使用 CNN 进行推文情绪预测

Question

I have a built a CNN model for tweet emotion detection and the final step is as follows:

tweets_emotion = model.predict(val_tweets, verbose= 0)

which gave me predicted output like this

array([[3.1052819e-01, 2.7634043e-01, 1.6270137e-03, 7.7674150e-01],
       [5.0230421e-02, 7.7430069e-01, 7.7313791e-09, 2.0278792e-01],
       [9.9952579e-01, 1.3450404e-03, 5.8804121e-20, 3.2991991e-07],
       ...,
       [3.9727339e-01, 2.8888196e-01, 1.9649005e-02, 2.1239746e-01],
       [1.2528910e-01, 3.2127723e-01, 3.2503495e-03, 5.5401272e-01],
       [5.8543805e-02, 4.5720499e-05, 2.9060062e-12, 9.3766922e-01]],
      dtype=float32)

my actual output should look like this:

array([[1., 0., 0., 0.],
       [1., 0., 0., 0.],
       [1., 0., 0., 0.],
       ...,
       [0., 0., 0., 1.],
       [0., 0., 0., 1.],
       [0., 0., 0., 1.]], dtype=float32)

Is there a way to convert my predicted output (tweets_emotion) to look like the output that I expected?

Answer 1

Using the 6 predictions example you show here:

import numpy as np

tweets_emotion = np.array([[3.1052819e-01, 2.7634043e-01, 1.6270137e-03, 7.7674150e-01],
                           [5.0230421e-02, 7.7430069e-01, 7.7313791e-09, 2.0278792e-01],
                           [9.9952579e-01, 1.3450404e-03, 5.8804121e-20, 3.2991991e-07],
                           [3.9727339e-01, 2.8888196e-01, 1.9649005e-02, 2.1239746e-01],
                           [1.2528910e-01, 3.2127723e-01, 3.2503495e-03, 5.5401272e-01],
                           [5.8543805e-02, 4.5720499e-05, 2.9060062e-12, 9.3766922e-01]])

tweets_emotion_class = np.argmax(tweets_emotion, axis=1)
tweets_emotion_class
# array([3, 1, 0, 0, 3, 3])

You should be able to verify by simple visual inspection that indeed, the maximum values of each array element are the ones shown in tweets_emotion_class .

Irrelevant to your issue but, as mentioned in the comments, sigmoid activation for the last network layer does not make sense in single-label multi-class settings with one-hot encoded labels, as your setting seems to be - you should change it to softmax .