Refactor code and configs, rename modules, adjust default LR schedule

configs/acoustic.yaml

@@ -63,21 +63,28 @@ use_tension_embed: false
 use_key_shift_embed: false
 use_speed_embed: false
 
+diffusion_type: reflow
+time_scale_factor: 1000
 timesteps: 1000
 max_beta: 0.02
 rel_pos: true
+sampling_algorithm: euler
+sampling_steps: 20
 diff_accelerator: ddim
 diff_speedup: 10
 hidden_size: 256
 residual_layers: 20
 residual_channels: 512
 dilation_cycle_length: 4  # *
-diff_decoder_type: 'wavenet'
-diff_loss_type: l2
+backbone_type: 'wavenet'
+main_loss_type: l2
+main_loss_log_norm: false
 schedule_type: 'linear'
 
 # shallow diffusion
 use_shallow_diffusion: true
+T_start: 0.4
+T_start_infer: 0.4
 K_step: 400
 K_step_infer: 400
 
@@ -100,20 +107,19 @@ num_sanity_val_steps: 1
 optimizer_args:
   lr: 0.0006
 lr_scheduler_args:
-  step_size: 30000
-  gamma: 0.5
+  step_size: 10000
+  gamma: 0.75
 max_batch_frames: 50000
 max_batch_size: 64
 dataset_size_key: 'lengths'
 val_with_vocoder: true
 val_check_interval: 2000
 num_valid_plots: 10
-max_updates: 200000
+max_updates: 160000
 num_ckpt_keep: 5
-permanent_ckpt_start: 120000
+permanent_ckpt_start: 80000
 permanent_ckpt_interval: 20000
 
-
 finetune_enabled: false
 finetune_ckpt_path: null
 

configs/templates/config_acoustic.yaml

@@ -53,7 +53,10 @@ residual_channels: 512
 residual_layers: 20
 
 # shallow diffusion
+diffusion_type: reflow
 use_shallow_diffusion: true
+T_start: 0.4
+T_start_infer: 0.4
 K_step: 300
 K_step_infer: 300
 shallow_diffusion_args:
@@ -73,11 +76,11 @@ optimizer_args:
   lr: 0.0006
 lr_scheduler_args:
   scheduler_cls: torch.optim.lr_scheduler.StepLR
-  step_size: 30000
-  gamma: 0.5
+  step_size: 10000
+  gamma: 0.75
 max_batch_frames: 50000
 max_batch_size: 64
-max_updates: 200000
+max_updates: 160000
 
 num_valid_plots: 10
 val_with_vocoder: true

configs/templates/config_variance.yaml

@@ -65,6 +65,8 @@ use_glide_embed: false
 glide_types: [up, down]
 glide_embed_scale: 11.313708498984760  # sqrt(128)
 
+diffusion_type: reflow
+
 pitch_prediction_args:
   pitd_norm_min: -8.0
   pitd_norm_max: 8.0
@@ -89,16 +91,16 @@ optimizer_args:
   lr: 0.0006
 lr_scheduler_args:
   scheduler_cls: torch.optim.lr_scheduler.StepLR
-  step_size: 12000
+  step_size: 10000
   gamma: 0.75
 max_batch_frames: 80000
 max_batch_size: 48
-max_updates: 288000
+max_updates: 160000
 
 num_valid_plots: 10
 val_check_interval: 2000
 num_ckpt_keep: 5
-permanent_ckpt_start: 180000
+permanent_ckpt_start: 80000
 permanent_ckpt_interval: 10000
 pl_trainer_devices: 'auto'
 pl_trainer_precision: '16-mixed'

configs/variance.yaml

@@ -97,12 +97,17 @@ lambda_dur_loss: 1.0
 lambda_pitch_loss: 1.0
 lambda_var_loss: 1.0
 
+diffusion_type: reflow  # ddpm
+time_scale_factor: 1000
 schedule_type: 'linear'
 K_step: 1000
 timesteps: 1000
 max_beta: 0.02
-diff_decoder_type: 'wavenet'
-diff_loss_type: l2
+backbone_type: 'wavenet'
+main_loss_type: l2
+main_loss_log_norm: true
+sampling_algorithm: euler
+sampling_steps: 20
 diff_accelerator: ddim
 diff_speedup: 10
 
@@ -111,16 +116,16 @@ num_sanity_val_steps: 1
 optimizer_args:
   lr: 0.0006
 lr_scheduler_args:
-  step_size: 12000
+  step_size: 10000
   gamma: 0.75
 max_batch_frames: 80000
 max_batch_size: 48
 dataset_size_key: 'lengths'
 val_check_interval: 2000
 num_valid_plots: 10
-max_updates: 288000
+max_updates: 160000
 num_ckpt_keep: 5
-permanent_ckpt_start: 180000
+permanent_ckpt_start: 80000
 permanent_ckpt_interval: 10000
 
 finetune_enabled: false