gluonnlp.model

GluonNLP Toolkit supplies models for common NLP tasks with pre-trained weights. By default, all requested pre-trained weights are downloaded from public repo and stored in ~/.mxnet/models/.

Model Registry

The model registry provides an easy interface to obtain pre-defined and pre-trained models.

get_model

Returns a pre-defined model by name.

The get_model function returns a pre-defined model given the name of a registered model. The following sections of this page present a list of registered names for each model category.

Language Modeling

Components

AWDRNN

AWD language model by salesforce.

BiLMEncoder

Bidirectional LM encoder.

LSTMPCellWithClip

Long-Short Term Memory Projected (LSTMP) network cell with cell clip and projection clip.

StandardRNN

Standard RNN language model.

BigRNN

Big language model with LSTMP for inference.

Pre-defined models

awd_lstm_lm_1150

3-layer LSTM language model with weight-drop, variational dropout, and tied weights.

awd_lstm_lm_600

3-layer LSTM language model with weight-drop, variational dropout, and tied weights.

standard_lstm_lm_200

Standard 2-layer LSTM language model with tied embedding and output weights.

standard_lstm_lm_650

Standard 2-layer LSTM language model with tied embedding and output weights.

standard_lstm_lm_1500

Standard 2-layer LSTM language model with tied embedding and output weights.

big_rnn_lm_2048_512

Big 1-layer LSTMP language model.

Machine Translation

Seq2SeqEncoder

Base class of the encoders in sequence to sequence learning models.

TransformerEncoder

Structure of the Transformer Encoder.

TransformerEncoderCell

Structure of the Transformer Encoder Cell.

PositionwiseFFN

Structure of the Positionwise Feed-Forward Neural Network for Transformer.

transformer_en_de_512

Transformer pretrained model.

Bidirectional Encoder Representations from Transformers

Components

BERTModel

Generic Model for BERT (Bidirectional Encoder Representations from Transformers).

BERTLayerNorm

BERT style Layer Normalization.

BERTEncoder

Structure of the BERT Encoder.

BERTEncoderCell

Structure of the Transformer Encoder Cell for BERT.

BERTPositionwiseFFN

Structure of the Positionwise Feed-Forward Neural Network for BERT.

Pre-defined models

bert_12_768_12

Generic BERT BASE model.

bert_24_1024_16

Generic BERT LARGE model.

Convolutional Encoder

ConvolutionalEncoder

Convolutional encoder.

ELMo

Components

ELMoBiLM

ELMo Bidirectional language model

ELMoCharacterEncoder

ELMo character encoder

Pre-defined models

elmo_2x1024_128_2048cnn_1xhighway

ELMo 2-layer BiLSTM with 1024 hidden units, 128 projection size, 1 highway layer.

elmo_2x2048_256_2048cnn_1xhighway

ELMo 2-layer BiLSTM with 2048 hidden units, 256 projection size, 1 highway layer.

elmo_2x4096_512_2048cnn_2xhighway

ELMo 2-layer BiLSTM with 4096 hidden units, 512 projection size, 2 highway layer.

Highway Network

Highway

Highway network.

Attention Cell

AttentionCell

Abstract class for attention cells.

MultiHeadAttentionCell

Multi-head Attention Cell.

MLPAttentionCell

Concat the query and the key and use a single-hidden-layer MLP to get the attention score.

DotProductAttentionCell

Dot product attention between the query and the key.

Sequence Sampling

BeamSearchScorer

Score function used in beam search.

BeamSearchSampler

Draw samples from the decoder by beam search.

SequenceSampler

Draw samples from the decoder according to the step-wise distribution.

Other Modeling Utilities

WeightDropParameter

A Container holding parameters (weights) of Blocks and performs dropout.

apply_weight_drop

Apply weight drop to the parameter of a block.

L2Normalization

Normalize the input array by dividing the L2 norm along the given axis.

GELU

Gaussian Error Linear Unit.

ISDense

Importance sampled Dense block, which computes sampled pred output and labels for importance sampled softmax loss during training.

NCEDense

Noise contrastive estimated Dense block, which computes sampled pred output and labels for noise contrastive estimation loss during training.

SparseISDense

Importance sampled Dense block with sparse weights, which computes sampled pred output and labels for importance sampled softmax loss during training.

SparseNCEDense

Noise contrastive estimated Dense block with sparse weights, which computes sampled pred output and labels for noise contrastive estimation loss during training.

API Reference

Module for pre-defined NLP models.

This module contains definitions for the following model architectures: - AWD

You can construct a model with random weights by calling its constructor. Because NLP models are tied to vocabularies, you can either specify a dataset name to load and use the vocabulary of that dataset:

import gluonnlp as nlp
awd, vocab = nlp.model.awd_lstm_lm_1150(dataset_name='wikitext-2')

or directly specify a vocabulary object:

awd, vocab = nlp.model.awd_lstm_lm_1150(None, vocab=custom_vocab)

We provide pre-trained models for all the listed models. These models can constructed by passing pretrained=True:

awd, vocab = nlp.model.awd_lstm_lm_1150(dataset_name='wikitext-2'
                                        pretrained=True)

You can construct a predefined ELMo model structure:

import gluonnlp as nlp
elmo = nlp.model.elmo_2x1024_128_2048cnn_1xhighway(dataset_name='gbw')

You can also get a ELMo model with pretrained parameters:

import gluonnlp as nlp
elmo = nlp.model.elmo_2x1024_128_2048cnn_1xhighway(dataset_name='gbw', pretrained=True)
class gluonnlp.model.AWDRNN(mode, vocab_size, embed_size, hidden_size, num_layers, tie_weights, dropout, weight_drop, drop_h, drop_i, drop_e, **kwargs)[source]

AWD language model by salesforce.

Reference: https://github.com/salesforce/awd-lstm-lm

License: BSD 3-Clause

Parameters
  • mode (str) – The type of RNN to use. Options are ‘lstm’, ‘gru’, ‘rnn_tanh’, ‘rnn_relu’.

  • vocab_size (int) – Size of the input vocabulary.

  • embed_size (int) – Dimension of embedding vectors.

  • hidden_size (int) – Number of hidden units for RNN.

  • num_layers (int) – Number of RNN layers.

  • tie_weights (bool, default False) – Whether to tie the weight matrices of output dense layer and input embedding layer.

  • dropout (float) – Dropout rate to use for encoder output.

  • weight_drop (float) – Dropout rate to use on encoder h2h weights.

  • drop_h (float) – Dropout rate to on the output of intermediate layers of encoder.

  • drop_i (float) – Dropout rate to on the output of embedding.

  • drop_e (float) – Dropout rate to use on the embedding layer.

hybrid_forward(F, inputs, begin_state=None)[source]

Implement forward computation.

Parameters
  • inputs (NDArray) – input tensor with shape (sequence_length, batch_size) when layout is “TNC”.

  • begin_state (list) – initial recurrent state tensor with length equals to num_layers. the initial state with shape (1, batch_size, num_hidden)

Returns

  • out (NDArray) – output tensor with shape (sequence_length, batch_size, input_size) when layout is “TNC”.

  • out_states (list) – output recurrent state tensor with length equals to num_layers. the state with shape (1, batch_size, num_hidden)

class gluonnlp.model.StandardRNN(mode, vocab_size, embed_size, hidden_size, num_layers, dropout, tie_weights, **kwargs)[source]

Standard RNN language model.

Parameters
  • mode (str) – The type of RNN to use. Options are ‘lstm’, ‘gru’, ‘rnn_tanh’, ‘rnn_relu’.

  • vocab_size (int) – Size of the input vocabulary.

  • embed_size (int) – Dimension of embedding vectors.

  • hidden_size (int) – Number of hidden units for RNN.

  • num_layers (int) – Number of RNN layers.

  • dropout (float) – Dropout rate to use for encoder output.

  • tie_weights (bool, default False) – Whether to tie the weight matrices of output dense layer and input embedding layer.

hybrid_forward(F, inputs, begin_state=None)[source]

Defines the forward computation. Arguments can be either NDArray or Symbol.

Parameters
  • inputs (NDArray) –

    input tensor with shape (sequence_length, batch_size)

    when layout is “TNC”.

  • begin_state (list) – initial recurrent state tensor with length equals to num_layers-1. the initial state with shape (num_layers, batch_size, num_hidden)

Returns

  • out (NDArray) –

    output tensor with shape (sequence_length, batch_size, input_size)

    when layout is “TNC”.

  • out_states (list) – output recurrent state tensor with length equals to num_layers-1. the state with shape (num_layers, batch_size, num_hidden)

class gluonnlp.model.BigRNN(vocab_size, embed_size, hidden_size, num_layers, projection_size, embed_dropout=0.0, encode_dropout=0.0, **kwargs)[source]

Big language model with LSTMP for inference.

Parameters
  • vocab_size (int) – Size of the input vocabulary.

  • embed_size (int) – Dimension of embedding vectors.

  • hidden_size (int) – Number of hidden units for LSTMP.

  • num_layers (int) – Number of LSTMP layers.

  • projection_size (int) – Number of projection units for LSTMP.

  • embed_dropout (float) – Dropout rate to use for embedding output.

  • encode_dropout (float) – Dropout rate to use for encoder output.

forward(inputs, begin_state)[source]

Implement forward computation.

Parameters
  • inputs (NDArray) – input tensor with shape (sequence_length, batch_size) when layout is “TNC”.

  • begin_state (list) – initial recurrent state tensor with length equals to num_layers*2. For each layer the two initial states have shape (batch_size, num_hidden) and (batch_size, num_projection)

Returns

  • out (NDArray) –

    output tensor with shape (sequence_length, batch_size, vocab_size)

    when layout is “TNC”.

  • out_states (list) – output recurrent state tensor with length equals to num_layers*2. For each layer the two initial states have shape (batch_size, num_hidden) and (batch_size, num_projection)

gluonnlp.model.awd_lstm_lm_1150(dataset_name=None, vocab=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

3-layer LSTM language model with weight-drop, variational dropout, and tied weights.

Embedding size is 400, and hidden layer size is 1150.

Parameters
  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. Options are ‘wikitext-2’. If specified, then the returned vocabulary is extracted from the training set of the dataset. If None, then vocab is required, for specifying embedding weight size, and is directly returned. The pre-trained model achieves 73.32/69.74 ppl on Val and Test of wikitext-2 respectively.

  • vocab (gluonnlp.Vocab or None, default None) – Vocab object to be used with the language model. Required when dataset_name is not specified.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block, gluonnlp.Vocab

gluonnlp.model.awd_lstm_lm_600(dataset_name=None, vocab=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

3-layer LSTM language model with weight-drop, variational dropout, and tied weights.

Embedding size is 200, and hidden layer size is 600.

Parameters
  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. Options are ‘wikitext-2’. If specified, then the returned vocabulary is extracted from the training set of the dataset. If None, then vocab is required, for specifying embedding weight size, and is directly returned. The pre-trained model achieves 84.61/80.96 ppl on Val and Test of wikitext-2 respectively.

  • vocab (gluonnlp.Vocab or None, default None) – Vocab object to be used with the language model. Required when dataset_name is not specified.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block, gluonnlp.Vocab

gluonnlp.model.standard_lstm_lm_200(dataset_name=None, vocab=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

Standard 2-layer LSTM language model with tied embedding and output weights.

Both embedding and hidden dimensions are 200.

Parameters
  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. Options are ‘wikitext-2’. If specified, then the returned vocabulary is extracted from the training set of the dataset. If None, then vocab is required, for specifying embedding weight size, and is directly returned. The pre-trained model achieves 108.25/102.26 ppl on Val and Test of wikitext-2 respectively.

  • vocab (gluonnlp.Vocab or None, default None) – Vocabulary object to be used with the language model. Required when dataset_name is not specified.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block, gluonnlp.Vocab

gluonnlp.model.standard_lstm_lm_650(dataset_name=None, vocab=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

Standard 2-layer LSTM language model with tied embedding and output weights.

Both embedding and hidden dimensions are 650.

Parameters
  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. Options are ‘wikitext-2’. If specified, then the returned vocabulary is extracted from the training set of the dataset. If None, then vocab is required, for specifying embedding weight size, and is directly returned. The pre-trained model achieves 98.96/93.90 ppl on Val and Test of wikitext-2 respectively.

  • vocab (gluonnlp.Vocab or None, default None) – Vocabulary object to be used with the language model. Required when dataset_name is not specified.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block, gluonnlp.Vocab

gluonnlp.model.standard_lstm_lm_1500(dataset_name=None, vocab=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

Standard 2-layer LSTM language model with tied embedding and output weights.

Both embedding and hidden dimensions are 1500.

Parameters
  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. Options are ‘wikitext-2’. If specified, then the returned vocabulary is extracted from the training set of the dataset. If None, then vocab is required, for specifying embedding weight size, and is directly returned. The pre-trained model achieves 98.29/92.83 ppl on Val and Test of wikitext-2 respectively.

  • vocab (gluonnlp.Vocab or None, default None) – Vocabulary object to be used with the language model. Required when dataset_name is not specified.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block, gluonnlp.Vocab

gluonnlp.model.big_rnn_lm_2048_512(dataset_name=None, vocab=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

Big 1-layer LSTMP language model.

Both embedding and projection size are 512. Hidden size is 2048.

Parameters
  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. Options are ‘gbw’. If specified, then the returned vocabulary is extracted from the training set of the dataset. If None, then vocab is required, for specifying embedding weight size, and is directly returned. The pre-trained model achieves 44.05 ppl on Test of GBW dataset.

  • vocab (gluonnlp.Vocab or None, default None) – Vocabulary object to be used with the language model. Required when dataset_name is not specified.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block, gluonnlp.Vocab

class gluonnlp.model.BeamSearchScorer(alpha=1.0, K=5.0, from_logits=True, **kwargs)[source]

Score function used in beam search.

Implements the length-penalized score function used in the GNMT paper:

scores = (log_probs + scores) / length_penalty
length_penalty = (K + length)^\alpha / (K + 1)^\alpha
Parameters
  • alpha (float, default 1.0) –

  • K (float, default 5.0) –

  • from_logits (bool, default True) – Whether input is a log probability (usually from log_softmax) instead of unnormalized numbers.

hybrid_forward(F, outputs, scores, step)[source]

Overrides to construct symbolic graph for this Block.

Parameters
  • x (Symbol or NDArray) – The first input tensor.

  • *args (list of Symbol or list of NDArray) – Additional input tensors.

class gluonnlp.model.BeamSearchSampler(beam_size, decoder, eos_id, scorer=BeamSearchScorer( ), max_length=100)[source]

Draw samples from the decoder by beam search.

Parameters
  • beam_size (int) – The beam size.

  • decoder (callable) –

    Function of the one-step-ahead decoder, should have the form:

    outputs, new_states = decoder(step_input, states)
    

    The outputs, input should follow these rules:

    • step_input has shape (batch_size,),

    • outputs has shape (batch_size, V),

    • states and new_states have the same structure and the leading dimension of the inner NDArrays is the batch dimension.

  • eos_id (int) – Id of the EOS token. No other elements will be appended to the sample if it reaches eos_id.

  • scorer (BeamSearchScorer, default BeamSearchScorer(alpha=1.0, K=5)) – The score function used in beam search.

  • max_length (int, default 100) – The maximum search length.

class gluonnlp.model.HybridBeamSearchSampler(batch_size, beam_size, decoder, eos_id, scorer=BeamSearchScorer( ), max_length=100, vocab_size=None, prefix=None, params=None)[source]

Draw samples from the decoder by beam search.

Parameters
  • batch_size (int) – The batch size.

  • beam_size (int) – The beam size.

  • decoder (callable, must be hybridizable) –

    Function of the one-step-ahead decoder, should have the form:

    outputs, new_states = decoder(step_input, states)
    

    The outputs, input should follow these rules:

    • step_input has shape (batch_size,),

    • outputs has shape (batch_size, V),

    • states and new_states have the same structure and the leading dimension of the inner NDArrays is the batch dimension.

  • eos_id (int) – Id of the EOS token. No other elements will be appended to the sample if it reaches eos_id.

  • scorer (BeamSearchScorer, default BeamSearchScorer(alpha=1.0, K=5), must be hybridizable) – The score function used in beam search.

  • max_length (int, default 100) – The maximum search length.

  • vocab_size (int, default None, meaning decoder._vocab_size) – The vocabulary size

hybrid_forward(F, inputs, states)[source]

Sample by beam search.

Parameters
  • F

  • inputs (NDArray or Symbol) – The initial input of the decoder. Shape is (batch_size,).

  • states (Object that contains NDArrays or Symbols) – The initial states of the decoder.

Returns

  • samples (NDArray or Symbol) – Samples draw by beam search. Shape (batch_size, beam_size, length). dtype is int32.

  • scores (NDArray or Symbol) – Scores of the samples. Shape (batch_size, beam_size). We make sure that scores[i, :] are in descending order.

  • valid_length (NDArray or Symbol) – The valid length of the samples. Shape (batch_size, beam_size). dtype will be int32.

class gluonnlp.model.SequenceSampler(beam_size, decoder, eos_id, max_length=100, temperature=1.0, top_k=None)[source]

Draw samples from the decoder according to the step-wise distribution.

Parameters
  • beam_size (int) – The beam size.

  • decoder (callable) –

    Function of the one-step-ahead decoder, should have the form:

    outputs, new_states = decoder(step_input, states)
    

    The outputs, input should follow these rules:

    • step_input has shape (batch_size,)

    • outputs is the unnormalized prediction before softmax with shape (batch_size, V)

    • states and new_states have the same structure and the leading dimension of the inner NDArrays is the batch dimension.

  • eos_id (int) – Id of the EOS token. No other elements will be appended to the sample if it reaches eos_id.

  • max_length (int, default 100) – The maximum search length.

  • temperature (float, default 1.0) – Softmax temperature.

  • top_k (int or None, default None) – Sample only from the top-k candidates. If None, all candidates are considered.

class gluonnlp.model.AttentionCell(prefix=None, params=None)[source]

Abstract class for attention cells. Extend the class to implement your own attention method. One typical usage is to define your own _compute_weight() function to calculate the weights:

cell = AttentionCell()
out = cell(query, key, value, mask)
cast(dtype)[source]

Cast this Block to use another data type.

Parameters

dtype (str or numpy.dtype) – The new data type.

forward(query, key, value=None, mask=None)[source]

Defines the forward computation. Arguments can be either NDArray or Symbol.

hybrid_forward(F, query, key, value, mask=None)[source]

Overrides to construct symbolic graph for this Block.

Parameters
  • x (Symbol or NDArray) – The first input tensor.

  • *args (list of Symbol or list of NDArray) – Additional input tensors.

class gluonnlp.model.MultiHeadAttentionCell(base_cell, query_units, key_units, value_units, num_heads, use_bias=True, weight_initializer=None, bias_initializer='zeros', prefix=None, params=None)[source]

Multi-head Attention Cell.

In the MultiHeadAttentionCell, the input query/key/value will be linearly projected for num_heads times with different projection matrices. Each projected key, value, query will be used to calculate the attention weights and values. The output of each head will be concatenated to form the final output.

The idea is first proposed in “[Arxiv2014] Neural Turing Machines” and is later adopted in “[NIPS2017] Attention is All You Need” to solve the Neural Machine Translation problem.

Parameters
  • base_cell (AttentionCell) –

  • query_units (int) – Total number of projected units for query. Must be divided exactly by num_heads.

  • key_units (int) – Total number of projected units for key. Must be divided exactly by num_heads.

  • value_units (int) – Total number of projected units for value. Must be divided exactly by num_heads.

  • num_heads (int) – Number of parallel attention heads

  • use_bias (bool, default True) – Whether to use bias when projecting the query/key/values

  • weight_initializer (str or Initializer or None, default None) – Initializer of the weights.

  • bias_initializer (str or Initializer, default ‘zeros’) – Initializer of the bias.

  • prefix (str or None, default None) – See document of Block.

  • params (str or None, default None) – See document of Block.

class gluonnlp.model.MLPAttentionCell(units, act=Activation(tanh), normalized=False, dropout=0.0, weight_initializer=None, bias_initializer='zeros', prefix=None, params=None)[source]

Concat the query and the key and use a single-hidden-layer MLP to get the attention score. We provide two mode, the standard mode and the normalized mode.

In the standard mode:

score = v tanh(W [h_q, h_k] + b)

In the normalized mode (Same as TensorFlow):

score = g v / ||v||_2 tanh(W [h_q, h_k] + b)

This type of attention is first proposed in

Parameters
  • units (int) –

  • act (Activation, default nn.Activation('tanh')) –

  • normalized (bool, default False) – Whether to normalize the weight that maps the embedded hidden states to the final score. This strategy can be interpreted as a type of “[NIPS2016] Weight Normalization”.

  • dropout (float, default 0.0) – Attention dropout.

  • weight_initializer (str or Initializer or None, default None) – Initializer of the weights.

  • bias_initializer (str or Initializer, default ‘zeros’) – Initializer of the bias.

  • prefix (str or None, default None) – See document of Block.

  • params (ParameterDict or None, default None) – See document of Block.

class gluonnlp.model.DotProductAttentionCell(units=None, luong_style=False, scaled=True, normalized=False, use_bias=True, dropout=0.0, weight_initializer=None, bias_initializer='zeros', prefix=None, params=None)[source]

Dot product attention between the query and the key.

Depending on parameters, defined as:

units is None:
    score = <h_q, h_k>
units is not None and luong_style is False:
    score = <W_q h_q, W_k h_k>
units is not None and luong_style is True:
    score = <W h_q, h_k>
Parameters
  • units (int or None, default None) –

    Project the query and key to vectors with units dimension before applying the attention. If set to None, the query vector and the key vector are directly used to compute the attention and should have the same dimension:

    If the units is None,
        score = <h_q, h_k>
    Else if the units is not None and luong_style is False:
        score = <W_q h_q, W_k h_k>
    Else if the units is not None and luong_style is True:
        score = <W h_q, h_k>
    

  • luong_style (bool, default False) –

    If turned on, the score will be:

    score = <W h_q, h_k>
    

    units must be the same as the dimension of the key vector

  • scaled (bool, default True) –

    Whether to divide the attention weights by the sqrt of the query dimension. This is first proposed in “[NIPS2017] Attention is all you need.”:

    score = <h_q, h_k> / sqrt(dim_q)
    

  • normalized (bool, default False) –

    If turned on, the cosine distance is used, i.e:

    score = <h_q / ||h_q||, h_k / ||h_k||>
    

  • use_bias (bool, default True) – Whether to use bias in the projection layers.

  • dropout (float, default 0.0) – Attention dropout

  • weight_initializer (str or Initializer or None, default None) – Initializer of the weights

  • bias_initializer (str or Initializer, default ‘zeros’) – Initializer of the bias

  • prefix (str or None, default None) – See document of Block.

  • params (str or None, default None) – See document of Block.

gluonnlp.model.apply_weight_drop(block, local_param_regex, rate, axes=(), weight_dropout_mode='training')[source]

Apply weight drop to the parameter of a block.

Parameters
  • block (Block or HybridBlock) – The block whose parameter is to be applied weight-drop.

  • local_param_regex (str) – The regex for parameter names used in the self.params.get(), such as ‘weight’.

  • rate (float) – Fraction of the input units to drop. Must be a number between 0 and 1.

  • axes (tuple of int, default ()) – The axes on which dropout mask is shared. If empty, regular dropout is applied.

  • weight_drop_mode ({'training', 'always'}, default 'training') – Whether the weight dropout should be applied only at training time, or always be applied.

Examples

>>> net = gluon.rnn.LSTM(10, num_layers=2, bidirectional=True)
>>> gluonnlp.model.apply_weight_drop(net, r'.*h2h_weight', 0.5)
>>> net.collect_params()
lstm0_ (
  Parameter lstm0_l0_i2h_weight (shape=(40, 0), dtype=float32)
  WeightDropParameter lstm0_l0_h2h_weight (shape=(40, 10), dtype=float32, rate=0.5, mode=training)
  Parameter lstm0_l0_i2h_bias (shape=(40,), dtype=float32)
  Parameter lstm0_l0_h2h_bias (shape=(40,), dtype=float32)
  Parameter lstm0_r0_i2h_weight (shape=(40, 0), dtype=float32)
  WeightDropParameter lstm0_r0_h2h_weight (shape=(40, 10), dtype=float32, rate=0.5, mode=training)
  Parameter lstm0_r0_i2h_bias (shape=(40,), dtype=float32)
  Parameter lstm0_r0_h2h_bias (shape=(40,), dtype=float32)
  Parameter lstm0_l1_i2h_weight (shape=(40, 20), dtype=float32)
  WeightDropParameter lstm0_l1_h2h_weight (shape=(40, 10), dtype=float32, rate=0.5, mode=training)
  Parameter lstm0_l1_i2h_bias (shape=(40,), dtype=float32)
  Parameter lstm0_l1_h2h_bias (shape=(40,), dtype=float32)
  Parameter lstm0_r1_i2h_weight (shape=(40, 20), dtype=float32)
  WeightDropParameter lstm0_r1_h2h_weight (shape=(40, 10), dtype=float32, rate=0.5, mode=training)
  Parameter lstm0_r1_i2h_bias (shape=(40,), dtype=float32)
  Parameter lstm0_r1_h2h_bias (shape=(40,), dtype=float32)
)
>>> ones = mx.nd.ones((3, 4, 5))
>>> net.initialize()
>>> with mx.autograd.train_mode():
...     net(ones).max().asscalar() != net(ones).max().asscalar()
True
class gluonnlp.model.WeightDropParameter(parameter, rate=0.0, mode='training', axes=())[source]

A Container holding parameters (weights) of Blocks and performs dropout.

Parameters
  • parameter (Parameter) – The parameter which drops out.

  • rate (float, default 0.0) – Fraction of the input units to drop. Must be a number between 0 and 1. Dropout is not applied if dropout_rate is 0.

  • mode (str, default 'training') – Whether to only turn on dropout during training or to also turn on for inference. Options are ‘training’ and ‘always’.

  • axes (tuple of int, default ()) – Axes on which dropout mask is shared.

data(ctx=None)[source]

Returns a copy of this parameter on one context. Must have been initialized on this context before.

Parameters

ctx (Context) – Desired context.

Returns

Return type

NDArray on ctx

class gluonnlp.model.RNNCellLayer(rnn_cell, layout='TNC', **kwargs)[source]

A block that takes an rnn cell and makes it act like rnn layer.

Parameters
  • rnn_cell (Cell) – The cell to wrap into a layer-like block.

  • layout (str, default 'TNC') – The output layout of the layer.

forward(inputs, states=None)[source]

Defines the forward computation. Arguments can be either NDArray or Symbol.

class gluonnlp.model.L2Normalization(axis=-1, eps=1e-06, **kwargs)[source]

Normalize the input array by dividing the L2 norm along the given axis.

..code

out = data / (sqrt(sum(data**2, axis)) + eps)

Parameters
  • axis (int, default -1) – The axis to compute the norm value.

  • eps (float, default 1E-6) – The epsilon value to avoid dividing zero

hybrid_forward(F, x)[source]

Overrides to construct symbolic graph for this Block.

Parameters
  • x (Symbol or NDArray) – The first input tensor.

  • *args (list of Symbol or list of NDArray) – Additional input tensors.

class gluonnlp.model.GELU(**kwargs)[source]

Gaussian Error Linear Unit.

This is a smoother version of the RELU. https://arxiv.org/abs/1606.08415

Parameters
  • Inputs

    • data: input tensor with arbitrary shape.

  • Outputs

    • out: output tensor with the same shape as data.

hybrid_forward(F, x)[source]

Overrides to construct symbolic graph for this Block.

Parameters
  • x (Symbol or NDArray) – The first input tensor.

  • *args (list of Symbol or list of NDArray) – Additional input tensors.

class gluonnlp.model.Highway(input_size, num_layers, activation='relu', highway_bias=<gluonnlp.initializer.initializer.HighwayBias object>, **kwargs)[source]

Highway network.

We implemented the highway network proposed in the following work:

@article{srivastava2015highway,
  title={Highway networks},
  author={Srivastava, Rupesh Kumar and Greff, Klaus and Schmidhuber, J{\"u}rgen},
  journal={arXiv preprint arXiv:1505.00387},
  year={2015}
}

The full version of the work:

@inproceedings{srivastava2015training,
 title={Training very deep networks},
 author={Srivastava, Rupesh K and Greff, Klaus and Schmidhuber, J{\"u}rgen},
 booktitle={Advances in neural information processing systems},
 pages={2377--2385},
 year={2015}
}

A Highway layer is defined as below:

\[y = (1 - t) * x + t * f(A(x))\]

which is a gated combination of a linear transform and a non-linear transform of its input, where \(x\) is the input tensor, \(A\) is a linear transformer, \(f\) is an element-wise non-linear transformer, and \(t\) is an element-wise transform gate, and \(1-t\) refers to carry gate.

Parameters
  • input_size (int) – The dimension of the input tensor. We assume the input has shape (batch_size, input_size).

  • num_layers (int) – The number of highway layers to apply to the input.

  • activation (str, default 'relu') – The non-linear activation function to use. If you don’t specify anything, no activation is applied (ie. “linear” activation: a(x) = x).

  • highway_bias (HighwayBias,) – default HighwayBias(nonlinear_transform_bias=0.0, transform_gate_bias=-2.0) The biases applied to the highway layer. We set the default according to the above original work.

hybrid_forward(F, inputs, **kwargs)[source]

Forward computation for highway layer

Parameters

inputs (NDArray) – The input tensor is of shape (…, input_size).

Returns

outputs – The output tensor is of the same shape with input tensor (…, input_size).

Return type

NDArray

class gluonnlp.model.ConvolutionalEncoder(embed_size=15, num_filters=(25, 50, 75, 100, 125, 150), ngram_filter_sizes=(1, 2, 3, 4, 5, 6), conv_layer_activation='tanh', num_highway=1, highway_layer_activation='relu', highway_bias=<gluonnlp.initializer.initializer.HighwayBias object>, output_size=None, **kwargs)[source]

Convolutional encoder.

We implement the convolutional encoder proposed in the following work:

@inproceedings{kim2016character,
 title={Character-Aware Neural Language Models.},
 author={Kim, Yoon and Jernite, Yacine and Sontag, David and Rush, Alexander M},
 booktitle={AAAI},
 pages={2741--2749},
 year={2016}
}
Parameters
  • embed_size (int, default 15) – The input dimension to the encoder. We set the default according to the original work’s experiments on PTB dataset with Char-small model setting.

  • num_filters (Tuple[int], default (25, 50, 75, 100, 125, 150)) – The output dimension for each convolutional layer according to the filter sizes, which are the number of the filters learned by the layers. We set the default according to the original work’s experiments on PTB dataset with Char-small model setting.

  • ngram_filter_sizes (Tuple[int], default (1, 2, 3, 4, 5, 6)) – The size of each convolutional layer, and len(ngram_filter_sizes) equals to the number of convolutional layers. We set the default according to the original work’s experiments on PTB dataset with Char-small model setting.

  • conv_layer_activation (str, default 'tanh') – Activation function to be used after convolutional layer. We set the default according to the original work’s experiments on PTB dataset with Char-small model setting.

  • num_highway (int, default '1') – The number of layers of the Highway layer. We set the default according to the original work’s experiments on PTB dataset with Char-small model setting.

  • highway_layer_activation (str, default 'relu') – Activation function to be used after highway layer. If you don’t specify anything, no activation is applied (ie. “linear” activation: a(x) = x). We set the default according to the original work’s experiments on PTB dataset with Char-small model setting.

  • highway_bias (HighwayBias,) – default HighwayBias(nonlinear_transform_bias=0.0, transform_gate_bias=-2.0) The biases applied to the highway layer. We set the default according to the above original work.

  • output_size (int, default None) – The output dimension after conducting the convolutions and max pooling, and applying highways, as well as linear projection.

hybrid_forward(F, inputs, mask=None)[source]

Forward computation for char_encoder

Parameters
  • inputs (NDArray) – The input tensor is of shape (seq_len, batch_size, embedding_size) TNC.

  • mask (NDArray) – The mask applied to the input of shape (seq_len, batch_size), the mask will be broadcasted along the embedding dimension.

Returns

output – The output of the encoder with shape (batch_size, output_size)

Return type

NDArray

class gluonnlp.model.ISDense(num_classes, num_sampled, in_unit, remove_accidental_hits=True, dtype='float32', weight_initializer=None, bias_initializer='zeros', sparse_grad=True, prefix=None, params=None)[source]

Importance sampled Dense block, which computes sampled pred output and labels for importance sampled softmax loss during training.

Reference:

Exploring the Limits of Language Modeling Jozefowicz, Rafal and Vinyals, Oriol and Schuster, Mike and Shazeer, Noam and Wu, Yonghui https://arxiv.org/pdf/1602.02410

Please use loss.SoftmaxCrossEntropyLoss for sampled softmax loss.

Note

If sparse_grad is set to True, the gradient w.r.t input and output embeddings will be sparse. Only a subset of optimizers support sparse gradients, including SGD, AdaGrad and Adam. By default lazy_update is turned on for these optimizers, which may perform differently from standard updates. For more details, please check the Optimization API at https://mxnet.incubator.apache.org/api/python/optimization/optimization.html

Example:

# network with importance sampling for training
encoder = Encoder(..)
decoder = ISDense(..)
train_net.add(encoder)
train_net.add(decoder)
loss = SoftmaxCrossEntropyLoss()

# training
for x, y, sampled_values in train_batches:
    pred, new_targets = train_net(x, sampled_values, y)
    l = loss(pred, new_targets)

# network for testing
test_net.add(encoder)
test_net.add(Dense(..., params=decoder.params))

# testing
for x, y in test_batches:
    pred = test_net(x)
    l = loss(pred, y)
Parameters
  • num_classes (int) – Number of possible classes.

  • num_sampled (int) – Number of classes randomly sampled for each batch.

  • in_unit (int) – Dimensionality of the input space.

  • remove_accidental_hits (bool, default True) – Whether to remove “accidental hits” when a sampled candidate is equal to one of the true classes.

  • dtype (str or np.dtype, default 'float32') – Data type of output embeddings.

  • weight_initializer (str or Initializer, optional) – Initializer for the kernel weights matrix.

  • bias_initializer (str or Initializer, optional) – Initializer for the bias vector.

  • sparse_grad (bool, default True.) – Whether to use sparse gradient.

  • Inputs

    • x: A tensor of shape (batch_size, in_unit). The forward activation of the input network.

    • sampled_values : A list of three tensors for sampled_classes with shape (num_samples,), expected_count_sampled with shape (num_samples,), and expected_count_true with shape (sequence_length, batch_size).

    • label: A tensor of shape (batch_size,1). The target classes.

  • Outputs

    • out: A tensor of shape (batch_size, 1+num_sampled). The output probability for the true class and sampled classes

    • new_targets: A tensor of shape (batch_size,). The new target classes.

class gluonnlp.model.NCEDense(num_classes, num_sampled, in_unit, remove_accidental_hits=False, dtype='float32', weight_initializer=None, bias_initializer='zeros', sparse_grad=True, prefix=None, params=None)[source]

Noise contrastive estimated Dense block, which computes sampled pred output and labels for noise contrastive estimation loss during training.

Reference:

Exploring the Limits of Language Modeling Jozefowicz, Rafal and Vinyals, Oriol and Schuster, Mike and Shazeer, Noam and Wu, Yonghui https://arxiv.org/pdf/1602.02410

Please use loss.SigmoidBinaryCrossEntropyLoss for noise contrastive estimation loss during training.

Note

If sparse_grad is set to True, the gradient w.r.t input and output embeddings will be sparse. Only a subset of optimizers support sparse gradients, including SGD, AdaGrad and Adam. By default lazy_update is turned on for these optimizers, which may perform differently from standard updates. For more details, please check the Optimization API at: https://mxnet.incubator.apache.org/api/python/optimization/optimization.html

Example:

# network with sampling for training
encoder = Encoder(..)
decoder = NCEDense(..)
train_net.add(encoder)
train_net.add(decoder)
loss_train = SigmoidBinaryCrossEntropyLoss()

# training
for x, y, sampled_values in train_batches:
    pred, new_targets = train_net(x, sampled_values, y)
    l = loss_train(pred, new_targets)

# network for testing
test_net.add(encoder)
test_net.add(Dense(..., params=decoder.params))
loss_test = SoftmaxCrossEntropyLoss()

# testing
for x, y in test_batches:
    pred = test_net(x)
    l = loss_test(pred, y)
Parameters
  • num_classes (int) – Number of possible classes.

  • num_sampled (int) – Number of classes randomly sampled for each batch.

  • in_unit (int) – Dimensionality of the input space.

  • remove_accidental_hits (bool, default False) – Whether to remove “accidental hits” when a sampled candidate is equal to one of the true classes.

  • dtype (str or np.dtype, default 'float32') – Data type of output embeddings.

  • weight_initializer (str or Initializer, optional) – Initializer for the kernel weights matrix.

  • bias_initializer (str or Initializer, optional) – Initializer for the bias vector.

  • sparse_grad (bool, default True.) – Whether to use sparse gradient.

  • Inputs

    • x: A tensor of shape (batch_size, in_unit). The forward activation of the input network.

    • sampled_values : A list of three tensors for sampled_classes with shape (num_samples,), expected_count_sampled with shape (num_samples,), and expected_count_true with shape (sequence_length, batch_size).

    • label: A tensor of shape (batch_size,1). The target classes.

  • Outputs

    • out: A tensor of shape (batch_size, 1+num_sampled). The output probability for the true class and sampled classes

    • new_targets: A tensor of shape (batch_size, 1+num_sampled). The new target classes.

class gluonnlp.model.SparseISDense(num_classes, num_sampled, in_unit, remove_accidental_hits=True, dtype='float32', weight_initializer=None, bias_initializer='zeros', prefix=None, params=None)[source]

Importance sampled Dense block with sparse weights, which computes sampled pred output and labels for importance sampled softmax loss during training.

Reference:

Exploring the Limits of Language Modeling Jozefowicz, Rafal and Vinyals, Oriol and Schuster, Mike and Shazeer, Noam and Wu, Yonghui https://arxiv.org/pdf/1602.02410

Please use loss.SoftmaxCrossEntropyLoss for sampled softmax loss.

The block is designed for distributed training with extremely large number of classes to reduce communication overhead and memory consumption. Both weight and gradient w.r.t. weight are RowSparseNDArray.

Note

Different from ISDense block, the weight parameter is stored in row_sparse format, which helps reduce memory consumption and communication overhead during multi-GPU training. However, sparse parameters cannot be shared with other blocks, nor could we hybridize a block containing sparse parameters. Therefore, the parameters have to be saved before they are used for testing.

Example:

# network with importance sampled softmax for training
encoder = Encoder(..)
train_net.add(encoder)
train_net.add(SparseISDense(.., prefix='decoder')))
loss = SoftmaxCrossEntropyLoss()

# training
for x, y, sampled_values in train_batches:
    pred, new_targets = train_net(x, sampled_values, y)
    l = loss(pred, new_targets)

# save params
train_net.save_parameters('net.params')

# network for testing
test_net.add(encoder)
test_net.add(Dense(..., prefix='decoder'))

# load params
test_net.load_parameters('net.params')

# testing
for x, y in test_batches:
    pred = test_net(x)
    l = loss(pred, y)
Parameters
  • num_classes (int) – Number of possible classes.

  • num_sampled (int) – Number of classes randomly sampled for each batch.

  • in_unit (int) – Dimensionality of the input space.

  • remove_accidental_hits (bool, default True) – Whether to remove “accidental hits” when a sampled candidate is equal to one of the true classes.

  • dtype (str or np.dtype, default 'float32') – Data type of output embeddings.

  • weight_initializer (str or Initializer, optional) – Initializer for the kernel weights matrix.

  • bias_initializer (str or Initializer, optional) – Initializer for the bias vector.

  • Inputs

    • x: A tensor of shape (batch_size, in_unit). The forward activation of the input network.

    • sampled_values : A list of three tensors for sampled_classes with shape (num_samples,), expected_count_sampled with shape (num_samples,), and expected_count_true with shape (sequence_length, batch_size).

    • label: A tensor of shape (batch_size,1). The target classes.

  • Outputs

    • out: A tensor of shape (batch_size, 1+num_sampled). The output probability for the true class and sampled classes

    • new_targets: A tensor of shape (batch_size,). The new target classes.

class gluonnlp.model.SparseNCEDense(num_classes, num_sampled, in_unit, remove_accidental_hits=True, dtype='float32', weight_initializer=None, bias_initializer='zeros', prefix=None, params=None)[source]

Noise contrastive estimated Dense block with sparse weights, which computes sampled pred output and labels for noise contrastive estimation loss during training.

Reference:

Exploring the Limits of Language Modeling Jozefowicz, Rafal and Vinyals, Oriol and Schuster, Mike and Shazeer, Noam and Wu, Yonghui https://arxiv.org/pdf/1602.02410

Please use loss.SigmoidBinaryCrossEntropyLoss for noise contrastive estimation loss during training.

The block is designed for distributed training with extremely large number of classes to reduce communication overhead and memory consumption. Both weight and gradient w.r.t. weight are RowSparseNDArray.

Note

Different from NCEDense block, the weight parameter is stored in row_sparse format, which helps reduce memory consumption and communication overhead during multi-GPU training. However, sparse parameters cannot be shared with other blocks, nor could we hybridize a block containing sparse parameters. Therefore, the parameters have to be saved before they are used for testing.

Example:

# network with importance sampled softmax for training
encoder = Encoder(..)
train_net.add(encoder)
train_net.add(SparseNCEDense(.., prefix='decoder')))
train_loss = SigmoidBinaryCrossEntropyLoss()

# training
for x, y, sampled_values in train_batches:
    pred, new_targets = train_net(x, sampled_values, y)
    l = train_loss(pred, new_targets)

# save params
train_net.save_parameters('net.params')

# network for testing
test_net.add(encoder)
test_net.add(Dense(..., prefix='decoder'))

# load params
test_net.load_parameters('net.params')
test_loss = SoftmaxCrossEntropyLoss()

# testing
for x, y in test_batches:
    pred = test_net(x)
    l = test_loss(pred, y)
Parameters
  • num_classes (int) – Number of possible classes.

  • num_sampled (int) – Number of classes randomly sampled for each batch.

  • in_unit (int) – Dimensionality of the input space.

  • remove_accidental_hits (bool, default True) – Whether to remove “accidental hits” when a sampled candidate is equal to one of the true classes.

  • dtype (str or np.dtype, default 'float32') – Data type of output embeddings.

  • weight_initializer (str or Initializer, optional) – Initializer for the kernel weights matrix.

  • bias_initializer (str or Initializer, optional) – Initializer for the bias vector.

  • Inputs

    • x: A tensor of shape (batch_size, in_unit). The forward activation of the input network.

    • sampled_values : A list of three tensors for sampled_classes with shape (num_samples,), expected_count_sampled with shape (num_samples,), and expected_count_true with shape (sequence_length, batch_size).

    • label: A tensor of shape (batch_size, 1+num_samples). The target classes.

  • Outputs

    • out: A tensor of shape (batch_size, 1+num_sampled). The output probability for the true class and sampled classes

    • new_targets: A tensor of shape (batch_size, 1+num_sampled). The new target classes.

gluonnlp.model.get_model(name, **kwargs)[source]

Returns a pre-defined model by name.

Parameters
  • name (str) – Name of the model.

  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. For language model, options are ‘wikitext-2’. For ELMo, Options are ‘gbw’ and ‘5bw’. ‘gbw’ represents 1 Billion Word Language Model Benchmark http://www.statmt.org/lm-benchmark/; ‘5bw’ represents a dataset of 5.5B tokens consisting of Wikipedia (1.9B) and all of the monolingual news crawl data from WMT 2008-2012 (3.6B). If specified, then the returned vocabulary is extracted from the training set of the dataset. If None, then vocab is required, for specifying embedding weight size, and is directly returned.

  • vocab (gluonnlp.Vocab or None, default None) – Vocabulary object to be used with the language model. Required when dataset_name is not specified. None Vocabulary object is required with the ELMo model.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models' with MXNET_HOME defaults to '~/.mxnet') – Location for keeping the model parameters.

Returns

Return type

gluon.Block, gluonnlp.Vocab, (optional) gluonnlp.Vocab

class gluonnlp.model.BiLMEncoder(mode, num_layers, input_size, hidden_size, dropout=0.0, skip_connection=True, proj_size=None, cell_clip=None, proj_clip=None, **kwargs)[source]

Bidirectional LM encoder.

We implement the encoder of the biLM proposed in the following work:

@inproceedings{Peters:2018,
author={Peters, Matthew E. and  Neumann, Mark and Iyyer, Mohit and Gardner, Matt and Clark,
Christopher and Lee, Kenton and Zettlemoyer, Luke},
title={Deep contextualized word representations},
booktitle={Proc. of NAACL},
year={2018}
}
Parameters
  • mode (str) – The type of RNN cell to use. Options are ‘lstmpc’, ‘rnn_tanh’, ‘rnn_relu’, ‘lstm’, ‘gru’.

  • num_layers (int) – The number of RNN cells in the encoder.

  • input_size (int) – The initial input size of in the RNN cell.

  • hidden_size (int) – The hidden size of the RNN cell.

  • dropout (float) – The dropout rate to use for encoder output.

  • skip_connection (bool) – Whether to add skip connections (add RNN cell input to output)

  • proj_size (int) – The projection size of each LSTMPCellWithClip cell

  • cell_clip (float) – Clip cell state between [-cellclip, cell_clip] in LSTMPCellWithClip cell

  • proj_clip (float) – Clip projection between [-projclip, projclip] in LSTMPCellWithClip cell

hybrid_forward(F, inputs, states=None, mask=None)[source]

Defines the forward computation for cache cell. Arguments can be either NDArray or Symbol.

Parameters
  • inputs (NDArray) – The input data layout=’TNC’.

  • states (Tuple[List[List[NDArray]]]) – The states. including: states[0] indicates the states used in forward layer, Each layer has a list of two initial tensors with shape (batch_size, proj_size) and (batch_size, hidden_size). states[1] indicates the states used in backward layer, Each layer has a list of two initial tensors with shape (batch_size, proj_size) and (batch_size, hidden_size).

Returns

  • out (NDArray) – The output data with shape (num_layers, seq_len, batch_size, 2*input_size).

  • [states_forward, states_backward] (List) – Including: states_forward: The out states from forward layer, which has the same structure with states[0]. states_backward: The out states from backward layer, which has the same structure with states[1].

class gluonnlp.model.LSTMPCellWithClip(hidden_size, projection_size, i2h_weight_initializer=None, h2h_weight_initializer=None, h2r_weight_initializer=None, i2h_bias_initializer='zeros', h2h_bias_initializer='zeros', input_size=0, cell_clip=None, projection_clip=None, prefix=None, params=None)[source]

Long-Short Term Memory Projected (LSTMP) network cell with cell clip and projection clip. Each call computes the following function:

\[\begin{split}\DeclareMathOperator{\sigmoid}{sigmoid} \begin{array}{ll} i_t = \sigmoid(W_{ii} x_t + b_{ii} + W_{ri} r_{(t-1)} + b_{ri}) \\ f_t = \sigmoid(W_{if} x_t + b_{if} + W_{rf} r_{(t-1)} + b_{rf}) \\ g_t = \tanh(W_{ig} x_t + b_{ig} + W_{rc} r_{(t-1)} + b_{rg}) \\ o_t = \sigmoid(W_{io} x_t + b_{io} + W_{ro} r_{(t-1)} + b_{ro}) \\ c_t = c_{\text{clip}}(f_t * c_{(t-1)} + i_t * g_t) \\ h_t = o_t * \tanh(c_t) \\ r_t = p_{\text{clip}}(W_{hr} h_t) \end{array}\end{split}\]

where \(c_{\text{clip}}\) is the cell clip applied on the next cell; \(r_t\) is the projected recurrent activation at time t, \(p_{\text{clip}}\) means apply projection clip on he projected output. math:h_t is the hidden state at time t, \(c_t\) is the cell state at time t, \(x_t\) is the input at time t, and \(i_t\), \(f_t\), \(g_t\), \(o_t\) are the input, forget, cell, and out gates, respectively.

Parameters
  • hidden_size (int) – Number of units in cell state symbol.

  • projection_size (int) – Number of units in output symbol.

  • i2h_weight_initializer (str or Initializer) – Initializer for the input weights matrix, used for the linear transformation of the inputs.

  • h2h_weight_initializer (str or Initializer) – Initializer for the recurrent weights matrix, used for the linear transformation of the hidden state.

  • h2r_weight_initializer (str or Initializer) – Initializer for the projection weights matrix, used for the linear transformation of the recurrent state.

  • i2h_bias_initializer (str or Initializer, default 'lstmbias') – Initializer for the bias vector. By default, bias for the forget gate is initialized to 1 while all other biases are initialized to zero.

  • h2h_bias_initializer (str or Initializer) – Initializer for the bias vector.

  • prefix (str) – Prefix for name of Block`s (and name of weight if params is `None).

  • params (Parameter or None) – Container for weight sharing between cells. Created if None.

  • cell_clip (float) – Clip cell state between [-cell_clip, cell_clip] in LSTMPCellWithClip cell

  • projection_clip (float) – Clip projection between [-projection_clip, projection_clip] in LSTMPCellWithClip cell

hybrid_forward(F, inputs, states, i2h_weight, h2h_weight, h2r_weight, i2h_bias, h2h_bias)[source]

Hybrid forward computation for Long-Short Term Memory Projected network cell with cell clip and projection clip.

Parameters
  • inputs (input tensor with shape (batch_size, input_size).) –

  • states (a list of two initial recurrent state tensors, with shape) – (batch_size, projection_size) and (batch_size, hidden_size) respectively.

Returns

  • out (output tensor with shape (batch_size, num_hidden).)

  • next_states (a list of two output recurrent state tensors. Each has) – the same shape as states.

class gluonnlp.model.ELMoBiLM(rnn_type, output_size, filters, char_embed_size, char_vocab_size, num_highway, conv_layer_activation, max_chars_per_token, input_size, hidden_size, proj_size, num_layers, cell_clip, proj_clip, skip_connection=True, **kwargs)[source]

ELMo Bidirectional language model

Run a pre-trained bidirectional language model, outputting the weighted ELMo representation.

We implement the ELMo Bidirectional language model (BiLm) proposed in the following work:

@inproceedings{Peters:2018,
author={Peters, Matthew E. and  Neumann, Mark and Iyyer, Mohit and Gardner,
Matt and Clark, Christopher and Lee, Kenton and Zettlemoyer, Luke},
title={Deep contextualized word representations},
booktitle={Proc. of NAACL},
year={2018}
}
Parameters
  • rnn_type (str) – The type of RNN cell to use. The option for pre-trained models is ‘lstmpc’.

  • output_size (int) – The output dimension after conducting the convolutions and max pooling, and applying highways, as well as linear projection.

  • filters (list of tuple) – List of tuples representing the settings for convolution layers. Each element is (ngram_filter_size, num_filters).

  • char_embed_size (int) – The input dimension to the encoder.

  • char_vocab_size (int) – Size of character-level vocabulary.

  • num_highway (int) – The number of layers of the Highway layer.

  • conv_layer_activation (str) – Activation function to be used after convolutional layer.

  • max_chars_per_token (int) – The maximum number of characters of a token.

  • input_size (int) – The initial input size of in the RNN cell.

  • hidden_size (int) – The hidden size of the RNN cell.

  • proj_size (int) – The projection size of each LSTMPCellWithClip cell

  • num_layers (int) – The number of RNN cells.

  • cell_clip (float) – Clip cell state between [-cellclip, cell_clip] in LSTMPCellWithClip cell

  • proj_clip (float) – Clip projection between [-projclip, projclip] in LSTMPCellWithClip cell

  • skip_connection (bool) – Whether to add skip connections (add RNN cell input to output)

hybrid_forward(F, inputs, states=None, mask=None)[source]
Parameters
  • inputs (NDArray) – Shape (batch_size, sequence_length, max_character_per_token) of character ids representing the current batch.

  • states ((list of list of NDArray, list of list of NDArray)) – The states. First tuple element is the forward layer states, while the second is the states from backward layer. Each is a list of states for each layer. The state of each layer has a list of two initial tensors with shape (batch_size, proj_size) and (batch_size, hidden_size).

  • mask (NDArray) – Shape (batch_size, sequence_length) with sequence mask.

Returns

  • output (list of NDArray) – A list of activations at each layer of the network, each of shape (batch_size, sequence_length, embedding_size)

  • states ((list of list of NDArray, list of list of NDArray)) – The states. First tuple element is the forward layer states, while the second is the states from backward layer. Each is a list of states for each layer. The state of each layer has a list of two initial tensors with shape (batch_size, proj_size) and (batch_size, hidden_size).

class gluonnlp.model.ELMoCharacterEncoder(output_size, filters, char_embed_size, num_highway, conv_layer_activation, max_chars_per_token, char_vocab_size, **kwargs)[source]

ELMo character encoder

Compute context-free character-based token representation with character-level convolution.

This encoder has input character ids of shape (batch_size, sequence_length, max_character_per_word) and returns (batch_size, sequence_length, embedding_size).

Parameters
  • output_size (int) – The output dimension after conducting the convolutions and max pooling, and applying highways, as well as linear projection.

  • filters (list of tuple) – List of tuples representing the settings for convolution layers. Each element is (ngram_filter_size, num_filters).

  • char_embed_size (int) – The input dimension to the encoder.

  • num_highway (int) – The number of layers of the Highway layer.

  • conv_layer_activation (str) – Activation function to be used after convolutional layer.

  • max_chars_per_token (int) – The maximum number of characters of a token.

  • char_vocab_size (int) – Size of character-level vocabulary.

hybrid_forward(F, inputs)[source]

Compute context insensitive token embeddings for ELMo representations.

Parameters

inputs (NDArray) – Shape (batch_size, sequence_length, max_character_per_token) of character ids representing the current batch.

Returns

token_embedding – Shape (batch_size, sequence_length, embedding_size) with context insensitive token representations.

Return type

NDArray

gluonnlp.model.elmo_2x1024_128_2048cnn_1xhighway(dataset_name=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

ELMo 2-layer BiLSTM with 1024 hidden units, 128 projection size, 1 highway layer.

Parameters
  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. Options are ‘gbw’.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block

gluonnlp.model.elmo_2x2048_256_2048cnn_1xhighway(dataset_name=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

ELMo 2-layer BiLSTM with 2048 hidden units, 256 projection size, 1 highway layer.

Parameters
  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. Options are ‘gbw’.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block

gluonnlp.model.elmo_2x4096_512_2048cnn_2xhighway(dataset_name=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

ELMo 2-layer BiLSTM with 4096 hidden units, 512 projection size, 2 highway layer.

Parameters
  • dataset_name (str or None, default None) – The dataset name on which the pre-trained model is trained. Options are ‘gbw’ and ‘5bw’.

  • pretrained (bool, default False) – Whether to load the pre-trained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pre-trained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block

class gluonnlp.model.Seq2SeqEncoder(prefix=None, params=None)[source]

Base class of the encoders in sequence to sequence learning models.

forward(inputs, valid_length=None, states=None)[source]

Overrides to implement forward computation using NDArray. Only accepts positional arguments.

Parameters

*args (list of NDArray) – Input tensors.

class gluonnlp.model.TransformerEncoder(attention_cell='multi_head', num_layers=2, units=512, hidden_size=2048, max_length=50, num_heads=4, scaled=True, dropout=0.0, use_residual=True, output_attention=False, weight_initializer=None, bias_initializer='zeros', prefix=None, params=None)[source]

Structure of the Transformer Encoder.

Parameters
  • attention_cell (AttentionCell or str, default 'multi_head') – Arguments of the attention cell. Can be ‘multi_head’, ‘scaled_luong’, ‘scaled_dot’, ‘dot’, ‘cosine’, ‘normed_mlp’, ‘mlp’

  • num_layers (int) – Number of attention layers.

  • units (int) – Number of units for the output.

  • hidden_size (int) – number of units in the hidden layer of position-wise feed-forward networks

  • max_length (int) – Maximum length of the input sequence

  • num_heads (int) – Number of heads in multi-head attention

  • scaled (bool) – Whether to scale the softmax input by the sqrt of the input dimension in multi-head attention

  • dropout (float) – Dropout probability of the attention probabilities.

  • use_residual (bool) –

  • output_attention (bool) – Whether to output the attention weights

  • weight_initializer (str or Initializer) – Initializer for the input weights matrix, used for the linear transformation of the inputs.

  • bias_initializer (str or Initializer) – Initializer for the bias vector.

  • prefix (str, default None.) – Prefix for name of Block`s. (and name of weight if params is `None).

  • params (Parameter or None) – Container for weight sharing between cells. Created if None.

  • Inputs

    • inputs : input sequence of shape (batch_size, length, C_in)

    • states : list of tensors for initial states and masks.

    • valid_lengthvalid lengths of each sequence. Usually used when part of sequence

      has been padded. Shape is (batch_size, )

  • Outputs

    • outputs : the output of the encoder. Shape is (batch_size, length, C_out)

    • additional_outputslist of tensors.

      Either be an empty list or contains the attention weights in this step. The attention weights will have shape (batch_size, length, mem_length) or (batch_size, num_heads, length, mem_length)

class gluonnlp.model.PositionwiseFFN(units=512, hidden_size=2048, dropout=0.0, use_residual=True, ffn1_dropout=False, weight_initializer=None, bias_initializer='zeros', prefix=None, params=None, activation='relu', layer_norm_eps=None)[source]

Structure of the Positionwise Feed-Forward Neural Network for Transformer.

Computes the positionwise encoding of the inputs.

Parameters
  • units (int) – Number of units for the output

  • hidden_size (int) – Number of units in the hidden layer of position-wise feed-forward networks

  • dropout (float) – Dropout probability for the output

  • use_residual (bool) – Add residual connection between the input and the output

  • ffn1_dropout (bool, default False) – If True, apply dropout both after the first and second Positionwise Feed-Forward Neural Network layers. If False, only apply dropout after the second.

  • weight_initializer (str or Initializer) – Initializer for the input weights matrix, used for the linear transformation of the inputs.

  • bias_initializer (str or Initializer) – Initializer for the bias vector.

  • prefix (str, default None) – Prefix for name of Block`s (and name of weight if params is `None).

  • params (Parameter or None) – Container for weight sharing between cells. Created if None.

  • activation (str, default 'relu') – Activation methods in PositionwiseFFN

  • layer_norm_eps (float, default None) – Epsilon for layer_norm

  • Inputs

    • inputs : input sequence of shape (batch_size, length, C_in).

  • Outputs

    • outputs : output encoding of shape (batch_size, length, C_out).

class gluonnlp.model.TransformerEncoderCell(attention_cell='multi_head', units=128, hidden_size=512, num_heads=4, scaled=True, dropout=0.0, use_residual=True, output_attention=False, weight_initializer=None, bias_initializer='zeros', prefix=None, params=None, activation='relu', layer_norm_eps=None)[source]

Structure of the Transformer Encoder Cell.

Parameters
  • attention_cell (AttentionCell or str, default 'multi_head') – Arguments of the attention cell. Can be ‘multi_head’, ‘scaled_luong’, ‘scaled_dot’, ‘dot’, ‘cosine’, ‘normed_mlp’, ‘mlp’

  • units (int) – Number of units for the output

  • hidden_size (int) – number of units in the hidden layer of position-wise feed-forward networks

  • num_heads (int) – Number of heads in multi-head attention

  • scaled (bool) – Whether to scale the softmax input by the sqrt of the input dimension in multi-head attention

  • dropout (float) –

  • use_residual (bool) –

  • output_attention (bool) – Whether to output the attention weights

  • weight_initializer (str or Initializer) – Initializer for the input weights matrix, used for the linear transformation of the inputs.

  • bias_initializer (str or Initializer) – Initializer for the bias vector.

  • prefix (str, default None) – Prefix for name of Block`s. (and name of weight if params is `None).

  • params (Parameter or None) – Container for weight sharing between cells. Created if None.

  • activation (str, default None) – Activation methods in PositionwiseFFN

  • layer_norm_eps (float, default None) – Epsilon for layer_norm

  • Inputs

    • inputs : input sequence. Shape (batch_size, length, C_in)

    • mask : mask for inputs. Shape (batch_size, length, length)

  • Outputs

    • outputs: output tensor of the transformer encoder cell.

      Shape (batch_size, length, C_out)

    • additional_outputs: the additional output of all the transformer encoder cell.

gluonnlp.model.transformer_en_de_512(dataset_name=None, src_vocab=None, tgt_vocab=None, pretrained=False, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

Transformer pretrained model.

Embedding size is 400, and hidden layer size is 1150.

Parameters
  • dataset_name (str or None, default None) –

  • src_vocab (gluonnlp.Vocab or None, default None) –

  • tgt_vocab (gluonnlp.Vocab or None, default None) –

  • pretrained (bool, default False) – Whether to load the pretrained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pretrained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

Returns

Return type

gluon.Block, gluonnlp.Vocab, gluonnlp.Vocab

class gluonnlp.model.BERTModel(encoder, vocab_size=None, token_type_vocab_size=None, units=None, embed_size=None, embed_dropout=0.0, embed_initializer=None, word_embed=None, token_type_embed=None, use_pooler=True, use_decoder=True, use_classifier=True, use_token_type_embed=True, prefix=None, params=None)[source]

Generic Model for BERT (Bidirectional Encoder Representations from Transformers).

Parameters
  • encoder (BERTEncoder) – Bidirectional encoder that encodes the input sentence.

  • vocab_size (int or None, default None) – The size of the vocabulary.

  • token_type_vocab_size (int or None, default None) – The vocabulary size of token types (number of segments).

  • units (int or None, default None) – Number of units for the final pooler layer.

  • embed_size (int or None, default None) – Size of the embedding vectors. It is used to generate the word and token type embeddings if word_embed and token_type_embed are None.

  • embed_dropout (float, default 0.0) – Dropout rate of the embedding weights. It is used to generate the source and target embeddings if word_embed and token_type_embed are None.

  • embed_initializer (Initializer, default None) – Initializer of the embedding weights. It is used to generate the source and target embeddings if word_embed and token_type_embed are None.

  • word_embed (Block or None, default None) – The word embedding. If set to None, word_embed will be constructed using embed_size and embed_dropout.

  • token_type_embed (Block or None, default None) – The token type embedding (segment embedding). If set to None and the token_type_embed will be constructed using embed_size and embed_dropout.

  • use_pooler (bool, default True) – Whether to include the pooler which converts the encoded sequence tensor of shape (batch_size, seq_length, units) to a tensor of shape (batch_size, units) for segment level classification task.

  • use_decoder (bool, default True) – Whether to include the decoder for masked language model prediction.

  • use_classifier (bool, default True) – Whether to include the classifier for next sentence classification.

  • use_token_type_embed (bool, default True) – Whether to include token type embedding (segment embedding).

  • prefix (str or None) – See document of mx.gluon.Block.

  • params (ParameterDict or None) – See document of mx.gluon.Block.

  • Inputs

    • inputs: input sequence tensor, shape (batch_size, seq_length)

    • token_types: optional input token type tensor, shape (batch_size, seq_length).

      If the inputs contain two sequences, then the token type of the first sequence differs from that of the second one.

    • valid_length: optional tensor of input sequence valid lengths, shape (batch_size,)

    • masked_positions: optional tensor of position of tokens for masked LM decoding,

      shape (batch_size, num_masked_positions).

  • Outputs

    • sequence_outputs: Encoded sequence, which can be either a tensor of the last

      layer of the Encoder, or a list of all sequence encodings of all layers. In both cases shape of the tensor(s) is/are (batch_size, seq_length, units).

    • attention_outputs: output list of all intermediate encodings per layer

      Returned only if BERTEncoder.output_attention is True. List of num_layers length of tensors of shape (batch_size, num_attention_heads, seq_length, seq_length)

    • pooled_output: output tensor of pooled representation of the first tokens.

      Returned only if use_pooler is True. Shape (batch_size, units)

    • next_sentence_classifier_output: output tensor of next sentence classification.

      Returned only if use_classifier is True. Shape (batch_size, 2)

    • masked_lm_outputs: output tensor of sequence decoding for masked language model

      prediction. Returned only if use_decoder True. Shape (batch_size, num_masked_positions, vocab_size)

hybrid_forward(F, inputs, token_types, valid_length=None, masked_positions=None)[source]

Generate the representation given the inputs.

This is used in training or fine-tuning a BERT model.

class gluonnlp.model.RoBERTaModel(encoder, vocab_size=None, units=None, embed_size=None, embed_dropout=0.0, embed_initializer=None, word_embed=None, use_decoder=True, prefix=None, params=None)[source]

Generic Model for BERT (Bidirectional Encoder Representations from Transformers).

Parameters
  • encoder (BERTEncoder) – Bidirectional encoder that encodes the input sentence.

  • vocab_size (int or None, default None) – The size of the vocabulary.

  • units (int or None, default None) – Number of units for the final pooler layer.

  • embed_size (int or None, default None) – Size of the embedding vectors. It is used to generate the word and token type embeddings if word_embed and token_type_embed are None.

  • embed_dropout (float, default 0.0) – Dropout rate of the embedding weights. It is used to generate the source and target embeddings if word_embed and token_type_embed are None.

  • embed_initializer (Initializer, default None) – Initializer of the embedding weights. It is used to generate the source and target embeddings if word_embed and token_type_embed are None.

  • word_embed (Block or None, default None) – The word embedding. If set to None, word_embed will be constructed using embed_size and embed_dropout.

  • use_decoder (bool, default True) – Whether to include the decoder for masked language model prediction.

  • prefix (str or None) – See document of mx.gluon.Block.

  • params (ParameterDict or None) – See document of mx.gluon.Block.

  • Inputs

    • inputs: input sequence tensor, shape (batch_size, seq_length)

    • valid_length: optional tensor of input sequence valid lengths, shape (batch_size,)

    • masked_positions: optional tensor of position of tokens for masked LM decoding,

      shape (batch_size, num_masked_positions).

  • Outputs

    • sequence_outputs: Encoded sequence, which can be either a tensor of the last

      layer of the Encoder, or a list of all sequence encodings of all layers. In both cases shape of the tensor(s) is/are (batch_size, seq_length, units).

    • attention_outputs: output list of all intermediate encodings per layer

      Returned only if BERTEncoder.output_attention is True. List of num_layers length of tensors of shape (num_masks, num_attention_heads, seq_length, seq_length)

    • masked_lm_outputs: output tensor of sequence decoding for masked language model

      prediction. Returned only if use_decoder True. Shape (batch_size, num_masked_positions, vocab_size)

class gluonnlp.model.BERTEncoder(attention_cell='multi_head', num_layers=2, units=512, hidden_size=2048, max_length=50, num_heads=4, scaled=True, dropout=0.0, use_residual=True, output_attention=False, output_all_encodings=False, weight_initializer=None, bias_initializer='zeros', prefix=None, params=None, activation='gelu', layer_norm_eps=None)[source]

Structure of the BERT Encoder.

Different from the original encoder for transformer, BERTEncoder uses learnable positional embedding, BERTPositionwiseFFN and BERTLayerNorm.

Parameters
  • attention_cell (AttentionCell or str, default 'multi_head') – Arguments of the attention cell. Can be ‘multi_head’, ‘scaled_luong’, ‘scaled_dot’, ‘dot’, ‘cosine’, ‘normed_mlp’, ‘mlp’

  • num_layers (int) – Number of attention layers.

  • units (int) – Number of units for the output.

  • hidden_size (int) – number of units in the hidden layer of position-wise feed-forward networks

  • max_length (int) – Maximum length of the input sequence

  • num_heads (int) – Number of heads in multi-head attention

  • scaled (bool) – Whether to scale the softmax input by the sqrt of the input dimension in multi-head attention

  • dropout (float) – Dropout probability of the attention probabilities.

  • use_residual (bool) –

  • output_attention (bool, default False) – Whether to output the attention weights

  • output_all_encodings (bool, default False) – Whether to output encodings of all encoder cells

  • weight_initializer (str or Initializer) – Initializer for the input weights matrix, used for the linear transformation of the inputs.

  • bias_initializer (str or Initializer) – Initializer for the bias vector.

  • prefix (str, default None.) – Prefix for name of Block`s. (and name of weight if params is `None).

  • params (Parameter or None) – Container for weight sharing between cells. Created if None.

  • activation (str, default 'gelu') – Activation methods in PositionwiseFFN

  • layer_norm_eps (float, default None) – Epsilon for layer_norm

  • Inputs

    • inputs : input sequence of shape (batch_size, length, C_in)

    • states : list of tensors for initial states and masks.

    • valid_lengthvalid lengths of each sequence. Usually used when part of sequence

      has been padded. Shape is (batch_size, )

  • Outputs

    • outputs : the output of the encoder. Shape is (batch_size, length, C_out)

    • additional_outputslist of tensors.

      Either be an empty list or contains the attention weights in this step. The attention weights will have shape (batch_size, num_heads, length, mem_length)

class gluonnlp.model.BERTEncoderCell(attention_cell='multi_head', units=128, hidden_size=512, num_heads=4, scaled=True, dropout=0.0, use_residual=True, output_attention=False, weight_initializer=None, bias_initializer='zeros', prefix=None, params=None, activation='gelu', layer_norm_eps=None)[source]

Structure of the Transformer Encoder Cell for BERT.

Different from the original encoder cell for transformer, BERTEncoderCell adds bias terms for attention and the projection on attention output. It also uses BERTPositionwiseFFN and BERTLayerNorm.

Parameters
  • attention_cell (AttentionCell or str, default 'multi_head') – Arguments of the attention cell. Can be ‘multi_head’, ‘scaled_luong’, ‘scaled_dot’, ‘dot’, ‘cosine’, ‘normed_mlp’, ‘mlp’

  • units (int) – Number of units for the output

  • hidden_size (int) – number of units in the hidden layer of position-wise feed-forward networks

  • num_heads (int) – Number of heads in multi-head attention

  • scaled (bool) – Whether to scale the softmax input by the sqrt of the input dimension in multi-head attention

  • dropout (float) –

  • use_residual (bool) –

  • output_attention (bool) – Whether to output the attention weights

  • weight_initializer (str or Initializer) – Initializer for the input weights matrix, used for the linear transformation of the inputs.

  • bias_initializer (str or Initializer) – Initializer for the bias vector.

  • prefix (str, default None) – Prefix for name of Block`s. (and name of weight if params is `None).

  • params (Parameter or None) – Container for weight sharing between cells. Created if None.

  • activation (str, default 'gelu') – Activation methods in PositionwiseFFN

  • layer_norm_eps (float, default None) – Epsilon for layer_norm

  • Inputs

    • inputs : input sequence. Shape (batch_size, length, C_in)

    • mask : mask for inputs. Shape (batch_size, length, length)

  • Outputs

    • outputs: output tensor of the transformer encoder cell.

      Shape (batch_size, length, C_out)

    • additional_outputs: the additional output of all the transformer encoder cell.

class gluonnlp.model.BERTPositionwiseFFN(units=512, hidden_size=2048, dropout=0.0, use_residual=True, weight_initializer=None, bias_initializer='zeros', prefix=None, params=None, activation='gelu', layer_norm_eps=None)[source]

Structure of the Positionwise Feed-Forward Neural Network for BERT.

Different from the original positionwise feed forward network for transformer, BERTPositionwiseFFN uses GELU for activation and BERTLayerNorm for layer normalization.

Parameters
  • units (int) – Number of units for the output

  • hidden_size (int) – Number of units in the hidden layer of position-wise feed-forward networks

  • dropout (float) – Dropout probability for the output

  • use_residual (bool) – Add residual connection between the input and the output

  • weight_initializer (str or Initializer) – Initializer for the input weights matrix, used for the linear transformation of the inputs.

  • bias_initializer (str or Initializer) – Initializer for the bias vector.

  • prefix (str, default None) – Prefix for name of Block`s (and name of weight if params is `None).

  • params (Parameter or None) – Container for weight sharing between cells. Created if None.

  • activation (str, default 'gelu') – Activation methods in PositionwiseFFN

  • layer_norm_eps (float, default None) – Epsilon for layer_norm

  • Inputs

    • inputs : input sequence of shape (batch_size, length, C_in).

  • Outputs

    • outputs : output encoding of shape (batch_size, length, C_out).

class gluonnlp.model.BERTLayerNorm(epsilon=1e-12, in_channels=0, prefix=None, params=None)[source]

BERT style Layer Normalization.

Epsilon is added inside the square root and set to 1e-12 by default.

Inputs:
  • data: input tensor with arbitrary shape.

  • out: output tensor with the same shape as data.

hybrid_forward(F, data, gamma, beta)[source]

forward computation.

class gluonnlp.model.BERTClassifier(bert, num_classes=2, dropout=0.0, prefix=None, params=None)[source]

Model for sentence (pair) classification task with BERT.

The model feeds token ids and token type ids into BERT to get the pooled BERT sequence representation, then apply a Dense layer for classification.

Parameters
  • bert (BERTModel) – Bidirectional encoder with transformer.

  • num_classes (int, default is 2) – The number of target classes.

  • dropout (float or None, default 0.0.) – Dropout probability for the bert output.

  • prefix (str or None) – See document of mx.gluon.Block.

  • params (ParameterDict or None) – See document of mx.gluon.Block.

hybrid_forward(F, inputs, token_types, valid_length=None)[source]

Generate the unnormalized score for the given the input sequences.

Parameters
  • inputs (NDArray or Symbol, shape (batch_size, seq_length)) – Input words for the sequences.

  • token_types (NDArray or Symbol, shape (batch_size, seq_length)) – Token types for the sequences, used to indicate whether the word belongs to the first sentence or the second one.

  • valid_length (NDArray or None, shape (batch_size)) – Valid length of the sequence. This is used to mask the padded tokens.

Returns

outputs – Shape (batch_size, num_classes)

Return type

NDArray

class gluonnlp.model.RoBERTaClassifier(roberta, num_classes=2, dropout=0.0, prefix=None, params=None)[source]

Model for sentence (pair) classification task with BERT.

The model feeds token ids and token type ids into BERT to get the pooled BERT sequence representation, then apply a Dense layer for classification.

Parameters
  • bert (RoBERTaModel) – The RoBERTa model.

  • num_classes (int, default is 2) – The number of target classes.

  • dropout (float or None, default 0.0.) – Dropout probability for the bert output.

  • prefix (str or None) – See document of mx.gluon.Block.

  • params (ParameterDict or None) – See document of mx.gluon.Block.

  • Inputs

    • inputs: input sequence tensor, shape (batch_size, seq_length)

    • valid_length: optional tensor of input sequence valid lengths.

      Shape (batch_size, num_classes).

  • Outputs

    • output: Regression output, shape (batch_size, num_classes)

hybrid_forward(F, inputs, valid_length=None)[source]

Generate the unnormalized score for the given the input sequences.

Parameters
  • inputs (NDArray or Symbol, shape (batch_size, seq_length)) – Input words for the sequences.

  • valid_length (NDArray or Symbol, or None, shape (batch_size)) – Valid length of the sequence. This is used to mask the padded tokens.

Returns

outputs – Shape (batch_size, num_classes)

Return type

NDArray or Symbol

gluonnlp.model.bert_12_768_12(dataset_name=None, vocab=None, pretrained=True, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', use_pooler=True, use_decoder=True, use_classifier=True, pretrained_allow_missing=False, **kwargs)[source]

Generic BERT BASE model.

The number of layers (L) is 12, number of units (H) is 768, and the number of self-attention heads (A) is 12.

Parameters
  • dataset_name (str or None, default None) – If not None, the dataset name is used to load a vocabulary for the dataset. If the pretrained argument is set to True, the dataset name is further used to select the pretrained parameters to load. The supported datasets are ‘book_corpus_wiki_en_cased’, ‘book_corpus_wiki_en_uncased’, ‘wiki_cn_cased’, ‘openwebtext_book_corpus_wiki_en_uncased’, ‘wiki_multilingual_uncased’, ‘wiki_multilingual_cased’, ‘scibert_scivocab_uncased’, ‘scibert_scivocab_cased’, ‘scibert_basevocab_uncased’, ‘scibert_basevocab_cased’, ‘biobert_v1.0_pmc’, ‘biobert_v1.0_pubmed’, ‘biobert_v1.0_pubmed_pmc’, ‘biobert_v1.1_pubmed’, ‘clinicalbert’

  • vocab (gluonnlp.vocab.BERTVocab or None, default None) – Vocabulary for the dataset. Must be provided if dataset_name is not specified. Ignored if dataset_name is specified.

  • pretrained (bool, default True) – Whether to load the pretrained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pretrained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

  • use_pooler (bool, default True) – Whether to include the pooler which converts the encoded sequence tensor of shape (batch_size, seq_length, units) to a tensor of shape (batch_size, units) for for segment level classification task.

  • use_decoder (bool, default True) – Whether to include the decoder for masked language model prediction. Note that ‘biobert_v1.0_pmc’, ‘biobert_v1.0_pubmed’, ‘biobert_v1.0_pubmed_pmc’, ‘biobert_v1.1_pubmed’, ‘clinicalbert’ do not include these parameters.

  • use_classifier (bool, default True) – Whether to include the classifier for next sentence classification. Note that ‘biobert_v1.0_pmc’, ‘biobert_v1.0_pubmed’, ‘biobert_v1.0_pubmed_pmc’, ‘biobert_v1.1_pubmed’ do not include these parameters.

  • pretrained_allow_missing (bool, default False) – Whether to ignore if any parameters for the BERTModel are missing in the pretrained weights for model. Some BERTModels for example do not provide decoder or classifier weights. In that case it is still possible to construct a BERTModel with use_decoder=True and/or use_classifier=True, but the respective parameters will be missing from the pretrained file. If pretrained_allow_missing=True, this will be ignored and the parameters will be left uninitialized. Otherwise AssertionError is raised.

  • pretrained parameters for dataset_name (The) –

  • were obtained by running the ('openwebtext_book_corpus_wiki_en_uncased') –

  • BERT pre-training script on OpenWebText. (GluonNLP) –

  • pretrained parameters for dataset_name 'scibert_scivocab_uncased', (The) –

  • 'scibert_basevocab_uncased', ('scibert_scivocab_cased',) –

  • were obtained by converting the parameters ('scibert_basevocab_cased') –

  • by "Beltagy, I., Cohan, A., & Lo, K. (2019) Scibert (published) –

  • embeddings for scientific text. arXiv preprint (contextualized) –

  • arXiv (1903.10676.") –

  • pretrained parameters for dataset_name 'biobert_v1.0_pmc', (The) –

  • 'biobert_v1.0_pubmed_pmc', 'biobert_v1.1_pubmed' ('biobert_v1.0_pubmed',) –

  • obtained by converting the parameters published by "Lee, J., Yoon, W., (were) –

  • S., Kim, D., Kim, S., So, C. H., & Kang, J. (2019) Biobert (Kim,) –

  • biomedical language representation model for biomedical text (pre-trained) –

  • arXiv preprint arXiv (mining.) –

  • pretrained parameters for dataset_name 'clinicalbert' were obtained by (The) –

  • the parameters published by "Huang, K., Altosaar, J., & (converting) –

  • R. (2019) ClinicalBERT (Ranganath,) –

  • Readmission. arXiv preprint arXiv (Hospital) –

Returns

Return type

BERTModel, gluonnlp.vocab.BERTVocab

gluonnlp.model.bert_24_1024_16(dataset_name=None, vocab=None, pretrained=True, ctx=cpu(0), use_pooler=True, use_decoder=True, use_classifier=True, root='/var/lib/jenkins/.mxnet/models', pretrained_allow_missing=False, **kwargs)[source]

Generic BERT LARGE model.

The number of layers (L) is 24, number of units (H) is 1024, and the number of self-attention heads (A) is 16.

Parameters
  • dataset_name (str or None, default None) – If not None, the dataset name is used to load a vocabulary for the dataset. If the pretrained argument is set to True, the dataset name is further used to select the pretrained parameters to load. Options include ‘book_corpus_wiki_en_uncased’ and ‘book_corpus_wiki_en_cased’.

  • vocab (gluonnlp.vocab.BERTVocab or None, default None) – Vocabulary for the dataset. Must be provided if dataset_name is not specified. Ignored if dataset_name is specified.

  • pretrained (bool, default True) – Whether to load the pretrained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pretrained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

  • use_pooler (bool, default True) – Whether to include the pooler which converts the encoded sequence tensor of shape (batch_size, seq_length, units) to a tensor of shape (batch_size, units) for for segment level classification task.

  • use_decoder (bool, default True) – Whether to include the decoder for masked language model prediction.

  • use_classifier (bool, default True) – Whether to include the classifier for next sentence classification.

  • pretrained_allow_missing (bool, default False) – Whether to ignore if any parameters for the BERTModel are missing in the pretrained weights for model. Some BERTModels for example do not provide decoder or classifier weights. In that case it is still possible to construct a BERTModel with use_decoder=True and/or use_classifier=True, but the respective parameters will be missing from the pretrained file. If pretrained_allow_missing=True, this will be ignored and the parameters will be left uninitialized. Otherwise AssertionError is raised.

Returns

Return type

BERTModel, gluonnlp.vocab.BERTVocab

gluonnlp.model.ernie_12_768_12(dataset_name=None, vocab=None, pretrained=True, ctx=cpu(0), root='/var/lib/jenkins/.mxnet/models', use_pooler=True, use_decoder=True, use_classifier=True, **kwargs)[source]

Baidu ERNIE model.

Reference: https://arxiv.org/pdf/1904.09223.pdf

The number of layers (L) is 12, number of units (H) is 768, and the number of self-attention heads (A) is 12.

Parameters
  • dataset_name (str or None, default None) – If not None, the dataset name is used to load a vocabulary for the dataset. If the pretrained argument is set to True, the dataset name is further used to select the pretrained parameters to load. The supported datasets are ‘baidu_ernie’

  • vocab (gluonnlp.vocab.BERTVocab or None, default None) – Vocabulary for the dataset. Must be provided if dataset_name is not specified. Ignored if dataset_name is specified.

  • pretrained (bool, default True) – Whether to load the pretrained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pretrained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

  • use_pooler (bool, default True) – Whether to include the pooler which converts the encoded sequence tensor of shape (batch_size, seq_length, units) to a tensor of shape (batch_size, units) for for segment level classification task.

  • use_decoder (bool, default True) – Whether to include the decoder for masked language model prediction.

  • use_classifier (bool, default True) – Whether to include the classifier for next sentence classification.

Returns

Return type

(BERTModel, gluonnlp.vocab.BERTVocab)

gluonnlp.model.roberta_12_768_12(dataset_name=None, vocab=None, pretrained=True, ctx=cpu(0), use_decoder=True, root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

Generic RoBERTa BASE model.

The number of layers (L) is 12, number of units (H) is 768, and the number of self-attention heads (A) is 12.

Parameters
  • dataset_name (str or None, default None) – If not None, the dataset name is used to load a vocabulary for the dataset. If the pretrained argument is set to True, the dataset name is further used to select the pretrained parameters to load. Options include ‘book_corpus_wiki_en_uncased’ and ‘book_corpus_wiki_en_cased’.

  • vocab (gluonnlp.vocab.Vocab or None, default None) – Vocabulary for the dataset. Must be provided if dataset_name is not specified. Ignored if dataset_name is specified.

  • pretrained (bool, default True) – Whether to load the pretrained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pretrained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

  • use_decoder (bool, default True) – Whether to include the decoder for masked language model prediction.

Returns

Return type

RoBERTaModel, gluonnlp.vocab.Vocab

gluonnlp.model.roberta_24_1024_16(dataset_name=None, vocab=None, pretrained=True, ctx=cpu(0), use_decoder=True, root='/var/lib/jenkins/.mxnet/models', **kwargs)[source]

Generic RoBERTa LARGE model.

The number of layers (L) is 24, number of units (H) is 1024, and the number of self-attention heads (A) is 16.

Parameters
  • dataset_name (str or None, default None) – If not None, the dataset name is used to load a vocabulary for the dataset. If the pretrained argument is set to True, the dataset name is further used to select the pretrained parameters to load. Options include ‘book_corpus_wiki_en_uncased’ and ‘book_corpus_wiki_en_cased’.

  • vocab (gluonnlp.vocab.Vocab or None, default None) – Vocabulary for the dataset. Must be provided if dataset_name is not specified. Ignored if dataset_name is specified.

  • pretrained (bool, default True) – Whether to load the pretrained weights for model.

  • ctx (Context, default CPU) – The context in which to load the pretrained weights.

  • root (str, default '$MXNET_HOME/models') – Location for keeping the model parameters. MXNET_HOME defaults to ‘~/.mxnet’.

  • use_decoder (bool, default True) – Whether to include the decoder for masked language model prediction.

Returns

Return type

RoBERTaModel, gluonnlp.vocab.Vocab