Function _top_k_sample

LanguageNetwork/GPT2/train/modeling.py:376–407

Does top-k sampling. if ignore_ids is on, then we will zero out those logits. :param logits: [batch_size, vocab_size] tensor :param ignore_ids: [vocab_size] one-hot representation of the indices we'd like to ignore and never predict, like padding maybe :param

(logits, ignore_ids=None, num_samples=1, k=10)

def _top_k_sample(logits, ignore_ids=None, num_samples=1, k=10):
    """
    Does top-k sampling. if ignore_ids is on, then we will zero out those logits.
    :param logits: [batch_size, vocab_size] tensor
    :param ignore_ids: [vocab_size] one-hot representation of the indices we'd like to ignore and never predict,
                        like padding maybe
    :param p: topp threshold to use, either a float or a [batch_size] vector
    :return: [batch_size, num_samples] samples
    # TODO FIGURE OUT HOW TO DO THIS ON TPUS. IT'S HELLA SLOW RIGHT NOW, DUE TO ARGSORT I THINK
    """
    with tf.variable_scope('top_p_sample'):
        batch_size, vocab_size = get_shape_list(logits, expected_rank=2)

        probs = tf.nn.softmax(logits if ignore_ids is None else logits - tf.cast(ignore_ids[None], tf.float32) * 1e10,
                              axis=-1)
        # [batch_size, vocab_perm]
        indices = tf.argsort(probs, direction='DESCENDING')

        # find the top pth index to cut off. careful we don't want to cutoff everything!
        # result will be [batch_size, vocab_perm]
        k_expanded = k if isinstance(k, int) else k[:, None]
        exclude_mask = tf.range(vocab_size)[None] >= k_expanded

        # OPTION A - sample in the sorted space, then unsort.
        logits_to_use = tf.batch_gather(logits, indices) - tf.cast(exclude_mask, tf.float32) * 1e10
        sample_perm = tf.random.categorical(logits=logits_to_use, num_samples=num_samples)
        sample = tf.batch_gather(indices, sample_perm)

    return {
        'probs': probs,
        'sample': sample,
    }


class GroverModel(object):