超解像可能に

4 files changed, 15 insertions, 8 deletions

diff --git a/become_yukarin/__init__.py b/become_yukarin/__init__.py
index 81bea87..810ea1f 100644
--- a/become_yukarin/__init__.py
+++ b/become_yukarin/__init__.py
@@ -1,4 +1,5 @@
 from . import config
 from . import dataset
 from . import param
+from .super_resolution import SuperResolution
 from .voice_changer import VoiceChanger
diff --git a/become_yukarin/dataset/dataset.py b/become_yukarin/dataset/dataset.py
index 38cf749..ab05568 100644
--- a/become_yukarin/dataset/dataset.py
+++ b/become_yukarin/dataset/dataset.py
@@ -75,7 +75,7 @@ class WaveFileLoadProcess(BaseDataProcess):
         self._pad_second = pad_second
         self._dtype = dtype
 
-    def __call__(self, data: str, test):
+    def __call__(self, data: str, test=None):
         wave = librosa.core.load(data, sr=self._sample_rate, dtype=self._dtype)[0]
         if self._top_db is not None:
             wave = librosa.effects.remix(wave, intervals=librosa.effects.split(wave, top_db=self._top_db))
@@ -92,7 +92,7 @@ class AcousticFeatureProcess(BaseDataProcess):
         self._alpha = alpha
         self._dtype = dtype
 
-    def __call__(self, data: Wave, test):
+    def __call__(self, data: Wave, test=None):
         x = data.wave.astype(numpy.float64)
         fs = data.sampling_rate
 
diff --git a/become_yukarin/model/sr_model.py b/become_yukarin/model/sr_model.py
index 8f9a34b..2e83526 100644
--- a/become_yukarin/model/sr_model.py
+++ b/become_yukarin/model/sr_model.py
@@ -116,6 +116,6 @@ def create_predictor_sr(config: SRModelConfig):
 
 
 def create_sr(config: SRModelConfig):
-    predictor = create_predictor_sr()
+    predictor = create_predictor_sr(config)
     discriminator = SRDiscriminator(in_ch=1, out_ch=1)
     return predictor, discriminator
diff --git a/become_yukarin/super_resolution.py b/become_yukarin/super_resolution.py
index 535af57..bdb2e61 100644
--- a/become_yukarin/super_resolution.py
+++ b/become_yukarin/super_resolution.py
@@ -1,6 +1,5 @@
 from functools import partial
 from pathlib import Path
-from typing import Optional
 
 import chainer
 import numpy
@@ -39,21 +38,28 @@ class SuperResolution(object):
 
     def convert(self, input: numpy.ndarray) -> numpy.ndarray:
         converter = partial(chainer.dataset.convert.concat_examples, padding=0)
-        inputs = converter([numpy.log(input)[:, :-1]])
+        pad = 128 - len(input) % 128
+        input = numpy.pad(input, [(0, pad), (0, 0)], mode='minimum')
+        input = numpy.log(input)[:, :-1]
+        input = input[numpy.newaxis]
+        inputs = converter([input])
 
         with chainer.using_config('train', False):
             out = self.model(inputs).data[0]
 
         out = out[0]
-        out[:, out.shape[1]] = out[:, -1]
+        out = numpy.pad(out, [(0, 0), (0, 1)], mode='edge')
+        out = numpy.exp(out)
+        out = out[:-pad]
         return out
 
     def convert_to_audio(
             self,
             input: numpy.ndarray,
             acoustic_feature: AcousticFeature,
-            sampling_rate: Optional[int] = None,
+            sampling_rate: int,
     ):
+        acoustic_feature = acoustic_feature.astype_only_float(numpy.float64)
         out = pyworld.synthesize(
             f0=acoustic_feature.f0.ravel(),
             spectrogram=input.astype(numpy.float64),
@@ -76,7 +82,7 @@ class SuperResolution(object):
             self,
             input: numpy.ndarray,
             acoustic_feature: AcousticFeature,
-            sampling_rate: Optional[int] = None,
+            sampling_rate: int,
     ):
         high = self.convert(input)
         return self.convert_to_audio(high, acoustic_feature=acoustic_feature, sampling_rate=sampling_rate)