refactored keep_all_best

TTS/bin/train_glow_tts.py

@@ -551,8 +551,8 @@ def main(args):  # pylint: disable=redefined-outer-name
         best_loss = torch.load(args.best_path,
                                map_location='cpu')['model_loss']
         print(f" > Starting with loaded last best loss {best_loss}.")
-    keep_best = c.get('keep_best', False)
+    keep_all_best = c.get('keep_all_best', False)
-    keep_after = c.get('keep_after', 10000)  # void if keep_best False
+    keep_after = c.get('keep_after', 10000)  # void if keep_all_best False
 
     # define dataloaders
     train_loader = setup_loader(ap, 1, is_val=False, verbose=True)
@@ -574,7 +574,7 @@ def main(args):  # pylint: disable=redefined-outer-name
             target_loss = eval_avg_loss_dict['avg_loss']
         best_loss = save_best_model(target_loss, best_loss, model, optimizer,
                                     global_step, epoch, c.r, OUT_PATH,
-                                    keep_best=keep_best, keep_after=keep_after)
+                                    keep_all_best=keep_all_best, keep_after=keep_after)
 
 
 if __name__ == '__main__':

TTS/bin/train_speedy_speech.py

@@ -515,8 +515,8 @@ def main(args):  # pylint: disable=redefined-outer-name
         best_loss = torch.load(args.best_path,
                                map_location='cpu')['model_loss']
         print(f" > Starting with loaded last best loss {best_loss}.")
-    keep_best = c.get('keep_best', False)
+    keep_all_best = c.get('keep_all_best', False)
-    keep_after = c.get('keep_after', 10000)  # void if keep_best False
+    keep_after = c.get('keep_after', 10000)  # void if keep_all_best False
 
     # define dataloaders
     train_loader = setup_loader(ap, 1, is_val=False, verbose=True)
@@ -536,7 +536,7 @@ def main(args):  # pylint: disable=redefined-outer-name
             target_loss = eval_avg_loss_dict['avg_loss']
         best_loss = save_best_model(target_loss, best_loss, model, optimizer,
                                     global_step, epoch, c.r, OUT_PATH,
-                                    keep_best=keep_best, keep_after=keep_after)
+                                    keep_all_best=keep_all_best, keep_after=keep_after)
 
 
 if __name__ == '__main__':

TTS/bin/train_tacotron.py

@@ -595,8 +595,8 @@ def main(args):  # pylint: disable=redefined-outer-name
         best_loss = torch.load(args.best_path,
                                map_location='cpu')['model_loss']
         print(f" > Starting with loaded last best loss {best_loss}.")
-    keep_best = c.get('keep_best', False)
+    keep_all_best = c.get('keep_all_best', False)
-    keep_after = c.get('keep_after', 10000)  # void if keep_best False
+    keep_after = c.get('keep_after', 10000)  # void if keep_all_best False
 
     # define data loaders
     train_loader = setup_loader(ap,
@@ -648,7 +648,7 @@ def main(args):  # pylint: disable=redefined-outer-name
             epoch,
             c.r,
             OUT_PATH,
-            keep_best=keep_best,
+            keep_all_best=keep_all_best,
             keep_after=keep_after,
             scaler=scaler.state_dict() if c.mixed_precision else None
         )

TTS/bin/train_vocoder_gan.py

@@ -555,8 +555,8 @@ def main(args):  # pylint: disable=redefined-outer-name
         best_loss = torch.load(args.best_path,
                                map_location='cpu')['model_loss']
         print(f" > Starting with best loss of {best_loss}.")
-    keep_best = c.get('keep_best', False)
+    keep_all_best = c.get('keep_all_best', False)
-    keep_after = c.get('keep_after', 10000)  # void if keep_best False
+    keep_after = c.get('keep_after', 10000)  # void if keep_all_best False
 
     global_step = args.restore_step
     for epoch in range(0, c.epochs):
@@ -581,7 +581,7 @@ def main(args):  # pylint: disable=redefined-outer-name
                                     global_step,
                                     epoch,
                                     OUT_PATH,
-                                    keep_best=keep_best,
+                                    keep_all_best=keep_all_best,
                                     keep_after=keep_after,
                                     model_losses=eval_avg_loss_dict,
                                     )

TTS/bin/train_vocoder_wavegrad.py

@@ -403,8 +403,8 @@ def main(args):  # pylint: disable=redefined-outer-name
         best_loss = torch.load(args.best_path,
                                map_location='cpu')['model_loss']
         print(f" > Starting with loaded last best loss {best_loss}.")
-    keep_best = c.get('keep_best', False)
+    keep_all_best = c.get('keep_all_best', False)
-    keep_after = c.get('keep_after', 10000)  # void if keep_best False
+    keep_after = c.get('keep_after', 10000)  # void if keep_all_best False
 
     global_step = args.restore_step
     for epoch in range(0, c.epochs):
@@ -426,7 +426,7 @@ def main(args):  # pylint: disable=redefined-outer-name
             global_step,
             epoch,
             OUT_PATH,
-            keep_best=keep_best,
+            keep_all_best=keep_all_best,
             keep_after=keep_after,
             model_losses=eval_avg_loss_dict,
             scaler=scaler.state_dict() if c.mixed_precision else None

TTS/bin/train_vocoder_wavernn.py

@@ -426,8 +426,8 @@ def main(args):  # pylint: disable=redefined-outer-name
         best_loss = torch.load(args.best_path,
                                map_location='cpu')['model_loss']
         print(f" > Starting with loaded last best loss {best_loss}.")
-    keep_best = c.get('keep_best', False)
+    keep_all_best = c.get('keep_all_best', False)
-    keep_after = c.get('keep_after', 10000)  # void if keep_best False
+    keep_after = c.get('keep_after', 10000)  # void if keep_all_best False
 
     global_step = args.restore_step
     for epoch in range(0, c.epochs):
@@ -450,7 +450,7 @@ def main(args):  # pylint: disable=redefined-outer-name
             global_step,
             epoch,
             OUT_PATH,
-            keep_best=keep_best,
+            keep_all_best=keep_all_best,
             keep_after=keep_after,
             model_losses=eval_avg_loss_dict,
             scaler=scaler.state_dict() if c.mixed_precision else None

TTS/tts/configs/config.json

@@ -121,8 +121,8 @@
     "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
     "save_step": 10000,      // Number of training steps expected to save traninpg stats and checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

TTS/tts/configs/glow_tts_gated_conv.json

@@ -93,8 +93,8 @@
     "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
     "save_step": 5000,      // Number of training steps expected to save traninpg stats and checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
     "apex_amp_level": null,
 

TTS/tts/configs/glow_tts_ljspeech.json

@@ -105,8 +105,8 @@
         "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
         "save_step": 5000,      // Number of training steps expected to save traninpg stats and checkpoints.
         "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-        "keep_best": false,  // If true, keeps all best_models after keep_after steps
+        "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-        "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+        "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
         "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
         // DATA LOADING

TTS/tts/configs/ljspeech_tacotron2_dynamic_conv_attn.json

@@ -121,8 +121,8 @@
     "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
     "save_step": 10000,      // Number of training steps expected to save traninpg stats and checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

TTS/tts/configs/speedy_speech_ljspeech.json

@@ -109,8 +109,8 @@
    "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
    "save_step": 5000,      // Number of training steps expected to save traninpg stats and checkpoints.
    "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-   "keep_best": false,  // If true, keeps all best_models after keep_after steps
+   "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-   "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+   "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
    "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.:set n
    "mixed_precision": false,
 

TTS/utils/arguments.py

@@ -67,7 +67,7 @@ def parse_arguments(argv):
     return parser.parse_args()
 
 
-def get_last_models(path):
+def get_last_checkpoint(path):
     """Get latest checkpoint or/and best model in path.
 
     It is based on globbing for `*.pth.tar` and the RegEx
@@ -144,7 +144,7 @@ def process_args(args, model_type):
     if args.continue_path:
         args.output_path = args.continue_path
         args.config_path = os.path.join(args.continue_path, "config.json")
-        args.restore_path, best_model = get_last_models(args.continue_path)
+        args.restore_path, best_model = get_last_checkpoint(args.continue_path)
         if not args.best_path:
             args.best_path = best_model
 

TTS/vocoder/configs/multiband-melgan_and_rwd_config.json

@@ -138,8 +138,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

TTS/vocoder/configs/multiband_melgan_config.json

@@ -128,8 +128,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

TTS/vocoder/configs/multiband_melgan_config_mozilla.json

@@ -141,8 +141,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

TTS/vocoder/configs/parallel_wavegan_config.json

@@ -130,8 +130,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

TTS/vocoder/configs/universal_fullband_melgan.json

@@ -124,8 +124,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

TTS/vocoder/configs/wavegrad_libritts.json

@@ -103,8 +103,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 5000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": true,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

TTS/vocoder/configs/wavernn_config.json

@@ -89,8 +89,8 @@
     "print_eval": false, // If True, it prints loss values for each step in eval run.
     "save_step": 25000, // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true, // If true, it saves checkpoints per "save_step"
-    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": false,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false, // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
 // DATA LOADING

TTS/vocoder/utils/io.py

@@ -64,7 +64,7 @@ def save_checkpoint(model, optimizer, scheduler, model_disc, optimizer_disc,
 
 def save_best_model(current_loss, best_loss, model, optimizer, scheduler,
                     model_disc, optimizer_disc, scheduler_disc, current_step,
-                    epoch, out_path, keep_best=False, keep_after=10000,
+                    epoch, out_path, keep_all_best=False, keep_after=10000,
                     **kwargs):
     if current_loss < best_loss:
         best_model_name = f'best_model_{current_step}.pth.tar'
@@ -82,7 +82,7 @@ def save_best_model(current_loss, best_loss, model, optimizer, scheduler,
                    model_loss=current_loss,
                    **kwargs)
         # only delete previous if current is saved successfully
-        if not keep_best or (current_step < keep_after):
+        if not keep_all_best or (current_step < keep_after):
             model_names = glob.glob(
                 os.path.join(out_path, 'best_model*.pth.tar'))
             for model_name in model_names:

tests/inputs/test_glow_tts.json

@@ -106,8 +106,8 @@
     "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
     "save_step": 5000,      // Number of training steps expected to save traninpg stats and checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": true,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": true,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
     "apex_amp_level": null,
 

tests/inputs/test_speedy_speech.json

@@ -111,8 +111,8 @@
    "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
    "save_step": 5000,      // Number of training steps expected to save traninpg stats and checkpoints.
    "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-   "keep_best": true,  // If true, keeps all best_models after keep_after steps
+   "keep_all_best": true,  // If true, keeps all best_models after keep_after steps
-   "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+   "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
    "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.:set n
    "mixed_precision": false,
 

tests/inputs/test_train_config.json

@@ -122,8 +122,8 @@
     "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
     "save_step": 10000,      // Number of training steps expected to save traninpg stats and checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": true,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": true,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

tests/inputs/test_vocoder_multiband_melgan_config.json

@@ -131,8 +131,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": true,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": true,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

tests/inputs/test_vocoder_wavegrad.json

@@ -101,8 +101,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 10000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": true,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": true,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": true,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING

tests/inputs/test_vocoder_wavernn_config.json

@@ -97,8 +97,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "keep_best": true,  // If true, keeps all best_models after keep_after steps
+    "keep_all_best": true,  // If true, keeps all best_models after keep_after steps
-    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "keep_after": 10000,    // Global step after which to keep best models if keep_all_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING