1

I am training a very simple 2D dataset with 2 features. Its tabular data and contains only numeric information. I tried using keras to train a neural network but the performance does not bulge. I wanted to let the network fully converge and overfit but it doesnt work. 

overfitCallback = EarlyStopping(monitor='loss', min_delta=0, patience = 20)
act= 'relu'
last_act = 'softmax' 
los = 'mse'
ep = 100000
batch = 8
opt = keras.optimizers.Adam(learning_rate=0.001)
# define the keras model
model = Sequential()
dense1= model.add(Dense(5, input_dim=2, activation=act,kernel_initializer='he_uniform'))
dense2= model.add(Dense(2, activation=act))
dense3= model.add(Dense(1, activation=last_act))
model.compile(loss=los, optimizer=opt, metrics=['accuracy'])
history = model.fit(data_train, target_train, epochs=ep, batch_size=batch,verbose=0,callbacks=[overfitCallback])
GotStranger
  • 11
  • 1

1 Answers1

1

Your model has only 22 trainable weights (if I did not miscount, you can check yourself with model.summary()). Try to increase those by increasing your layer sizes e.g. Dense(256, activation="relu").

Felix
  • 41
  • 2