What is broken cannot be reforged

xperiments/xp_petofi.py

@@ -6,22 +6,31 @@
 from brainforge.optimization import RMSprop
 
 from keras.models import Sequential
-from keras.layers import LSTM as kLSTM, Dense as kDense
+from keras.layers import LSTM as kLSTM, Dense as kDense, BatchNormalization
 
-data = LazyText(os.path.expanduser("~/tmp/RIS.txt.txt"), n_gram=1, timestep=5)
+data = LazyText(os.path.expanduser("~/Prog/data/txt/scripts.txt"), n_gram=1, timestep=10)
 inshape, outshape = data.neurons_required
-net = BackpropNetwork(input_shape=inshape, layerstack=[
-    LSTM(60, activation="tanh"),
-    DenseLayer(60, activation="tanh"),
-    DenseLayer(outshape, activation="softmax")
-], cost="xent", optimizer=RMSprop(eta=0.01))
-
-net.fit_generator(data.batchgen(20), lessons_per_epoch=data.N)
-
-keras = Sequential([
-    kLSTM(60, input_shape=inshape, activation="tanh"),
-    kDense(60, activation="tanh"),
-    kDense(outshape, activation="softmax")
-])
-keras.compile(optimizer="rmsprop", loss="categorical_crossentropy", metrics=["acc"])
-keras.fit_generator(data.batchgen(20), steps_per_epoch=data.N, epochs=30)
+
+
+def run_brainforge():
+    net = BackpropNetwork(input_shape=inshape, layerstack=[
+        LSTM(60, activation="tanh"),
+        DenseLayer(60, activation="tanh"),
+        DenseLayer(outshape, activation="softmax")
+    ], cost="xent", optimizer=RMSprop(eta=0.01))
+
+    net.fit_generator(data.batchgen(20), lessons_per_epoch=data.N)
+
+
+def run_keras():
+    keras = Sequential([
+        kLSTM(120, input_shape=inshape, activation="relu"), BatchNormalization(),
+        kDense(60, activation="relu"), BatchNormalization(),
+        kDense(outshape, activation="softmax")
+    ])
+    keras.compile(optimizer="adam", loss="categorical_crossentropy", metrics=["acc"])
+    keras.fit_generator(data.batchgen(50), steps_per_epoch=data.N, epochs=30)
+
+
+if __name__ == '__main__':
+    run_keras()