時間方向のブラー増強

3 files changed, 36 insertions, 3 deletions

diff --git a/become_yukarin/config/sr_config.py b/become_yukarin/config/sr_config.py
index 266ea04..4f980a2 100644
--- a/become_yukarin/config/sr_config.py
+++ b/become_yukarin/config/sr_config.py
@@ -14,6 +14,7 @@ class SRDatasetConfig(NamedTuple):
     train_crop_size: int
     input_global_noise: float
     input_local_noise: float
+    blur_size_factor: float
     seed: int
     num_test: int
 
@@ -79,6 +80,7 @@ def create_from_json(s: Union[str, Path]):
             train_crop_size=d['dataset']['train_crop_size'],
             input_global_noise=d['dataset']['input_global_noise'],
             input_local_noise=d['dataset']['input_local_noise'],
+            blur_size_factor=d['dataset']['blur_size_factor'],
             seed=d['dataset']['seed'],
             num_test=d['dataset']['num_test'],
         ),
@@ -102,4 +104,5 @@ def create_from_json(s: Union[str, Path]):
 
 
 def backward_compatible(d: Dict):
-    pass
+    if 'blur_size_factor' not in d['dataset']:
+        d['dataset']['blur_size_factor'] = 0
diff --git a/become_yukarin/dataset/dataset.py b/become_yukarin/dataset/dataset.py
index 7e3acbf..b584e44 100644
--- a/become_yukarin/dataset/dataset.py
+++ b/become_yukarin/dataset/dataset.py
@@ -8,6 +8,7 @@ from typing import Any
 from typing import Callable
 from typing import Dict
 from typing import List
+import scipy.ndimage
 
 import chainer
 import librosa
@@ -418,6 +419,19 @@ class AddNoiseProcess(BaseDataProcess):
         return data + g + l
 
 
+class RandomBlurProcess(BaseDataProcess):
+    def __init__(self, blur_size_factor: float, time_axis: int = 1) -> None:
+        assert time_axis == 1
+        self._blur_size_factor = blur_size_factor
+        self._time_axis = time_axis
+
+    def __call__(self, data: numpy.ndarray, test=None):
+        assert not test
+
+        blur_size = numpy.abs(numpy.random.randn()) * self._blur_size_factor
+        return scipy.ndimage.gaussian_filter(data, (0, blur_size))
+
+
 class DataProcessDataset(chainer.dataset.DatasetMixin):
     def __init__(self, data: typing.List, data_process: BaseDataProcess) -> None:
         self._data = data
@@ -559,6 +573,17 @@ def create_sr(config: SRDatasetConfig):
 
     data_process_train = copy.deepcopy(data_process_base)
 
+    # blur
+    data_process_train.append(SplitProcess(dict(
+        input=ChainProcess([
+            LambdaProcess(lambda d, test: d['input']),
+            RandomBlurProcess(blur_size_factor=config.blur_size_factor),
+        ]),
+        target=ChainProcess([
+            LambdaProcess(lambda d, test: d['target']),
+        ]),
+    )))
+
     # cropping
     if config.train_crop_size is not None:
         def add_seed():
diff --git a/scripts/voice_conversion_test.py b/scripts/voice_conversion_test.py
index d96d5ce..27e588b 100644
--- a/scripts/voice_conversion_test.py
+++ b/scripts/voice_conversion_test.py
@@ -16,11 +16,13 @@ parser.add_argument('model_names', nargs='+')
 parser.add_argument('-md', '--model_directory', type=Path, default=Path('/mnt/dwango/hiroshiba/become-yukarin/'))
 parser.add_argument('-iwd', '--input_wave_directory', type=Path,
                     default=Path('/mnt/dwango/hiroshiba/become-yukarin/dataset/hiho-wave/hiho-pause-atr503-subset/'))
+parser.add_argument('-it', '--iteration', type=int)
 parser.add_argument('-g', '--gpu', type=int)
 args = parser.parse_args()
 
 model_directory = args.model_directory  # type: Path
 input_wave_directory = args.input_wave_directory  # type: Path
+it = args.iteration
 gpu = args.gpu
 
 paths_test = list(Path('./test_data/').glob('*.wav'))
@@ -53,8 +55,11 @@ for model_name in args.model_names:
     path_train = input_paths[0]
     path_test = input_paths[-1]
 
-    model_paths = base_model.glob('predictor*.npz')
-    model_path = list(sorted(model_paths, key=extract_number))[-1]
+    if it is not None:
+        model_path = base_model / 'predictor_{}.npz'.format(it)
+    else:
+        model_paths = base_model.glob('predictor_*.npz')
+        model_path = list(sorted(model_paths, key=extract_number))[-1]
     print(model_path)
     acoustic_converter = AcousticConverter(config, model_path, gpu=gpu)