modeling

class BertModel(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, pool_act='tanh')[source]

Bases: paddlenlp.transformers.bert.modeling.BertPretrainedModel

The bare BERT Model transformer outputting raw hidden-states.

This model inherits from PretrainedModel. Refer to the superclass documentation for the generic methods.

This model is also a Paddle paddle.nn.Layer subclass. Use it as a regular Paddle Layer and refer to the Paddle documentation for all matter related to general usage and behavior.

Parameters

  • vocab_size (int) – Vocabulary size of inputs_ids in BertModel. Also is the vocab size of token embedding matrix. Defines the number of different tokens that can be represented by the inputs_ids passed when calling BertModel.

  • hidden_size (int, optional) – Dimensionality of the embedding layer, encoder layer and pooler layer. Defaults to 768.

  • num_hidden_layers (int, optional) – Number of hidden layers in the Transformer encoder. Defaults to 12.

  • num_attention_heads (int, optional) – Number of attention heads for each attention layer in the Transformer encoder. Defaults to 12.

  • intermediate_size (int, optional) – Dimensionality of the feed-forward (ff) layer in the encoder. Input tensors to ff layers are firstly projected from hidden_size to intermediate_size, and then projected back to hidden_size. Typically intermediate_size is larger than hidden_size. Defaults to 3072.

  • hidden_act (str, optional) – The non-linear activation function in the feed-forward layer. "gelu", "relu" and any other paddle supported activation functions are supported. Defaults to "gelu".

  • hidden_dropout_prob (float, optional) – The dropout probability for all fully connected layers in the embeddings and encoder. Defaults to 0.1.

  • attention_probs_dropout_prob (float, optional) – The dropout probability used in MultiHeadAttention in all encoder layers to drop some attention target. Defaults to 0.1.

  • max_position_embeddings (int, optional) – The maximum value of the dimensionality of position encoding, which dictates the maximum supported length of an input sequence. Defaults to 512.

  • type_vocab_size (int, optional) – The vocabulary size of token_type_ids. Defaults to 16.

  • initializer_range (float, optional) –

    The standard deviation of the normal initializer. Defaults to 0.02.

    Note

    A normal_initializer initializes weight matrices as normal distributions. See BertPretrainedModel.init_weights() for how weights are initialized in BertModel.

  • pad_token_id (int, optional) – The index of padding token in the token vocabulary. Defaults to 0.

  • pooled_act (str, optional) – The non-linear activation function in the pooling layer. Defaults to "tanh".

forward(input_ids, token_type_ids=None, position_ids=None, attention_mask=None, output_hidden_states=False)[source]

The BertModel forward method, overrides the __call__() special method.

Parameters

  • input_ids (Tensor) – Indices of input sequence tokens in the vocabulary. They are numerical representations of tokens that build the input sequence. Its data type should be int64 and it has a shape of [batch_size, sequence_length].

  • token_type_ids (Tensor, optional) –

    Segment token indices to indicate different portions of the inputs. Selected in the range [0, type_vocab_size - 1]. If type_vocab_size is 2, which means the inputs have two portions. Indices can either be 0 or 1:

    • 0 corresponds to a sentence A token,

    • 1 corresponds to a sentence B token.

    Its data type should be int64 and it has a shape of [batch_size, sequence_length]. Defaults to None, which means we don’t add segment embeddings.

  • position_ids (Tensor, optional) – Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, max_position_embeddings - 1]. Shape as (batch_size, num_tokens) and dtype as int64. Defaults to None.

  • attention_mask (Tensor, optional) – Mask used in multi-head attention to avoid performing attention on to some unwanted positions, usually the paddings or the subsequent positions. Its data type can be int, float and bool. When the data type is bool, the masked tokens have False values and the others have True values. When the data type is int, the masked tokens have 0 values and the others have 1 values. When the data type is float, the masked tokens have -INF values and the others have 0 values. It is a tensor with shape broadcasted to [batch_size, num_attention_heads, sequence_length, sequence_length]. Defaults to None, which means nothing needed to be prevented attention to.

  • output_hidden_states (bool, optional) – Whether to return the output of each hidden layers. Defaults to False.

Returns

Returns tuple (sequence_output, pooled_output) or (encoder_outputs, pooled_output).

With the fields:

  • sequence_output (Tensor):

    Sequence of hidden-states at the last layer of the model. It’s data type should be float32 and its shape is [batch_size, sequence_length, hidden_size].

  • pooled_output (Tensor):

    The output of first token ([CLS]) in sequence. We “pool” the model by simply taking the hidden state corresponding to the first token. Its data type should be float32 and its shape is [batch_size, hidden_size].

  • encoder_outputs (List(Tensor)):

    A list of Tensor containing hidden-states of the model at each hidden layer in the Transformer encoder. The length of the list is num_hidden_layers. Each Tensor has a data type of float32 and its shape is [batch_size, sequence_length, hidden_size].

Return type

tuple

Example

import paddle
from paddlenlp.transformers import BertModel, BertTokenizer

tokenizer = BertTokenizer.from_pretrained('bert-wwm-chinese')
model = BertModel.from_pretrained('bert-wwm-chinese')

inputs = tokenizer("欢迎使用百度飞桨!")
inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()}
output = model(**inputs)
class BertPretrainedModel(name_scope=None, dtype='float32')[source]

Bases: paddlenlp.transformers.model_utils.PretrainedModel

An abstract class for pretrained BERT models. It provides BERT 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.

init_weights(layer)[source]

Initialization hook

base_model_class

alias of paddlenlp.transformers.bert.modeling.BertModel

class BertForPretraining(bert)[source]

Bases: paddlenlp.transformers.bert.modeling.BertPretrainedModel

Bert Model with pretraining tasks on top.

Parameters

bert (BertModel) – An instance of BertModel.

forward(input_ids, token_type_ids=None, position_ids=None, attention_mask=None, masked_positions=None)[source]
Parameters
Returns

Returns tuple (prediction_scores, seq_relationship_score).

With the fields:

  • prediction_scores (Tensor):

    The scores of masked token prediction. Its data type should be float32. If masked_positions is None, its shape is [batch_size, sequence_length, vocab_size]. Otherwise, its shape is [batch_size, mask_token_num, vocab_size].

  • seq_relationship_score (Tensor):

    The scores of next sentence prediction. Its data type should be float32 and its shape is [batch_size, 2].

Return type

tuple

class BertPretrainingCriterion(vocab_size)[source]

Bases: paddle.fluid.dygraph.layers.Layer

Parameters

vocab_size (int) – Vocabulary size of inputs_ids in BertModel. Defines the number of different tokens that can be represented by the inputs_ids passed when calling BertModel.

forward(prediction_scores, seq_relationship_score, masked_lm_labels, next_sentence_labels, masked_lm_scale)[source]
Parameters

  • prediction_scores (Tensor) – The scores of masked token prediction. Its data type should be float32. If masked_positions is None, its shape is [batch_size, sequence_length, vocab_size]. Otherwise, its shape is [batch_size, mask_token_num, vocab_size]

  • seq_relationship_score (Tensor) – The scores of next sentence prediction. Its data type should be float32 and its shape is [batch_size, 2]

  • masked_lm_labels (Tensor) – The labels of the masked language modeling, its dimensionality is equal to prediction_scores. Its data type should be int64. If masked_positions is None, its shape is [batch_size, sequence_length, 1]. Otherwise, its shape is [batch_size, mask_token_num, 1]

  • next_sentence_labels (Tensor) – The labels of the next sentence prediction task, the dimensionality of next_sentence_labels is equal to seq_relation_labels. Its data type should be int64 and its shape is [batch_size, 1]

  • masked_lm_scale (Tensor or int) – The scale of masked tokens. Used for the normalization of masked language modeling loss. If it is a Tensor, its data type should be int64 and its shape is equal to prediction_scores.

Returns

The pretraining loss, equals to the sum of masked_lm_loss plus the mean of next_sentence_loss. Its data type should be float32 and its shape is [1].

Return type

Tensor

class BertPretrainingHeads(hidden_size, vocab_size, activation, embedding_weights=None)[source]

Bases: paddle.fluid.dygraph.layers.Layer

Perform language modeling task and next sentence classification task.

Parameters

  • hidden_size (int) – See BertModel.

  • vocab_size (int) – See BertModel.

  • activation (str) – Activation function used in the language modeling task.

  • embedding_weights (Tensor, optional) – Decoding weights used to map hidden_states to logits of the masked token prediction. Its data type should be float32 and its shape is [vocab_size, hidden_size]. Defaults to None, which means use the same weights of the embedding layer.

forward(sequence_output, pooled_output, masked_positions=None)[source]
Parameters

  • sequence_output (Tensor) – Sequence of hidden-states at the last layer of the model. It’s data type should be float32 and its shape is [batch_size, sequence_length, hidden_size].

  • pooled_output (Tensor) – The output of first token ([CLS]) in sequence. We “pool” the model by simply taking the hidden state corresponding to the first token. Its data type should be float32 and its shape is [batch_size, hidden_size].

  • masked_positions (Tensor, optional) – A tensor indicates positions to be masked in the position embedding. Its data type should be int64 and its shape is [batch_size, mask_token_num]. mask_token_num is the number of masked tokens. It should be no bigger than sequence_length. Defaults to None, which means we output hidden-states of all tokens in masked token prediction.

Returns

Returns tuple (prediction_scores, seq_relationship_score).

With the fields:

  • prediction_scores (Tensor):

    The scores of masked token prediction. Its data type should be float32. If masked_positions is None, its shape is [batch_size, sequence_length, vocab_size]. Otherwise, its shape is [batch_size, mask_token_num, vocab_size].

  • seq_relationship_score (Tensor):

    The scores of next sentence prediction. Its data type should be float32 and its shape is [batch_size, 2].

Return type

tuple

class BertForSequenceClassification(bert, num_classes=2, dropout=None)[source]

Bases: paddlenlp.transformers.bert.modeling.BertPretrainedModel

Bert Model with a linear layer on top of the output layer, designed for sequence classification/regression tasks like GLUE tasks.

Parameters

  • bert (BertModel) – An instance of BertModel.

  • num_classes (int, optional) – The number of classes. Defaults to 2.

  • dropout (float, optional) – The dropout probability for output of BERT. If None, use the same value as hidden_dropout_prob of BertModel instance bert. Defaults to None.

forward(input_ids, token_type_ids=None, position_ids=None, attention_mask=None)[source]

The BertForSequenceClassification forward method, overrides the __call__() special method.

Parameters

  • input_ids (Tensor) – See BertModel.

  • token_type_ids (Tensor, optional) – See BertModel.

  • position_ids (Tensor, optional) – See BertModel.

  • attention_mask (list, optional) – See BertModel.

Returns

Returns tensor logits, a tensor of the input text classification logits. Shape as [batch_size, num_classes] and dtype as float32.

Return type

Tensor

Example

import paddle
from paddlenlp.transformers.bert.modeling import BertForSequenceClassification
from paddlenlp.transformers.bert.tokenizer import BertTokenizer

tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
model = BertForSequenceClassification.from_pretrained('bert-base-cased', num_classes=2)

inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!")
inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()}

logits = model(**inputs)
print(logits.shape)
# [1, 2]
class BertForTokenClassification(bert, num_classes=2, dropout=None)[source]

Bases: paddlenlp.transformers.bert.modeling.BertPretrainedModel

Bert Model with a linear layer on top of the hidden-states output layer, designed for token classification tasks like NER tasks.

Parameters

  • bert (BertModel) – An instance of BertModel.

  • num_classes (int, optional) – The number of classes. Defaults to 2.

  • dropout (float, optional) – The dropout probability for output of BERT. If None, use the same value as hidden_dropout_prob of BertModel instance bert. Defaults to None.

forward(input_ids, token_type_ids=None, position_ids=None, attention_mask=None)[source]

The BertForTokenClassification forward method, overrides the __call__() special method.

Parameters

  • input_ids (Tensor) – See BertModel.

  • token_type_ids (Tensor, optional) – See BertModel.

  • position_ids (Tensor, optional) – See BertModel.

  • attention_mask (list, optional) – See BertModel.

Returns

Returns tensor logits, a tensor of the input token classification logits. Shape as [batch_size, sequence_length, num_classes] and dtype as float32.

Return type

Tensor

Example

import paddle
from paddlenlp.transformers.bert.modeling import BertForTokenClassification
from paddlenlp.transformers.bert.tokenizer import BertTokenizer

tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
model = BertForTokenClassification.from_pretrained('bert-base-cased', num_classes=2)

inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!")
inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()}

logits = model(**inputs)
print(logits.shape)
# [1, 13, 2]
class BertForQuestionAnswering(bert, dropout=None)[source]

Bases: paddlenlp.transformers.bert.modeling.BertPretrainedModel

Bert Model with a linear layer on top of the hidden-states output to compute span_start_logits and span_end_logits, designed for question-answering tasks like SQuAD.

Parameters

  • bert (BertModel) – An instance of BertModel.

  • dropout (float, optional) – The dropout probability for output of BERT. If None, use the same value as hidden_dropout_prob of BertModel instance bert. Defaults to None.

forward(input_ids, token_type_ids=None, position_ids=None, attention_mask=None)[source]

The BertForQuestionAnswering forward method, overrides the __call__() special method.

Parameters

  • input_ids (Tensor) – See BertModel.

  • token_type_ids (Tensor, optional) – See BertModel.

Returns

Returns tuple (start_logits, end_logits).

With the fields:

  • start_logits (Tensor):

    A tensor of the input token classification logits, indicates the start position of the labelled span. Its data type should be float32 and its shape is [batch_size, sequence_length].

  • end_logits (Tensor):

    A tensor of the input token classification logits, indicates the end position of the labelled span. Its data type should be float32 and its shape is [batch_size, sequence_length].

Return type

tuple

Example

import paddle
from paddlenlp.transformers.bert.modeling import BertForQuestionAnswering
from paddlenlp.transformers.bert.tokenizer import BertTokenizer

tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
model = BertForQuestionAnswering.from_pretrained('bert-base-cased')

inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!")
inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()}
outputs = model(**inputs)

start_logits = outputs[0]
end_logits = outputs[1]
class BertForMultipleChoice(bert, num_choices=2, dropout=None)[source]

Bases: paddlenlp.transformers.bert.modeling.BertPretrainedModel

Bert Model with a linear layer on top of the hidden-states output layer, designed for multiple choice tasks like RocStories/SWAG tasks.

Parameters

  • bert (BertModel) – An instance of BertModel.

  • num_choices (int, optional) – The number of choices. Defaults to 2.

  • dropout (float, optional) – The dropout probability for output of Bert. If None, use the same value as hidden_dropout_prob of BertModel instance bert. Defaults to None.

forward(input_ids, token_type_ids=None, position_ids=None, attention_mask=None)[source]

The BertForMultipleChoice forward method, overrides the __call__() special method.

Parameters

  • input_ids (Tensor) – See BertModel and shape as [batch_size, num_choice, sequence_length].

  • token_type_ids (Tensor, optional) – See BertModel and shape as [batch_size, num_choice, sequence_length].

  • position_ids (Tensor, optional) – See BertModel and shape as [batch_size, num_choice, sequence_length].

  • attention_mask (list, optional) – See BertModel and shape as [batch_size, num_choice, sequence_length].

Returns

Returns tensor reshaped_logits, a tensor of the multiple choice classification logits. Shape as [batch_size, num_choice] and dtype as float32.

Return type

Tensor

Example

import paddle
from paddlenlp.transformers import BertForMultipleChoice, BertTokenizer
from paddlenlp.data import Pad, Dict

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = BertForMultipleChoice.from_pretrained('bert-base-uncased', num_choices=2)

data = [
    {
        "question": "how do you turn on an ipad screen?",
        "answer1": "press the volume button.",
        "answer2": "press the lock button.",
        "label": 1,
    },
    {
        "question": "how do you indent something?",
        "answer1": "leave a space before starting the writing",
        "answer2": "press the spacebar",
        "label": 0,
    },
]

text = []
text_pair = []
for d in data:
    text.append(d["question"])
    text_pair.append(d["answer1"])
    text.append(d["question"])
    text_pair.append(d["answer2"])

inputs = tokenizer(text, text_pair)
batchify_fn = lambda samples, fn=Dict(
    {
        "input_ids": Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input_ids
        "token_type_ids": Pad(
            axis=0, pad_val=tokenizer.pad_token_type_id
        ),  # token_type_ids
    }
): fn(samples)
inputs = batchify_fn(inputs)

reshaped_logits = model(
    input_ids=paddle.to_tensor(inputs[0], dtype="int64"),
    token_type_ids=paddle.to_tensor(inputs[1], dtype="int64"),
)
print(reshaped_logits.shape)
# [2, 2]
class BertForMaskedLM(bert)[source]

Bases: paddlenlp.transformers.bert.modeling.BertPretrainedModel

Bert Model with a masked language modeling head on top.

Parameters

bert (BertModel) – An instance of BertModel.

forward(input_ids, token_type_ids=None, position_ids=None, attention_mask=None)[source]
Parameters

  • input_ids (Tensor) – See BertModel.

  • token_type_ids (Tensor, optional) – See BertModel.

  • position_ids (Tensor, optional) – See BertModel.

  • attention_mask (Tensor, optional) – See BertModel.

Returns

Returns tensor prediction_scores, The scores of masked token prediction. Its data type should be float32 and shape is [batch_size, sequence_length, vocab_size].

Return type

Tensor

Example

import paddle
from paddlenlp.transformers import BertForMaskedLM, BertTokenizer

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = BertForMaskedLM.from_pretrained('bert-base-uncased')

inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!")
inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()}

logits = model(**inputs)
print(logits.shape)
# [1, 13, 30522]