modeling_albert.py

# coding=utf-8
# Copyright 2018 Google AI, Google Brain and the HuggingFace Inc. team.
#
# 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 ALBERT model. """

import math
import os
from dataclasses import dataclass
from typing import Optional, Tuple

import torch
import torch.nn as nn
from torch.nn import CrossEntropyLoss, MSELoss

from ...activations import ACT2FN
from ...file_utils import (
    ModelOutput,
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    replace_return_docstrings,
)
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPooling,
    MaskedLMOutput,
    MultipleChoiceModelOutput,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutput,
    TokenClassifierOutput,
)
from ...modeling_utils import (
    PreTrainedModel,
    apply_chunking_to_forward,
    find_pruneable_heads_and_indices,
    prune_linear_layer,
)
from ...utils import logging
from .configuration_albert import AlbertConfig


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "AlbertConfig"
_TOKENIZER_FOR_DOC = "AlbertTokenizer"


ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "albert-base-v1",
    "albert-large-v1",
    "albert-xlarge-v1",
    "albert-xxlarge-v1",
    "albert-base-v2",
    "albert-large-v2",
    "albert-xlarge-v2",
    "albert-xxlarge-v2",
    # See all ALBERT models at https://huggingface.co/models?filter=albert
]


def load_tf_weights_in_albert(model, config, tf_checkpoint_path):
    """ Load tf checkpoints in a pytorch model."""
    try:
        import re

        import numpy as np
        import tensorflow as tf
    except ImportError:
        logger.error(
            "Loading a TensorFlow model 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)
    logger.info("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:
        logger.info("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):
        print(name)

    for name, array in zip(names, arrays):
        original_name = name

        # If saved from the TF HUB module
        name = name.replace("module/", "")

        # Renaming and simplifying
        name = name.replace("ffn_1", "ffn")
        name = name.replace("bert/", "albert/")
        name = name.replace("attention_1", "attention")
        name = name.replace("transform/", "")
        name = name.replace("LayerNorm_1", "full_layer_layer_norm")
        name = name.replace("LayerNorm", "attention/LayerNorm")
        name = name.replace("transformer/", "")

        # The feed forward layer had an 'intermediate' step which has been abstracted away
        name = name.replace("intermediate/dense/", "")
        name = name.replace("ffn/intermediate/output/dense/", "ffn_output/")

        # ALBERT attention was split between self and output which have been abstracted away
        name = name.replace("/output/", "/")
        name = name.replace("/self/", "/")

        # The pooler is a linear layer
        name = name.replace("pooler/dense", "pooler")

        # The classifier was simplified to predictions from cls/predictions
        name = name.replace("cls/predictions", "predictions")
        name = name.replace("predictions/attention", "predictions")

        # Naming was changed to be more explicit
        name = name.replace("embeddings/attention", "embeddings")
        name = name.replace("inner_group_", "albert_layers/")
        name = name.replace("group_", "albert_layer_groups/")

        # Classifier
        if len(name.split("/")) == 1 and ("output_bias" in name or "output_weights" in name):
            name = "classifier/" + name

        # No ALBERT model currently handles the next sentence prediction task
        if "seq_relationship" in name:
            name = name.replace("seq_relationship/output_", "sop_classifier/classifier/")
            name = name.replace("weights", "weight")

        name = name.split("/")

        # Ignore the gradients applied by the LAMB/ADAM optimizers.
        if (
            "adam_m" in name
            or "adam_v" in name
            or "AdamWeightDecayOptimizer" in name
            or "AdamWeightDecayOptimizer_1" in name
            or "global_step" in name
        ):
            logger.info("Skipping {}".format("/".join(name)))
            continue

        pointer = model
        for m_name in name:
            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
                scope_names = re.split(r"_(\d+)", m_name)
            else:
                scope_names = [m_name]

            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
                pointer = getattr(pointer, "weight")
            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
                pointer = getattr(pointer, "bias")
            elif scope_names[0] == "output_weights":
                pointer = getattr(pointer, "weight")
            elif scope_names[0] == "squad":
                pointer = getattr(pointer, "classifier")
            else:
                try:
                    pointer = getattr(pointer, scope_names[0])
                except AttributeError:
                    logger.info("Skipping {}".format("/".join(name)))
                    continue
            if len(scope_names) >= 2:
                num = int(scope_names[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
            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
        except AssertionError as e:
            e.args += (pointer.shape, array.shape)
            raise
        print("Initialize PyTorch weight {} from {}".format(name, original_name))
        pointer.data = torch.from_numpy(array)

    return model


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

    def __init__(self, config):
        super().__init__()
        self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.embedding_size)
        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_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 = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))

    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.forward
    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
        if input_ids is not None:
            input_shape = input_ids.size()
        else:
            input_shape = inputs_embeds.size()[:-1]

        seq_length = input_shape[1]

        if position_ids is None:
            position_ids = self.position_ids[:, :seq_length]

        if token_type_ids is None:
            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)

        if inputs_embeds is None:
            inputs_embeds = self.word_embeddings(input_ids)
        position_embeddings = self.position_embeddings(position_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)

        embeddings = inputs_embeds + position_embeddings + token_type_embeddings
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


class AlbertAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
            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.num_attention_heads = config.num_attention_heads
        self.hidden_size = config.hidden_size
        self.attention_head_size = 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.attention_dropout = nn.Dropout(config.attention_probs_dropout_prob)
        self.output_dropout = nn.Dropout(config.hidden_dropout_prob)
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.pruned_heads = set()

    # Copied from transformers.models.bert.modeling_bert.BertSelfAttention.transpose_for_scores
    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 prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(
            heads, self.num_attention_heads, self.attention_head_size, self.pruned_heads
        )

        # Prune linear layers
        self.query = prune_linear_layer(self.query, index)
        self.key = prune_linear_layer(self.key, index)
        self.value = prune_linear_layer(self.value, index)
        self.dense = prune_linear_layer(self.dense, index, dim=1)

        # Update hyper params and store pruned heads
        self.num_attention_heads = self.num_attention_heads - len(heads)
        self.all_head_size = self.attention_head_size * self.num_attention_heads
        self.pruned_heads = self.pruned_heads.union(heads)

    def forward(self, input_ids, attention_mask=None, head_mask=None, output_attentions=False):
        mixed_query_layer = self.query(input_ids)
        mixed_key_layer = self.key(input_ids)
        mixed_value_layer = self.value(input_ids)

        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)

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        if attention_mask is not None:
            # 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 = nn.Softmax(dim=-1)(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.attention_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, value_layer)

        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()

        # Should find a better way to do this
        w = (
            self.dense.weight.t()
            .view(self.num_attention_heads, self.attention_head_size, self.hidden_size)
            .to(context_layer.dtype)
        )
        b = self.dense.bias.to(context_layer.dtype)

        projected_context_layer = torch.einsum("bfnd,ndh->bfh", context_layer, w) + b
        projected_context_layer_dropout = self.output_dropout(projected_context_layer)
        layernormed_context_layer = self.LayerNorm(input_ids + projected_context_layer_dropout)
        return (layernormed_context_layer, attention_probs) if output_attentions else (layernormed_context_layer,)


class AlbertLayer(nn.Module):
    def __init__(self, config):
        super().__init__()

        self.config = config
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1
        self.full_layer_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.attention = AlbertAttention(config)
        self.ffn = nn.Linear(config.hidden_size, config.intermediate_size)
        self.ffn_output = nn.Linear(config.intermediate_size, config.hidden_size)
        self.activation = ACT2FN[config.hidden_act]
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(
        self, hidden_states, attention_mask=None, head_mask=None, output_attentions=False, output_hidden_states=False
    ):
        attention_output = self.attention(hidden_states, attention_mask, head_mask, output_attentions)

        ffn_output = apply_chunking_to_forward(
            self.ff_chunk,
            self.chunk_size_feed_forward,
            self.seq_len_dim,
            attention_output[0],
        )
        hidden_states = self.full_layer_layer_norm(ffn_output + attention_output[0])

        return (hidden_states,) + attention_output[1:]  # add attentions if we output them

    def ff_chunk(self, attention_output):
        ffn_output = self.ffn(attention_output)
        ffn_output = self.activation(ffn_output)
        ffn_output = self.ffn_output(ffn_output)
        return ffn_output


class AlbertLayerGroup(nn.Module):
    def __init__(self, config):
        super().__init__()

        self.albert_layers = nn.ModuleList([AlbertLayer(config) for _ in range(config.inner_group_num)])

    def forward(
        self, hidden_states, attention_mask=None, head_mask=None, output_attentions=False, output_hidden_states=False
    ):
        layer_hidden_states = ()
        layer_attentions = ()

        for layer_index, albert_layer in enumerate(self.albert_layers):
            layer_output = albert_layer(hidden_states, attention_mask, head_mask[layer_index], output_attentions)
            hidden_states = layer_output[0]

            if output_attentions:
                layer_attentions = layer_attentions + (layer_output[1],)

            if output_hidden_states:
                layer_hidden_states = layer_hidden_states + (hidden_states,)

        outputs = (hidden_states,)
        if output_hidden_states:
            outputs = outputs + (layer_hidden_states,)
        if output_attentions:
            outputs = outputs + (layer_attentions,)
        return outputs  # last-layer hidden state, (layer hidden states), (layer attentions)


class AlbertTransformer(nn.Module):
    def __init__(self, config):
        super().__init__()

        self.config = config
        self.embedding_hidden_mapping_in = nn.Linear(config.embedding_size, config.hidden_size)
        self.albert_layer_groups = nn.ModuleList([AlbertLayerGroup(config) for _ in range(config.num_hidden_groups)])

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        output_attentions=False,
        output_hidden_states=False,
        return_dict=True,
    ):
        hidden_states = self.embedding_hidden_mapping_in(hidden_states)

        all_hidden_states = (hidden_states,) if output_hidden_states else None
        all_attentions = () if output_attentions else None

        for i in range(self.config.num_hidden_layers):
            # Number of layers in a hidden group
            layers_per_group = int(self.config.num_hidden_layers / self.config.num_hidden_groups)

            # Index of the hidden group
            group_idx = int(i / (self.config.num_hidden_layers / self.config.num_hidden_groups))

            layer_group_output = self.albert_layer_groups[group_idx](
                hidden_states,
                attention_mask,
                head_mask[group_idx * layers_per_group : (group_idx + 1) * layers_per_group],
                output_attentions,
                output_hidden_states,
            )
            hidden_states = layer_group_output[0]

            if output_attentions:
                all_attentions = all_attentions + layer_group_output[-1]

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
        return BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
        )


class AlbertPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = AlbertConfig
    base_model_prefix = "albert"
    _keys_to_ignore_on_load_missing = [r"position_ids"]

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, (nn.Linear, nn.Embedding)):
            # 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 isinstance(module, (nn.Linear)) and module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


@dataclass
class AlbertForPreTrainingOutput(ModelOutput):
    """
    Output type of :class:`~transformers.AlbertForPreTraining`.

    Args:
        loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
            Total loss as the sum of the masked language modeling loss and the next sequence prediction
            (classification) loss.
        prediction_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        sop_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
            before SoftMax).
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
            of shape :obj:`(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(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.
    """

    loss: Optional[torch.FloatTensor] = None
    prediction_logits: torch.FloatTensor = None
    sop_logits: torch.FloatTensor = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None


ALBERT_START_DOCSTRING = r"""

    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
    pruning heads etc.)

    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
    general usage and behavior.

    Args:
        config (:class:`~transformers.AlbertConfig`): 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:`~transformers.PreTrainedModel.from_pretrained` method to load the model
            weights.
"""

ALBERT_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
            Indices of input sequence tokens in the vocabulary.

            Indices can be obtained using :class:`~transformers.AlbertTokenizer`. See
            :meth:`transformers.PreTrainedTokenizer.__call__` and :meth:`transformers.PreTrainedTokenizer.encode` for
            details.

            `What are input IDs? <../glossary.html#input-ids>`__
        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({0})`, `optional`):
            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            `What are attention masks? <../glossary.html#attention-mask>`__
        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
            Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
            1]``:

            - 0 corresponds to a `sentence A` token,
            - 1 corresponds to a `sentence B` token.

            `What are token type IDs? <../glossary.html#token-type-ids>`_
        position_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
            config.max_position_embeddings - 1]``.

            `What are position IDs? <../glossary.html#position-ids>`_
        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
            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**.

        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
            vectors than the model's internal embedding lookup matrix.
        output_attentions (:obj:`bool`, `optional`):
            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
            tensors for more detail.
        output_hidden_states (:obj:`bool`, `optional`):
            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
            more detail.
        return_dict (:obj:`bool`, `optional`):
            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""


@add_start_docstrings(
    "The bare ALBERT Model transformer outputting raw hidden-states without any specific head on top.",
    ALBERT_START_DOCSTRING,
)
class AlbertModel(AlbertPreTrainedModel):

    config_class = AlbertConfig
    load_tf_weights = load_tf_weights_in_albert
    base_model_prefix = "albert"

    def __init__(self, config, add_pooling_layer=True):
        super().__init__(config)

        self.config = config
        self.embeddings = AlbertEmbeddings(config)
        self.encoder = AlbertTransformer(config)
        if add_pooling_layer:
            self.pooler = nn.Linear(config.hidden_size, config.hidden_size)
            self.pooler_activation = nn.Tanh()
        else:
            self.pooler = None
            self.pooler_activation = None

        self.init_weights()

    def get_input_embeddings(self):
        return self.embeddings.word_embeddings

    def set_input_embeddings(self, value):
        self.embeddings.word_embeddings = value

    def _resize_token_embeddings(self, new_num_tokens):
        old_embeddings = self.embeddings.word_embeddings
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
        self.embeddings.word_embeddings = new_embeddings
        return self.embeddings.word_embeddings

    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} ALBERT has
        a different architecture in that its layers are shared across groups, which then has inner groups. If an ALBERT
        model has 12 hidden layers and 2 hidden groups, with two inner groups, there is a total of 4 different layers.

        These layers are flattened: the indices [0,1] correspond to the two inner groups of the first hidden layer,
        while [2,3] correspond to the two inner groups of the second hidden layer.

        Any layer with in index other than [0,1,2,3] will result in an error. See base class PreTrainedModel for more
        information about head pruning
        """
        for layer, heads in heads_to_prune.items():
            group_idx = int(layer / self.config.inner_group_num)
            inner_group_idx = int(layer - group_idx * self.config.inner_group_num)
            self.encoder.albert_layer_groups[group_idx].albert_layers[inner_group_idx].attention.prune_heads(heads)

    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="albert-base-v2",
        output_type=BaseModelOutputWithPooling,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        device = input_ids.device if input_ids is not None else inputs_embeds.device

        if attention_mask is None:
            attention_mask = torch.ones(input_shape, device=device)
        if token_type_ids is None:
            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)

        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)

        embedding_output = self.embeddings(
            input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
        )
        encoder_outputs = self.encoder(
            embedding_output,
            extended_attention_mask,
            head_mask=head_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = encoder_outputs[0]

        pooled_output = self.pooler_activation(self.pooler(sequence_output[:, 0])) if self.pooler is not None else None

        if not return_dict:
            return (sequence_output, pooled_output) + encoder_outputs[1:]

        return BaseModelOutputWithPooling(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )


@add_start_docstrings(
    """
    Albert Model with two heads on top as done during the pre-training: a `masked language modeling` head and a
    `sentence order prediction (classification)` head.
    """,
    ALBERT_START_DOCSTRING,
)
class AlbertForPreTraining(AlbertPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.albert = AlbertModel(config)
        self.predictions = AlbertMLMHead(config)
        self.sop_classifier = AlbertSOPHead(config)

        self.init_weights()

    def get_output_embeddings(self):
        return self.predictions.decoder

    def get_input_embeddings(self):
        return self.albert.embeddings.word_embeddings

    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @replace_return_docstrings(output_type=AlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        sentence_order_label=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (``torch.LongTensor`` of shape ``(batch_size, sequence_length)``, `optional`):
            Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
            config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
        sentence_order_label (``torch.LongTensor`` of shape ``(batch_size,)``, `optional`):
            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair
            (see :obj:`input_ids` docstring) Indices should be in ``[0, 1]``. ``0`` indicates original order (sequence
            A, then sequence B), ``1`` indicates switched order (sequence B, then sequence A).

        Returns:

        Example::

            >>> from transformers import AlbertTokenizer, AlbertForPreTraining
            >>> import torch

            >>> tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
            >>> model = AlbertForPreTraining.from_pretrained('albert-base-v2')

            >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
            >>> outputs = model(input_ids)

            >>> prediction_logits = outputs.prediction_logits
            >>> sop_logits = outputs.sop_logits

        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.albert(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output, pooled_output = outputs[:2]

        prediction_scores = self.predictions(sequence_output)
        sop_scores = self.sop_classifier(pooled_output)

        total_loss = None
        if labels is not None and sentence_order_label is not None:
            loss_fct = CrossEntropyLoss()
            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
            sentence_order_loss = loss_fct(sop_scores.view(-1, 2), sentence_order_label.view(-1))
            total_loss = masked_lm_loss + sentence_order_loss

        if not return_dict:
            output = (prediction_scores, sop_scores) + outputs[2:]
            return ((total_loss,) + output) if total_loss is not None else output

        return AlbertForPreTrainingOutput(
            loss=total_loss,
            prediction_logits=prediction_scores,
            sop_logits=sop_scores,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


class AlbertMLMHead(nn.Module):
    def __init__(self, config):
        super().__init__()

        self.LayerNorm = nn.LayerNorm(config.embedding_size)
        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
        self.dense = nn.Linear(config.hidden_size, config.embedding_size)
        self.decoder = nn.Linear(config.embedding_size, config.vocab_size)
        self.activation = ACT2FN[config.hidden_act]

        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
        self.decoder.bias = self.bias

    def forward(self, hidden_states):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.activation(hidden_states)
        hidden_states = self.LayerNorm(hidden_states)
        hidden_states = self.decoder(hidden_states)

        prediction_scores = hidden_states

        return prediction_scores


class AlbertSOPHead(nn.Module):
    def __init__(self, config):
        super().__init__()

        self.dropout = nn.Dropout(config.classifier_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

    def forward(self, pooled_output):
        dropout_pooled_output = self.dropout(pooled_output)
        logits = self.classifier(dropout_pooled_output)
        return logits


@add_start_docstrings(
    "Albert Model with a `language modeling` head on top.",
    ALBERT_START_DOCSTRING,
)
class AlbertForMaskedLM(AlbertPreTrainedModel):

    _keys_to_ignore_on_load_unexpected = [r"pooler"]

    def __init__(self, config):
        super().__init__(config)

        self.albert = AlbertModel(config, add_pooling_layer=False)
        self.predictions = AlbertMLMHead(config)

        self.init_weights()

    def get_output_embeddings(self):
        return self.predictions.decoder

    def get_input_embeddings(self):
        return self.albert.embeddings.word_embeddings

    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="albert-base-v2",
        output_type=MaskedLMOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
            config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.albert(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_outputs = outputs[0]

        prediction_scores = self.predictions(sequence_outputs)

        masked_lm_loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))

        if not return_dict:
            output = (prediction_scores,) + outputs[2:]
            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output

        return MaskedLMOutput(
            loss=masked_lm_loss,
            logits=prediction_scores,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """
    Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
    output) e.g. for GLUE tasks.
    """,
    ALBERT_START_DOCSTRING,
)
class AlbertForSequenceClassification(AlbertPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.albert = AlbertModel(config)
        self.dropout = nn.Dropout(config.classifier_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)

        self.init_weights()

    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="albert-base-v2",
        output_type=SequenceClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for computing the sequence classification/regression loss. Indices should be in ``[0, ...,
            config.num_labels - 1]``. If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.albert(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        pooled_output = outputs[1]

        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)

        loss = None
        if labels is not None:
            if self.num_labels == 1:
                #  We are doing regression
                loss_fct = MSELoss()
                loss = loss_fct(logits.view(-1), labels.view(-1))
            else:
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """
    Albert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
    Named-Entity-Recognition (NER) tasks.
    """,
    ALBERT_START_DOCSTRING,
)
class AlbertForTokenClassification(AlbertPreTrainedModel):

    _keys_to_ignore_on_load_unexpected = [r"pooler"]

    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.albert = AlbertModel(config, add_pooling_layer=False)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)

        self.init_weights()

    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="albert-base-v2",
        output_type=TokenClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Labels for computing the token classification loss. Indices should be in ``[0, ..., config.num_labels -
            1]``.
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.albert(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]

        sequence_output = self.dropout(sequence_output)
        logits = self.classifier(sequence_output)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            # Only keep active parts of the loss
            if attention_mask is not None:
                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))

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """
    Albert 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`).
    """,
    ALBERT_START_DOCSTRING,
)
class AlbertForQuestionAnswering(AlbertPreTrainedModel):

    _keys_to_ignore_on_load_unexpected = [r"pooler"]

    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.albert = AlbertModel(config, add_pooling_layer=False)
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)

        self.init_weights()

    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="albert-base-v2",
        output_type=QuestionAnsweringModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        start_positions=None,
        end_positions=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
            sequence are not taken into account for computing the loss.
        end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
            sequence are not taken into account for computing the loss.
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.albert(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]

        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)

        total_loss = None
        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

        if not return_dict:
            output = (start_logits, end_logits) + outputs[2:]
            return ((total_loss,) + output) if total_loss is not None else output

        return QuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """
    Albert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
    softmax) e.g. for RocStories/SWAG tasks.
    """,
    ALBERT_START_DOCSTRING,
)
class AlbertForMultipleChoice(AlbertPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.albert = AlbertModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, 1)

        self.init_weights()

    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
    @add_code_sample_docstrings(
        tokenizer_class=_TOKENIZER_FOR_DOC,
        checkpoint="albert-base-v2",
        output_type=MultipleChoiceModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for computing the multiple choice classification loss. Indices should be in ``[0, ...,
            num_choices-1]`` where `num_choices` is the size of the second dimension of the input tensors. (see
            `input_ids` above)
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]

        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
        inputs_embeds = (
            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
            if inputs_embeds is not None
            else None
        )
        outputs = self.albert(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        pooled_output = outputs[1]

        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        reshaped_logits = logits.view(-1, num_choices)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(reshaped_logits, labels)

        if not return_dict:
            output = (reshaped_logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return MultipleChoiceModelOutput(
            loss=loss,
            logits=reshaped_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )