fengshen/models/zen2/modeling.py

# coding: utf-8
# Copyright 2019 Sinovation Ventures AI Institute
#
# 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.
#
# This file is partially derived from the code at
# https://github.com/huggingface/transformers/tree/master/transformers
#
# Original copyright notice:
#
# Copyright 2018 The Google AI Language Team 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.
"""PyTorch ZEN2 model classes."""

from __future__ import absolute_import, division, print_function, unicode_literals

import copy
import logging
import math
import os

import torch
from torch import nn
from torch.nn import CrossEntropyLoss
from dataclasses import dataclass
from typing import Optional
from transformers import PreTrainedModel

from fengshen.models.zen2.configuration_zen2 import ZenConfig

logger = logging.getLogger(__name__)

PRETRAINED_MODEL_ARCHIVE_MAP = {
    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-pytorch_model.bin",
    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-pytorch_model.bin",
    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-pytorch_model.bin",
    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-pytorch_model.bin",
    'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-pytorch_model.bin",
    'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-pytorch_model.bin",
    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-pytorch_model.bin",
    'bert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-cased-pytorch_model.bin",
    'bert-large-uncased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-pytorch_model.bin",
    'bert-large-cased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-pytorch_model.bin",
    'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-pytorch_model.bin",
    'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-pytorch_model.bin",
    'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
    'IDEA-CCNL/Erlangshen-ZEN2-345M-Chinese': 'https://huggingface.co/IDEA-CCNL/Erlangshen-ZEN2-345M-Chinese/resolve/main/pytorch_model.bin',
    'IDEA-CCNL/Erlangshen-ZEN2-668M-Chinese': 'https://huggingface.co/IDEA-CCNL/Erlangshen-ZEN2-668M-Chinese/resolve/main/pytorch_model.bin',
}
PRETRAINED_CONFIG_ARCHIVE_MAP = {
    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-config.json",
    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-config.json",
    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-config.json",
    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-config.json",
    'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-config.json",
    'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-config.json",
    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-config.json",
    'bert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-cased-config.json",
    'bert-large-uncased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-config.json",
    'bert-large-cased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-config.json",
    'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-config.json",
    'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-config.json",
    'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
    'IDEA-CCNL/Erlangshen-ZEN2-345M-Chinese': 'https://huggingface.co/IDEA-CCNL/Erlangshen-ZEN2-345M-Chinese/resolve/main/config.json',
    'IDEA-CCNL/Erlangshen-ZEN2-668M-Chinese': 'https://huggingface.co/IDEA-CCNL/Erlangshen-ZEN2-668M-Chinese/resolve/main/config.json',
}
BERT_CONFIG_NAME = 'bert_config.json'
TF_WEIGHTS_NAME = 'model.ckpt'


def prune_linear_layer(layer, index, dim=0):
    """ Prune a linear layer (a model parameters) to keep only entries in index.
        Return the pruned layer as a new layer with requires_grad=True.
        Used to remove heads.
    """
    index = index.to(layer.weight.device)
    W = layer.weight.index_select(dim, index).clone().detach()
    if layer.bias is not None:
        if dim == 1:
            b = layer.bias.clone().detach()
        else:
            b = layer.bias[index].clone().detach()
    new_size = list(layer.weight.size())
    new_size[dim] = len(index)
    new_layer = nn.Linear(new_size[1], new_size[0], bias=layer.bias is not None).to(layer.weight.device)
    new_layer.weight.requires_grad = False
    new_layer.weight.copy_(W.contiguous())
    new_layer.weight.requires_grad = True
    if layer.bias is not None:
        new_layer.bias.requires_grad = False
        new_layer.bias.copy_(b.contiguous())
        new_layer.bias.requires_grad = True
    return new_layer


def load_tf_weights_in_bert(model, tf_checkpoint_path):
    """ Load tf checkpoints in a pytorch model
    """
    try:
        import re
        import numpy as np
        import tensorflow as tf
    except ImportError:
        print("Loading a TensorFlow models in PyTorch, requires TensorFlow to be installed. Please see "
              "https://www.tensorflow.org/install/ for installation instructions.")
        raise
    tf_path = os.path.abspath(tf_checkpoint_path)
    print("Converting TensorFlow checkpoint from {}".format(tf_path))
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path)
    names = []
    arrays = []
    for name, shape in init_vars:
        print("Loading TF weight {} with shape {}".format(name, shape))
        array = tf.train.load_variable(tf_path, name)
        names.append(name)
        arrays.append(array)

    for name, array in zip(names, arrays):
        name = name.split('/')
        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
        # which are not required for using pretrained model
        if any(n in ["adam_v", "adam_m", "global_step"] for n in name):
            print("Skipping {}".format("/".join(name)))
            continue
        pointer = model
        for m_name in name:
            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
                name_lists = re.split(r'_(\d+)', m_name)
            else:
                name_lists = [m_name]
            if name_lists[0] == 'kernel' or name_lists[0] == 'gamma':
                pointer = getattr(pointer, 'weight')
            elif name_lists[0] == 'output_bias' or name_lists[0] == 'beta':
                pointer = getattr(pointer, 'bias')
            elif name_lists[0] == 'output_weights':
                pointer = getattr(pointer, 'weight')
            elif name_lists[0] == 'squad':
                pointer = getattr(pointer, 'classifier')
            else:
                try:
                    pointer = getattr(pointer, name_lists[0])
                except AttributeError:
                    print("Skipping {}".format("/".join(name)))
                    continue
            if len(name_lists) >= 2:
                num = int(name_lists[1])
                pointer = pointer[num]
        if m_name[-11:] == '_embeddings':
            pointer = getattr(pointer, 'weight')
        elif m_name == 'kernel':
            array = np.transpose(array)
        try:
            assert pointer.shape == array.shape
        except AssertionError as e:
            e.args += (pointer.shape, array.shape)
            raise
        print("Initialize PyTorch weight {}".format(name))
        pointer.data = torch.from_numpy(array)
    return model


def gelu(x):
    """Implementation of the gelu activation function.
        For information: OpenAI GPT's gelu is slightly different (and gives slightly different results):
        0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
        Also see https://arxiv.org/abs/1606.08415
    """
    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))


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


ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish}


try:
    # from apex.normalization.fused_layer_norm import FusedLayerNorm as BertLayerNorm
    from torch.nn import LayerNorm as BertLayerNorm
except ImportError:
    logger.info("Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex .")

    class BertLayerNorm(nn.Module):
        def __init__(self, hidden_size, eps=1e-12):
            """Construct a layernorm module in the TF style (epsilon inside the square root).
            """
            super(BertLayerNorm, self).__init__()
            self.weight = nn.Parameter(torch.ones(hidden_size))
            self.bias = nn.Parameter(torch.zeros(hidden_size))
            self.variance_epsilon = eps

        def forward(self, x):
            u = x.mean(-1, keepdim=True)
            s = (x - u).pow(2).mean(-1, keepdim=True)
            x = (x - u) / torch.sqrt(s + self.variance_epsilon)
            return self.weight * x + self.bias


class BertEmbeddings(nn.Module):
    """Construct the embeddings from word, position and token_type embeddings.
    """

    def __init__(self, config):
        super(BertEmbeddings, self).__init__()
        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=0)
        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)

        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
        # any TensorFlow checkpoint file
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, input_ids, token_type_ids=None):
        if token_type_ids is None:
            token_type_ids = torch.zeros_like(input_ids)

        words_embeddings = self.word_embeddings(input_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)

        embeddings = words_embeddings + token_type_embeddings
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


class BertWordEmbeddings(nn.Module):
    """Construct the embeddings from ngram, position and token_type embeddings.
    """

    def __init__(self, config):
        super(BertWordEmbeddings, self).__init__()
        self.word_embeddings = nn.Embedding(config.word_size, config.hidden_size, padding_idx=0)
        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)

        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
        # any TensorFlow checkpoint file
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, input_ids, token_type_ids=None):
        if token_type_ids is None:
            token_type_ids = torch.zeros_like(input_ids)

        words_embeddings = self.word_embeddings(input_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)

        embeddings = words_embeddings + token_type_embeddings
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


class RelativeSinusoidalPositionalEmbedding(nn.Module):
    """This module produces sinusoidal positional embeddings of any length.
    Padding symbols are ignored.
    """

    def __init__(self, embedding_dim, padding_idx, init_size=1568):
        """

        :param embedding_dim: 每个位置的dimension
        :param padding_idx:
        :param init_size:
        """
        super().__init__()
        self.embedding_dim = embedding_dim
        self.padding_idx = padding_idx
        assert init_size % 2 == 0
        weights = self.get_embedding(
            init_size+1,
            embedding_dim,
            padding_idx,
        )
        self.register_buffer('weights', weights)
        self.register_buffer('_float_tensor', torch.FloatTensor(1))

    def get_embedding(self, num_embeddings, embedding_dim, padding_idx=None):
        """Build sinusoidal embeddings.
        This matches the implementation in tensor2tensor, but differs slightly
        from the description in Section 3.5 of "Attention Is All You Need".
        """
        half_dim = embedding_dim // 2
        emb = math.log(10000) / (half_dim - 1)
        emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb)
        emb = torch.arange(-num_embeddings//2, num_embeddings//2, dtype=torch.float).unsqueeze(1) * emb.unsqueeze(0)
        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1)
        if embedding_dim % 2 == 1:
            # zero pad
            emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1)
        if padding_idx is not None:
            emb[padding_idx, :] = 0
        self.origin_shift = num_embeddings//2 + 1
        return emb

    def forward(self, input):
        """Input is expected to be of size [bsz x seqlen].
        """
        bsz, _, _, seq_len = input.size()
        max_pos = self.padding_idx + seq_len
        if max_pos > self.origin_shift:
            # recompute/expand embeddings if needed
            weights = self.get_embedding(
                max_pos*2,
                self.embedding_dim,
                self.padding_idx,
            )
            weights = weights.to(self._float_tensor)
            del self.weights
            self.origin_shift = weights.size(0)//2
            self.register_buffer('weights', weights)

        positions = torch.arange(-seq_len, seq_len).to(input.device).long() + self.origin_shift  # 2*seq_len
        embed = self.weights.index_select(0, positions.long()).detach()
        return embed


class BertSelfAttention(nn.Module):
    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(BertSelfAttention, self).__init__()
        if config.hidden_size % config.num_attention_heads != 0:
            raise ValueError(
                "The hidden size (%d) is not a multiple of the number of attention "
                "heads (%d)" % (config.hidden_size, config.num_attention_heads))
        self.output_attentions = output_attentions
        self.keep_multihead_output = keep_multihead_output
        self.multihead_output = None

        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size)
        self.key = nn.Linear(config.hidden_size, self.all_head_size)
        self.value = nn.Linear(config.hidden_size, self.all_head_size)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
        self.softmax = nn.Softmax(dim=-1)

        self.position_embedding = RelativeSinusoidalPositionalEmbedding(self.attention_head_size, 0, 1200)
        self.r_r_bias = nn.Parameter(
            nn.init.xavier_normal_(torch.zeros(self.num_attention_heads, self.attention_head_size)))
        self.r_w_bias = nn.Parameter(
            nn.init.xavier_normal_(torch.zeros(self.num_attention_heads, self.attention_head_size)))

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward(self, hidden_states, attention_mask, head_mask=None):
        position_embedding = self.position_embedding(attention_mask)

        mixed_query_layer = self.query(hidden_states)
        mixed_key_layer = self.key(hidden_states)
        mixed_value_layer = self.value(hidden_states)

        query_layer = self.transpose_for_scores(mixed_query_layer)
        key_layer = self.transpose_for_scores(mixed_key_layer)
        value_layer = self.transpose_for_scores(mixed_value_layer)

        rw_head_q = query_layer + self.r_r_bias[:, None]
        AC = torch.einsum('bnqd,bnkd->bnqk', [rw_head_q.float(), key_layer.float()])  # b x n x l x d, n是head

        D_ = torch.einsum('nd,ld->nl', self.r_w_bias.float(), position_embedding.float())[None, :,
                                                                                          None]  # head x 2max_len, 每个head对位置的bias
        B_ = torch.einsum('bnqd,ld->bnql', query_layer.float(),
                          position_embedding.float())  # bsz x head  x max_len x 2max_len，每个query对每个shift的偏移
        BD = B_ + D_  # bsz x head x max_len x 2max_len, 要转换为bsz x head x max_len x max_len
        BD = self._shift(BD)
        attention_scores = AC + BD
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
        attention_scores = attention_scores + attention_mask

        # Normalize the attention scores to probabilities.
        attention_probs = self.softmax(attention_scores)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs = self.dropout(attention_probs)

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

        context_layer = torch.matmul(attention_probs.type_as(value_layer), value_layer)
        if self.keep_multihead_output:
            self.multihead_output = context_layer
            self.multihead_output.retain_grad()
        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)
        if self.output_attentions:
            return attention_probs, context_layer
        return context_layer

    def _shift(self, BD):
        """
        类似
        -3 -2 -1 0 1 2
        -3 -2 -1 0 1 2
        -3 -2 -1 0 1 2

        转换为
        0   1  2
        -1  0  1
        -2 -1  0

        :param BD: batch_size x n_head x max_len x 2max_len
        :return: batch_size x n_head x max_len x max_len
        """
        bsz, n_head, max_len, _ = BD.size()
        zero_pad = BD.new_zeros(bsz, n_head, max_len, 1)
        BD = torch.cat([BD, zero_pad], dim=-1).view(bsz, n_head, -1, max_len)  # bsz x n_head x (2max_len+1) x max_len
        BD = BD[:, :, :-1].view(bsz, n_head, max_len, -1)  # bsz x n_head x 2max_len x max_len
        BD = BD[:, :, :, max_len:]
        return BD


class BertSelfOutput(nn.Module):
    def __init__(self, config):
        super(BertSelfOutput, self).__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states


class BertAttention(nn.Module):
    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(BertAttention, self).__init__()
        self.output_attentions = output_attentions
        self.self = BertSelfAttention(config, output_attentions=output_attentions,
                                      keep_multihead_output=keep_multihead_output)
        self.output = BertSelfOutput(config)

    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        mask = torch.ones(self.self.num_attention_heads, self.self.attention_head_size)
        for head in heads:
            mask[head] = 0
        mask = mask.view(-1).contiguous().eq(1)
        index = torch.arange(len(mask))[mask].long()
        # Prune linear layers
        self.self.query = prune_linear_layer(self.self.query, index)
        self.self.key = prune_linear_layer(self.self.key, index)
        self.self.value = prune_linear_layer(self.self.value, index)
        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
        # Update hyper params
        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads

    def forward(self, input_tensor, attention_mask, head_mask=None):
        self_output = self.self(input_tensor, attention_mask, head_mask)
        if self.output_attentions:
            attentions, self_output = self_output
        attention_output = self.output(self_output, input_tensor)
        if self.output_attentions:
            return attentions, attention_output
        return attention_output


class BertIntermediate(nn.Module):
    def __init__(self, config):
        super(BertIntermediate, self).__init__()
        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
        # if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
        if isinstance(config.hidden_act, str):
            self.intermediate_act_fn = ACT2FN[config.hidden_act]
        else:
            self.intermediate_act_fn = config.hidden_act

    def forward(self, hidden_states):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)
        return hidden_states


class BertOutput(nn.Module):
    def __init__(self, config):
        super(BertOutput, self).__init__()
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states


class BertLayer(nn.Module):
    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(BertLayer, self).__init__()
        self.output_attentions = output_attentions
        self.attention = BertAttention(config, output_attentions=output_attentions,
                                       keep_multihead_output=keep_multihead_output)
        self.intermediate = BertIntermediate(config)
        self.output = BertOutput(config)

    def forward(self, hidden_states, attention_mask, head_mask=None):
        attention_output = self.attention(hidden_states, attention_mask, head_mask)
        if self.output_attentions:
            attentions, attention_output = attention_output
        intermediate_output = self.intermediate(attention_output)
        layer_output = self.output(intermediate_output, attention_output)
        if self.output_attentions:
            return attentions, layer_output
        return layer_output


class ZenEncoder(nn.Module):
    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(ZenEncoder, self).__init__()
        self.output_attentions = output_attentions
        layer = BertLayer(config, output_attentions=output_attentions,
                          keep_multihead_output=keep_multihead_output)
        self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.num_hidden_layers)])
        self.word_layers = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.num_hidden_word_layers)])
        self.num_hidden_word_layers = config.num_hidden_word_layers

    def forward(self, hidden_states, ngram_hidden_states, ngram_position_matrix, attention_mask,
                ngram_attention_mask,
                output_all_encoded_layers=True, head_mask=None):
        # Need to check what is the attention masking doing here
        all_encoder_layers = []
        all_attentions = []
        num_hidden_ngram_layers = self.num_hidden_word_layers
        for i, layer_module in enumerate(self.layer):
            hidden_states = layer_module(hidden_states, attention_mask, head_mask[i])
            if i < num_hidden_ngram_layers:
                ngram_hidden_states = self.word_layers[i](ngram_hidden_states, ngram_attention_mask, head_mask[i])
                if self.output_attentions:
                    ngram_attentions, ngram_hidden_states = ngram_hidden_states
                    all_attentions.append(ngram_attentions)
            if self.output_attentions:
                attentions, hidden_states = hidden_states
                all_attentions.append(attentions)
            hidden_states += torch.bmm(ngram_position_matrix.float(), ngram_hidden_states.float())
            if output_all_encoded_layers:
                all_encoder_layers.append(hidden_states)
        if not output_all_encoded_layers:
            all_encoder_layers.append(hidden_states)
        if self.output_attentions:
            return all_attentions, all_encoder_layers
        return all_encoder_layers


class BertPooler(nn.Module):
    def __init__(self, config):
        super(BertPooler, self).__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = nn.Tanh()

    def forward(self, hidden_states):
        # We "pool" the model by simply taking the hidden state corresponding
        # to the first token.
        first_token_tensor = hidden_states[:, 0]
        pooled_output = self.dense(first_token_tensor)
        pooled_output = self.activation(pooled_output)
        return pooled_output


class BertPredictionHeadTransform(nn.Module):
    def __init__(self, config):
        super(BertPredictionHeadTransform, self).__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        # if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
        if isinstance(config.hidden_act, str):
            self.transform_act_fn = ACT2FN[config.hidden_act]
        else:
            self.transform_act_fn = config.hidden_act
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(self, hidden_states):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.transform_act_fn(hidden_states)
        hidden_states = self.LayerNorm(hidden_states)
        return hidden_states


class BertLMPredictionHead(nn.Module):
    def __init__(self, config, bert_model_embedding_weights):
        super(BertLMPredictionHead, self).__init__()
        self.transform = BertPredictionHeadTransform(config)

        # The output weights are the same as the input embeddings, but there is
        # an output-only bias for each token.
        self.decoder = nn.Linear(bert_model_embedding_weights.size(1),
                                 bert_model_embedding_weights.size(0),
                                 bias=False)
        self.decoder.weight = bert_model_embedding_weights
        self.bias = nn.Parameter(torch.zeros(bert_model_embedding_weights.size(0)))

    def forward(self, hidden_states):
        hidden_states = self.transform(hidden_states)
        hidden_states = self.decoder(hidden_states) + self.bias
        return hidden_states


class ZenOnlyMLMHead(nn.Module):
    def __init__(self, config, bert_model_embedding_weights):
        super(ZenOnlyMLMHead, self).__init__()
        self.predictions = BertLMPredictionHead(config, bert_model_embedding_weights)

    def forward(self, sequence_output):
        prediction_scores = self.predictions(sequence_output)
        return prediction_scores


class ZenOnlyNSPHead(nn.Module):
    def __init__(self, config):
        super(ZenOnlyNSPHead, self).__init__()
        self.seq_relationship = nn.Linear(config.hidden_size, 2)

    def forward(self, pooled_output):
        seq_relationship_score = self.seq_relationship(pooled_output)
        return seq_relationship_score


class ZenPreTrainingHeads(nn.Module):
    def __init__(self, config, bert_model_embedding_weights):
        super(ZenPreTrainingHeads, self).__init__()
        self.predictions = BertLMPredictionHead(config, bert_model_embedding_weights)
        self.seq_relationship = nn.Linear(config.hidden_size, 2)

    def forward(self, sequence_output, pooled_output):
        prediction_scores = self.predictions(sequence_output)
        seq_relationship_score = self.seq_relationship(pooled_output)
        return prediction_scores, seq_relationship_score


class ZenPreTrainedModel(PreTrainedModel):
    """ An abstract class to handle weights initialization and
        a simple interface for dowloading and loading pretrained models.
    """
    config_class = ZenConfig
    supports_gradient_checkpointing = True
    _keys_to_ignore_on_load_missing = [r"position_ids"]

    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, nn.Linear):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(
                mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(
                mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


class ZenModel(ZenPreTrainedModel):
    """ZEN model ("BERT-based Chinese (Z) text encoder Enhanced by N-gram representations").

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `output_all_encoded_layers`: boolean which controls the content of the `encoded_layers` output as described below. Default: `True`.
        `head_mask`: an optional torch.Tensor of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
            It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.
        `input_ngram_ids`: input_ids of ngrams.
        `ngram_token_type_ids`: token_type_ids of ngrams.
        `ngram_attention_mask`: attention_mask of ngrams.
        `ngram_position_matrix`: position matrix of ngrams.


    Outputs: Tuple of (encoded_layers, pooled_output)
        `encoded_layers`: controled by `output_all_encoded_layers` argument:
            - `output_all_encoded_layers=True`: outputs a list of the full sequences of encoded-hidden-states at the end
                of each attention block (i.e. 12 full sequences for BERT-base, 24 for BERT-large), each
                encoded-hidden-state is a torch.FloatTensor of size [batch_size, sequence_length, hidden_size],
            - `output_all_encoded_layers=False`: outputs only the full sequence of hidden-states corresponding
                to the last attention block of shape [batch_size, sequence_length, hidden_size],
        `pooled_output`: a torch.FloatTensor of size [batch_size, hidden_size] which is the output of a
            classifier pretrained on top of the hidden state associated to the first character of the
            input (`CLS`) to train on the Next-Sentence task (see BERT's paper).

    """

    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(ZenModel, self).__init__(config)
        self.output_attentions = output_attentions
        self.embeddings = BertEmbeddings(config)
        self.word_embeddings = BertWordEmbeddings(config)
        self.encoder = ZenEncoder(config, output_attentions=output_attentions,
                                  keep_multihead_output=keep_multihead_output)
        self.pooler = BertPooler(config)
        self.init_weights()

    def prune_heads(self, heads_to_prune):
        """ Prunes heads of the model.
            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

    def get_multihead_outputs(self):
        """ Gather all multi-head outputs.
            Return: list (layers) of multihead module outputs with gradients
        """
        return [layer.attention.self.multihead_output for layer in self.encoder.layer]

    def forward(self, input_ids,
                input_ngram_ids,
                ngram_position_matrix,
                token_type_ids=None,
                ngram_token_type_ids=None,
                attention_mask=None,
                ngram_attention_mask=None,
                output_all_encoded_layers=True,
                head_mask=None):
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids)
        if token_type_ids is None:
            token_type_ids = torch.zeros_like(input_ids)

        if ngram_attention_mask is None:
            ngram_attention_mask = torch.ones_like(input_ngram_ids)
        if ngram_token_type_ids is None:
            ngram_token_type_ids = torch.zeros_like(input_ngram_ids)

        # We create a 3D attention mask from a 2D tensor mask.
        # Sizes are [batch_size, 1, 1, to_seq_length]
        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
        # this attention mask is more simple than the triangular masking of causal attention
        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
        extended_ngram_attention_mask = ngram_attention_mask.unsqueeze(1).unsqueeze(2)

        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
        # masked positions, this operation will create a tensor which is 0.0 for
        # positions we want to attend and -10000.0 for masked positions.
        # Since we are adding it to the raw scores before the softmax, this is
        # effectively the same as removing these entirely.
        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0

        extended_ngram_attention_mask = extended_ngram_attention_mask.to(dtype=next(self.parameters()).dtype)
        extended_ngram_attention_mask = (1.0 - extended_ngram_attention_mask) * -10000.0

        # 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]
        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
        if head_mask is not None:
            if head_mask.dim() == 1:
                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
                head_mask = head_mask.expand_as(self.config.num_hidden_layers, -1, -1, -1, -1)
            elif head_mask.dim() == 2:
                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(
                    -1)  # We can specify head_mask for each layer
            head_mask = head_mask.to(
                dtype=next(self.parameters()).dtype)  # switch to fload if need + fp16 compatibility
        else:
            head_mask = [None] * self.config.num_hidden_layers

        embedding_output = self.embeddings(input_ids, token_type_ids)
        ngram_embedding_output = self.word_embeddings(input_ngram_ids, ngram_token_type_ids)

        encoded_layers = self.encoder(embedding_output,
                                      ngram_embedding_output,
                                      ngram_position_matrix,
                                      extended_attention_mask,
                                      extended_ngram_attention_mask,
                                      output_all_encoded_layers=output_all_encoded_layers,
                                      head_mask=head_mask)
        if self.output_attentions:
            all_attentions, encoded_layers = encoded_layers
        sequence_output = encoded_layers[-1]
        pooled_output = self.pooler(sequence_output)
        if not output_all_encoded_layers:
            encoded_layers = encoded_layers[-1]
        if self.output_attentions:
            return all_attentions, encoded_layers, pooled_output
        return encoded_layers, pooled_output


class ZenForPreTraining(ZenPreTrainedModel):
    """ZEN model with pre-training heads.
    This module comprises the ZEN model followed by the two pre-training heads:
        - the masked language modeling head, and
        - the next sentence classification head.

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `masked_lm_labels`: optional masked language modeling labels: torch.LongTensor of shape [batch_size, sequence_length]
            with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss
            is only computed for the labels set in [0, ..., vocab_size]
        `next_sentence_label`: optional next sentence classification loss: torch.LongTensor of shape [batch_size]
            with indices selected in [0, 1].
            0 => next sentence is the continuation, 1 => next sentence is a random sentence.
        `head_mask`: an optional torch.Tensor of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
            It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.
        `input_ngram_ids`: input_ids of ngrams.
        `ngram_token_type_ids`: token_type_ids of ngrams.
        `ngram_attention_mask`: attention_mask of ngrams.
        `ngram_position_matrix`: position matrix of ngrams.

    Outputs:
        if `masked_lm_labels` and `next_sentence_label` are not `None`:
            Outputs the total_loss which is the sum of the masked language modeling loss and the next
            sentence classification loss.
        if `masked_lm_labels` or `next_sentence_label` is `None`:
            Outputs a tuple comprising
            - the masked language modeling logits of shape [batch_size, sequence_length, vocab_size], and
            - the next sentence classification logits of shape [batch_size, 2].

    """

    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(ZenForPreTraining, self).__init__(config)
        self.output_attentions = output_attentions
        self.bert = ZenModel(config, output_attentions=output_attentions,
                             keep_multihead_output=keep_multihead_output)
        self.cls = ZenPreTrainingHeads(config, self.bert.embeddings.word_embeddings.weight)
        self.init_weights()

    def forward(self, input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids=None,
                ngram_token_type_ids=None,
                attention_mask=None,
                ngram_attention_mask=None,
                masked_lm_labels=None,
                next_sentence_label=None, head_mask=None):
        outputs = self.bert(input_ids,
                            input_ngram_ids,
                            ngram_position_matrix,
                            token_type_ids,
                            ngram_token_type_ids,
                            attention_mask,
                            ngram_attention_mask,
                            output_all_encoded_layers=False, head_mask=head_mask)
        if self.output_attentions:
            all_attentions, sequence_output, pooled_output = outputs
        else:
            sequence_output, pooled_output = outputs
        prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)

        if masked_lm_labels is not None and next_sentence_label is not None:
            loss_fct = CrossEntropyLoss(ignore_index=-1)
            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
            total_loss = masked_lm_loss + next_sentence_loss
            return total_loss
        elif self.output_attentions:
            return all_attentions, prediction_scores, seq_relationship_score
        return prediction_scores, seq_relationship_score


class ZenForMaskedLM(ZenPreTrainedModel):
    """ZEN model with the masked language modeling head.
    This module comprises the ZEN model followed by the masked language modeling head.

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `masked_lm_labels`: masked language modeling labels: torch.LongTensor of shape [batch_size, sequence_length]
            with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss
            is only computed for the labels set in [0, ..., vocab_size]
        `head_mask`: an optional torch.LongTensor of shape [num_heads] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `head_mask`: an optional torch.Tensor of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
            It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.
        `input_ngram_ids`: input_ids of ngrams.
        `ngram_token_type_ids`: token_type_ids of ngrams.
        `ngram_attention_mask`: attention_mask of ngrams.
        `ngram_position_matrix`: position matrix of ngrams.

    Outputs:
        if `masked_lm_labels` is  not `None`:
            Outputs the masked language modeling loss.
        if `masked_lm_labels` is `None`:
            Outputs the masked language modeling logits of shape [batch_size, sequence_length, vocab_size].

    """

    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(ZenForMaskedLM, self).__init__(config)
        self.output_attentions = output_attentions
        self.bert = ZenModel(config, output_attentions=output_attentions,
                             keep_multihead_output=keep_multihead_output)
        self.cls = ZenOnlyMLMHead(config, self.bert.embeddings.word_embeddings.weight)
        self.init_weights()

    def forward(self, input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids=None, attention_mask=None, masked_lm_labels=None, head_mask=None):
        outputs = self.bert(input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids, attention_mask,
                            output_all_encoded_layers=False,
                            head_mask=head_mask)
        if self.output_attentions:
            all_attentions, sequence_output, _ = outputs
        else:
            sequence_output, _ = outputs
        prediction_scores = self.cls(sequence_output)

        if masked_lm_labels is not None:
            loss_fct = CrossEntropyLoss(ignore_index=-1)
            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
            return masked_lm_loss
        elif self.output_attentions:
            return all_attentions, prediction_scores
        return prediction_scores


class ZenForNextSentencePrediction(ZenPreTrainedModel):
    """ZEN model with next sentence prediction head.
    This module comprises the ZEN model followed by the next sentence classification head.

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `next_sentence_label`: next sentence classification loss: torch.LongTensor of shape [batch_size]
            with indices selected in [0, 1].
            0 => next sentence is the continuation, 1 => next sentence is a random sentence.
        `head_mask`: an optional torch.Tensor of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
            It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.
        `input_ngram_ids`: input_ids of ngrams.
        `ngram_token_type_ids`: token_type_ids of ngrams.
        `ngram_attention_mask`: attention_mask of ngrams.
        `ngram_position_matrix`: position matrix of ngrams.

    Outputs:
        if `next_sentence_label` is not `None`:
            Outputs the total_loss which is the sum of the masked language modeling loss and the next
            sentence classification loss.
        if `next_sentence_label` is `None`:
            Outputs the next sentence classification logits of shape [batch_size, 2].

    """

    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(ZenForNextSentencePrediction, self).__init__(config)
        self.output_attentions = output_attentions
        self.bert = ZenModel(config, output_attentions=output_attentions,
                             keep_multihead_output=keep_multihead_output)
        self.cls = ZenOnlyNSPHead(config)
        self.init_weights()

    def forward(self, input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids=None, attention_mask=None, next_sentence_label=None, head_mask=None):
        outputs = self.bert(input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids, attention_mask,
                            output_all_encoded_layers=False,
                            head_mask=head_mask)
        if self.output_attentions:
            all_attentions, _, pooled_output = outputs
        else:
            _, pooled_output = outputs
        seq_relationship_score = self.cls(pooled_output)

        if next_sentence_label is not None:
            loss_fct = CrossEntropyLoss(ignore_index=-1)
            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
            return next_sentence_loss
        elif self.output_attentions:
            return all_attentions, seq_relationship_score
        return seq_relationship_score


class ZenForSequenceClassification(ZenPreTrainedModel):
    """ZEN model for classification.
    This module is composed of the ZEN model with a linear layer on top of
    the pooled output.

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False
        `num_labels`: the number of classes for the classifier. Default = 2.

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary. Items in the batch should begin with the special "CLS" token. (see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `labels`: labels for the classification output: torch.LongTensor of shape [batch_size]
            with indices selected in [0, ..., num_labels].
        `head_mask`: an optional torch.Tensor of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
            It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.
        `input_ngram_ids`: input_ids of ngrams.
        `ngram_token_type_ids`: token_type_ids of ngrams.
        `ngram_attention_mask`: attention_mask of ngrams.
        `ngram_position_matrix`: position matrix of ngrams.

    Outputs:
        if `labels` is not `None`:
            Outputs the CrossEntropy classification loss of the output with the labels.
        if `labels` is `None`:
            Outputs the classification logits of shape [batch_size, num_labels].

    """

    def __init__(self, config, num_labels=2, output_attentions=False, keep_multihead_output=False):
        super(ZenForSequenceClassification, self).__init__(config)
        self.output_attentions = output_attentions
        self.num_labels = config.num_labels
        self.bert = ZenModel(config, output_attentions=output_attentions,
                             keep_multihead_output=keep_multihead_output)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, self.num_labels)
        self.init_weights()

    def forward(self, input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids=None, attention_mask=None, labels=None, head_mask=None):
        outputs = self.bert(input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids,
                            attention_mask=attention_mask,
                            output_all_encoded_layers=False,
                            head_mask=head_mask)
        if self.output_attentions:
            all_attentions, _, pooled_output = outputs
        else:
            _, pooled_output = outputs
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            return loss, logits
        elif self.output_attentions:
            return all_attentions, logits
        return loss, logits


@dataclass
class TokenClassifierOutput:
    """
    Base class for outputs of token classification models.
    """

    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None


class ZenForTokenClassification(ZenPreTrainedModel):
    """ZEN model for token-level classification.
    This module is composed of the ZEN model with a linear layer on top of
    the full hidden state of the last layer.

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False
        `num_labels`: the number of classes for the classifier. Default = 2.

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `labels`: labels for the classification output: torch.LongTensor of shape [batch_size, sequence_length]
            with indices selected in [0, ..., num_labels].
        `head_mask`: an optional torch.Tensor of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
            It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.
        `input_ngram_ids`: input_ids of ngrams.
        `ngram_token_type_ids`: token_type_ids of ngrams.
        `ngram_attention_mask`: attention_mask of ngrams.
        `ngram_position_matrix`: position matrix of ngrams.

    Outputs:
        if `labels` is not `None`:
            Outputs the CrossEntropy classification loss of the output with the labels.
        if `labels` is `None`:
            Outputs the classification logits of shape [batch_size, sequence_length, num_labels].

    """

    def __init__(self, config, num_labels=2, output_attentions=False, keep_multihead_output=False):
        super(ZenForTokenClassification, self).__init__(config)
        self.output_attentions = output_attentions
        self.num_labels = config.num_labels
        self.bert = ZenModel(config, output_attentions=output_attentions,
                             keep_multihead_output=keep_multihead_output)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, self.num_labels)
        self.init_weights()

    def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None, valid_ids=None,
                input_ngram_ids=None, ngram_position_matrix=None, head_mask=None, b_use_valid_filter=False):
        outputs = self.bert(input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids,
                            attention_mask=attention_mask, output_all_encoded_layers=False, head_mask=head_mask)
        if self.output_attentions:
            all_attentions, sequence_output, _ = outputs
        else:
            sequence_output, _ = outputs

        # if b_use_valid_filter:
        #     batch_size, max_len, feat_dim = sequence_output.shape
        #     valid_output = torch.zeros(batch_size, max_len, feat_dim, dtype=sequence_output.dtype,
        #                                device=input_ids.device)
        #     for i in range(batch_size):
        #         temp = sequence_output[i][valid_ids[i] == 1]
        #         valid_output[i][:temp.size(0)] = temp
        # else:
        #     valid_output = sequence_output
        valid_output = sequence_output

        sequence_output = self.dropout(valid_output)
        logits = self.classifier(sequence_output)
        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss(ignore_index=0)
            # Only keep active parts of the loss
            # attention_mask_label = None
            # if attention_mask_label is not None:
            if attention_mask is not None:
                # active_loss = attention_mask_label.view(-1) == 1
                active_loss = attention_mask.view(-1) == 1
                active_logits = logits.view(-1, self.num_labels)[active_loss]
                active_labels = labels.view(-1)[active_loss]
                loss = loss_fct(active_logits, active_labels)
            else:
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            return TokenClassifierOutput(loss, logits)
        else:
            return TokenClassifierOutput(loss, logits)


class ZenForQuestionAnswering(ZenPreTrainedModel):
    """BERT model for Question Answering (span extraction).
    This module is composed of the BERT model with a linear layer on top of
    the sequence output that computes start_logits and end_logits

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `start_positions`: position of the first token for the labeled span: torch.LongTensor of shape [batch_size].
            Positions are clamped to the length of the sequence and position outside of the sequence are not taken
            into account for computing the loss.
        `end_positions`: position of the last token for the labeled span: torch.LongTensor of shape [batch_size].
            Positions are clamped to the length of the sequence and position outside of the sequence are not taken
            into account for computing the loss.
        `head_mask`: an optional torch.Tensor of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
            It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.

    Outputs:
        if `start_positions` and `end_positions` are not `None`:
            Outputs the total_loss which is the sum of the CrossEntropy loss for the start and end token positions.
        if `start_positions` or `end_positions` is `None`:
            Outputs a tuple of start_logits, end_logits which are the logits respectively for the start and end
            position tokens of shape [batch_size, sequence_length].

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    model = BertForQuestionAnswering(config)
    start_logits, end_logits = model(input_ids, token_type_ids, input_mask)
    ```
    """

    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(ZenForQuestionAnswering, self).__init__(config)
        self.output_attentions = output_attentions
        self.bert = ZenModel(config, output_attentions=output_attentions,
                             keep_multihead_output=keep_multihead_output)
        self.qa_outputs = nn.Linear(config.hidden_size, 2)
        self.init_weights()

    def forward(self, input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids=None, attention_mask=None, start_positions=None,
                end_positions=None, head_mask=None):
        outputs = self.bert(input_ids, input_ngram_ids, ngram_position_matrix, token_type_ids,
                            attention_mask=attention_mask,
                            output_all_encoded_layers=False,
                            head_mask=head_mask)
        if self.output_attentions:
            all_attentions, sequence_output, _ = outputs
        else:
            sequence_output, _ = outputs
        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1)
        end_logits = end_logits.squeeze(-1)

        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions.clamp_(0, ignored_index)
            end_positions.clamp_(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2
            return total_loss
        elif self.output_attentions:
            return all_attentions, start_logits, end_logits
        return start_logits, end_logits