model.predict() is not producing the expected labels?

user_12000601 0
Viewed 860

I am doing a simple binary text classification. The steps go roughly like this:

  1. preprocess training data with CountVectorizer()
  2. build a keras Sequential() model
  3. model.fit(x_train, y_train)
  4. model.evaluate(x_val, y_val)
  5. model.predict(x_test)

I am stuck on step 5 - when I print the predicted values, I get a numpy array of:

 [0.9434484 ]
 [0.3787447 ]
 ...
 [0.87870705]
 [0.7575223 ]
 [0.39714795]]

Since I am doing a binary classification, and my labels are 0 and 1, I expected the prediction output to be the same? Now it seems like it predicts the probability between the labels 0 and 1, which is not what I wanted. Do I need to encode the prediction output somehow so that it returns the proper labels or have I done something wrong in the steps before??

keras classification python
2 Answers

One solution is to use simple statistical interpretation where we will be using 0.5 cutoff. Thus everything above 0.5 will be treated as 1 and below as 0.

import numpy as np

pred =  np.array([[0.9434484 ]
 ,[0.3787447 ]
 ,[0.87870705]
 ,[0.7575223 ]
 ,[0.39714795]])

np.round(pred)
Out[37]: 
array([[1.],
       [0.],
       [1.],
       [1.],
       [0.]])

If results is not a probabilities then sth like :

def sigmoid(x):
  return 1 / (1 + math.exp(-x))

have to be used to scale it to 0-1 scale.

user_54425270

The step 5 model.predict(x_test) can be replaced by:

model.predict_classes(x_test)

to predict classes in sequential model. In case you ever need this in functional model in future, this is the solution:

y_prob = model.predict(x_test) 
y_classes = y_prob.argmax(axis=-1)
user_31574280
Related