keras model 总是预测类别而不是概率

Question

I am trying to do a binary classification using MLP by outputting a 2 neuron Dense layer with softmax as its activation function. But i keep getting predicted classes instead.

So what am i doing wrong here?

Please note that i this is a shorter version of my main NN.

inp = tf.keras.layers.Input(shape = (len(feat_cols),))
x = tf.keras.layers.Dense(32, activation = "relu")(inp)
out = tf.keras.layers.Dense(2, activation= "softmax")(x)
model = tf.keras.models.Model(inputs = inp, outputs = out)
model.compile(loss='binary_crossentropy',
          optimizer=tf.keras.optimizers.RMSprop(),
          metrics=['accuracy',"roc_auc"])
model.fit(temp[feat_cols].values.astype('float64'), np_utils.to_categorical(temp[target].values.astype('float64')), epochs = 1)
model.predict(test_df[feat_cols].values.astype("float64"))

My output for the above model is:

750/750 [==============================] - 2s 2ms/step - loss: 948.7341 - accuracy: 0.5005
array([[0., 1.],
       [0., 1.],
       [1., 0.],
       ...,
       [1., 0.],
       [1., 0.],
       [1., 0.]], dtype=float32)

Edit:

The target is 0 or 1 labels and they are provided to the model as one hot encoded labels. And i am using my sklearn roc_auc metric passed as a py_function. I have also tried by removing the.values.astype("float64") part...still so difference.

Answer 1

You use binary_crossentropy loss and 2 neuron output. usually, binary classification means you have 1 output and tries to classify between 0 and 1. try categorical_crossentropy as your loss, you'll get the probability of prediction.