分类器如何预测一个类？ （对于张量流/keras）

Question

I want to implement the model by myself and thus have to know how does it classify the data. I build a model for 12-class classifier and it predicts fine. But the last conv layer just outputs 12 floating point value and I don't know how it suddenly predicts the right class.

Can someone explain for me? Like is it depend on some threshold or it chooses the max value or something? Thanks!

Answer 1

According to the documentation of SparseCategoricalAccuracy , it's equivalent to this computation:

acc = np.dot(sample_weight, np.equal(y_true, np.argmax(y_pred, axis=1))

This means that it calculates the frequency with which the maximum value per row of y_pred matches y_true . For instance:

m = tf.keras.metrics.SparseCategoricalAccuracy()
m.update_state([[2], [1]], [[0.1, 0.6, 0.3],  # max at 1 
                            [0.05, 0.95, 0]]) # max at 1 
m.result().numpy()

0.5

Because [2] != [1] and [1] == [1] so 0.5 of the time, they are equal.

Answer 2

you need to add a flatten layer to your model followed by a dense layer. The dense layer should have 12 nodes and use the softmax activation as shown below' Your model will now output a list of 12 probability values for each image.

flatten=tf.keras.layers.Flatten()(last_conv_layer)
output = Dense(12, activation='softmax')(flatten)
#after you train you can evaluate your model on your test set using model.evaluate() 
#to make Predictions use model.predict()
predictions=model.predict(.....
#you can get the index of the predicted class for the images you predict with
for p in predictions:
    predicted_index=argmax(p)
    print (predicted_index)

documentation for model.evaluate and model.predict is here. Don't forget to recompile your model after you add the two layers.