Get top 3 prediction of LSTM instead of only the top-CodePudding

I have a LSTM model trained on text content. And now I want to use that model to generate some sentences. But instead of always picking the best option, i want it to select from for example the top 3, so that it can produce different sentences with the same input, because now I get the same answer for almost every input. How do i modify this code so that is possible, I know I need to remove the np.argmax but i don't know how to the return the index of the top 3 highest values.

Current code:

def prediction(seed_text, next_words): 
  for _ in range(next_words):
    token_list = tokenizer.texts_to_sequences([seed_text])[0]
    token_list = pad_sequences([token_list], maxlen=max_seq_length-1, padding='pre')
    predicted = np.argmax(model.predict(token_list, verbose=0), axis=-1)
    ouput_word = ""
    for word, index in tokenizer.word_index.items():
      if index == predicted:
        output_word = word
        break
  
    seed_text  = ' ' output_word
  return seed_text

CodePudding user response：

np.argsort will give you the indices of the items in an array in the order that sorts them small to large: https://numpy.org/doc/stable/reference/generated/numpy.argsort.html

Here's an example using argsort. Note that the one with the lowest prediction (index 2, "c" with the predicted value of 0.05) is left out of what is printed.

import numpy as np

word_index = {'a': 0, 'b': 1, 'c': 2, 'd': 3}

predictions = np.array([0.1, 0.7, 0.05, 0.15])

# add negative to sort large to small; slice to select just up to 3rd index
top_3 = np.argsort(-predictions)[:3]

for word, index in word_index.items():
    if index in top_3:
        print(word)
#> a
#> b
#> d