Training accuracy remains zero and Validation accuracy remains constant

Question

I am trying to make a time series deep learning model, but I do not know why the training accuracy remains at zero and the validation accuracy remains constant. Here are the code and training and accuracy figures. Would appreciate it if someone helps me. train_x.shape=(28,14,9) and train_y.shape=(28,)

optimizer = keras.optimizers.Adam(lr=.0000001)
model = Sequential()

model.add(LSTM(units=16, activation='relu', return_sequences=True, input_shape=(train_x.shape[1], train_x.shape[2])))

model.add(Dropout(0.2))

model.add(Dense(units=8))
model.add(Dropout(0.2))
model.add(Dense(units=1))
model.compile(optimizer=optimizer, loss='mae', metrics=['accuracy'])

history = model.fit(train_x, train_y, epochs=10, batch_size=16, validation_split=0.28, verbose=1)

Epoch 1/10
2/2 [==============================] - 2s 256ms/step - loss: 1.0722 - accuracy: 0.0000e+00 - val_loss: 0.2569 - val_accuracy: 0.1250
Epoch 2/10
2/2 [==============================] - 0s 23ms/step - loss: 1.0573 - accuracy: 0.0000e+00 - val_loss: 0.2568 - val_accuracy: 0.1250
Epoch 3/10
2/2 [==============================] - 0s 24ms/step - loss: 1.0742 - accuracy: 0.0000e+00 - val_loss: 0.2568 - val_accuracy: 0.1250
Epoch 4/10
2/2 [==============================] - 0s 26ms/step - loss: 1.0680 - accuracy: 0.0000e+00 - val_loss: 0.2568 - val_accuracy: 0.1250
Epoch 5/10
2/2 [==============================] - 0s 24ms/step - loss: 1.0826 - accuracy: 0.0000e+00 - val_loss: 0.2568 - val_accuracy: 0.1250
Epoch 6/10
2/2 [==============================] - 0s 23ms/step - loss: 1.0595 - accuracy: 0.0000e+00 - val_loss: 0.2568 - val_accuracy: 0.1250
Epoch 7/10
2/2 [==============================] - 0s 28ms/step - loss: 1.0163 - accuracy: 0.0000e+00 - val_loss: 0.2568 - val_accuracy: 0.1250
Epoch 8/10
2/2 [==============================] - 0s 27ms/step - loss: 1.0415 - accuracy: 0.0000e+00 - val_loss: 0.2568 - val_accuracy: 0.1250
Epoch 9/10
2/2 [==============================] - 0s 30ms/step - loss: 1.0222 - accuracy: 0.0000e+00 - val_loss: 0.2568 - val_accuracy: 0.1250
Epoch 10/10
2/2 [==============================] - 0s 26ms/step - loss: 1.0404 - accuracy: 0.0000e+00 - val_loss: 0.2568 - val_accuracy: 0.1250

Your Adam `lr` is quite small. Try to make it bigger. – Lior Pollak May 18 '22 at 18:12 — Lior Pollak, May 18 '22 at 18:12
Also accuracy is meaningless for a regression task. – Frightera May 18 '22 at 18:15 — Frightera, May 18 '22 at 18:15

score 1 · Answer 1 · answered May 18 '22 at 19:15

Accuracy between target y and prediction p is

  #{ y_i == p_i }
 ----------------
       #y

Given that you are working with regression what is an accuracy between

y = 1, 1, 1, 1

and

p = 1.000000000000001, 1.000000000000001, 1.000000000000001, 1.000000000000001

?

To solve: change your metric to mae, or change your problem to actual classification before using accuracy.

Training accuracy remains zero and Validation accuracy remains constant

1 Answers1