如何在损失正常的情况下修复深度学习中的零精度

Question

I have a dataset of 50x22 which includes 22 features. The target is to classify the target which is scaled from 1 to 5, equivalently 5 classes. I used random forest with 98% accuracy but the validation is 63% which is not satisfiable. That's why I decided to create a deep model and I created a model with 3 layers. The result of loss is satisfiable around 6.7*10e-4 but the accuracy is fixed with zero. I think there is some thing wrong in my code. So, what's the problem?

def build_and_compile_model(norm):
    model = keras.Sequential([
    norm,
    layers.Dense(32, activation='relu'),
    layers.Dense(32, activation='relu'),
    layers.Dense(1,activation='sigmoid')
    ])
 model.compile(optimizer='sgd',
          loss='binary_crossentropy',
          metrics=[tf.keras.metrics.Accuracy()])
return model

def plot_acc(history):
    plt.plot(history.history['accuracy'], label='accuracy')
    plt.plot(history.history['val_accuracy'], label='val_accuracy')
    plt.ylim([0, 1])
    plt.xlabel('Epoch')
    plt.ylabel('Accuracy [GSR]')
    plt.legend()
    plt.grid(True)

dnn_qoe_model = build_and_compile_model(feature_normalizer)
dnn_qoe_model.summary()

history = dnn_qoe_model.fit(
          train_features[:22], train_labels,
          validation_split=0.2,
          verbose=0, epochs=100)
plot_acc(history)

the loss plot

Answer 1

You are using loss='binary_crossentropy' and layers.Dense(1,activation='sigmoid') , which are used for binary classification problems.

Since you are looking to predict one of 5 classes, you are looking at a multi class problem.

If your target is one hot encoded which would look like so: [0,1,0,0,0] for one class, you should use layers.Dense(5,activation='softmax') and loss='categorical_crossentropy' .

If your target isn't one hot encoded, which means the response is an integer referring to the class number, which would be [1] (position of the positive class) in the previous example, you should use layers.Dense(5,activation='softmax') , and change the loss function to loss='sparse_categorical_crossentropy' , as your target variable is encoded as a sparse vector (refers to the index of the item containing a 1 in a vector of zeros)