Python - Keras Deep Regressor

import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
import keras
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation
from keras.optimizers import SGD
from keras.callbacks import LearningRateScheduler,EarlyStopping, ModelCheckpoint
from keras import backend as K

epochs = 2000
learning_rate = 0.01
decay_rate = 5e-6
momentum = 0.9
reg=0.01

sd=[]
class LossHistory(keras.callbacks.Callback):
    def on_train_begin(self, logs={}):
        self.losses = [1,1]

    def on_epoch_end(self, batch, logs={}):
        self.losses.append(logs.get('loss'))
        sd.append(step_decay(len(self.losses)))
        print('learning rate:', step_decay(len(self.losses)))
        print('derivative of loss:', 2*np.sqrt((self.losses[-1])))

def my_init(shape, name=None):
    value = np.random.random(shape)
    return K.variable(value, name=name)

def step_decay(losses):
    if float(2*np.sqrt(np.array(history.losses[-1])))<1:
        lrate=0.01*1/(1+0.1*len(history.losses))
        momentum=0.2
        decay_rate=0.0
        return lrate
    else:
        lrate=0.01
        return lrate

model = Sequential()
model.add(Dense(6, input_dim=3, init=my_init))
model.add(Activation('tanh'))
model.add(Dropout(0.5))
model.add(Dense(2, init='uniform'))
model.add(Activation('softmax'))
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.97, nesterov=True)
model.compile(loss='categorical_crossentropy',
              optimizer=sgd,metrics=['accuracy'])

aa=pd.read_csv('questao1NN.csv',sep=',',header=0)
df=aa[0:1800]
df2=aa[1801:1999]

y=np.array(df[[1]])
y_train=[item for sublist in y for item in sublist]
y_train=pd.get_dummies(y_train)
y_train=np.array(y_train)
y_train=pd.DataFrame(y_train).values

x=np.array(df)
x1=x.T

for i in range(1,14):
    x1[i]=np.array([float(i) for i in x1[i]])

x2=[x1[6],x1[7],x1[9]]
x3=np.array(x2).T
X_train=x3
X_train.shape

y2=np.array(df2[[1]])
y_train2=[item for sublist in y2 for item in sublist]
y_train20=pd.get_dummies(y_train2)
y_train2=np.array(y_train20)
y_test=pd.DataFrame(y_train2).values

x0=np.array(df2)
x10=x0.T
x20=[x10[1],x10[2]]
x30=np.array(x20).T
X_test=x30

history=LossHistory()
lrate=LearningRateScheduler(step_decay)

model.fit(X_train, y_train,nb_epoch=20,batch_size=1000,callbacks=[history,lrate])

model.evaluate(X_train, y_train, batch_size=16)
model.predict(X_train, batch_size=32, verbose=0)

b=[float(i) for i in history.losses]
plt.plot(b,color='r')
plt.title("LOSS HISTORY")
plt.show()