bug in sampling function

[BenjaminWegener]

text generation via nuleus sampling is buggy, sampling quality is poor

-> python implementation as reference:
`import numpy as np

def generate(model,
bpe,
texts,
length=100,
top_k=1,
top_p=0.9,
temperature=1.0):
"""Generate text after the given contexts.
:param model: The trained model.
:param bpe: Byte pair encoding object.
:param texts: A list of texts.
:param length: The length of following texts to be generated.
:param top_k: Choose the next token from top K.
:param temperature: Randomness in boltzmann distribution.
:return: A list of generated texts.
"""
batch_size = len(texts)
encodes = [bpe.encode(text) for text in texts]
text_lens = [len(encode) for encode in encodes]
max_len = max(text_lens)
input_data = [encode + [0] * (max_len - len(encode)) for encode in encodes]
for shift in range(length):
output_data = model.predict(np.array(input_data))
for index in range(batch_size):
probs = [(prob, i) for i, prob in enumerate(output_data[index, text_lens[index] + shift - 1])]
probs.sort(reverse=True)
probs = probs[:top_k]
indices, probs = list(map(lambda x: x[1], probs)), list(map(lambda x: x[0], probs))
probs = np.array(probs) / temperature
probs = probs - np.max(probs)
probs = np.exp(probs)
probs = probs / np.sum(probs)
if top_p > 0.0:
-np.sort(-probs)
while np.sum(probs) > top_p:
probs = probs[:-1]
indices = indices[:-1]
probs /= probs.sum()
next_token = np.random.choice(indices, p=probs)
input_data[index].append(0)
input_data[index][text_lens[index] + shift] = next_token
outputs = [bpe.decode(input_data[index][:text_lens[index] + length]) for index in range(batch_size)]
return outputs`