split and process valid set (#25)

Prepare Valid set for LibriTTS and LJSpeech

bins/tts/preprocess.py

@@ -25,7 +25,7 @@
 )
 
 
-def extract_acoustic_features(dataset, output_path, cfg, n_workers=1):
+def extract_acoustic_features(dataset, output_path, cfg, dataset_types, n_workers=1):
     """Extract acoustic features of utterances in the dataset
 
     Args:
@@ -34,17 +34,9 @@ def extract_acoustic_features(dataset, output_path, cfg, n_workers=1):
         cfg (dict): dictionary that stores configurations
         n_workers (int, optional): num of processes to extract features in parallel. Defaults to 1.
     """
-    # types = ["train", "test"] if "eval" not in dataset else ["test"]
-    types = list()
-    types.append((cfg.preprocess.train_file).split(".")[0])
-    types.append((cfg.preprocess.valid_file).split(".")[0])
-    if "test" not in types:
-        types.append("test")
-    if "eval" in dataset:
-        types = ["test"]
-    print("types: ", types)
+
     metadata = []
-    for dataset_type in types:
+    for dataset_type in dataset_types:
         dataset_output = os.path.join(output_path, dataset)
         dataset_file = os.path.join(dataset_output, "{}.json".format(dataset_type))
         with open(dataset_file, "r") as f:
@@ -58,26 +50,17 @@ def extract_acoustic_features(dataset, output_path, cfg, n_workers=1):
     )
 
 
-def extract_content_features(dataset, output_path, cfg, num_workers=1):
+def extract_content_features(dataset, output_path, cfg, dataset_types, num_workers=1):
     """Extract content features of utterances in the dataset
 
     Args:
         dataset (str): name of dataset, e.g. opencpop
         output_path (str): directory that stores train, test and feature files of datasets
         cfg (dict): dictionary that stores configurations
     """
-    # types = ["train", "test"] if "eval" not in dataset else ["test"]
-
-    types = list()
-    types.append((cfg.preprocess.train_file).split(".")[0])
-    types.append((cfg.preprocess.valid_file).split(".")[0])
-    if "test" not in types:
-        types.append("test")
-    if "eval" in dataset:
-        types = ["test"]
 
     metadata = []
-    for dataset_type in types:
+    for dataset_type in dataset_types:
         dataset_output = os.path.join(output_path, dataset)
         # dataset_file = os.path.join(dataset_output, "{}.json".format(dataset_type))
         dataset_file = os.path.join(dataset_output, "{}.json".format(dataset_type))
@@ -89,7 +72,7 @@ def extract_content_features(dataset, output_path, cfg, num_workers=1):
     )
 
 
-def extract_phonme_sequences(dataset, output_path, cfg):
+def extract_phonme_sequences(dataset, output_path, cfg, dataset_types):
     """Extract phoneme features of utterances in the dataset
 
     Args:
@@ -98,18 +81,9 @@ def extract_phonme_sequences(dataset, output_path, cfg):
         cfg (dict): dictionary that stores configurations
 
     """
-    # types = ["train", "test"] if "eval" not in dataset else ["test"]
-
-    types = list()
-    types.append((cfg.preprocess.train_file).split(".")[0])
-    types.append((cfg.preprocess.valid_file).split(".")[0])
-    if "test" not in types:
-        types.append("test")
-    if "eval" in dataset:
-        types = ["test"]
 
     metadata = []
-    for dataset_type in types:
+    for dataset_type in dataset_types:
         dataset_output = os.path.join(output_path, dataset)
         dataset_file = os.path.join(dataset_output, "{}.json".format(dataset_type))
         with open(dataset_file, "r") as f:
@@ -162,8 +136,17 @@ def preprocess(cfg, args):
     except:
         print("No Data Augmentation.")
 
+    # json files
+    dataset_types = list()
+    dataset_types.append((cfg.preprocess.train_file).split(".")[0])
+    dataset_types.append((cfg.preprocess.valid_file).split(".")[0])
+    if "test" not in dataset_types:
+        dataset_types.append("test")
+    if "eval" in dataset:
+        dataset_types = ["test"]
+
     # Dump metadata of datasets (singers, train/test durations, etc.)
-    cal_metadata(cfg)
+    cal_metadata(cfg, dataset_types)
 
     # Prepare the acoustic features
     for dataset in cfg.dataset:
@@ -180,7 +163,9 @@ def preprocess(cfg, args):
                 dataset, args.num_workers
             )
         )
-        extract_acoustic_features(dataset, output_path, cfg, args.num_workers)
+        extract_acoustic_features(
+            dataset, output_path, cfg, dataset_types, args.num_workers
+        )
         # Calculate the statistics of acoustic features
         if cfg.preprocess.mel_min_max_norm:
             acoustic_extractor.cal_mel_min_max(dataset, output_path, cfg)
@@ -229,13 +214,15 @@ def preprocess(cfg, args):
     # Prepare the content features
     for dataset in cfg.dataset:
         print("Extracting content features for {}...".format(dataset))
-        extract_content_features(dataset, output_path, cfg, args.num_workers)
+        extract_content_features(
+            dataset, output_path, cfg, dataset_types, args.num_workers
+        )
 
     # Prepare the phenome squences
     if cfg.preprocess.extract_phone:
         for dataset in cfg.dataset:
             print("Extracting phoneme sequence for {}...".format(dataset))
-            extract_phonme_sequences(dataset, output_path, cfg)
+            extract_phonme_sequences(dataset, output_path, cfg, dataset_types)
 
 
 def main():

preprocessors/libritts.py

@@ -63,6 +63,7 @@ def main(output_path, dataset_path):
     os.makedirs(save_dir, exist_ok=True)
     train_output_file = os.path.join(save_dir, "train.json")
     test_output_file = os.path.join(save_dir, "test.json")
+    valid_output_file = os.path.join(save_dir, "valid.json")
     singer_dict_file = os.path.join(save_dir, "singers.json")
     utt2singer_file = os.path.join(save_dir, "utt2singer")
     if has_existed(train_output_file):
@@ -79,12 +80,15 @@ def main(output_path, dataset_path):
     # We select pharases of standard spekaer as test songs
     train = []
     test = []
+    valid = []
 
     train_index_count = 0
     test_index_count = 0
+    valid_index_count = 0
 
     train_total_duration = 0
     test_total_duration = 0
+    valid_total_duration = 0
 
     for distribution, speakers2pharases2utts in tqdm(
         distribution2speakers2pharases2utts.items()
@@ -107,27 +111,49 @@ def main(output_path, dataset_path):
                     res["Path"] = os.path.join(libritts_path, res["Path"])
                     assert os.path.exists(res["Path"])
 
+                    text_file_path = os.path.join(
+                        libritts_path,
+                        distribution,
+                        speaker,
+                        chosen_pharase,
+                        chosen_uid + ".normalized.txt",
+                    )
+                    with open(text_file_path, "r") as f:
+                        lines = f.readlines()
+                        assert len(lines) == 1
+                        text = lines[0].strip()
+                        res["Text"] = text
+
                     waveform, sample_rate = torchaudio.load(res["Path"])
                     duration = waveform.size(-1) / sample_rate
                     res["Duration"] = duration
 
-                    if not "train" in distribution:
+                    if "test" in distribution:
                         res["index"] = test_index_count
                         test_total_duration += duration
                         test.append(res)
                         test_index_count += 1
-                    else:
+                    elif "train" in distribution:
                         res["index"] = train_index_count
                         train_total_duration += duration
                         train.append(res)
                         train_index_count += 1
+                    elif "dev" in distribution:
+                        res["index"] = valid_index_count
+                        valid_total_duration += duration
+                        valid.append(res)
+                        valid_index_count += 1
 
                     utt2singer.write("{}\t{}\n".format(res["Uid"], res["Singer"]))
 
-    print("#Train = {}, #Test = {}".format(len(train), len(test)))
     print(
-        "#Train hours= {}, #Test hours= {}".format(
-            train_total_duration / 3600, test_total_duration / 3600
+        "#Train = {}, #Test = {}, #Valid = {}".format(len(train), len(test), len(valid))
+    )
+    print(
+        "#Train hours= {}, #Test hours= {}, #Valid hours= {}".format(
+            train_total_duration / 3600,
+            test_total_duration / 3600,
+            valid_total_duration / 3600,
         )
     )
 
@@ -136,6 +162,8 @@ def main(output_path, dataset_path):
         json.dump(train, f, indent=4, ensure_ascii=False)
     with open(test_output_file, "w") as f:
         json.dump(test, f, indent=4, ensure_ascii=False)
+    with open(valid_output_file, "w") as f:
+        json.dump(valid, f, indent=4, ensure_ascii=False)
 
     # Save singers.json
     singer_lut = {name: i for i, name in enumerate(unique_speakers)}

preprocessors/ljspeech.py

@@ -89,19 +89,26 @@ def get_uid2utt(ljspeech_path, dataset, cfg):
     return uid2utt, total_duration / 3600
 
 
-def split_dataset(lines, test_rate=0.05, test_size=None):
+def split_dataset(
+    lines, test_rate=0.05, valid_rate=0.05, test_size=None, valid_size=None
+):
     if test_size == None:
         test_size = int(len(lines) * test_rate)
+    if valid_size == None:
+        valid_size = int(len(lines) * valid_rate)
     random.shuffle(lines)
 
     train_set = []
     test_set = []
+    valid_set = []
 
     for line in lines[:test_size]:
         test_set.append(line)
-    for line in lines[test_size:]:
+    for line in lines[test_size : test_size + valid_size]:
+        valid_set.append(line)
+    for line in lines[test_size + valid_size :]:
         train_set.append(line)
-    return train_set, test_set
+    return train_set, test_set, valid_set
 
 
 max_wav_value = 32768.0
@@ -162,6 +169,7 @@ def main(output_path, dataset_path, cfg):
 
     train_output_file = os.path.join(save_dir, "train.json")
     test_output_file = os.path.join(save_dir, "test.json")
+    valid_output_file = os.path.join(save_dir, "valid.json")
     singer_dict_file = os.path.join(save_dir, "singers.json")
 
     speaker = "LJSpeech"
@@ -170,13 +178,17 @@ def main(output_path, dataset_path, cfg):
     with open(singer_dict_file, "w") as f:
         json.dump(singer_lut, f, indent=4, ensure_ascii=False)
 
-    if has_existed(train_output_file) and has_existed(test_output_file):
+    if (
+        has_existed(train_output_file)
+        and has_existed(test_output_file)
+        and has_existed(valid_output_file)
+    ):
         return
 
     meta_file = os.path.join(ljspeech_path, "metadata.csv")
     lines = get_lines(meta_file)
 
-    train_set, test_set = split_dataset(lines)
+    train_set, test_set, valid_set = split_dataset(lines)
 
     res, hours = get_uid2utt(ljspeech_path, train_set, cfg)
 
@@ -195,3 +207,10 @@ def main(output_path, dataset_path, cfg):
         json.dump(res, f, indent=4, ensure_ascii=False)
 
     print("Test_hours= {}".format(hours))
+
+    # Save valid
+    os.makedirs(save_dir, exist_ok=True)
+    with open(valid_output_file, "w") as f:
+        json.dump(res, f, indent=4, ensure_ascii=False)
+
+    print("Valid_hours= {}".format(hours))