modeling_tf_xlnet.py

# coding=utf-8
# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University Authors and the HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" TF 2.0 XLNet model.
"""
from __future__ import absolute_import, division, print_function, unicode_literals

import json
import logging
import math
import os
import sys
from io import open

import numpy as np
import tensorflow as tf

from .configuration_xlnet import XLNetConfig
from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list
from .file_utils import add_start_docstrings
from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model


logger = logging.getLogger(__name__)

TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP = {
    'xlnet-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/xlnet-base-cased-tf_model.h5",
    'xlnet-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/xlnet-large-cased-tf_model.h5",
}


def load_xlnet_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
    tf_inputs = tf.constant(inputs_list)
    tfo = tf_model(tf_inputs, training=False)  # build the network
    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)


def gelu(x):
    """ Implementation of the gelu activation function.
        XLNet is using OpenAI GPT's gelu
        Also see https://arxiv.org/abs/1606.08415
    """
    cdf = 0.5 * (1.0 + tf.tanh(
        (np.sqrt(2 / np.pi) * (x + 0.044715 * tf.pow(x, 3)))))
    return x * cdf


def swish(x):
    return x * tf.sigmoid(x)


ACT2FN = {"gelu": tf.keras.layers.Activation(gelu),
          "relu": tf.keras.activations.relu,
          "swish": tf.keras.layers.Activation(swish)}


class TFXLNetRelativeAttention(tf.keras.layers.Layer):
    def __init__(self, config, **kwargs):
        super(TFXLNetRelativeAttention, self).__init__(**kwargs)
        self.output_attentions = config.output_attentions

        if config.d_model % config.n_head != 0:
            raise ValueError(
                "The hidden size (%d) is not a multiple of the number of attention "
                "heads (%d)" % (config.d_model, config.n_head))

        self.n_head = config.n_head
        self.d_head = config.d_head
        self.d_model = config.d_model
        self.scale = 1 / (config.d_head ** 0.5)
        self.initializer_range = config.initializer_range

        self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name='layer_norm')
        self.dropout = tf.keras.layers.Dropout(config.dropout)

    def build(self, input_shape):
        initializer = tf.random_normal_initializer(mean=0., stddev=self.initializer_range)
        self.q = self.add_weight(shape=(self.d_model, self.n_head, self.d_head),
                                 initializer=initializer,
                                 trainable=True, name='q')
        self.k = self.add_weight(shape=(self.d_model, self.n_head, self.d_head),
                                 initializer=initializer,
                                 trainable=True, name='k')
        self.v = self.add_weight(shape=(self.d_model, self.n_head, self.d_head),
                                 initializer=initializer,
                                 trainable=True, name='v')
        self.o = self.add_weight(shape=(self.d_model, self.n_head, self.d_head),
                                 initializer=initializer,
                                 trainable=True, name='o')
        self.r = self.add_weight(shape=(self.d_model, self.n_head, self.d_head),
                                 initializer=initializer,
                                 trainable=True, name='r')
        self.r_r_bias = self.add_weight(shape=(self.n_head, self.d_head),
                                        initializer=initializer,
                                        trainable=True, name='r_r_bias')
        self.r_s_bias = self.add_weight(shape=(self.n_head, self.d_head),
                                        initializer=initializer,
                                        trainable=True, name='r_s_bias')
        self.r_w_bias = self.add_weight(shape=(self.n_head, self.d_head),
                                        initializer=initializer,
                                        trainable=True, name='r_w_bias')
        self.seg_embed = self.add_weight(shape=(2, self.n_head, self.d_head),
                                        initializer=initializer,
                                        trainable=True, name='seg_embed')
        super(TFXLNetRelativeAttention, self).build(input_shape)

    def prune_heads(self, heads):
        raise NotImplementedError

    @staticmethod
    def rel_shift(x, klen=-1):
        """perform relative shift to form the relative attention score."""
        x_size = shape_list(x)

        x = tf.reshape(x, (x_size[1], x_size[0], x_size[2], x_size[3]))
        x = x[1:, ...]
        x = tf.reshape(x, (x_size[0], x_size[1] - 1, x_size[2], x_size[3]))
        x = x[:, 0:klen, :, :]
        # x = torch.index_select(x, 1, torch.arange(klen, device=x.device, dtype=torch.long))

        return x

    def rel_attn_core(self, inputs, training=False):
        """Core relative positional attention operations."""

        q_head, k_head_h, v_head_h, k_head_r, seg_mat, attn_mask, head_mask = inputs

        # content based attention score
        ac = tf.einsum('ibnd,jbnd->ijbn', q_head + self.r_w_bias, k_head_h)

        # position based attention score
        bd = tf.einsum('ibnd,jbnd->ijbn', q_head + self.r_r_bias, k_head_r)
        bd = self.rel_shift(bd, klen=ac.shape[1])

        # segment based attention score
        if seg_mat is None:
            ef = 0
        else:
            ef = tf.einsum('ibnd,snd->ibns', q_head + self.r_s_bias, self.seg_embed)
            ef = tf.einsum('ijbs,ibns->ijbn', seg_mat, ef)

        # merge attention scores and perform masking
        attn_score = (ac + bd + ef) * self.scale
        if attn_mask is not None:
            # attn_score = attn_score * (1 - attn_mask) - 1e30 * attn_mask
            if attn_mask.dtype == tf.float16:
                attn_score = attn_score - 65500 * attn_mask
            else:
                attn_score = attn_score - 1e30 * attn_mask

        # attention probability
        attn_prob = tf.nn.softmax(attn_score, axis=1)

        attn_prob = self.dropout(attn_prob, training=training)

        # Mask heads if we want to
        if head_mask is not None:
            attn_prob = attn_prob * head_mask

        # attention output
        attn_vec = tf.einsum('ijbn,jbnd->ibnd', attn_prob, v_head_h)

        if self.output_attentions:
            return attn_vec, attn_prob

        return attn_vec

    def post_attention(self, inputs, residual=True, training=False):
        """Post-attention processing."""
        # post-attention projection (back to `d_model`)
        h, attn_vec = inputs

        attn_out = tf.einsum('ibnd,hnd->ibh', attn_vec, self.o)

        attn_out = self.dropout(attn_out, training=training)

        if residual:
            attn_out = attn_out + h
        output = self.layer_norm(attn_out)

        return output

    def call(self, inputs, training=False):
        (h, g, attn_mask_h, attn_mask_g,
         r, seg_mat, mems, target_mapping, head_mask) = inputs

        if g is not None:
            ###### Two-stream attention with relative positional encoding.
            # content based attention score
            if mems is not None and mems.shape.ndims > 1:
                cat = tf.concat([mems, h], axis=0)
            else:
                cat = h

            # content-based key head
            k_head_h = tf.einsum('ibh,hnd->ibnd', cat, self.k)

            # content-based value head
            v_head_h = tf.einsum('ibh,hnd->ibnd', cat, self.v)

            # position-based key head
            k_head_r = tf.einsum('ibh,hnd->ibnd', r, self.r)

            ##### h-stream
            # content-stream query head
            q_head_h = tf.einsum('ibh,hnd->ibnd', h, self.q)

            # core attention ops
            attn_vec_h = self.rel_attn_core(
                [q_head_h, k_head_h, v_head_h, k_head_r, seg_mat, attn_mask_h, head_mask],
                training=training)

            if self.output_attentions:
                attn_vec_h, attn_prob_h = attn_vec_h

            # post processing
            output_h = self.post_attention([h, attn_vec_h], training=training)

            ##### g-stream
            # query-stream query head
            q_head_g = tf.einsum('ibh,hnd->ibnd', g, self.q)

            # core attention ops
            if target_mapping is not None:
                q_head_g = tf.einsum('mbnd,mlb->lbnd', q_head_g, target_mapping)
                attn_vec_g = self.rel_attn_core(
                    [q_head_g, k_head_h, v_head_h, k_head_r, seg_mat, attn_mask_g, head_mask],
                    training=training)

                if self.output_attentions:
                    attn_vec_g, attn_prob_g = attn_vec_g

                attn_vec_g = tf.einsum('lbnd,mlb->mbnd', attn_vec_g, target_mapping)
            else:
                attn_vec_g = self.rel_attn_core(
                    [q_head_g, k_head_h, v_head_h, k_head_r, seg_mat, attn_mask_g, head_mask],
                    training=training)

                if self.output_attentions:
                    attn_vec_g, attn_prob_g = attn_vec_g

            # post processing
            output_g = self.post_attention([g, attn_vec_g], training=training)

            if self.output_attentions:
                attn_prob = attn_prob_h, attn_prob_g

        else:
            ###### Multi-head attention with relative positional encoding
            if mems is not None and mems.shape.ndims > 1:
                cat = tf.concat([mems, h], axis=0)
            else:
                cat = h

            # content heads
            q_head_h = tf.einsum('ibh,hnd->ibnd', h, self.q)
            k_head_h = tf.einsum('ibh,hnd->ibnd', cat, self.k)
            v_head_h = tf.einsum('ibh,hnd->ibnd', cat, self.v)

            # positional heads
            k_head_r = tf.einsum('ibh,hnd->ibnd', r, self.r)

            # core attention ops
            attn_vec = self.rel_attn_core(
                [q_head_h, k_head_h, v_head_h, k_head_r, seg_mat, attn_mask_h, head_mask],
                training=training)

            if self.output_attentions:
                attn_vec, attn_prob = attn_vec

            # post processing
            output_h = self.post_attention([h, attn_vec], training=training)
            output_g = None

        outputs = (output_h, output_g)
        if self.output_attentions:
            outputs = outputs + (attn_prob,)
        return outputs

class TFXLNetFeedForward(tf.keras.layers.Layer):
    def __init__(self, config, **kwargs):
        super(TFXLNetFeedForward, self).__init__(**kwargs)
        self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name='layer_norm')
        self.layer_1 = tf.keras.layers.Dense(config.d_inner, name='layer_1')
        self.layer_2 = tf.keras.layers.Dense(config.d_model, name='layer_2')
        self.dropout = tf.keras.layers.Dropout(config.dropout)
        if isinstance(config.ff_activation, str) or \
                (sys.version_info[0] == 2 and isinstance(config.ff_activation, unicode)):
            self.activation_function = ACT2FN[config.ff_activation]
        else:
            self.activation_function = config.ff_activation

    def call(self, inp, training=False):
        output = inp
        output = self.layer_1(output)
        output = self.activation_function(output)
        output = self.dropout(output, training=training)
        output = self.layer_2(output)
        output = self.dropout(output, training=training)
        output = self.layer_norm(output + inp)
        return output

class TFXLNetLayer(tf.keras.layers.Layer):
    def __init__(self, config, **kwargs):
        super(TFXLNetLayer, self).__init__(**kwargs)
        self.rel_attn = TFXLNetRelativeAttention(config, name='rel_attn')
        self.ff = TFXLNetFeedForward(config, name='ff')
        self.dropout = tf.keras.layers.Dropout(config.dropout)

    def call(self, inputs, training=False):
        outputs = self.rel_attn(inputs, training=training)
        output_h, output_g = outputs[:2]

        if output_g is not None:
            output_g = self.ff(output_g, training=training)
        output_h = self.ff(output_h, training=training)

        outputs = (output_h, output_g) + outputs[2:]  # Add again attentions if there are there
        return outputs


class TFXLNetLMHead(tf.keras.layers.Layer):
    def __init__(self, config, input_embeddings, **kwargs):
        super(TFXLNetLMHead, self).__init__(**kwargs)
        self.vocab_size = config.vocab_size
        # The output weights are the same as the input embeddings, but there is
        # an output-only bias for each token.
        self.input_embeddings = input_embeddings

    def build(self, input_shape):
        self.bias = self.add_weight(shape=(self.vocab_size,),
                                    initializer='zeros',
                                    trainable=True,
                                    name='bias')
        super(TFXLNetLMHead, self).build(input_shape)

    def call(self, hidden_states):
        hidden_states = self.input_embeddings(hidden_states, mode="linear")
        hidden_states = hidden_states + self.bias
        return hidden_states


class TFXLNetMainLayer(tf.keras.layers.Layer):
    def __init__(self, config, **kwargs):
        super(TFXLNetMainLayer, self).__init__(**kwargs)
        self.output_attentions = config.output_attentions
        self.output_hidden_states = config.output_hidden_states

        self.mem_len = config.mem_len
        self.reuse_len = config.reuse_len
        self.d_model = config.d_model
        self.same_length = config.same_length
        self.attn_type = config.attn_type
        self.bi_data = config.bi_data
        self.clamp_len = config.clamp_len
        self.n_layer = config.n_layer
        self.use_bfloat16 = config.use_bfloat16
        self.initializer_range = config.initializer_range

        self.word_embedding = TFSharedEmbeddings(config.n_token, config.d_model, initializer_range=config.initializer_range, name='word_embedding')
        self.layer = [TFXLNetLayer(config, name='layer_._{}'.format(i)) for i in range(config.n_layer)]
        self.dropout = tf.keras.layers.Dropout(config.dropout)

    def build(self, input_shape):
        initializer = tf.random_normal_initializer(mean=0., stddev=self.initializer_range)
        self.mask_emb = self.add_weight(shape=(1, 1, self.d_model),
                                 initializer=initializer,
                                 trainable=True, name='mask_emb')

    def _resize_token_embeddings(self, new_num_tokens):
        raise NotImplementedError

    def _prune_heads(self, heads_to_prune):
        raise NotImplementedError

    def create_mask(self, qlen, mlen, dtype=tf.float32):
        """
        Creates causal attention mask. Float mask where 1.0 indicates masked, 0.0 indicates not-masked.

        Args:
            qlen: TODO Lysandre didn't fill
            mlen: TODO Lysandre didn't fill

        ::

                  same_length=False:      same_length=True:
                  <mlen > <  qlen >       <mlen > <  qlen >
               ^ [0 0 0 0 0 1 1 1 1]     [0 0 0 0 0 1 1 1 1]
                 [0 0 0 0 0 0 1 1 1]     [1 0 0 0 0 0 1 1 1]
            qlen [0 0 0 0 0 0 0 1 1]     [1 1 0 0 0 0 0 1 1]
                 [0 0 0 0 0 0 0 0 1]     [1 1 1 0 0 0 0 0 1]
               v [0 0 0 0 0 0 0 0 0]     [1 1 1 1 0 0 0 0 0]

        """
        attn_mask = tf.ones([qlen, qlen], dtype=dtype)
        mask_u = tf.matrix_band_part(attn_mask, 0, -1)
        mask_dia = tf.matrix_band_part(attn_mask, 0, 0)
        attn_mask_pad = tf.zeros([qlen, mlen], dtype=dtype)
        ret = tf.concat([attn_mask_pad, mask_u - mask_dia], 1)
        if self.same_length:
            mask_l = tf.matrix_band_part(attn_mask, -1, 0)
            ret = tf.concat([ret[:, :qlen] + mask_l - mask_dia, ret[:, qlen:]], 1)
        return ret

    def cache_mem(self, curr_out, prev_mem):
        """cache hidden states into memory."""
        if self.mem_len is None or self.mem_len == 0:
            return None
        else:
            if self.reuse_len is not None and self.reuse_len > 0:
                curr_out = curr_out[:self.reuse_len]

            if prev_mem is None:
                new_mem = curr_out[-self.mem_len:]
            else:
                new_mem = tf.concat([prev_mem, curr_out], 0)[-self.mem_len:]

        return tf.stop_gradient(new_mem)

    @staticmethod
    def positional_embedding(pos_seq, inv_freq, bsz=None):
        sinusoid_inp = tf.einsum('i,d->id', pos_seq, inv_freq)
        pos_emb = tf.concat([tf.sin(sinusoid_inp), tf.cos(sinusoid_inp)], axis=-1)
        pos_emb = pos_emb[:, None, :]

        if bsz is not None:
            pos_emb = tf.tile(pos_emb, [1, bsz, 1])

        return pos_emb

    def relative_positional_encoding(self, qlen, klen, bsz=None, dtype=None):
        """create relative positional encoding."""
        freq_seq = tf.range(0, self.d_model, 2.0)
        if dtype is not None and dtype != tf.float32:
            freq_seq = tf.cast(freq_seq, dtype=dtype)
        inv_freq = 1 / (10000 ** (freq_seq / self.d_model))

        if self.attn_type == 'bi':
            # beg, end = klen - 1, -qlen
            beg, end = klen, -qlen
        elif self.attn_type == 'uni':
            # beg, end = klen - 1, -1
            beg, end = klen, -1
        else:
            raise ValueError('Unknown `attn_type` {}.'.format(self.attn_type))

        if self.bi_data:
            fwd_pos_seq = tf.range(beg, end, -1.0)
            bwd_pos_seq = tf.range(-beg, -end, 1.0)

            if dtype is not None and dtype != tf.float32:
                fwd_pos_seq = tf.cast(fwd_pos_seq, dtype=dtype)
                bwd_pos_seq = tf.cast(bwd_pos_seq, dtype=dtype)

            if self.clamp_len > 0:
                fwd_pos_seq = tf.clip_by_value(fwd_pos_seq, -self.clamp_len, self.clamp_len)
                bwd_pos_seq = tf.clip_by_value(bwd_pos_seq, -self.clamp_len, self.clamp_len)

            if bsz is not None:
                # With bi_data, the batch size should be divisible by 2.
                assert bsz%2 == 0
                fwd_pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq, bsz//2)
                bwd_pos_emb = self.positional_embedding(bwd_pos_seq, inv_freq, bsz//2)
            else:
                fwd_pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq)
                bwd_pos_emb = self.positional_embedding(bwd_pos_seq, inv_freq)

            pos_emb = tf.concat([fwd_pos_emb, bwd_pos_emb], axis=1)
        else:
            fwd_pos_seq = tf.range(beg, end, -1.0)
            if dtype is not None and dtype != tf.float32:
                fwd_pos_seq = tf.cast(fwd_pos_seq, dtype=dtype)
            if self.clamp_len > 0:
                fwd_pos_seq = tf.clip_by_value(fwd_pos_seq, -clamp_len, clamp_len)
            pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq, bsz)

        return pos_emb

    def call(self, inputs, attention_mask=None, mems=None, perm_mask=None, target_mapping=None,
            token_type_ids=None, input_mask=None, head_mask=None, training=False):
        if isinstance(inputs, (tuple, list)):
            input_ids = inputs[0]
            attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
            mems = inputs[2] if len(inputs) > 2 else mems
            perm_mask = inputs[3] if len(inputs) > 3 else perm_mask
            target_mapping = inputs[4] if len(inputs) > 4 else target_mapping
            token_type_ids = inputs[5] if len(inputs) > 5 else token_type_ids
            input_mask = inputs[6] if len(inputs) > 6 else input_mask
            head_mask = inputs[7] if len(inputs) > 7 else head_mask
            assert len(inputs) <= 8, "Too many inputs."
        elif isinstance(inputs, dict):
            input_ids = inputs.get('input_ids')
            attention_mask = inputs.get('attention_mask', attention_mask)
            mems = inputs.get('mems', mems)
            perm_mask = inputs.get('perm_mask', perm_mask)
            target_mapping = inputs.get('target_mapping', target_mapping)
            token_type_ids = inputs.get('token_type_ids', token_type_ids)
            input_mask = inputs.get('input_mask', input_mask)
            head_mask = inputs.get('head_mask', head_mask)
            assert len(inputs) <= 8, "Too many inputs."
        else:
            input_ids = inputs

        # the original code for XLNet uses shapes [len, bsz] with the batch dimension at the end
        # but we want a unified interface in the library with the batch size on the first dimension
        # so we move here the first dimension (batch) to the end

        input_ids = tf.transpose(input_ids, perm=(1, 0))
        token_type_ids = tf.transpose(token_type_ids, perm=(1, 0)) if token_type_ids is not None else None
        input_mask = tf.transpose(input_mask, perm=(1, 0)) if input_mask is not None else None
        attention_mask = tf.transpose(attention_mask, perm=(1, 0)) if attention_mask is not None else None
        perm_mask = tf.transpose(perm_mask, perm=(1, 2, 0)) if perm_mask is not None else None
        target_mapping = tf.transpose(target_mapping, perm=(1, 2, 0)) if target_mapping is not None else None

        qlen, bsz = shape_list(input_ids)[:2]
        mlen = shape_list(mems[0])[0] if mems is not None and mems[0] is not None else 0
        klen = mlen + qlen

        dtype_float = tf.bfloat16 if self.use_bfloat16 else tf.float32

        ##### Attention mask
        # causal attention mask
        if self.attn_type == 'uni':
            attn_mask = self.create_mask(qlen, mlen)
            attn_mask = attn_mask[:, :, None, None]
        elif self.attn_type == 'bi':
            attn_mask = None
        else:
            raise ValueError('Unsupported attention type: {}'.format(self.attn_type))

        # data mask: input mask & perm mask
        assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) "
        "or attention_mask (uses 0 for padding, added for compatbility with BERT). Please choose one."
        if input_mask is None and attention_mask is not None:
            input_mask = 1.0 - attention_mask
        if input_mask is not None and perm_mask is not None:
            data_mask = input_mask[None] + perm_mask
        elif input_mask is not None and perm_mask is None:
            data_mask = input_mask[None]
        elif input_mask is None and perm_mask is not None:
            data_mask = perm_mask
        else:
            data_mask = None

        if data_mask is not None:
            # all mems can be attended to
            mems_mask = tf.zeros([tf.shape(data_mask)[0], mlen, bsz],
                                dtype=dtype_float)
            data_mask = tf.concat([mems_mask, data_mask], axis=1)
            if attn_mask is None:
                attn_mask = data_mask[:, :, :, None]
            else:
                attn_mask += data_mask[:, :, :, None]

        if attn_mask is not None:
            attn_mask = tf.cast(attn_mask > 0, dtype=dtype_float)

        if attn_mask is not None:
            non_tgt_mask = -tf.eye(qlen, dtype=dtype_float)
            non_tgt_mask = tf.concat([tf.zeros([qlen, mlen], dtype=dtype_float), non_tgt_mask], axis=-1)
            non_tgt_mask = tf.cast((attn_mask + non_tgt_mask[:, :, None, None]) > 0, dtype=dtype_float)
        else:
            non_tgt_mask = None

        ##### Word embeddings and prepare h & g hidden states
        word_emb_k = self.word_embedding(input_ids)
        output_h = self.dropout(word_emb_k, training=training)
        if target_mapping is not None:
            word_emb_q = tf.tile(self.mask_emb, [tf.shape(target_mapping)[0], bsz, 1])
        # else:  # We removed the inp_q input which was same as target mapping
        #     inp_q_ext = inp_q[:, :, None]
        #     word_emb_q = inp_q_ext * self.mask_emb + (1 - inp_q_ext) * word_emb_k
            output_g = self.dropout(word_emb_q, training=training)
        else:
            output_g = None

        ##### Segment embedding
        if token_type_ids is not None:
            # Convert `token_type_ids` to one-hot `seg_mat`
            mem_pad = tf.zeros([mlen, bsz], dtype=tf.int32)
            cat_ids = tf.concat([mem_pad, token_type_ids], 0)

            # `1` indicates not in the same segment [qlen x klen x bsz]
            seg_mat = tf.cast(
                tf.logical_not(tf.equal(token_type_ids[:, None], cat_ids[None, :])),
                tf.int32)
            seg_mat = tf.one_hot(seg_mat, 2, dtype=dtype_float)
        else:
            seg_mat = None

        ##### Positional encoding
        pos_emb = self.relative_positional_encoding(qlen, klen, bsz=bsz, dtype=dtype_float)
        pos_emb = self.dropout(pos_emb, training=training)

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads] (a head_mask for each layer)
        # and head_mask is converted to shape [num_hidden_layers x qlen x klen x bsz x n_head]
        if head_mask is not None:
            if head_mask.dim() == 1:
                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(0).unsqueeze(0)
                head_mask = head_mask.expand(self.n_layer, -1, -1, -1, -1)
            elif head_mask.dim() == 2:
                head_mask = head_mask.unsqueeze(1).unsqueeze(1).unsqueeze(1)
            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
        else:
            head_mask = [None] * self.n_layer

        new_mems = ()
        if mems is None:
            mems = [None] * len(self.layer)

        attentions = []
        hidden_states = []
        for i, layer_module in enumerate(self.layer):
            # cache new mems
            new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
            if self.output_hidden_states:
                hidden_states.append((output_h, output_g) if output_g is not None else output_h)

            outputs = layer_module([output_h, output_g, non_tgt_mask, attn_mask,
                                    pos_emb, seg_mat, mems[i], target_mapping,
                                    head_mask[i]], training=training)
            output_h, output_g = outputs[:2]
            if self.output_attentions:
                attentions.append(outputs[2])

        # Add last hidden state
        if self.output_hidden_states:
            hidden_states.append((output_h, output_g) if output_g is not None else output_h)

        output = self.dropout(output_g if output_g is not None else output_h, training=training)

        # Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
        outputs = (tf.transpose(output, perm=(1, 0, 2)), new_mems)
        if self.output_hidden_states:
            if output_g is not None:
                hidden_states = tuple(tf.transpose(h, perm=(1, 0, 2)) for hs in hidden_states for h in hs)
            else:
                hidden_states = tuple(tf.transpose(hs, perm=(1, 0, 2)) for hs in hidden_states)
            outputs = outputs + (hidden_states,)
        if self.output_attentions:
            attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
            outputs = outputs + (attentions,)

        return outputs  # outputs, new_mems, (hidden_states), (attentions)


class TFXLNetPreTrainedModel(TFPreTrainedModel):
    """ An abstract class to handle weights initialization and
        a simple interface for dowloading and loading pretrained models.
    """
    config_class = XLNetConfig
    pretrained_model_archive_map = TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
    load_pt_weights = load_xlnet_pt_weights_in_tf2
    base_model_prefix = "transformer"


XLNET_START_DOCSTRING = r"""    The XLNet model was proposed in
    `XLNet: Generalized Autoregressive Pretraining for Language Understanding`_
    by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
    XLnet is an extension of the Transformer-XL model pre-trained using an autoregressive method
    to learn bidirectional contexts by maximizing the expected likelihood over all permutations
    of the input sequence factorization order.

    The specific attention pattern can be controlled at training and test time using the `perm_mask` input.

    Do to the difficulty of training a fully auto-regressive model over various factorization order,
    XLNet is pretrained using only a sub-set of the output tokens as target which are selected
    with the `target_mapping` input.

    To use XLNet for sequential decoding (i.e. not in fully bi-directional setting), use the `perm_mask` and
    `target_mapping` inputs to control the attention span and outputs (see examples in `examples/run_generation.py`)

    This model is a tf.keras.Model `tf.keras.Model`_ sub-class. Use it as a regular TF 2.0 Keras Model and
    refer to the TF 2.0 documentation for all matter related to general usage and behavior.

    .. _`XLNet: Generalized Autoregressive Pretraining for Language Understanding`:
        http://arxiv.org/abs/1906.08237

    .. _`tf.keras.Model`:
        https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model

    Note on the model inputs:
        TF 2.0 models accepts two formats as inputs:

            - having all inputs as keyword arguments (like PyTorch models), or
            - having all inputs as a list, tuple or dict in the first positional arguments.

        This second option is usefull when using `tf.keras.Model.fit()` method which currently requires having all the tensors in the first argument of the model call function: `model(inputs)`.

        If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :

        - a single Tensor with input_ids only and nothing else: `model(inputs_ids)
        - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
            `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
        - a dictionary with one or several input Tensors associaed to the input names given in the docstring:
            `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`

    Parameters:
        config (:class:`~pytorch_transformers.XLNetConfig`): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the configuration.
            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
"""

XLNET_INPUTS_DOCSTRING = r"""
    Inputs:
        **input_ids**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
            Indices of input sequence tokens in the vocabulary.
            XLNet is a model with relative position embeddings so you can either pad the inputs on
            the right or on the left.
            Indices can be obtained using :class:`pytorch_transformers.XLNetTokenizer`.
            See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
            :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
        **attention_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
            Mask to avoid performing attention on padding token indices.
            Mask values selected in ``[0, 1]``:
            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
        **mems**: (`optional`)
            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer):
            that contains pre-computed hidden-states (key and values in the attention blocks) as output by the model
            (see `mems` output below). Can be used to speed up sequential decoding and attend to longer context.
            To activate mems you need to set up config.mem_len to a positive value which will be the max number of tokens in
            the memory output by the model. E.g. `model = XLNetModel.from_pretrained('xlnet-base-case, mem_len=1024)` will
            instantiate a model which can use up to 1024 tokens of memory (in addition to the input it self).
        **perm_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, sequence_length)``:
            Mask to indicate the attention pattern for each input token with values selected in ``[0, 1]``:
            If ``perm_mask[k, i, j] = 0``, i attend to j in batch k;
            if ``perm_mask[k, i, j] = 1``, i does not attend to j in batch k.
            If None, each token attends to all the others (full bidirectional attention).
            Only used during pretraining (to define factorization order) or for sequential decoding (generation).
        **target_mapping**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, num_predict, sequence_length)``:
            Mask to indicate the output tokens to use.
            If ``target_mapping[k, i, j] = 1``, the i-th predict in batch k is on the j-th token.
            Only used during pretraining for partial prediction or for sequential decoding (generation).
        **token_type_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
            A parallel sequence of tokens (can be used to indicate various portions of the inputs).
            The type indices in XLNet are NOT selected in the vocabulary, they can be arbitrary numbers and
            the important thing is that they should be different for tokens which belong to different segments.
            The model will compute relative segment differences from the given type indices:
            0 if the segment id of two tokens are the same, 1 if not.
        **input_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
            Mask to avoid performing attention on padding token indices.
            Negative of `attention_mask`, i.e. with 0 for real tokens and 1 for padding.
            Kept for compatibility with the original code base.
            You can only uses one of `input_mask` and `attention_mask`
            Mask values selected in ``[0, 1]``:
            ``1`` for tokens that are MASKED, ``0`` for tokens that are NOT MASKED.
        **head_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
            Mask to nullify selected heads of the self-attention modules.
            Mask values selected in ``[0, 1]``:
            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
"""

@add_start_docstrings("The bare XLNet Model transformer outputing raw hidden-states without any specific head on top.",
                      XLNET_START_DOCSTRING, XLNET_INPUTS_DOCSTRING)
class TFXLNetModel(TFXLNetPreTrainedModel):
    r"""
    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
            Sequence of hidden-states at the last layer of the model.
        **mems**:
            list of ``tf.Tensor`` (one for each layer):
            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
            See details in the docstring of the `mems` input above.
        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
            list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
            of shape ``(batch_size, sequence_length, hidden_size)``:
            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

    Examples::

        import tensorflow as tf
        from pytorch_transformers import XLNetTokenizer, TFXLNetModel

        tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
        model = TFXLNetModel.from_pretrained('xlnet-large-cased')
        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
        outputs = model(input_ids)
        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple

    """
    def __init__(self, config, *inputs, **kwargs):
        super(TFXLNetModel, self).__init__(config, *inputs, **kwargs)
        self.transformer = TFXLNetMainLayer(config, name='transformer')

    def call(self, inputs, **kwargs):
        outputs = self.transformer(inputs, **kwargs)
        return outputs


@add_start_docstrings("""XLNet Model with a language modeling head on top
    (linear layer with weights tied to the input embeddings). """,
    XLNET_START_DOCSTRING, XLNET_INPUTS_DOCSTRING)
class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
    r"""
    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **prediction_scores**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        **mems**:
            list of ``tf.Tensor`` (one for each layer):
            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
            See details in the docstring of the `mems` input above.
        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
            list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
            of shape ``(batch_size, sequence_length, hidden_size)``:
            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

    Examples::

        import tensorflow as tf
        from pytorch_transformers import XLNetTokenizer, TFXLNetLMHeadModel

        tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
        model = TFXLNetLMHeadModel.from_pretrained('xlnet-large-cased')

        # We show how to setup inputs to predict a next token using a bi-directional context.
        input_ids = tf.constant(tokenizer.encode("Hello, my dog is very <mask>"))[None, :]  # We will predict the masked token
        perm_mask = tf.zeros((1, input_ids.shape[1], input_ids.shape[1]))
        perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
        target_mapping = tf.zeros((1, 1, input_ids.shape[1]))  # Shape [1, 1, seq_length] => let's predict one token
        target_mapping[0, 0, -1] = 1.0  # Our first (and only) prediction will be the last token of the sequence (the masked token)
        outputs = model(input_ids, perm_mask=perm_mask, target_mapping=target_mapping)

        next_token_logits = outputs[0]  # Output has shape [target_mapping.size(0), target_mapping.size(1), config.vocab_size]

    """
    def __init__(self, config, *inputs, **kwargs):
        super(TFXLNetLMHeadModel, self).__init__(config, *inputs, **kwargs)
        self.transformer = TFXLNetMainLayer(config, name='transformer')
        self.lm_loss = TFXLNetLMHead(config, self.transformer.word_embedding, name='lm_loss')

    def call(self, inputs, **kwargs):
        transformer_outputs = self.transformer(inputs, **kwargs)
        hidden_state = transformer_outputs[0]
        logits = self.lm_loss(hidden_state)

        outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it

        return outputs  # return logits, mems, (hidden states), (attentions)


@add_start_docstrings("""XLNet Model with a sequence classification/regression head on top (a linear layer on top of
    the pooled output) e.g. for GLUE tasks. """,
    XLNET_START_DOCSTRING, XLNET_INPUTS_DOCSTRING)
class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel):
    r"""
    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **logits**: ``tf.Tensor`` of shape ``(batch_size, config.num_labels)``
            Classification (or regression if config.num_labels==1) scores (before SoftMax).
        **mems**:
            list of ``tf.Tensor`` (one for each layer):
            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
            See details in the docstring of the `mems` input above.
        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
            list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
            of shape ``(batch_size, sequence_length, hidden_size)``:
            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

    Examples::

        import tensorflow as tf
        from pytorch_transformers import XLNetTokenizer, TFXLNetForSequenceClassification

        tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
        model = TFXLNetForSequenceClassification.from_pretrained('xlnet-large-cased')
        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
        outputs = model(input_ids)
        logits = outputs[0]

    """
    def __init__(self, config, *inputs, **kwargs):
        super(TFXLNetForSequenceClassification, self).__init__(config, *inputs, **kwargs)
        self.num_labels = config.num_labels

        self.transformer = TFXLNetMainLayer(config, name='transformer')
        self.sequence_summary = TFSequenceSummary(config, name='sequence_summary')
        self.logits_proj = tf.keras.layers.Dense(config.num_labels, name='logits_proj')

    def call(self, inputs, **kwargs):
        transformer_outputs = self.transformer(inputs, **kwargs)
        output = transformer_outputs[0]

        output = self.sequence_summary(output)
        logits = self.logits_proj(output)

        outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it

        return outputs  # return logits, mems, (hidden states), (attentions)


# @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
#     the hidden-states output to compute `span start logits` and `span end logits`). """,
#     XLNET_START_DOCSTRING, XLNET_INPUTS_DOCSTRING)
# class TFXLNetForQuestionAnswering(TFXLNetPreTrainedModel):
class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
    r"""
    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **start_scores**: ``tf.Tensor`` of shape ``(batch_size, sequence_length,)``
            Span-start scores (before SoftMax).
        **end_scores**: ``tf.Tensor`` of shape ``(batch_size, sequence_length,)``
            Span-end scores (before SoftMax).
        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
            list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
            of shape ``(batch_size, sequence_length, hidden_size)``:
            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

    Examples::

        import tensorflow as tf
        from pytorch_transformers import XLNetTokenizer, TFXLNetForQuestionAnsweringSimple

        tokenizer = XLNetTokenizer.from_pretrained('xlnet-base-cased')
        model = TFXLNetForQuestionAnsweringSimple.from_pretrained('xlnet-base-cased')
        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
        outputs = model(input_ids)
        start_scores, end_scores = outputs[:2]

    """
    def __init__(self, config, *inputs, **kwargs):
        super(TFXLNetForQuestionAnsweringSimple, self).__init__(config, *inputs, **kwargs)
        self.transformer = TFXLNetMainLayer(config, name='transformer')
        self.qa_outputs = tf.keras.layers.Dense(config.num_labels, name='qa_outputs')

    def call(self, inputs, **kwargs):
        transformer_outputs = self.transformer(inputs, **kwargs)

        sequence_output = transformer_outputs[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = tf.split(logits, 2, axis=-1)
        start_logits = tf.squeeze(start_logits, axis=-1)
        end_logits = tf.squeeze(end_logits, axis=-1)

        outputs = (start_logits, end_logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it

        return outputs  # start_logits, end_logits, (hidden_states), (attentions)

# @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
#     the hidden-states output to compute `span start logits` and `span end logits`). """,
#     XLNET_START_DOCSTRING, XLNET_INPUTS_DOCSTRING)
# class TFXLNetForQuestionAnswering(TFXLNetPreTrainedModel):
#     r"""
#     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
#         **start_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
#             ``tf.Tensor`` of shape ``(batch_size, config.start_n_top)``
#             Log probabilities for the top config.start_n_top start token possibilities (beam-search).
#         **start_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
#             ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top)``
#             Indices for the top config.start_n_top start token possibilities (beam-search).
#         **end_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
#             ``tf.Tensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
#             Log probabilities for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
#         **end_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
#             ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
#             Indices for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
#         **cls_logits**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
#             ``tf.Tensor`` of shape ``(batch_size,)``
#             Log probabilities for the ``is_impossible`` label of the answers.
#         **mems**:
#             list of ``tf.Tensor`` (one for each layer):
#             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
#             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
#             See details in the docstring of the `mems` input above.
#         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
#             list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
#             of shape ``(batch_size, sequence_length, hidden_size)``:
#             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
#         **attentions**: (`optional`, returned when ``config.output_attentions=True``)
#             list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
#             Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

#     Examples::

#         tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
#         model = XLMForQuestionAnswering.from_pretrained('xlnet-large-cased')
#         input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
#         start_positions = tf.constant([1])
#         end_positions = tf.constant([3])
#         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
#         loss, start_scores, end_scores = outputs[:2]

#     """
#     def __init__(self, config, *inputs, **kwargs):
#         super(TFXLNetForQuestionAnswering, self).__init__(config, *inputs, **kwargs)
#         self.start_n_top = config.start_n_top
#         self.end_n_top = config.end_n_top

#         self.transformer = TFXLNetMainLayer(config, name='transformer')
#         self.start_logits = TFPoolerStartLogits(config, name='start_logits')
#         self.end_logits = TFPoolerEndLogits(config, name='end_logits')
#         self.answer_class = TFPoolerAnswerClass(config, name='answer_class')

#     def call(self, inputs, training=False):
#         transformer_outputs = self.transformer(inputs, training=training)
#         hidden_states = transformer_outputs[0]
#         start_logits = self.start_logits(hidden_states, p_mask=p_mask)

#         outputs = transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it

#         if start_positions is not None and end_positions is not None:
#             # If we are on multi-GPU, let's remove the dimension added by batch splitting
#             for x in (start_positions, end_positions, cls_index, is_impossible):
#                 if x is not None and x.dim() > 1:
#                     x.squeeze_(-1)

#             # during training, compute the end logits based on the ground truth of the start position
#             end_logits = self.end_logits(hidden_states, start_positions=start_positions, p_mask=p_mask)

#             loss_fct = CrossEntropyLoss()
#             start_loss = loss_fct(start_logits, start_positions)
#             end_loss = loss_fct(end_logits, end_positions)
#             total_loss = (start_loss + end_loss) / 2

#             if cls_index is not None and is_impossible is not None:
#                 # Predict answerability from the representation of CLS and START
#                 cls_logits = self.answer_class(hidden_states, start_positions=start_positions, cls_index=cls_index)
#                 loss_fct_cls = nn.BCEWithLogitsLoss()
#                 cls_loss = loss_fct_cls(cls_logits, is_impossible)

#                 # note(zhiliny): by default multiply the loss by 0.5 so that the scale is comparable to start_loss and end_loss
#                 total_loss += cls_loss * 0.5

#             outputs = (total_loss,) + outputs

#         else:
#             # during inference, compute the end logits based on beam search
#             bsz, slen, hsz = hidden_states.size()
#             start_log_probs = F.softmax(start_logits, dim=-1) # shape (bsz, slen)

#             start_top_log_probs, start_top_index = torch.topk(start_log_probs, self.start_n_top, dim=-1) # shape (bsz, start_n_top)
#             start_top_index_exp = start_top_index.unsqueeze(-1).expand(-1, -1, hsz) # shape (bsz, start_n_top, hsz)
#             start_states = torch.gather(hidden_states, -2, start_top_index_exp) # shape (bsz, start_n_top, hsz)
#             start_states = start_states.unsqueeze(1).expand(-1, slen, -1, -1) # shape (bsz, slen, start_n_top, hsz)

#             hidden_states_expanded = hidden_states.unsqueeze(2).expand_as(start_states) # shape (bsz, slen, start_n_top, hsz)
#             p_mask = p_mask.unsqueeze(-1) if p_mask is not None else None
#             end_logits = self.end_logits(hidden_states_expanded, start_states=start_states, p_mask=p_mask)
#             end_log_probs = F.softmax(end_logits, dim=1) # shape (bsz, slen, start_n_top)

#             end_top_log_probs, end_top_index = torch.topk(end_log_probs, self.end_n_top, dim=1) # shape (bsz, end_n_top, start_n_top)
#             end_top_log_probs = end_top_log_probs.view(-1, self.start_n_top * self.end_n_top)
#             end_top_index = end_top_index.view(-1, self.start_n_top * self.end_n_top)

#             start_states = torch.einsum("blh,bl->bh", hidden_states, start_log_probs)  # get the representation of START as weighted sum of hidden states
#             cls_logits = self.answer_class(hidden_states, start_states=start_states, cls_index=cls_index)  # Shape (batch size,): one single `cls_logits` for each sample

#             outputs = (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits) + outputs

#         # return start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits
#         # or (if labels are provided) (total_loss,)
#         return outputs