+        # cross-attention related
+        n_heads: &n_heads 4
+        slope: &slope  [1, 1, 1, 1]
+        learnable_slope: &learnable_slope false
+        learnable_pos_bias: &learnable_pos_bias true
+        mask_outside_bands: &mask_outside_bands false
+        
+
+      decoder:
+        _target_: spectral_feature_compression.core.model.crossattn_enc_dec.CrossAttnDecoder
+        # general SFC config
+        d_inner: *d_inner
+        d_model: ${emb_dim}
+        n_src: ${n_src}
+        n_chan: ${n_chan}
+        sample_rate: ${sr}
+        n_fft: ${n_fft}
+        n_bands: *n_bands
+        band_config: *band_config
+        query_type: *query_type
+        # cross-attention related
+        n_heads: *n_heads
+        slope: *slope
+        learnable_slope: *learnable_slope
+        learnable_pos_bias: *learnable_pos_bias
+        mask_outside_bands: *mask_outside_bands
+ 
+  loss:
+    _target_: spectral_feature_compression.core.loss.snr.ThresSNRLossWithInactiveSource
+    solve_perm: false
+    n_src: ${n_src}
+    zeroref_weight: 0.1
+    only_denominator: false
+
+
+  optimizer_config:
+    _target_: aiaccel.torch.lightning.OptimizerConfig
+    optimizer_generator:
+      _partial_: True
+      _target_: torch.optim.AdamW
+      lr: 1.e-3
+      weight_decay: 1.e-2
+      betas:
+        - 0.9
+        - 0.999
+    scheduler_generator:
+      _partial_: True
+      _target_: spectral_feature_compression.common.lr_schedulers.warmup.WarmUpStepLR
+      warmup_steps: 5000
+      decay_start_step: 101550  # 1354 steps x 75 epochs
+      decay_stop_step: 203100  # 1354 steps x 150 epochs
+      step_size: 1354  # 1 epoch
+      decay: 0.96
+    scheduler_interval: step