update preprocessors and utils

config/VCTK/preprocess.yaml

@@ -1,7 +1,7 @@
 dataset: "VCTK"
 
 path:
-  corpus_path: "/mnt/nfs2/speech-datasets/en/VCTK-Corpus-92"
+  corpus_path: "/mnt/nfs2/speech-datasets/en/VCTK-Corpus-0.92"
   wav_tag: "mic1"
   wav_dir: "wav48_silence_trimmed"
   txt_dir: "txt"

preprocessor/ljspeech.py

@@ -40,13 +40,27 @@ def __init__(self, config):
             config["preprocessing"]["mel"]["mel_fmin"],
             config["preprocessing"]["mel"]["mel_fmax"],
         )
+        self.val_prior = self.val_prior_names(os.path.join(self.out_dir, "val.txt"))
+
+    def val_prior_names(self, val_prior_path):
+        val_prior_names = set()
+        if os.path.isfile(val_prior_path):
+            print("Load pre-defined validation set...")
+            with open(val_prior_path, "r", encoding="utf-8") as f:
+                for m in f.readlines():
+                    val_prior_names.add(m.split("|")[0])
+            return list(val_prior_names)
+        else:
+            return None
 
     def build_from_path(self):
         os.makedirs((os.path.join(self.out_dir, "text")), exist_ok=True)
         os.makedirs((os.path.join(self.out_dir, "mel")), exist_ok=True)
 
         print("Processing Data ...")
         out = list()
+        train = list()
+        val = list()
         n_frames = 0
         mel_min = float('inf')
         mel_max = -float('inf')
@@ -65,7 +79,14 @@ def build_from_path(self):
                     continue
                 else:
                     info, n, m_min, m_max = ret
-                out.append(info)
+
+                if self.val_prior is not None:
+                    if basename not in self.val_prior:
+                        train.append(info)
+                    else:
+                        val.append(info)
+                else:
+                    out.append(info)
 
                 if mel_min > m_min:
                     mel_min = m_min
@@ -93,15 +114,24 @@ def build_from_path(self):
             )
         )
 
-        random.shuffle(out)
-        out = [r for r in out if r is not None]
+        if self.val_prior is not None:
+            assert len(out) == 0
+            random.shuffle(train)
+            train = [r for r in train if r is not None]
+            val = [r for r in val if r is not None]
+        else:
+            assert len(train) == 0 and len(val) == 0
+            random.shuffle(out)
+            out = [r for r in out if r is not None]
+            train = out[self.val_size :]
+            val = out[: self.val_size]
 
         # Write metadata
         with open(os.path.join(self.out_dir, "train.txt"), "w", encoding="utf-8") as f:
-            for m in out[self.val_size :]:
+            for m in train:
                 f.write(m + "\n")
         with open(os.path.join(self.out_dir, "val.txt"), "w", encoding="utf-8") as f:
-            for m in out[: self.val_size]:
+            for m in val:
                 f.write(m + "\n")
 
         return out

preprocessor/vctk.py

@@ -46,10 +46,22 @@ def __init__(self, config):
             config["preprocessing"]["mel"]["mel_fmin"],
             config["preprocessing"]["mel"]["mel_fmax"],
         )
+        self.val_prior = self.val_prior_names(os.path.join(self.out_dir, "val.txt"))
         self.speaker_emb = None
         if config["preprocessing"]["speaker_embedder"] != "none":
             self.speaker_emb = PreDefinedEmbedder(config)
 
+    def val_prior_names(self, val_prior_path):
+        val_prior_names = set()
+        if os.path.isfile(val_prior_path):
+            print("Load pre-defined validation set...")
+            with open(val_prior_path, "r", encoding="utf-8") as f:
+                for m in f.readlines():
+                    val_prior_names.add(m.split("|")[0])
+            return list(val_prior_names)
+        else:
+            return None
+
     def _init_spker_embeds(self, spkers):
         spker_embeds = dict()
         for spker in spkers:
@@ -101,10 +113,16 @@ def build_from_path(self):
                     info, n, m_min, m_max, spker_embed = ret
                 # out.append(info)
 
-                if i == 0 or i == 1: 
-                    val.append(info)
+                if self.val_prior is not None:
+                    if basename not in self.val_prior:
+                        train.append(info)
+                    else:
+                        val.append(info)
                 else:
-                    train.append(info)
+                    if i == 0 or i == 1: 
+                        val.append(info)
+                    else:
+                        train.append(info)
 
                 if self.speaker_emb is not None:
                     spker_embeds[speaker].append(spker_embed)
@@ -152,16 +170,16 @@ def build_from_path(self):
 
         # random.shuffle(out)
         # out = [r for r in out if r is not None]
+        if self.val_prior is None:
+            random.shuffle(train)
         train = [r for r in train if r is not None]
         val = [r for r in val if r is not None]
 
         # Write metadata
         with open(os.path.join(self.out_dir, "train.txt"), "w", encoding="utf-8") as f:
-            # for m in out[self.val_size :]:
             for m in train:
                 f.write(m + "\n")
         with open(os.path.join(self.out_dir, "val.txt"), "w", encoding="utf-8") as f:
-            # for m in out[: self.val_size]:
             for m in val:
                 f.write(m + "\n")
 

utils/tools.py

@@ -207,7 +207,8 @@ def infer_one_sample(targets, predictions, vocoder, mel_stats, model_config, pre
         attention=True,
     )
     plt.savefig(os.path.join(
-        path, str(args.restore_step), "{}_{}.png".format(basename, args.speaker_id) if multi_speaker else "{}.png".format(basename)))
+        path, str(args.restore_step), "{}_{}.png".format(basename, args.speaker_id)\
+             if multi_speaker and args.mode == "single" else "{}.png".format(basename)))
     plt.close()
 
     from .model import vocoder_infer
@@ -219,7 +220,8 @@ def infer_one_sample(targets, predictions, vocoder, mel_stats, model_config, pre
 
     sampling_rate = preprocess_config["preprocessing"]["audio"]["sampling_rate"]
     wavfile.write(os.path.join(
-        path, str(args.restore_step), "{}_{}.wav".format(basename, args.speaker_id) if multi_speaker else "{}.wav".format(basename)),
+        path, str(args.restore_step), "{}_{}.wav".format(basename, args.speaker_id)\
+             if multi_speaker and args.mode == "single" else "{}.wav".format(basename)),
         sampling_rate, wav_predictions[0])