This tutorial will guide you how to configure recurrent neural network in PaddlePaddle. PaddlePaddle supports highly flexible and efficient recurrent neural network configuration. In this tutorial, you will learn how to:
We will use vanilla recurrent neural network, and sequence to sequence model to guide you through these steps. The code of sequence to sequence model can be found at `book/08.machine_translation <https://github.com/PaddlePaddle/book/tree/develop/08.machine_translation>`_ .
And the data preparation of this model can be found at `python/paddle/v2/dataset/wmt14.py <https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/dataset/wmt14.py>`_
Generally speaking, a recurrent network perform the following operations from :math:`t=1` to :math:`t=T`, or reversely from :math:`t=T` to :math:`t=1`.
where :math:`f_x(.)` is called **step function**, and :math:`f_y(.)` is called **output function**. In vanilla recurrent neural network, both of the step function and output function are very simple. However, PaddlePaddle supports the configuration of very complex architectures by modifying these two functions. We will use the sequence to sequence model with attention as an example to demonstrate how you can configure complex recurrent neural network models. In this section, we will use a simple vanilla recurrent neural network as an example of configuring simple recurrent neural network using :code:`recurrent_group`. Notice that if you only need to use simple RNN, GRU, or LSTM, then :code:`grumemory` and :code:`lstmemory` is recommended because they are more computationally efficient than :code:`recurrent_group`.
where :math:`x_t` is the RNN state, and :math:`I_t` is the input, :math:`W_x` and :math:`W_i` are transformation matrices for RNN states and inputs, respectively. :math:`b` is the bias.
:code:`recurrent_group` is the most important tools for constructing recurrent neural networks. It defines the **step function**, **output function** and the inputs of the recurrent neural network. Notice that the :code:`step` argument of this function implements both the :code:`step function` and the :code:`output function`:
PaddlePaddle uses memory to construct step function. **Memory** is the most important concept when constructing recurrent neural networks in PaddlePaddle. A memory is a state that is used recurrently in step functions, such as :math:`x_{t+1} = f_x(x_t)`. One memory contains an **output** and a **input**. The output of memory at the current time step is utilized as the input of the memory at the next time step. A memory can also has a **boot layer**, whose output is utilized as the initial value of the memory. In our case, the output of the gated recurrent unit is employed as the output memory. Notice that the name of the layer :code:`rnn_out` is the same as the name of :code:`out_mem`. This means the output of the layer :code:`rnn_out` (:math:`x_{t+1}`) is utilized as the **output** of :code:`out_mem` memory.
A memory can also be a sequence. In this case, at each time step, we have a sequence as the state of the recurrent neural network. This can be useful when constructing very complex recurrent neural network. Other advanced functions include defining multiple memories, and defining hierarchical recurrent neural network architecture using sub-sequence.
We return :code:`rnn_out` at the end of the function. It means that the output of the layer :code:`rnn_out` is utilized as the **output** function of the gated recurrent neural network.
We will use the sequence to sequence model with attention as an example to demonstrate how you can configure complex recurrent neural network models. An illustration of the sequence to sequence model with attention is shown in the following figure.
In this model, the source sequence :math:`S = \{s_1, \dots, s_T\}` is encoded with a bidirectional gated recurrent neural networks. The hidden states of the bidirectional gated recurrent neural network :math:`H_S = \{H_1, \dots, H_T\}` is called *encoder vector* The decoder is a gated recurrent neural network. When decoding each token :math:`y_t`, the gated recurrent neural network generates a set of weights :math:`W_S^t = \{W_1^t, \dots, W_T^t\}`, which are used to compute a weighted sum of the encoder vector. The weighted sum of the encoder vector is utilized to condition the generation of the token :math:`y_t`.
The encoder part of the model is listed below. It calls :code:`grumemory` to represent gated recurrent neural network. It is the recommended way of using recurrent neural network if the network architecture is simple, because it is faster than :code:`recurrent_group`. We have implemented most of the commonly used recurrent neural network architectures, you can refer to :ref:`api_trainer_config_helpers_layers` for more details.
We also project the encoder vector to :code:`decoder_size` dimensional space, get the first instance of the backward recurrent network, and project it to :code:`decoder_size` dimensional space:
The decoder uses :code:`recurrent_group` to define the recurrent neural network. The step and output functions are defined in :code:`gru_decoder_with_attention`:
The implementation of the step function is listed as below. First, it defines the **memory** of the decoder network. Then it defines attention, gated recurrent unit step function, and the output function:
After training the model, we can use it to generate sequences. A common practice is to use **beam search** to generate sequences. The following code snippets defines a beam search algorithm. Notice that :code:`beam_search` function assumes the output function of the :code:`step` returns a softmax normalized probability vector of the next token. We made the following changes to the model.
* use :code:`GeneratedInput` for trg_embedding. :code:`GeneratedInput` computes the embedding of the generated token at the last time step for the input at the current time step.
* use :code:`beam_search` function. This function needs to set:
Notice that this generation technique is only useful for decoder like generation process. If you are working on sequence tagging tasks, please refer to `book/06.understand_sentiment <https://github.com/PaddlePaddle/book/tree/develop/06.understand_sentiment>`_ for more details.
The full configuration file is located at `book/08.machine_translation/train.py <https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/train.py>`_ .
