Keras - Loss Nan and 0.333 accuracy throughout the training

Question

I am running an experiment which has the goal to classify EEG time series data in 3 classes. However, whenever I run training, my Loss is NaN and the accuracy is 0.0.

My data is 150 steps long and has 4 channels. It is all normalized between 0 and 1.

Data Channels

I am feeding them into the following model.

model = Sequential()
model.add(Conv1D(8, kernel_size=(2,), strides=(1,),
                 activation='relu',
                 input_shape=(input_width, num_channels)))
model.add(MaxPooling1D(pool_size=2, strides=(2,), padding='same'))
model.add(Dropout(0.25))

model.add(Conv1D(9, kernel_size=(2,), strides=(1,),
                 activation='relu'))
model.add(MaxPooling1D(pool_size=2, strides=(2,), padding='same'))
model.add(Dropout(0.25))

model.add(Conv1D(18, kernel_size=(2,), strides=(1,),
                 activation='relu'))
model.add(MaxPooling1D(pool_size=2, strides=(2,), padding='same'))
model.add(Dropout(0.25))

model.add(Conv1D(36, kernel_size=(2,), strides=(1,),
                 activation='relu'))
model.add(MaxPooling1D(pool_size=2, strides=(2,), padding='same'))
model.add(Dropout(0.25))

model.add(Conv1D(72, kernel_size=(2,), strides=(1,),
                 activation='relu'))
model.add(MaxPooling1D(pool_size=2, strides=(2,), padding='same'))
model.add(Dropout(0.25))

model.add(Flatten())
model.add(Dropout(0.5))
model.add(Dense(128, activation='tanh'),)
model.add(Dense(num_labels, activation='softmax'))

and then run it

optimizer = Adam(lr=0.0001)

model.summary()

model.compile(optimizer=optimizer,
              loss='categorical_crossentropy',
              metrics=['accuracy'])

model.fit(X, labels,
          epochs=100,
          batch_size=32)

However, the result is this:

Epoch 1/100
3855/3855 [==============================] - 24s 6ms/step - loss: nan - acc: 0.3331
Epoch 2/100
3855/3855 [==============================] - 25s 7ms/step - loss: nan - acc: 0.3331
.....
Epoch 100/100
3855/3855 [==============================] - 25s 7ms/step - loss: nan - acc: 0.3331

check out this answer https://stackoverflow.com/questions/40050397/tensorflow-nan-loss-reasons — thefifthjack005, Apr 04 '18 at 11:12

Wei Zhang · Answer 1 · 2018-04-03T22:13:40.660

1

I generated synthetic data and trained with your codes. The NaN issue doesn't occur to me. You probably need to check your data to see if there're any corruptions. Another thing to try is to only keep one Conv/Pooling/Dropout layer and see if the issue still occurs.

edited Apr 03 '18 at 22:13

answered Apr 03 '18 at 16:12

Wei Zhang

334
1
5

Thanks! What should I look for specifically? My inputs are currently 150 steps between 0 and 1 with 4 channels. My output is an array with 3 categories, so [0, 0, 1] or [0, 1, 0] etc. Could it be the hyperparams? – Dominic Monn Apr 03 '18 at 22:02
I am not sure... But before tuning the hyperparameter, one thing to try is to gradually simplify the model to locate the bug, for example keep one Conv/Pooling/Dropout layer and see if the issue still occurs. – Wei Zhang Apr 03 '18 at 22:15

Keras - Loss Nan and 0.333 accuracy throughout the training

1 Answers1