I am trying to segment an image by first dividing it into superpixel, and than classifying those superpixel to merge them together.
If I am doing it for example with a SVM or a Decision Tree, it works fairly well, but is still doing some mistakes.
Now I wanted to use an Neural Network, to do the task, but after a few iterations it almost always gets stuck in the local minima and "maps every input to the same class".
It behaves in general really weird, because sometimes the accuracy first increases, but than steadily decreases again, till it reaches one of the minima from above (I am talking about the train accuracy, not the test accuracy, so overfitting is not an explanation, I also don't use Stochastic gradient descent but use all train data for an weight update, so fluctuations cant really explain the decreasing accuracy)
That looks like this:
617/617 [==============================] - 1s - loss: 1.1174 - acc: 0.2690
Epoch 2/1000
617/617 [==============================] - 0s - loss: 0.8662 - acc: 0.8169
Epoch 3/1000
617/617 [==============================] - 0s - loss: 0.5829 - acc: 0.8136
Epoch 4/1000
617/617 [==============================] - 0s - loss: 6.3685 - acc: 0.9060
Epoch 5/1000
617/617 [==============================] - 0s - loss: 9.0494 - acc: 0.4036
Epoch 6/1000
617/617 [==============================] - 0s - loss: 4.4840 - acc: 0.1297
Epoch 7/1000
617/617 [==============================] - 0s - loss: 1.9398 - acc: 0.1167
Epoch 8/1000
617/617 [==============================] - 0s - loss: 1.8809 - acc: 0.1167
(It stays at the same loss and acc from there)
The NN I am using looks like this (even though I also tried many others):
from keras.models import Sequential
from keras.layers import Dense, Dropout
from keras.optimizers import SGD
enc = OneHotEncoder()
labelOH = enc.fit_transform(label.reshape(-1,1)).A
clfN = Sequential()
clfN.add(Dense(32, input_dim=feature_array.shape[1]))
clfN.add(Dense(64, activation="relu"))
clfN.add(Dense(64, activation="relu"))
clfN.add(Dense(32, activation="relu"))
clfN.add(Dense(labelOH.shape[1]))
sgd = SGD(lr=0.0001, decay=1e-10, momentum=0.9, nesterov=True)
clfN.compile(loss='categorical_crossentropy',
optimizer=sgd,
metrics=['accuracy'])
clfN.summary()
clfN.fit(feature_array, labelOH, epochs=1000, batch_size=617)
What could be the cause for this behaviour, and what could I do about it?
