replay_buffer.py 文件源码

def sample(self, amount, temperature=1):
        priorities = self.buffer.buffers['priority'].value()[:self.size()]
        logprobs = tf.log(priorities / tf.reduce_sum(priorities)) / temperature
        positions = tf.multinomial(logprobs[None, ...], amount)[0]
        return [ tf.gather(b, positions) for key,b in self.buffer.buffers.items() if key != 'priority' ]