pythontensorflowkeras

Why my neural network's val_accuracy is always same?


I have started studying ML using keras and tensorflow and i wanted to train neural network with TSV-file. However, my val_accuracy is always same. TSV-file have 1000 arguments for using in prediction and 1 output that can be only 0 or 1.

I've tried using different optimizers and losses, varying the number of Dense layers, units, and activation functions. Here is my model

model = keras.Sequential([
    keras.Input((1000,)),
    normalize,
    keras.layers.Dense(1000, activation=keras.activations.relu),
    keras.layers.Dense(200, activation=keras.activations.relu),
    keras.layers.Dense(50, activation=keras.activations.relu),
    keras.layers.Dense(1, )
])

model.compile(optimizer='adam', loss=keras.losses.BinaryCrossentropy(from_logits=True), metrics=['accuracy'])

losses = model.fit(train_data, train_labels, epochs=80, validation_split=0.2)

after model.fit i have this result:

Epoch 77/80
4/4 ━━━━━━━━━━━━━━━━━━━━ 0s 36ms/step - accuracy: 0.3374 - loss: nan - val_accuracy: 0.1786 - val_loss: nan
Epoch 78/80
4/4 ━━━━━━━━━━━━━━━━━━━━ 0s 24ms/step - accuracy: 0.3603 - loss: nan - val_accuracy: 0.1786 - val_loss: nan
Epoch 79/80
4/4 ━━━━━━━━━━━━━━━━━━━━ 0s 25ms/step - accuracy: 0.3041 - loss: nan - val_accuracy: 0.1786 - val_loss: nan
Epoch 80/80
4/4 ━━━━━━━━━━━━━━━━━━━━ 0s 25ms/step - accuracy: 0.3645 - loss: nan - val_accuracy: 0.1786 - val_loss: nan

where val_accuracy is always 0.1786 and loss nan. Maybe i dont understand how to use Dense layers. So how can i fix this? For additional information u may check my code on Google Colab with TSV-file


Solution

  • The output from model.fit indicates that it cannot calculate the loss. If the loss remains as nan, the training process will not adjust the parameters, and consequently, the model will not improve its val_accuracy.

    I noticed in your source code that you are replacing "?" with None:

    train_data = np.array(data.map(lambda x: None if x == '?' else int(x)))
    

    With these None values, the model will be unable to calculate the output. I recommend filtering out lines with None before using them for training.