commit c2fe311e6b3ce615c086b600714af49913f91ffc
Author: Simon Kluettermann <Simon.Kluettermann@cs.tu-dortmund.de>
Date:   Mon Nov 21 11:02:58 2022 +0100

    initial push

diff --git a/model.py b/model.py
new file mode 100644
index 0000000..91c7d22
--- /dev/null
+++ b/model.py
@@ -0,0 +1,44 @@
+import numpy as np
+
+import tensorflow as tf
+from tensorflow import keras
+from tensorflow.keras import backend as K
+
+
+from disttf import GaussianRealisation,BiasedRealisation,BoxRealisation
+from disttf import MixtureLayer, MotioLayer, ScaleLayer, SplitLayer
+from disttf import NonLinearityLayer, SeperateLayer, RecombineLayer
+
+
+def gen_model(inputs, splits=10, realisation="gauss", mixture=0, nonlin=False):
+
+    i=keras.layers.Input(shape=(inputs,))
+    q=i
+    v=tf.constant(1.0,dtype=tf.float32)
+
+    if splits>1:
+        q,v=SplitLayer(splits)([q,v])
+    q,v=MotioLayer()([q,v])
+    q,v=ScaleLayer()([q,v])
+    for j in range(mixture):
+        q=MixtureLayer()(q)
+        if nonlin:
+            q=NonLinearityLayer()(q)
+    if realisation=="gauss":
+        q=GaussianRealisation()([q,v])
+    elif realisation=="biased":
+        q=BiasedRealisation()([q,v])
+    elif realisation=="box":
+        q=BoxRealisation()([q,v])
+    else:
+        raise ValueError("Unknown realisation type: "+realisation)
+
+    model=keras.Model(inputs=i,outputs=q)
+
+    loss=K.log(K.abs(q)+1e-6)
+    loss=-K.mean(loss)
+    model.add_loss(loss)
+
+    return model
+    
+
diff --git a/train.py b/train.py
new file mode 100644
index 0000000..ca827ac
--- /dev/null
+++ b/train.py
@@ -0,0 +1,81 @@
+import numpy as np
+
+import tensorflow as tf
+from tensorflow import keras
+from tensorflow.keras import backend as K
+
+from model import gen_model
+
+from sklearn.metrics import roc_auc_score
+
+from time import time
+
+
+def with_evaluation_cost(func):
+    def wrapper(*args, **kwargs):
+        t0=time()
+        dic = func(*args, **kwargs)
+        t1=time()
+        dic["evaluation_cost"]=t1-t0
+        return dic
+    return wrapper
+
+
+@with_evaluation_cost
+def train_one(x,tx,ty,seed=0,lr_epoch_modulo=10,lr_factor=0.8,lr_minima=1e-4,initial_lr=0.001,epochs=1000,batch_size=30,shall_early=True,patience=10,*args,**kwargs):
+   
+    np.random.seed(seed)
+    tf.random.set_seed(seed)
+
+    model=gen_model(int(x.shape[1]),*args,**kwargs)
+    model.compile(optimizer=keras.optimizers.Adam(initial_lr))
+    
+    
+    def shedule(epoch,lr):
+        if epoch%lr_epoch_modulo==0 and epoch>0 and lr>lr_minima:
+            return lr*lr_factor
+        return lr
+    
+
+    callbacks=[]
+    if shall_early:callbacks.append(keras.callbacks.EarlyStopping(patience=patience,restore_best_weights=True))
+    callbacks.append(keras.callbacks.LearningRateScheduler(shedule))
+    callbacks.append(keras.callbacks.TerminateOnNaN())
+
+    model.fit(x,None,
+                epochs=epochs,
+                batch_size=batch_size,
+                validation_split=0.2,
+                callbacks=callbacks)
+    
+    p=model.predict(tx)
+    try:
+        loss=model.evaluate(x)
+        auc=roc_auc_score(ty,-p)
+    except:
+        loss=1000000000.0
+        auc=-1.0
+    return {"loss":loss,"auc":auc}
+
+def dic_mean(dic):
+    ret={k:np.mean([d[k] for d in dic]) for k in dic[0]}
+    ret["min_loss"]=np.min([d["loss"] for d in dic])
+    ret["opt_auc"]=np.mean([d["auc"] for d in dic if d["loss"]==ret["min_loss"]])
+    return ret
+
+@with_evaluation_cost
+def train_many(x,tx,ty,count=10,*args,**kwargs):
+
+    dics=[]
+    for seed in range(count):
+        dics.append(train_one(x,tx,ty,seed=seed,*args,**kwargs))
+        if dics[-1]["auc"]<0.0:
+            break
+
+    return dic_mean(dics)
+
+if __name__=="__main__":
+    f=np.load("cardio.npz")
+    x,tx,ty=f["x"],f["tx"],f["ty"]
+    print(train_one(x,tx,ty,epochs=1000,shall_early=True,patience=10,lr_epoch_modulo=10,lr_factor=0.8,lr_minima=1e-4,initial_lr=0.001,batch_size=30,splits=1))
+