1

I am using Keras to tell the difference between background and signal.

A Background Example A Second Background Example A Signal Example

The model used here is

model = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(472,696)),#flatten the data
    tf.keras.layers.Dense(128, activation='relu'),#decide how many neurons to use
    tf.keras.layers.Dense(2, activation = 'sigmoid')])#how many classes you have

model.compile(optimizer='adam', loss=tf.keras.losses.SparseCategoricalCrossentropy(), metrics=['accuracy'])

My training is

model.fit(train_images, train_labels, epochs=6)#train with data

Often I just get a local minimum with accuracy approximately as 0.5.

Epoch 1/6 - loss: 45.3655 - accuracy: 0.4833

Epoch 2/6 - loss: 0.6934 - accuracy: 0.4855

Epoch 3/6 - loss: 0.6932 - accuracy: 0.4959

Epoch 4/6 - loss: 0.6931 - accuracy: 0.5145

Epoch 5/6 - loss: 0.6930 - accuracy: 0.5145

Epoch 6/6 - loss: 0.6929 - accuracy: 0.5145

Am I using the correct model?

I have tried to change the acitivation from "relu" to "LeakyReLU" and change optimizer to "sgd". But no significant improvement.

DW_T
  • 11
  • 2
  • You do not know if its a local minimum or not, and the activation at the output should be softmax, not sigmoid. – Dr. Snoopy Jan 22 '23 at 18:32
  • @Dr.Snoopy Thanks! If the loss doesn't change, does it not mean it's a local minimum? Also, I have changed the output layer and the result is improving already! – DW_T Jan 22 '23 at 18:58

1 Answers1

0

You need to experiment to see if its a local minimum or not. Add/remove layers, change number of neurons. Also use validation data to see if loss decreases do you overfit or not.

Yağızhan Arslan Akıncı
  • 26
  • 4