Deep Neural Network Accuracy Doesn't Change

đź’¬ Support & Discussions
,
1
  • MNE-Python version: 0.21.0
  • operating system: Windows 10

I am trying to create a Deep Neural Network using LSTM but I get always the same accuracy. I am using EEG data from a fif file with up to 9 motor tasks and 60 channels. When I first run it with data from 60 patients, 9 motor tasks, and 60 channels I get a maximum accuracy of 0.4828 and it doesn’t go up. I tried reducing the motor tasks to 5 and 3 and still got the same accuracy. I tried reducing from 60 channels to 9 channels and still got the same accuracy. I tried using data from just one patient and still got the same accuracy. I tried filtering data by frequency but still I got the same accuracy. The model reaches that accuracy really quickly and doesn’t go up even after many epochs (the maximum I tried were 5000 epochs). All what I have done was meant to change the accuracy but it didn’t.

I think I am feeding the wrong input to the model. The duration of an event is 4 seconds and a single run is 2 minutes (with multiple tasks). Can anybody help me with this?

events, _ = events_from_annotations(raw)

picks = pick_types(raw.info, meg=False, eeg=True, stim=False, eog=False,
                   exclude='bads')

epoch = Epochs(raw, events, tmax = 4, proj=True, picks = picks, preload=True)
epoch.filter(7.5,12)

X = epoch.get_data()
trX = X.transpose(0,2,1)
y = epoch.events[:, 2]
classLabels=to_categorical(y-1)
n_outputs=classLabels.shape[1]

from tensorflow import keras
#from sklearn.utils import compute_class_weight
sgd = keras.optimizers.SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)
#classWeight = compute_class_weight('balanced', numpy.unique(y), y) 

model = Sequential()
model.add(LSTM(10,input_shape=(trX.shape[1],trX.shape[2])))
model.add(Dropout(0.25))
model.add(Dense(10, activation='relu'))
model.add(Dense(n_outputs, activation='sigmoid'))
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
# fit network
model.summary()

batch_size = 64
epochs = 1000
verbose = 1

fitted=model.fit(trX, classLabels, epochs=epochs, batch_size=batch_size, verbose=verbose)

This is what I got when I train the model:

Epoch 1/1000
3/3 [==============================] - 0s 113ms/step - loss: 1.6093 - accuracy: 0.3966
Epoch 2/1000
3/3 [==============================] - 0s 112ms/step - loss: 1.6087 - accuracy: 0.4828
Epoch 3/1000
3/3 [==============================] - 0s 112ms/step - loss: 1.6077 - accuracy: 0.4828
Epoch 4/1000
3/3 [==============================] - 0s 108ms/step - loss: 1.6064 - accuracy: 0.4828
Epoch 5/1000
3/3 [==============================] - 0s 107ms/step - loss: 1.6048 - accuracy: 0.4828
Epoch 6/1000
3/3 [==============================] - 0s 104ms/step - loss: 1.6030 - accuracy: 0.4828
Epoch 7/1000
3/3 [==============================] - 0s 97ms/step - loss: 1.6012 - accuracy: 0.4828
Epoch 8/1000
3/3 [==============================] - 0s 95ms/step - loss: 1.5994 - accuracy: 0.4828
Epoch 9/1000
3/3 [==============================] - 0s 101ms/step - loss: 1.5979 - accuracy: 0.4828
Epoch 10/1000
3/3 [==============================] - 0s 98ms/step - loss: 1.5959 - accuracy: 0.4828
2

Hey @Mario unfortunately I cannot replicate your example because:

  1. some code is missing
  2. I don’t have the data

I suggest you share a complete minimal working example and use openly available data, so that others can try to help you debugging the issue. From just looking at it, I don’t know what the problem is.

If you solved the problem in the meantime, please reply about what was wrong and mark your reply as the solution to this thread. Thanks!

3

Hi @sappelhoff,

Thank you for your reply. I didn’t find any solutions for that but what I did was using an open source code that uses CNN to extract spatial information and LSTM to extract temporal information. I have attached the repository below.