Accuracy and Validation Accuracy stay unchanged while both losses reduce. Tried everything I could find, still doesn't work

Question

So, I am trying to code a multivariate LSTM for time series forecasting, and in my model, the losses decrease but accuracy metrics do not change at all. I tried changing number of neurons, layers, learning rate, early stopping, activation function on the output layer, and l2 regularization but nothing works. I am a beginner in machine learning, and so any help would be appreciated.Most of my efforts were like throwing stones in the dark. I am attaching a the GitHub link to my code, as well as a few of the training epochs.

# Importing the Keras libraries and packages
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from keras.layers import Dropout
from keras.regularizers import l2
from keras.optimizers import Adam
from keras.callbacks import EarlyStopping

model = Sequential()
model.add(LSTM(64,activation='sigmoid',return_sequences=True,input_shape = (trainX.shape[1],trainX.shape[2])))
model.add(LSTM(32,activation='sigmoid',return_sequences=False))
model.add(Dropout(0.3))
model.add(Dense(trainY.shape[1]))
opt = Adam(learning_rate= 1e-3)
model.compile(optimizer='adam',loss = 'mse', metrics=['accuracy'])
model.summary()

Model: "sequential_3"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm_6 (LSTM)                (None, 200, 64)           19200     
_________________________________________________________________
lstm_7 (LSTM)                (None, 32)                12416     
_________________________________________________________________
dropout_3 (Dropout)          (None, 32)                0         
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 33        
=================================================================
Total params: 31,649
Trainable params: 31,649
Non-trainable params: 0

es_callback = EarlyStopping(monitor='val_loss', patience=3)
history = model.fit(trainX,trainY,epochs=40,batch_size= 32,verbose=1,validation_split=0.2, callbacks= [es_callback])

Epoch 1/40
214/214 [==============================] - 58s 169ms/step - loss: 0.1663 - accuracy: 0.0000e+00 - val_loss: 0.0483 - val_accuracy: 5.8617e-04
Epoch 2/40
214/214 [==============================] - 35s 164ms/step - loss: 0.0497 - accuracy: 0.0000e+00 - val_loss: 0.0446 - val_accuracy: 5.8617e-04
Epoch 3/40
214/214 [==============================] - 35s 164ms/step - loss: 0.0309 - accuracy: 0.0000e+00 - val_loss: 0.0092 - val_accuracy: 5.8617e-04
Epoch 4/40
214/214 [==============================] - 35s 163ms/step - loss: 0.0143 - accuracy: 0.0000e+00 - val_loss: 0.0230 - val_accuracy: 5.8617e-04
Epoch 5/40
214/214 [==============================] - 35s 163ms/step - loss: 0.0115 - accuracy: 0.0000e+00 - val_loss: 0.0160 - val_accuracy: 5.8617e-04
Epoch 6/40
214/214 [==============================] - 35s 163ms/step - loss: 0.0099 - accuracy: 0.0000e+00 - val_loss: 0.0172 - val_accuracy: 5.8617e-04

My code: https://github.com/RiddhimanRaut/Deep-Learning-based-CPR-estimation/blob/main/CPR_prediction_multivariate_LSTM_tobetrialled_1.ipynb

Thank you!

Answer 1

Accuracy is the metric for classification tasks. To measure if a regression model is good or not, measurement such as MSE can be applied. I think the discussion here can provide more information.