Persephone asr refactor

egs/librispeech/v1/conf/reverb_noise20dB_aug.yaml

@@ -0,0 +1,34 @@
+reverb_aug:
+  reverb_prob: 0.45
+  max_reverb_context: 0.5
+  rir_types: 
+    smallroom:
+      weight: 1
+      rir_path: scp:data/rirs_smallroom/rirs.scp
+      rir_norm: max
+    mediumroom:
+      weight: 1
+      rir_path: scp:data/rirs_mediumroom/rirs.scp
+      rir_norm: max
+    realroom:
+      weight: 1
+      rir_path: scp:data/rirs_real/rirs.scp
+      rir_norm: max
+noise_aug:
+  noise_prob: 0.7
+  noise_types: 
+    noise:
+      weight: 1
+      noise_path: data/musan_noise_proc_audio/wav.scp
+      min_snr: 10
+      max_snr: 20
+    music:
+      weight: 1
+      noise_path: data/musan_music_proc_audio/wav.scp
+      min_snr: 10
+      max_snr: 20
+    babble:
+      weight: 1
+      noise_path: data/musan_speech_babble/wav.scp
+      min_snr: 10
+      max_snr: 20

egs/librispeech/v1/conf/train_wav2vec2base_conf_rnnt_k2_pruned_stage1_v1.2.yaml

@@ -0,0 +1,74 @@
+data:
+  train:
+    dataset:
+      wav_scale: 1
+      aug_cfgs: 
+        - conf/reverb_noise20dB_aug.yaml
+      return_segment_info:
+        - text
+    sampler:
+      sampler_type: bucketing_seg_sampler
+      max_batch_length: 28.
+      min_batch_size: 1
+      drop_last: false
+    data_loader:
+      num_workers: 4
+  val:
+    dataset:
+      aug_cfgs: 
+        - conf/reverb_noise20dB_aug.yaml
+      wav_scale: 1
+      return_segment_info:
+        - text
+    sampler:
+      sampler_type: bucketing_seg_sampler
+      max_batch_length: 28.
+      min_batch_size: 1
+      drop_last: true
+    data_loader:
+      num_workers: 4
+model: 
+  hf_feats:
+    pretrained_model_path: facebook/wav2vec2-base-960h
+  transducer:
+    encoder:
+      d_model: 512
+      num_heads: 8
+      num_blocks: 1
+      d_ff: 2048
+      in_layer_type: linear
+    decoder:
+      rnnt_loss: k2_pruned
+      simple_loss_scale: 0.2
+      predictor:
+        embed_dim: 1024
+        num_layers: 2
+        hid_feats: 512
+        embed_dropout_rate: 0.4
+        rnn_dropout_rate: 0.4
+        rnn_type: lstm
+      joiner:
+        hid_feats: 512
+  feat_fusion_method: weighted-avg
+  feat_fusion_start: 2
+trainer:
+  optim:
+    opt_type: adamw
+    lr: 0.001
+    beta1: 0.9
+    beta2: 0.99
+    weight_decay: 4e-4
+  lrsched:
+    lrsch_type: exp_lr
+    decay_rate: 0.5
+    decay_steps: 4200
+    hold_steps: 1500
+    min_lr: 4e-5
+    warmup_steps: 1500
+    update_lr_on_opt_step: true
+  grad_clip: 100
+  use_amp: true
+  log_interval: 1000
+  epochs: 120
+  eff_batch_size: 128
+  train_mode: hf-feats-frozen-nograd

egs/librispeech/v1/conf/train_wav2vec2base_conf_rnnt_k2_pruned_stage1_v3.0.yaml

@@ -0,0 +1,76 @@
+data:
+  train:
+    dataset:
+      wav_scale: 1
+      aug_cfgs: 
+        - conf/reverb_noise20dB_aug.yaml
+      return_segment_info:
+        - text
+    sampler:
+      sampler_type: bucketing_seg_sampler
+      max_batch_length: 28.
+      min_batch_size: 1
+      drop_last: false
+    data_loader:
+      num_workers: 4
+  val:
+    dataset:
+      aug_cfgs: 
+        - conf/reverb_noise20dB_aug.yaml
+      wav_scale: 1
+      return_segment_info:
+        - text
+    sampler:
+      sampler_type: bucketing_seg_sampler
+      max_batch_length: 28.
+      min_batch_size: 1
+      drop_last: true
+    data_loader:
+      num_workers: 4
+model: 
+  hf_feats:
+    pretrained_model_path: facebook/wav2vec2-base-960h
+  transducer:
+    encoder:
+      att_type: local-scaled-dot-prod-v1
+      att_context: 64
+      d_model: 512
+      num_heads: 8
+      num_blocks: 1
+      d_ff: 2048
+      in_layer_type: linear
+    decoder:
+      rnnt_loss: k2_pruned
+      simple_loss_scale: 0.2
+      predictor:
+        embed_dim: 1024
+        num_layers: 2
+        hid_feats: 512
+        embed_dropout_rate: 0.4
+        rnn_dropout_rate: 0.4
+        rnn_type: lstm
+      joiner:
+        hid_feats: 512
+  feat_fusion_method: weighted-avg
+  feat_fusion_start: 2
+trainer:
+  optim:
+    opt_type: adamw
+    lr: 0.001
+    beta1: 0.9
+    beta2: 0.99
+    weight_decay: 4e-4
+  lrsched:
+    lrsch_type: exp_lr
+    decay_rate: 0.5
+    decay_steps: 4200
+    hold_steps: 1500
+    min_lr: 4e-5
+    warmup_steps: 1500
+    update_lr_on_opt_step: true
+  grad_clip: 100
+  use_amp: true
+  log_interval: 1000
+  epochs: 120
+  eff_batch_size: 128
+  train_mode: hf-feats-frozen-nograd

egs/librispeech/v1/conf/train_wav2vec2base_conf_rnnt_k2_pruned_stage1_v3.1.yaml

@@ -0,0 +1,77 @@
+data:
+  train:
+    dataset:
+      wav_scale: 1
+      aug_cfgs: 
+        - conf/reverb_noise20dB_aug.yaml
+      return_segment_info:
+        - text
+    sampler:
+      sampler_type: bucketing_seg_sampler
+      max_batch_length: 28.
+      min_batch_size: 1
+      drop_last: false
+    data_loader:
+      num_workers: 4
+  val:
+    dataset:
+      aug_cfgs: 
+        - conf/reverb_noise20dB_aug.yaml
+      wav_scale: 1
+      return_segment_info:
+        - text
+    sampler:
+      sampler_type: bucketing_seg_sampler
+      max_batch_length: 28.
+      min_batch_size: 1
+      drop_last: true
+    data_loader:
+      num_workers: 4
+model: 
+  hf_feats:
+    pretrained_model_path: facebook/wav2vec2-base-960h
+  transducer:
+    encoder:
+      att_type: local-scaled-dot-prod-v1
+      att_context: 64
+      d_model: 512
+      num_heads: 8
+      num_blocks: 1
+      d_ff: 2048
+      in_layer_type: linear
+      pos_enc_type: abs
+    decoder:
+      rnnt_loss: k2_pruned
+      simple_loss_scale: 0.2
+      predictor:
+        embed_dim: 1024
+        num_layers: 2
+        hid_feats: 512
+        embed_dropout_rate: 0.4
+        rnn_dropout_rate: 0.4
+        rnn_type: lstm
+      joiner:
+        hid_feats: 512
+  feat_fusion_method: weighted-avg
+  feat_fusion_start: 2
+trainer:
+  optim:
+    opt_type: adamw
+    lr: 0.001
+    beta1: 0.9
+    beta2: 0.99
+    weight_decay: 4e-4
+  lrsched:
+    lrsch_type: exp_lr
+    decay_rate: 0.5
+    decay_steps: 4200
+    hold_steps: 1500
+    min_lr: 4e-5
+    warmup_steps: 1500
+    update_lr_on_opt_step: true
+  grad_clip: 100
+  use_amp: true
+  log_interval: 1000
+  epochs: 120
+  eff_batch_size: 128
+  train_mode: hf-feats-frozen-nograd

egs/librispeech/v1/conf/train_wav2vec2base_conf_rnnt_k2_pruned_stage1_v3.10.yaml

@@ -0,0 +1,76 @@
+data:
+  train:
+    dataset:
+      wav_scale: 1
+      aug_cfgs: 
+        - conf/reverb_noise20dB_aug.yaml
+      return_segment_info:
+        - text
+    sampler:
+      sampler_type: bucketing_seg_sampler
+      max_batch_length: 28.
+      min_batch_size: 1
+      drop_last: false
+    data_loader:
+      num_workers: 4
+  val:
+    dataset:
+      aug_cfgs: 
+        - conf/reverb_noise20dB_aug.yaml
+      wav_scale: 1
+      return_segment_info:
+        - text
+    sampler:
+      sampler_type: bucketing_seg_sampler
+      max_batch_length: 28.
+      min_batch_size: 1
+      drop_last: true
+    data_loader:
+      num_workers: 4
+model: 
+  hf_feats:
+    pretrained_model_path: facebook/wav2vec2-base-960h
+  transducer:
+    encoder:
+      att_type: block-scaled-dot-prod-v1
+      att_context: 1
+      d_model: 512
+      num_heads: 8
+      num_blocks: 1
+      d_ff: 2048
+      in_layer_type: linear
+    decoder:
+      rnnt_loss: k2_pruned
+      simple_loss_scale: 0.2
+      predictor:
+        embed_dim: 1024
+        num_layers: 2
+        hid_feats: 512
+        embed_dropout_rate: 0.4
+        rnn_dropout_rate: 0.4
+        rnn_type: lstm
+      joiner:
+        hid_feats: 512
+  feat_fusion_method: weighted-avg
+  feat_fusion_start: 2
+trainer:
+  optim:
+    opt_type: adamw
+    lr: 0.001
+    beta1: 0.9
+    beta2: 0.99
+    weight_decay: 4e-4
+  lrsched:
+    lrsch_type: exp_lr
+    decay_rate: 0.5
+    decay_steps: 4200
+    hold_steps: 1500
+    min_lr: 4e-5
+    warmup_steps: 1500
+    update_lr_on_opt_step: true
+  grad_clip: 100
+  use_amp: true
+  log_interval: 1000
+  epochs: 120
+  eff_batch_size: 128
+  train_mode: hf-feats-frozen-nograd