encoder#

class BoWEncoder(emb_dim)[source]#

Bases: Layer

A BoWEncoder takes as input a sequence of vectors and returns a single vector, which simply sums the embeddings of a sequence across the time dimension. The input to this encoder is of shape (batch_size, num_tokens, emb_dim), and the output is of shape (batch_size, emb_dim).

Parameters:

emb_dim (int) – The dimension of each vector in the input sequence.

Example

import paddle
import paddle.nn as nn
import paddlenlp as nlp

class BoWModel(nn.Layer):
    def __init__(self,
                vocab_size,
                num_classes,
                emb_dim=128,
                padding_idx=0,
                hidden_size=128,
                fc_hidden_size=96):
        super().__init__()
        self.embedder = nn.Embedding(
            vocab_size, emb_dim, padding_idx=padding_idx)
        self.bow_encoder = nlp.seq2vec.BoWEncoder(emb_dim)
        self.fc1 = nn.Linear(self.bow_encoder.get_output_dim(), hidden_size)
        self.fc2 = nn.Linear(hidden_size, fc_hidden_size)
        self.output_layer = nn.Linear(fc_hidden_size, num_classes)

    def forward(self, text):
        # Shape: (batch_size, num_tokens, embedding_dim)
        embedded_text = self.embedder(text)

        # Shape: (batch_size, embedding_dim)
        summed = self.bow_encoder(embedded_text)
        encoded_text = paddle.tanh(summed)

        # Shape: (batch_size, hidden_size)
        fc1_out = paddle.tanh(self.fc1(encoded_text))
        # Shape: (batch_size, fc_hidden_size)
        fc2_out = paddle.tanh(self.fc2(fc1_out))
        # Shape: (batch_size, num_classes)
        logits = self.output_layer(fc2_out)
        return logits

model = BoWModel(vocab_size=100, num_classes=2)

text = paddle.randint(low=1, high=10, shape=[1,10], dtype='int32')
logits = model(text)
get_input_dim()[source]#

Returns the dimension of the vector input for each element in the sequence input to a BoWEncoder. This is not the shape of the input tensor, but the last element of that shape.

get_output_dim()[source]#

Returns the dimension of the final vector output by this BoWEncoder. This is not the shape of the returned tensor, but the last element of that shape.

forward(inputs, mask=None)[source]#

It simply sums the embeddings of a sequence across the time dimension.

Parameters:
  • inputs (Tensor) – Shape as (batch_size, num_tokens, emb_dim) and dtype as float32 or float64. The sequence length of the input sequence.

  • mask (Tensor, optional) – Shape same as inputs. Its each elements identify whether the corresponding input token is padding or not. If True, not padding token. If False, padding token. Defaults to None.

Returns:

Returns tensor summed, the result vector of BagOfEmbedding. Its data type is same as inputs and its shape is [batch_size, emb_dim].

Return type:

Tensor

class CNNEncoder(emb_dim, num_filter, ngram_filter_sizes=(2, 3, 4, 5), conv_layer_activation=Tanh(), output_dim=None, **kwargs)[source]#

Bases: Layer

A CNNEncoder takes as input a sequence of vectors and returns a single vector, a combination of multiple convolution layers and max pooling layers. The input to this encoder is of shape (batch_size, num_tokens, emb_dim), and the output is of shape (batch_size, output_dim) or (batch_size, len(ngram_filter_sizes) * num_filter).

The CNN has one convolution layer for each ngram filter size. Each convolution operation gives out a vector of size num_filter. The number of times a convolution layer will be used is num_tokens - ngram_size + 1. The corresponding maxpooling layer aggregates all these outputs from the convolution layer and outputs the max.

This operation is repeated for every ngram size passed, and consequently the dimensionality of the output after maxpooling is len(ngram_filter_sizes) * num_filter. This then gets (optionally) projected down to a lower dimensional output, specified by output_dim.

We then use a fully connected layer to project in back to the desired output_dim. For more details, refer to A Sensitivity Analysis of (and Practitioners’ Guide to) Convolutional Neural Networks for Sentence Classification , Zhang and Wallace 2016, particularly Figure 1.

Parameters:
  • emb_dim (int) – The dimension of each vector in the input sequence.

  • num_filter (int) – This is the output dim for each convolutional layer, which is the number of “filters” learned by that layer.

  • ngram_filter_sizes (Tuple[int], optional) – This specifies both the number of convolutional layers we will create and their sizes. The default of (2, 3, 4, 5) will have four convolutional layers, corresponding to encoding ngrams of size 2 to 5 with some number of filters.

  • conv_layer_activation (Layer, optional) – Activation to use after the convolution layers. Defaults to paddle.nn.Tanh().

  • output_dim (int, optional) – After doing convolutions and pooling, we’ll project the collected features into a vector of this size. If this value is None, we will just return the result of the max pooling, giving an output of shape len(ngram_filter_sizes) * num_filter. Defaults to None.

Example

import paddle
import paddle.nn as nn
import paddlenlp as nlp

class CNNModel(nn.Layer):
    def __init__(self,
                vocab_size,
                num_classes,
                emb_dim=128,
                padding_idx=0,
                num_filter=128,
                ngram_filter_sizes=(3, ),
                fc_hidden_size=96):
        super().__init__()
        self.embedder = nn.Embedding(
            vocab_size, emb_dim, padding_idx=padding_idx)
        self.encoder = nlp.seq2vec.CNNEncoder(
            emb_dim=emb_dim,
            num_filter=num_filter,
            ngram_filter_sizes=ngram_filter_sizes)
        self.fc = nn.Linear(self.encoder.get_output_dim(), fc_hidden_size)
        self.output_layer = nn.Linear(fc_hidden_size, num_classes)

    def forward(self, text):
        # Shape: (batch_size, num_tokens, embedding_dim)
        embedded_text = self.embedder(text)
        # Shape: (batch_size, len(ngram_filter_sizes)*num_filter)
        encoder_out = self.encoder(embedded_text)
        encoder_out = paddle.tanh(encoder_out)
        # Shape: (batch_size, fc_hidden_size)
        fc_out = self.fc(encoder_out)
        # Shape: (batch_size, num_classes)
        logits = self.output_layer(fc_out)
        return logits

model = CNNModel(vocab_size=100, num_classes=2)

text = paddle.randint(low=1, high=10, shape=[1,10], dtype='int32')
logits = model(text)
get_input_dim()[source]#

Returns the dimension of the vector input for each element in the sequence input to a CNNEncoder. This is not the shape of the input tensor, but the last element of that shape.

get_output_dim()[source]#

Returns the dimension of the final vector output by this CNNEncoder. This is not the shape of the returned tensor, but the last element of that shape.

forward(inputs, mask=None)[source]#

The combination of multiple convolution layers and max pooling layers.

Parameters:
  • inputs (Tensor) – Shape as (batch_size, num_tokens, emb_dim) and dtype as float32 or float64. Tensor containing the features of the input sequence.

  • mask (Tensor, optional) – Shape should be same as inputs and dtype as int32, int64, float32 or float64. Its each elements identify whether the corresponding input token is padding or not. If True, not padding token. If False, padding token. Defaults to None.

Returns:

Returns tensor result. If output_dim is None, the result shape is of (batch_size, output_dim) and dtype is float; If not, the result shape is of (batch_size, len(ngram_filter_sizes) * num_filter).

Return type:

Tensor

class GRUEncoder(input_size, hidden_size, num_layers=1, direction='forward', dropout=0.0, pooling_type=None, **kwargs)[source]#

Bases: Layer

A GRUEncoder takes as input a sequence of vectors and returns a single vector, which is a combination of multiple paddle.nn.GRU subclass. The input to this encoder is of shape (batch_size, num_tokens, input_size), The output is of shape (batch_size, hidden_size * 2) if GRU is bidirection; If not, output is of shape (batch_size, hidden_size).

Paddle’s GRU have two outputs: the hidden state for every time step at last layer, and the hidden state at the last time step for every layer. If pooling_type is not None, we perform the pooling on the hidden state of every time step at last layer to create a single vector. If None, we use the hidden state of the last time step at last layer as a single output (shape of (batch_size, hidden_size)); And if direction is bidirection, the we concat the hidden state of the last forward gru and backward gru layer to create a single vector (shape of (batch_size, hidden_size * 2)).

Parameters:
  • input_size (int) – The number of expected features in the input (the last dimension).

  • hidden_size (int) – The number of features in the hidden state.

  • num_layers (int, optional) – Number of recurrent layers. E.g., setting num_layers=2 would mean stacking two GRUs together to form a stacked GRU, with the second GRU taking in outputs of the first GRU and computing the final results. Defaults to 1.

  • direction (str, optional) – The direction of the network. It can be “forward” and “bidirect” (it means bidirection network). If “bidirect”, it is a bidirectional GRU, and returns the concat output from both directions. Defaults to “forward”.

  • dropout (float, optional) – If non-zero, introduces a Dropout layer on the outputs of each GRU layer except the last layer, with dropout probability equal to dropout. Defaults to 0.0.

  • pooling_type (str, optional) – If pooling_type is None, then the GRUEncoder will return the hidden state of the last time step at last layer as a single vector. If pooling_type is not None, it must be one of “sum”, “max” and “mean”. Then it will be pooled on the GRU output (the hidden state of every time step at last layer) to create a single vector. Defaults to None

Example

import paddle
import paddle.nn as nn
import paddlenlp as nlp

class GRUModel(nn.Layer):
    def __init__(self,
                vocab_size,
                num_classes,
                emb_dim=128,
                padding_idx=0,
                gru_hidden_size=198,
                direction='forward',
                gru_layers=1,
                dropout_rate=0.0,
                pooling_type=None,
                fc_hidden_size=96):
        super().__init__()
        self.embedder = nn.Embedding(
            num_embeddings=vocab_size,
            embedding_dim=emb_dim,
            padding_idx=padding_idx)
        self.gru_encoder = nlp.seq2vec.GRUEncoder(
            emb_dim,
            gru_hidden_size,
            num_layers=gru_layers,
            direction=direction,
            dropout=dropout_rate,
            pooling_type=pooling_type)
        self.fc = nn.Linear(self.gru_encoder.get_output_dim(), fc_hidden_size)
        self.output_layer = nn.Linear(fc_hidden_size, num_classes)

    def forward(self, text, seq_len):
        # Shape: (batch_size, num_tokens, embedding_dim)
        embedded_text = self.embedder(text)
        # Shape: (batch_size, num_tokens, num_directions*gru_hidden_size)
        # num_directions = 2 if direction is 'bidirect'
        # if not, num_directions = 1
        text_repr = self.gru_encoder(embedded_text, sequence_length=seq_len)
        # Shape: (batch_size, fc_hidden_size)
        fc_out = paddle.tanh(self.fc(text_repr))
        # Shape: (batch_size, num_classes)
        logits = self.output_layer(fc_out)
        return logits

model = GRUModel(vocab_size=100, num_classes=2)

text = paddle.randint(low=1, high=10, shape=[1,10], dtype='int32')
seq_len = paddle.to_tensor([10])
logits = model(text, seq_len)
get_input_dim()[source]#

Returns the dimension of the vector input for each element in the sequence input to a GRUEncoder. This is not the shape of the input tensor, but the last element of that shape.

get_output_dim()[source]#

Returns the dimension of the final vector output by this GRUEncoder. This is not the shape of the returned tensor, but the last element of that shape.

forward(inputs, sequence_length)[source]#

GRUEncoder takes the a sequence of vectors and returns a single vector, which is a combination of multiple GRU layers. The input to this encoder is of shape (batch_size, num_tokens, input_size), The output is of shape (batch_size, hidden_size * 2) if GRU is bidirection; If not, output is of shape (batch_size, hidden_size).

Parameters:
  • inputs (Tensor) – Shape as (batch_size, num_tokens, input_size). Tensor containing the features of the input sequence.

  • sequence_length (Tensor) – Shape as (batch_size). The sequence length of the input sequence.

Returns:

Returns tensor output, the hidden state at the last time step for every layer. Its data type is float and its shape is [batch_size, hidden_size].

Return type:

Tensor

class LSTMEncoder(input_size, hidden_size, num_layers=1, direction='forward', dropout=0.0, pooling_type=None, **kwargs)[source]#

Bases: Layer

An LSTMEncoder takes as input a sequence of vectors and returns a single vector, which is a combination of multiple paddle.nn.LSTM subclass. The input to this encoder is of shape (batch_size, num_tokens, input_size). The output is of shape (batch_size, hidden_size * 2) if LSTM is bidirection; If not, output is of shape (batch_size, hidden_size).

Paddle’s LSTM have two outputs: the hidden state for every time step at last layer, and the hidden state and cell at the last time step for every layer. If pooling_type is not None, we perform the pooling on the hidden state of every time step at last layer to create a single vector. If None, we use the hidden state of the last time step at last layer as a single output (shape of (batch_size, hidden_size)); And if direction is bidirection, the we concat the hidden state of the last forward lstm and backward lstm layer to create a single vector (shape of (batch_size, hidden_size * 2)).

Parameters:
  • input_size (int) – The number of expected features in the input (the last dimension).

  • hidden_size (int) – The number of features in the hidden state.

  • num_layers (int, optional) – Number of recurrent layers. E.g., setting num_layers=2 would mean stacking two LSTMs together to form a stacked LSTM, with the second LSTM taking in outputs of the first LSTM and computing the final results. Defaults to 1.

  • direction (str, optional) – The direction of the network. It can be “forward” or “bidirect” (it means bidirection network). If “bidirect”, it is a bidirectional LSTM, and returns the concat output from both directions. Defaults to “forward”.

  • dropout (float, optional) – If non-zero, introduces a Dropout layer on the outputs of each LSTM layer except the last layer, with dropout probability equal to dropout. Defaults to 0.0 .

  • pooling_type (str, optional) – If pooling_type is None, then the LSTMEncoder will return the hidden state of the last time step at last layer as a single vector. If pooling_type is not None, it must be one of “sum”, “max” and “mean”. Then it will be pooled on the LSTM output (the hidden state of every time step at last layer) to create a single vector. Defaults to None.

Example

import paddle
import paddle.nn as nn
import paddlenlp as nlp

class LSTMModel(nn.Layer):
    def __init__(self,
                vocab_size,
                num_classes,
                emb_dim=128,
                padding_idx=0,
                lstm_hidden_size=198,
                direction='forward',
                lstm_layers=1,
                dropout_rate=0.0,
                pooling_type=None,
                fc_hidden_size=96):
        super().__init__()
        self.embedder = nn.Embedding(
            num_embeddings=vocab_size,
            embedding_dim=emb_dim,
            padding_idx=padding_idx)
        self.lstm_encoder = nlp.seq2vec.LSTMEncoder(
            emb_dim,
            lstm_hidden_size,
            num_layers=lstm_layers,
            direction=direction,
            dropout=dropout_rate,
            pooling_type=pooling_type)
        self.fc = nn.Linear(self.lstm_encoder.get_output_dim(), fc_hidden_size)
        self.output_layer = nn.Linear(fc_hidden_size, num_classes)

    def forward(self, text, seq_len):
        # Shape: (batch_size, num_tokens, embedding_dim)
        embedded_text = self.embedder(text)
        # Shape: (batch_size, num_tokens, num_directions*lstm_hidden_size)
        # num_directions = 2 if direction is 'bidirect'
        # if not, num_directions = 1
        text_repr = self.lstm_encoder(embedded_text, sequence_length=seq_len)
        # Shape: (batch_size, fc_hidden_size)
        fc_out = paddle.tanh(self.fc(text_repr))
        # Shape: (batch_size, num_classes)
        logits = self.output_layer(fc_out)
        return logits

model = LSTMModel(vocab_size=100, num_classes=2)

text = paddle.randint(low=1, high=10, shape=[1,10], dtype='int32')
seq_len = paddle.to_tensor([10])
logits = model(text, seq_len)
get_input_dim()[source]#

Returns the dimension of the vector input for each element in the sequence input to a LSTMEncoder. This is not the shape of the input tensor, but the last element of that shape.

get_output_dim()[source]#

Returns the dimension of the final vector output by this LSTMEncoder. This is not the shape of the returned tensor, but the last element of that shape.

forward(inputs, sequence_length)[source]#

LSTMEncoder takes the a sequence of vectors and returns a single vector, which is a combination of multiple LSTM layers. The input to this encoder is of shape (batch_size, num_tokens, input_size), The output is of shape (batch_size, hidden_size * 2) if LSTM is bidirection; If not, output is of shape (batch_size, hidden_size).

Parameters:
  • inputs (Tensor) – Shape as (batch_size, num_tokens, input_size). Tensor containing the features of the input sequence.

  • sequence_length (Tensor) – Shape as (batch_size). The sequence length of the input sequence.

Returns:

Returns tensor output, the hidden state at the last time step for every layer. Its data type is float and its shape is [batch_size, hidden_size].

Return type:

Tensor

class RNNEncoder(input_size, hidden_size, num_layers=1, direction='forward', dropout=0.0, pooling_type=None, **kwargs)[source]#

Bases: Layer

A RNNEncoder takes as input a sequence of vectors and returns a single vector, which is a combination of multiple paddle.nn.RNN subclass. The input to this encoder is of shape (batch_size, num_tokens, input_size), The output is of shape (batch_size, hidden_size * 2) if RNN is bidirection; If not, output is of shape (batch_size, hidden_size).

Paddle’s RNN have two outputs: the hidden state for every time step at last layer, and the hidden state at the last time step for every layer. If pooling_type is not None, we perform the pooling on the hidden state of every time step at last layer to create a single vector. If None, we use the hidden state of the last time step at last layer as a single output (shape of (batch_size, hidden_size)); And if direction is bidirection, the we concat the hidden state of the last forward rnn and backward rnn layer to create a single vector (shape of (batch_size, hidden_size * 2)).

Parameters:
  • input_size (int) – The number of expected features in the input (the last dimension).

  • hidden_size (int) – The number of features in the hidden state.

  • num_layers (int, optional) – Number of recurrent layers. E.g., setting num_layers=2 would mean stacking two RNNs together to form a stacked RNN, with the second RNN taking in outputs of the first RNN and computing the final results. Defaults to 1.

  • direction (str, optional) – The direction of the network. It can be “forward” and “bidirect” (it means bidirection network). If “bidirect”, it is a bidirectional RNN, and returns the concat output from both directions. Defaults to “forward”

  • dropout (float, optional) – If non-zero, introduces a Dropout layer on the outputs of each RNN layer except the last layer, with dropout probability equal to dropout. Defaults to 0.0.

  • pooling_type (str, optional) – If pooling_type is None, then the RNNEncoder will return the hidden state of the last time step at last layer as a single vector. If pooling_type is not None, it must be one of “sum”, “max” and “mean”. Then it will be pooled on the RNN output (the hidden state of every time step at last layer) to create a single vector. Defaults to None.

Example

import paddle
import paddle.nn as nn
import paddlenlp as nlp

class RNNModel(nn.Layer):
    def __init__(self,
                vocab_size,
                num_classes,
                emb_dim=128,
                padding_idx=0,
                rnn_hidden_size=198,
                direction='forward',
                rnn_layers=1,
                dropout_rate=0.0,
                pooling_type=None,
                fc_hidden_size=96):
        super().__init__()
        self.embedder = nn.Embedding(
            num_embeddings=vocab_size,
            embedding_dim=emb_dim,
            padding_idx=padding_idx)
        self.rnn_encoder = nlp.seq2vec.RNNEncoder(
            emb_dim,
            rnn_hidden_size,
            num_layers=rnn_layers,
            direction=direction,
            dropout=dropout_rate,
            pooling_type=pooling_type)
        self.fc = nn.Linear(self.rnn_encoder.get_output_dim(), fc_hidden_size)
        self.output_layer = nn.Linear(fc_hidden_size, num_classes)

    def forward(self, text, seq_len):
        # Shape: (batch_size, num_tokens, embedding_dim)
        embedded_text = self.embedder(text)
        # Shape: (batch_size, num_tokens, num_directions*rnn_hidden_size)
        # num_directions = 2 if direction is 'bidirect'
        # if not, num_directions = 1
        text_repr = self.rnn_encoder(embedded_text, sequence_length=seq_len)
        # Shape: (batch_size, fc_hidden_size)
        fc_out = paddle.tanh(self.fc(text_repr))
        # Shape: (batch_size, num_classes)
        logits = self.output_layer(fc_out)
        return logits

model = RNNModel(vocab_size=100, num_classes=2)

text = paddle.randint(low=1, high=10, shape=[1,10], dtype='int32')
seq_len = paddle.to_tensor([10])
logits = model(text, seq_len)
get_input_dim()[source]#

Returns the dimension of the vector input for each element in the sequence input to a RNNEncoder. This is not the shape of the input tensor, but the last element of that shape.

get_output_dim()[source]#

Returns the dimension of the final vector output by this RNNEncoder. This is not the shape of the returned tensor, but the last element of that shape.

forward(inputs, sequence_length)[source]#

RNNEncoder takes the a sequence of vectors and returns a single vector, which is a combination of multiple RNN layers. The input to this encoder is of shape (batch_size, num_tokens, input_size). The output is of shape (batch_size, hidden_size * 2) if RNN is bidirection; If not, output is of shape (batch_size, hidden_size).

Parameters:
  • inputs (Tensor) – Shape as (batch_size, num_tokens, input_size). Tensor containing the features of the input sequence.

  • sequence_length (Tensor) – Shape as (batch_size). The sequence length of the input sequence.

Returns:

Returns tensor output, the hidden state at the last time step for every layer. Its data type is float and its shape is [batch_size, hidden_size].

Return type:

Tensor

class TCNEncoder(input_size, num_channels, kernel_size=2, dropout=0.2)[source]#

Bases: Layer

A TCNEncoder takes as input a sequence of vectors and returns a single vector, which is the last one time step in the feature map. The input to this encoder is of shape (batch_size, num_tokens, input_size), and the output is of shape (batch_size, num_channels[-1]) with a receptive filed:

\[receptive filed = 2 * \sum_{i=0}^{len(num\_channels)-1}2^i(kernel\_size-1).\]

Temporal Convolutional Networks is a simple convolutional architecture. It outperforms canonical recurrent networks such as LSTMs in many tasks. See https://arxiv.org/pdf/1803.01271.pdf for more details.

Parameters:
  • input_size (int) – The number of expected features in the input (the last dimension).

  • num_channels (list) – The number of channels in different layer.

  • kernel_size (int) – The kernel size. Defaults to 2.

  • dropout (float) – The dropout probability. Defaults to 0.2.

get_input_dim()[source]#

Returns the dimension of the vector input for each element in the sequence input to a TCNEncoder. This is not the shape of the input tensor, but the last element of that shape.

get_output_dim()[source]#

Returns the dimension of the final vector output by this TCNEncoder. This is not the shape of the returned tensor, but the last element of that shape.

forward(inputs)[source]#

TCNEncoder takes as input a sequence of vectors and returns a single vector, which is the last one time step in the feature map. The input to this encoder is of shape (batch_size, num_tokens, input_size), and the output is of shape (batch_size, num_channels[-1]) with a receptive filed:

\[receptive filed = 2 * \sum_{i=0}^{len(num\_channels)-1}2^i(kernel\_size-1).\]
Parameters:

inputs (Tensor) – The input tensor with shape [batch_size, num_tokens, input_size].

Returns:

Returns tensor output with shape [batch_size, num_channels[-1]].

Return type:

Tensor