Added parameters to SimpleGan.

library/generators/SimpleGan.py

@@ -16,23 +16,29 @@ from keras.layers import Dense, Dropout, Input
 from keras.models import Model, Sequential
 from keras.layers.advanced_activations import LeakyReLU
 from tensorflow.keras.optimizers import Adam
+import tensorflow as tf
 
 
 class SimpleGan(GanBaseClass):
     """
     A class for a simple GAN.
     """
-    def __init__(self, numOfFeatures=786, noiseSize=100, epochs=3, batchSize=128):
+    def __init__(self, numOfFeatures=786, noiseSize=None, epochs=3, batchSize=128, withTanh=False, gLayers=None, dLayers=None):
         self.isTrained = False
-        self.noiseSize = noiseSize
+        self.noiseSize = noiseSize if noiseSize is not None else numOfFeatures
         self.numOfFeatures = numOfFeatures
         self.epochs = epochs
         self.batchSize = batchSize
+        self.scaler = 1.0
+        self.withTanh = withTanh
+        self.dLayers = dLayers if dLayers is not None else [1024, 512, 256]
+        self.gLayers = gLayers if gLayers is not None else [256, 512, 1024]
 
     def reset(self):
         """
         Resets the trained GAN to an random state.
         """
+        self.scaler = 1.0
         self.generator = self._createGenerator(self.numOfFeatures, self.noiseSize)
         self.discriminator = self._createDiscriminator(self.numOfFeatures)
         self.gan = self._createGan(self.noiseSize)
@@ -52,32 +58,34 @@ class SimpleGan(GanBaseClass):
 
     def _createGenerator(self, numOfFeatures, noiseSize):
         generator=Sequential()
-        generator.add(Dense(units=256, input_dim=noiseSize))
-        generator.add(LeakyReLU(0.2))
+        for (n, size) in enumerate(self.dLayers):
+            if n == 0:
+                generator.add(Dense(units=size, input_dim=noiseSize))
+                generator.add(LeakyReLU(0.2))
+            else:
+                generator.add(Dense(units=size))
+                generator.add(LeakyReLU(0.2))
 
-        generator.add(Dense(units=512))
-        generator.add(LeakyReLU(0.2))
 
-        generator.add(Dense(units=1024))
-        generator.add(LeakyReLU(0.2))
-
-        generator.add(Dense(units=numOfFeatures, activation='tanh'))
+        if self.withTanh:
+            generator.add(Dense(units=numOfFeatures, activation='tanh'))
+        else:
+            generator.add(Dense(units=numOfFeatures, activation='softsign'))
 
         generator.compile(loss='binary_crossentropy', optimizer=self._adamOptimizer())
         return generator
 
     def _createDiscriminator(self, numOfFeatures):
         discriminator=Sequential()
-        discriminator.add(Dense(units=1024, input_dim=numOfFeatures))
-        discriminator.add(LeakyReLU(0.2))
-        discriminator.add(Dropout(0.3))
-
-        discriminator.add(Dense(units=512))
-        discriminator.add(LeakyReLU(0.2))
-        discriminator.add(Dropout(0.3))
 
-        discriminator.add(Dense(units=256))
-        discriminator.add(LeakyReLU(0.2))
+        for (n, size) in enumerate(self.dLayers):
+            if n == 0:
+                discriminator.add(Dense(units=size, input_dim=numOfFeatures))
+                discriminator.add(LeakyReLU(0.2))
+            else:
+                discriminator.add(Dropout(0.3))
+                discriminator.add(Dense(units=size))
+                discriminator.add(LeakyReLU(0.2))
 
         discriminator.add(Dense(units=1, activation='sigmoid'))
 
@@ -91,6 +99,13 @@ class SimpleGan(GanBaseClass):
         if trainDataSize <= 0:
             raise AttributeError("Train GAN: Expected data class 1 to contain at least one point.")
 
+        if self.withTanh:
+            self.scaler = 1.0
+            scaleDown = 1.0
+        else:
+            self.scaler = max(1.0, 1.1 * tf.reduce_max(tf.abs(trainData)).numpy())
+            scaleDown = 1.0 / self.scaler
+
         for e in range(self.epochs):
             print(f"Epoch {e + 1}/{self.epochs}")
             for _ in range(self.batchSize):
@@ -98,15 +113,15 @@ class SimpleGan(GanBaseClass):
                 noise= np.random.normal(0, 1, [self.batchSize, self.noiseSize])
 
                 # Generate fake MNIST images from noised input
-                generatedImages = self.generator.predict(noise)
+                syntheticBatch = self.generator.predict(noise)
 
                 # Get a random set of  real images
-                image_batch = dataset.data1[
+                realBatch = dataset.data1[
                     np.random.randint(low=0, high=trainDataSize, size=self.batchSize)
                     ]
 
                 #Construct different batches of  real and fake data
-                X = np.concatenate([image_batch, generatedImages])
+                X = np.concatenate([scaleDown * realBatch, syntheticBatch])
 
                 # Labels for generated and real data
                 y_dis=np.zeros(2 * self.batchSize)
@@ -139,4 +154,4 @@ class SimpleGan(GanBaseClass):
         noise = np.random.normal(0, 1, [numOfSamples, self.noiseSize])
 
         # Generate fake MNIST images from noised input
-        return self.generator.predict(noise)
+        return self.scaler * self.generator.predict(noise)