Merge pull request #1 from Hiroshiba/pix2pix-expand-size

pix2pixモデルのフィルタサイズを可変にした

1 files changed, 63 insertions, 31 deletions

diff --git a/become_yukarin/model/model.py b/become_yukarin/model/model.py
index 56870d9..42333d0 100644
--- a/become_yukarin/model/model.py
+++ b/become_yukarin/model/model.py
@@ -52,8 +52,10 @@ class CBR(chainer.Chain):
         with self.init_scope():
             if sample == 'down':
                 self.c = Convolution1D(ch0, ch1, 4, 2, 1, initialW=w)
-            else:
+            elif sample == 'up':
                 self.c = Deconvolution1D(ch0, ch1, 4, 2, 1, initialW=w)
+            else:
+                self.c = Convolution1D(ch0, ch1, 1, 1, 0, initialW=w)
             if bn:
                 self.batchnorm = L.BatchNormalization(ch1)
 
@@ -69,18 +71,23 @@ class CBR(chainer.Chain):
 
 
 class Encoder(chainer.Chain):
-    def __init__(self, in_ch) -> None:
+    def __init__(self, in_ch, base=64, extensive_layers=8) -> None:
         super().__init__()
         w = chainer.initializers.Normal(0.02)
         with self.init_scope():
-            self.c0 = Convolution1D(in_ch, 64, 3, 1, 1, initialW=w)
-            self.c1 = CBR(64, 128, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c2 = CBR(128, 256, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c3 = CBR(256, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c4 = CBR(512, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c5 = CBR(512, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c6 = CBR(512, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c7 = CBR(512, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
+            if extensive_layers > 0:
+                self.c0 = Convolution1D(in_ch, base * 1, 3, 1, 1, initialW=w)
+            else:
+                self.c0 = Convolution1D(in_ch, base * 1, 1, 1, 0, initialW=w)
+
+            _choose = lambda i: 'down' if i < extensive_layers else 'same'
+            self.c1 = CBR(base * 1, base * 2, bn=True, sample=_choose(1), activation=F.leaky_relu, dropout=False)
+            self.c2 = CBR(base * 2, base * 4, bn=True, sample=_choose(2), activation=F.leaky_relu, dropout=False)
+            self.c3 = CBR(base * 4, base * 8, bn=True, sample=_choose(3), activation=F.leaky_relu, dropout=False)
+            self.c4 = CBR(base * 8, base * 8, bn=True, sample=_choose(4), activation=F.leaky_relu, dropout=False)
+            self.c5 = CBR(base * 8, base * 8, bn=True, sample=_choose(5), activation=F.leaky_relu, dropout=False)
+            self.c6 = CBR(base * 8, base * 8, bn=True, sample=_choose(6), activation=F.leaky_relu, dropout=False)
+            self.c7 = CBR(base * 8, base * 8, bn=True, sample=_choose(7), activation=F.leaky_relu, dropout=False)
 
     def __call__(self, x):
         hs = [F.leaky_relu(self.c0(x))]
@@ -90,18 +97,23 @@ class Encoder(chainer.Chain):
 
 
 class Decoder(chainer.Chain):
-    def __init__(self, out_ch) -> None:
+    def __init__(self, out_ch, base=64, extensive_layers=8) -> None:
         super().__init__()
         w = chainer.initializers.Normal(0.02)
         with self.init_scope():
-            self.c0 = CBR(512, 512, bn=True, sample='up', activation=F.relu, dropout=True)
-            self.c1 = CBR(1024, 512, bn=True, sample='up', activation=F.relu, dropout=True)
-            self.c2 = CBR(1024, 512, bn=True, sample='up', activation=F.relu, dropout=True)
-            self.c3 = CBR(1024, 512, bn=True, sample='up', activation=F.relu, dropout=False)
-            self.c4 = CBR(1024, 256, bn=True, sample='up', activation=F.relu, dropout=False)
-            self.c5 = CBR(512, 128, bn=True, sample='up', activation=F.relu, dropout=False)
-            self.c6 = CBR(256, 64, bn=True, sample='up', activation=F.relu, dropout=False)
-            self.c7 = Convolution1D(128, out_ch, 3, 1, 1, initialW=w)
+            _choose = lambda i: 'up' if i >= 8 - extensive_layers else 'same'
+            self.c0 = CBR(base * 8, base * 8, bn=True, sample=_choose(0), activation=F.relu, dropout=True)
+            self.c1 = CBR(base * 16, base * 8, bn=True, sample=_choose(1), activation=F.relu, dropout=True)
+            self.c2 = CBR(base * 16, base * 8, bn=True, sample=_choose(2), activation=F.relu, dropout=True)
+            self.c3 = CBR(base * 16, base * 8, bn=True, sample=_choose(3), activation=F.relu, dropout=False)
+            self.c4 = CBR(base * 16, base * 4, bn=True, sample=_choose(4), activation=F.relu, dropout=False)
+            self.c5 = CBR(base * 8, base * 2, bn=True, sample=_choose(5), activation=F.relu, dropout=False)
+            self.c6 = CBR(base * 4, base * 1, bn=True, sample=_choose(6), activation=F.relu, dropout=False)
+
+            if extensive_layers > 0:
+                self.c7 = Convolution1D(base * 2, out_ch, 3, 1, 1, initialW=w)
+            else:
+                self.c7 = Convolution1D(base * 2, out_ch, 1, 1, 0, initialW=w)
 
     def __call__(self, hs):
         h = self.c0(hs[-1])
@@ -115,27 +127,32 @@ class Decoder(chainer.Chain):
 
 
 class Predictor(chainer.Chain):
-    def __init__(self, in_ch, out_ch) -> None:
+    def __init__(self, in_ch, out_ch, base=64, extensive_layers=8) -> None:
         super().__init__()
         with self.init_scope():
-            self.encoder = Encoder(in_ch)
-            self.decoder = Decoder(out_ch)
+            self.encoder = Encoder(in_ch, base=base, extensive_layers=extensive_layers)
+            self.decoder = Decoder(out_ch, base=base, extensive_layers=extensive_layers)
 
     def __call__(self, x):
         return self.decoder(self.encoder(x))
 
 
 class Discriminator(chainer.Chain):
-    def __init__(self, in_ch, out_ch) -> None:
+    def __init__(self, in_ch, out_ch, base=32, extensive_layers=5, is_weak=False) -> None:
         super().__init__()
         w = chainer.initializers.Normal(0.02)
         with self.init_scope():
-            self.c0_0 = CBR(in_ch, 32, bn=False, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c0_1 = CBR(out_ch, 32, bn=False, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c1 = CBR(64, 128, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c2 = CBR(128, 256, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c3 = CBR(256, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
-            self.c4 = Convolution1D(512, 1, 3, 1, 1, initialW=w)
+            _choose = lambda i: 'down' if i < extensive_layers else 'same'
+            self.c0_0 = CBR(in_ch, base * 1, bn=False, sample=_choose(0), activation=F.leaky_relu, dropout=is_weak)
+            self.c0_1 = CBR(out_ch, base * 1, bn=False, sample=_choose(0), activation=F.leaky_relu, dropout=is_weak)
+            self.c1 = CBR(base * 2, base * 4, bn=True, sample=_choose(1), activation=F.leaky_relu, dropout=is_weak)
+            self.c2 = CBR(base * 4, base * 8, bn=True, sample=_choose(2), activation=F.leaky_relu, dropout=is_weak)
+            self.c3 = CBR(base * 8, base * 16, bn=True, sample=_choose(3), activation=F.leaky_relu, dropout=is_weak)
+
+            if extensive_layers > 4:
+                self.c4 = Convolution1D(base * 16, 1, 3, 1, 1, initialW=w)
+            else:
+                self.c4 = Convolution1D(base * 16, 1, 1, 1, 0, initialW=w)
 
     def __call__(self, x_0, x_1):
         h = F.concat([self.c0_0(x_0), self.c0_1(x_1)])
@@ -148,10 +165,25 @@ class Discriminator(chainer.Chain):
 
 
 def create_predictor(config: ModelConfig):
-    return Predictor(in_ch=config.in_channels, out_ch=config.out_channels)
+    return Predictor(
+        in_ch=config.in_channels,
+        out_ch=config.out_channels,
+        base=config.generator_base_channels,
+        extensive_layers=config.generator_extensive_layers,
+    )
+
+
+def create_discriminator(config: ModelConfig):
+    return Discriminator(
+        in_ch=config.in_channels,
+        out_ch=config.out_channels,
+        base=config.discriminator_base_channels,
+        extensive_layers=config.discriminator_extensive_layers,
+        is_weak=config.weak_discriminator,
+    )
 
 
 def create(config: ModelConfig):
     predictor = create_predictor(config)
-    discriminator = Discriminator(in_ch=config.in_channels, out_ch=config.out_channels)
+    discriminator = create_discriminator(config)
     return predictor, discriminator