modeling.py

# Copyright (c) 2020 PaddlePaddle Authors. 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.

import paddle
import paddle.nn as nn

from .. import PretrainedModel, register_base_model

__all__ = [
    'ErnieModel', 'ErniePretrainedModel', 'ErnieForSequenceClassification',
    'ErnieForTokenClassification', 'ErnieForQuestionAnswering',
    'ErnieForPretraining', 'ErniePretrainingCriterion'
]


class ErnieEmbeddings(nn.Layer):
    """
    Include embeddings from word, position and token_type embeddings
    """

    def __init__(self,
                 vocab_size,
                 hidden_size=768,
                 hidden_dropout_prob=0.1,
                 max_position_embeddings=512,
                 type_vocab_size=2,
                 pad_token_id=0):
        super(ErnieEmbeddings, self).__init__()
        self.word_embeddings = nn.Embedding(
            vocab_size, hidden_size, padding_idx=pad_token_id)
        self.position_embeddings = nn.Embedding(max_position_embeddings,
                                                hidden_size)
        self.token_type_embeddings = nn.Embedding(type_vocab_size, hidden_size)
        self.layer_norm = nn.LayerNorm(hidden_size)
        self.dropout = nn.Dropout(hidden_dropout_prob)

    def forward(self, input_ids, token_type_ids=None, position_ids=None):
        if position_ids is None:
            # maybe need use shape op to unify static graph and dynamic graph
            #seq_length = input_ids.shape[1]
            ones = paddle.ones_like(input_ids, dtype="int64")
            seq_length = paddle.cumsum(ones, axis=1)
            position_ids = seq_length - ones
            position_ids.stop_gradient = True
        if token_type_ids is None:
            token_type_ids = paddle.zeros_like(input_ids, dtype="int64")

        input_embedings = self.word_embeddings(input_ids)
        position_embeddings = self.position_embeddings(position_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)

        embeddings = input_embedings + position_embeddings + token_type_embeddings
        embeddings = self.layer_norm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


class ErniePooler(nn.Layer):
    """
    """

    def __init__(self, hidden_size):
        super(ErniePooler, self).__init__()
        self.dense = nn.Linear(hidden_size, 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 ErniePretrainedModel(PretrainedModel):
    """
    An abstract class for pretrained ERNIE models. It provides ERNIE related
    `model_config_file`, `resource_files_names`, `pretrained_resource_files_map`,
    `pretrained_init_configuration`, `base_model_prefix` for downloading and
    loading pretrained models. See `PretrainedModel` for more details.
    """

    model_config_file = "model_config.json"
    pretrained_init_configuration = {
        "ernie-1.0": {
            "attention_probs_dropout_prob": 0.1,
            "hidden_act": "relu",
            "hidden_dropout_prob": 0.1,
            "hidden_size": 768,
            "initializer_range": 0.02,
            "max_position_embeddings": 513,
            "num_attention_heads": 12,
            "num_hidden_layers": 12,
            "type_vocab_size": 2,
            "vocab_size": 18000,
            "pad_token_id": 0,
        },
        "ernie-tiny": {
            "attention_probs_dropout_prob": 0.1,
            "hidden_act": "relu",
            "hidden_dropout_prob": 0.1,
            "hidden_size": 1024,
            "initializer_range": 0.02,
            "intermediate_size": 4096,
            "max_position_embeddings": 600,
            "num_attention_heads": 16,
            "num_hidden_layers": 3,
            "type_vocab_size": 2,
            "vocab_size": 50006,
            "pad_token_id": 0,
        },
        "ernie-2.0-en": {
            "attention_probs_dropout_prob": 0.1,
            "hidden_act": "gelu",
            "hidden_dropout_prob": 0.1,
            "hidden_size": 768,
            "initializer_range": 0.02,
            "max_position_embeddings": 512,
            "num_attention_heads": 12,
            "num_hidden_layers": 12,
            "type_vocab_size": 4,
            "vocab_size": 30522,
            "pad_token_id": 0,
        },
        "ernie-2.0-large-en": {
            "attention_probs_dropout_prob": 0.1,
            "intermediate_size": 4096,  # special for ernie-2.0-large-en
            "hidden_act": "gelu",
            "hidden_dropout_prob": 0.1,
            "hidden_size": 1024,
            "initializer_range": 0.02,
            "max_position_embeddings": 512,
            "num_attention_heads": 16,
            "num_hidden_layers": 24,
            "type_vocab_size": 4,
            "vocab_size": 30522,
            "pad_token_id": 0,
        },
    }
    resource_files_names = {"model_state": "model_state.pdparams"}
    pretrained_resource_files_map = {
        "model_state": {
            "ernie-1.0":
            "https://paddlenlp.bj.bcebos.com/models/transformers/ernie/ernie_v1_chn_base.pdparams",
            "ernie-tiny":
            "https://paddlenlp.bj.bcebos.com/models/transformers/ernie_tiny/ernie_tiny.pdparams",
            "ernie-2.0-en":
            "https://paddlenlp.bj.bcebos.com/models/transformers/ernie_v2_base/ernie_v2_eng_base.pdparams",
            "ernie-2.0-large-en":
            "https://paddlenlp.bj.bcebos.com/models/transformers/ernie_v2_large/ernie_v2_eng_large.pdparams",
        }
    }
    base_model_prefix = "ernie"

    def init_weights(self, layer):
        """ Initialization hook """
        if isinstance(layer, (nn.Linear, nn.Embedding)):
            # only support dygraph, use truncated_normal and make it inplace
            # and configurable later
            if isinstance(layer.weight, paddle.Tensor):
                layer.weight.set_value(
                    paddle.tensor.normal(
                        mean=0.0,
                        std=self.initializer_range
                        if hasattr(self, "initializer_range") else
                        self.ernie.config["initializer_range"],
                        shape=layer.weight.shape))


@register_base_model
class ErnieModel(ErniePretrainedModel):
    """
    """

    def __init__(self,
                 vocab_size,
                 hidden_size=768,
                 num_hidden_layers=12,
                 num_attention_heads=12,
                 intermediate_size=3072,
                 hidden_act="gelu",
                 hidden_dropout_prob=0.1,
                 attention_probs_dropout_prob=0.1,
                 max_position_embeddings=512,
                 type_vocab_size=16,
                 initializer_range=0.02,
                 pad_token_id=0):
        super(ErnieModel, self).__init__()
        self.pad_token_id = pad_token_id
        self.initializer_range = initializer_range
        self.embeddings = ErnieEmbeddings(
            vocab_size, hidden_size, hidden_dropout_prob,
            max_position_embeddings, type_vocab_size, pad_token_id)
        encoder_layer = nn.TransformerEncoderLayer(
            hidden_size,
            num_attention_heads,
            intermediate_size,
            dropout=hidden_dropout_prob,
            activation=hidden_act,
            attn_dropout=attention_probs_dropout_prob,
            act_dropout=0)
        self.encoder = nn.TransformerEncoder(encoder_layer, num_hidden_layers)
        self.pooler = ErniePooler(hidden_size)
        self.apply(self.init_weights)

    def forward(self,
                input_ids,
                token_type_ids=None,
                position_ids=None,
                attention_mask=None):
        if attention_mask is None:
            attention_mask = paddle.unsqueeze(
                (input_ids == self.pad_token_id
                 ).astype(self.pooler.dense.weight.dtype) * -1e9,
                axis=[1, 2])
        embedding_output = self.embeddings(
            input_ids=input_ids,
            position_ids=position_ids,
            token_type_ids=token_type_ids)
        encoder_outputs = self.encoder(embedding_output, attention_mask)
        sequence_output = encoder_outputs
        pooled_output = self.pooler(sequence_output)
        return sequence_output, pooled_output


class ErnieForSequenceClassification(ErniePretrainedModel):
    """
    Model for sentence (pair) classification task with ERNIE.
    Args:
        ernie (ErnieModel): An instance of `ErnieModel`.
        num_classes (int, optional): The number of classes. Default 2
        dropout (float, optional): The dropout probability for output of ERNIE.
            If None, use the same value as `hidden_dropout_prob` of `ErnieModel`
            instance `Ernie`. Default None
    """

    def __init__(self, ernie, num_classes=2, dropout=None):
        super(ErnieForSequenceClassification, self).__init__()
        self.num_classes = num_classes
        self.ernie = ernie  # allow ernie to be config
        self.dropout = nn.Dropout(dropout if dropout is not None else
                                  self.ernie.config["hidden_dropout_prob"])
        self.classifier = nn.Linear(self.ernie.config["hidden_size"],
                                    num_classes)
        self.apply(self.init_weights)

    def forward(self,
                input_ids,
                token_type_ids=None,
                position_ids=None,
                attention_mask=None):
        _, pooled_output = self.ernie(
            input_ids,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            attention_mask=attention_mask)

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


class ErnieForQuestionAnswering(ErniePretrainedModel):
    def __init__(self, ernie):
        super(ErnieForQuestionAnswering, self).__init__()
        self.ernie = ernie  # allow ernie to be config
        self.classifier = nn.Linear(self.ernie.config["hidden_size"], 2)
        self.apply(self.init_weights)

    def forward(self,
                input_ids,
                token_type_ids=None,
                position_ids=None,
                attention_mask=None):
        sequence_output, _ = self.ernie(
            input_ids,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            attention_mask=attention_mask)

        logits = self.classifier(sequence_output)
        logits = paddle.transpose(logits, perm=[2, 0, 1])
        start_logits, end_logits = paddle.unstack(x=logits, axis=0)

        return start_logits, end_logits


class ErnieForTokenClassification(ErniePretrainedModel):
    def __init__(self, ernie, num_classes=2, dropout=None):
        super(ErnieForTokenClassification, self).__init__()
        self.num_classes = num_classes
        self.ernie = ernie  # allow ernie to be config
        self.dropout = nn.Dropout(dropout if dropout is not None else
                                  self.ernie.config["hidden_dropout_prob"])
        self.classifier = nn.Linear(self.ernie.config["hidden_size"],
                                    num_classes)
        self.apply(self.init_weights)

    def forward(self,
                input_ids,
                token_type_ids=None,
                position_ids=None,
                attention_mask=None):
        sequence_output, _ = self.ernie(
            input_ids,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            attention_mask=attention_mask)

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


class ErnieForPretraining(ErniePretrainedModel):
    def __init__(self, ernie):
        super(ErnieForPretraining, self).__init__()
        self.ernie = ernie
        self.cls = ErniePretrainingHeads(
            self.ernie.config["hidden_size"],
            self.ernie.config["vocab_size"],
            self.ernie.config["hidden_act"],
            embedding_weights=self.ernie.embeddings.word_embeddings.weight)

        self.apply(self.init_weights)

    def forward(self,
                input_ids,
                token_type_ids=None,
                position_ids=None,
                attention_mask=None,
                masked_positions=None):
        outputs = self.ernie(
            input_ids,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            attention_mask=attention_mask)
        sequence_output, pooled_output = outputs[:2]
        prediction_scores, seq_relationship_score = self.cls(
            sequence_output, pooled_output, masked_positions)
        return prediction_scores, seq_relationship_score


class ErniePretrainingCriterion(paddle.nn.Layer):
    def __init__(self, vocab_size):
        super(ErniePretrainingCriterion, self).__init__()
        self.loss_fn = paddle.nn.loss.CrossEntropyLoss(ignore_index=-1)
        self.vocab_size = vocab_size

    def forward(self, prediction_scores, seq_relationship_score,
                masked_lm_labels, next_sentence_labels, masked_lm_scale):
        masked_lm_loss = paddle.nn.functional.softmax_with_cross_entropy(
            prediction_scores, masked_lm_labels, ignore_index=-1)
        masked_lm_loss = masked_lm_loss / masked_lm_scale
        next_sentence_loss = paddle.nn.functional.softmax_with_cross_entropy(
            seq_relationship_score, next_sentence_labels)
        return paddle.sum(masked_lm_loss) + paddle.mean(next_sentence_loss)