LM_generate_sample.py

def generate_sample(lm, index2word, max_tokens=25):
    """
    Generates a string, sample a word from the distribution at each time step.
    :param lm - the language model
    :param index2word - a mapping from the index of a word in the vocabulary to the word itself
    """ 
    generated_sentence = '<s>'
    generated_tokens = 0
    curr_token = None

    while curr_token != '</s>' and generated_tokens < max_tokens:
        curr_distribution = lm(generated_sentence)  # vector of probabilities
        selected_index = np.random.choice(range(len(vocab)), p=curr_distribution)
        curr_token = index2word[int(selected_index)]
        generated_sentence += ' ' + curr_token
        generated_tokens += 1

    return generated_sentence


generated_str = generate_sample(stupid_lm, vocab)
print(generated_str)