mean var normalization

bin/normalize.sh

@@ -4,11 +4,12 @@ conda=/share/mini1/sw/std/python/anaconda3-2019.07/v3.7
 conda_env=torch_1.7
 source $conda/bin/activate $conda_env
 
-
-python preprocess/normalize.py \
-        --stats_path dump/libritts/train_nodev_clean/mel/train_nodev_clean.npy \
-        --dump_dir dump/libritts/ \
-        --split train_nodev \
-        --metadata data/libritts/train_nodev_clean/metadata.csv
+for split in train_nodev_clean dev_clean ;do
+    python preprocess/normalize.py \
+            --stats_path dump/libritts/train_nodev_clean/mel/train_nodev_clean.npy \
+            --dump_dir dump/libritts/ \
+            --split $split \
+            --metadata data/libritts/$split/metadata.csv
+done            
 
 

dataset.py

@@ -96,16 +96,16 @@ def __init__(self, config, metadata_csv, split):
                 # remove utterances that are too long for training.
                 if config['rm_long_utt']:
                     _duration = row['duration']
-                    if float(_duration) <= config['max_utt_duration']:
+                    if float(_duration) < config['max_utt_duration']:
                         self.metadata.append(row)
             f.close()    
 
-
+        print(f'{split} data samples {len(self.metadata)}')
         self.batch_size = config['batch_size']
         self.drop_last = config['drop_last']
         self.sort = config['sort']
         # feature dirs
-        self.mel_dir = os.path.join(config['dump_dir'], config['dataset'], split, 'mel')
+        self.mel_dir = os.path.join(config['dump_dir'], config['dataset'], split, 'norm_mel')
 
         self.ling_enc = config['ling_enc']
         self.ling_rep_dir = os.path.join(config['dump_dir'], config['dataset'], split, self.ling_enc)

inference.py

@@ -70,7 +70,7 @@ def load_wav(path, sample_rate = 16000):
 
     print(f'generating {len(eval_list)} samples')
     # split eval_list by sge_job_idx
-    n_per_task = round(len(eval_list) / args.sge_n_tasks, 0)    
+    n_per_task = np.ceil(len(eval_list) / args.sge_n_tasks)    
     start = int(( args.sge_task_id -1 ) * n_per_task)
     if int( args.sge_task_id * n_per_task) >= len(eval_list):
         end = len(eval_list) -1

preprocess/normalize.py

@@ -32,8 +32,8 @@
 
     # load stats
     scaler = StandardScaler()
-    scaler.mean_ = np.load(args.stats)[0]
-    scaler.scale_ = np.load(args.stats)[1]        
+    scaler.mean_ = np.load(args.stats_path)[0]
+    scaler.scale_ = np.load(args.stats_path)[1]        
     scaler.n_features_in_ = scaler.mean_.shape[0]
 
 
@@ -43,13 +43,13 @@
         spk = _meta['spk']
         mel_path = os.path.join(args.dump_dir, args.split, 'mel', spk, ID + '.npy')
         norm_path = os.path.join(args.dump_dir, args.split, 'norm_mel', spk, ID + '.npy')
-        os.makedirs(os.path.basename(norm_path), exists_ok = True)
+        os.makedirs(os.path.dirname(norm_path), exist_ok = True)
 
         mel = np.load(mel_path)
         norm_mel = scaler.transform(mel)
         np.save(
                 norm_path,
-                mel.astype(np.float32),
+                norm_mel.astype(np.float32),
                 allow_pickle=False,
                )
 

submit_train.sh

@@ -5,26 +5,26 @@ conda=/share/mini1/sw/std/python/anaconda3-2019.07/v3.7
 conda_env=torch_1.9
 
 #exp setup
-#ling=vqw2v
+ling=vqw2v
 #ling=conformerppg
-ling=hubertsoft
+#ling=hubertsoft
 spk=uttdvec
 pros=none
-#dec=fastspeech2
+dec=fastspeech2
 #dec=tacoar
-dec=tacomol
+#dec=tacomol
 
-exp_name=first_train
+exp_name=libritts_24khz_10ms
 config=configs/${ling}_${spk}_${pros}_${dec}.yaml
 exp_dir=exp
 model_name=${ling}_${spk}_${pros}_${dec}
 exp=$exp_dir/$model_name/$exp_name
-njobs=24
+njobs=1
 ngpus=2
 slots=8
 #gputypes="GeForceRTX3060|GeForceRTX3090"
-#gputypes="GeForceRTX3090"
-gputypes="GeForceGTXTITANX|GeForceGTX1080Ti|GeForceRTX3060"
+gputypes="GeForceRTX3090"
+#gputypes="GeForceGTXTITANX|GeForceGTX1080Ti|GeForceRTX3060"
 
 # create exp dir
 [ ! -e $exp ] && mkdir -p $exp