added kusal preprocess code

.gitignore

@@ -3,3 +3,4 @@ __pycache__/
 *.pyc
 .DS_Store
 run*.sh
+.vscode

datasets/kusal.py

@@ -0,0 +1,67 @@
+from concurrent.futures import ProcessPoolExecutor
+from functools import partial
+import glob
+import librosa
+import numpy as np
+import os
+
+from hparams import hparams
+from util import audio
+
+
+def build_from_path(in_dir, out_dir, num_workers=1, tqdm=lambda x: x):
+  '''Preprocesses the Amy dataset from a given input path into a given output directory.'''
+  # executor = ProcessPoolExecutor(max_workers=num_workers)
+  futures = []
+  count = 0
+  len_files = 0
+  # Read all of the .wav files:
+  paths = {}
+  for path in glob.glob(os.path.join(in_dir, 'audio', '*.wav')):
+    prompt_id = os.path.basename(path).split('-')[-2]
+    paths[prompt_id] = path
+
+  # Read the prompts file:
+  with open(os.path.join(in_dir, 'prompts.txt'), encoding='utf-8') as f:
+    for line in f:
+      parts = line.strip().split('\t')
+      if len(parts) == 4 and parts[0] in paths:
+        path = paths[parts[0]]
+        text = parts[2]
+        # futures.append(executor.submit(partial(_process_utterance, out_dir, parts[0], path, text)))
+        futures.append(partial(_process_utterance, out_dir, parts[0], path, text))
+        len_files += 1
+  # return [future.result() for future in futures]
+    # return [future() for future in futures]
+  metadata = []
+  for future in tqdm(futures):
+    try:
+      data = future()
+      metadata.append(data)
+    except:
+      count += 1
+      print("failed to process" , count, "/", len_files)
+  return metadata
+
+def _process_utterance(out_dir, prompt_id, wav_path, text):
+  # Load the audio to a numpy array:
+  wav = audio.load_wav(wav_path)
+  # Trim leading and trailing silence:
+  margin = int(hparams.sample_rate * 0.1)
+  wav = wav[margin:-margin]
+  wav, _ = librosa.effects.trim(wav, top_db=40, frame_length=1024, hop_length=256)
+
+  # Compute the linear-scale spectrogram from the wav:
+  spectrogram = audio.spectrogram(wav).astype(np.float32)
+  n_frames = spectrogram.shape[1]
+
+  # Compute a mel-scale spectrogram from the wav:
+  mel_spectrogram = audio.melspectrogram(wav).astype(np.float32)
+
+  # Write the spectrograms to disk:
+  spectrogram_filename = 'kusal-spec-%s.npy' % prompt_id
+  mel_filename = 'kusal-mel-%s.npy' % prompt_id
+  np.save(os.path.join(out_dir, spectrogram_filename), spectrogram.T, allow_pickle=False)
+  np.save(os.path.join(out_dir, mel_filename), mel_spectrogram.T, allow_pickle=False)
+  # Return a tuple describing this training example:
+  return (spectrogram_filename, mel_filename, n_frames, text)

eval.py

@@ -43,7 +43,10 @@ def main():
   parser.add_argument('--checkpoint', required=True, help='Path to model checkpoint')
   parser.add_argument('--hparams', default='',
     help='Hyperparameter overrides as a comma-separated list of name=value pairs')
+  parser.add_argument('--force_cpu', default=False, help='Force synthesize with cpu')
   args = parser.parse_args()
+  if args.force_cpu:
+    os.environ['CUDA_VISIBLE_DEVICES'] = ''
   os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
   hparams.parse(args.hparams)
   run_eval(args)

preprocess.py

@@ -2,7 +2,7 @@ import argparse
 import os
 from multiprocessing import cpu_count
 from tqdm import tqdm
-from datasets import amy, blizzard, ljspeech
+from datasets import amy, blizzard, ljspeech, kusal
 from hparams import hparams
 
 
@@ -30,6 +30,14 @@ def preprocess_amy(args):
   write_metadata(metadata, out_dir)
 
 
+def preprocess_kusal(args):
+  in_dir = os.path.join(args.base_dir, 'kusal')
+  out_dir = os.path.join(args.base_dir, args.output)
+  os.makedirs(out_dir, exist_ok=True)
+  metadata = kusal.build_from_path(in_dir, out_dir, args.num_workers, tqdm=tqdm)
+  write_metadata(metadata, out_dir)
+
+
 def write_metadata(metadata, out_dir):
   with open(os.path.join(out_dir, 'train.txt'), 'w', encoding='utf-8') as f:
     for m in metadata:
@@ -45,7 +53,7 @@ def main():
   parser = argparse.ArgumentParser()
   parser.add_argument('--base_dir', default=os.path.expanduser('~/tacotron'))
   parser.add_argument('--output', default='training')
-  parser.add_argument('--dataset', required=True, choices=['amy', 'blizzard', 'ljspeech'])
+  parser.add_argument('--dataset', required=True, choices=['amy', 'blizzard', 'ljspeech', 'kusal'])
   parser.add_argument('--num_workers', type=int, default=cpu_count())
   args = parser.parse_args()
   if args.dataset == 'amy':
@@ -54,6 +62,8 @@ def main():
     preprocess_blizzard(args)
   elif args.dataset == 'ljspeech':
     preprocess_ljspeech(args)
+  elif args.dataset == 'kusal':
+    preprocess_kusal(args)
 
 
 if __name__ == "__main__":