8.5. Implementation of Recurrent Neural Networks from Scratch

In this section we implement a language model introduce in Section 8 from scratch. It is based on a character-level recurrent neural network trained on H. G. Wells’ The Time Machine. As before, we start by reading the data set first, which is introduced in Section 8.3.

%matplotlib inline
import d2l
import math
from mxnet import autograd, nd, gluon

batch_size, num_steps = 32, 35
train_iter, vocab = d2l.load_data_time_machine(batch_size, num_steps)

8.5.1. One-hot Encoding

Remember that each token is presented as a numerical index in train_iter. Feeding these indices directly to the neural network might make it hard to learn. We often present each token as a more expressive feature vector. The easiest presentation is called one-hot encoding.

In a nutshell, we map each index to a different unit vector: assume that the number of different toeksn in the vocabulary is \(N\) (the len(vocab)) and the token indices range from 0 to \(N-1\). If the index of a token is the integer \(i\), then we create a vector \(\mathbf{e}_i\) of all 0s with a length of \(N\) and set the element at position \(i\) to 1. This vector is the one-hot vector of the original token. The one-hot vectors with indices 0 and 2 are shown below.

nd.one_hot(nd.array([0, 2]), len(vocab))
[[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0.]
 [0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 0.]]
<NDArray 2x28 @cpu(0)>

The shape of the mini-batch we sample each time is (batch size, time step). The one_hot function transforms such a mini-batch into a 3-D tensor with the last dimension equals to the vocabulary size. We often transpose the input so that we will obtain a (time step, batch size, vocabulary size) output that fits into a sequence model easier.

X = nd.arange(10).reshape((2, 5))
nd.one_hot(X.T, 28).shape
(5, 2, 28)

8.5.2. Initializing the Model Parameters

Next, we initialize the model parameters for a RNN model. The number of hidden units num_hiddens is a tunable parameter.

def get_params(vocab_size, num_hiddens, ctx):
    num_inputs = num_outputs = vocab_size
    normal = lambda shape: nd.random.normal(
        scale=0.01, shape=shape, ctx=ctx)
    # Hidden layer parameters
    W_xh = normal((num_inputs, num_hiddens))
    W_hh = normal((num_hiddens, num_hiddens))
    b_h = nd.zeros(num_hiddens, ctx=ctx)
    # Output layer parameters
    W_hq = normal((num_hiddens, num_outputs))
    b_q = nd.zeros(num_outputs, ctx=ctx)
    # Attach a gradient
    params = [W_xh, W_hh, b_h, W_hq, b_q]
    for param in params: param.attach_grad()
    return params

8.5.3. RNN Model

First, we need an init_rnn_state function to return the hidden state at initialization. It returns a tuple consisting of an NDArray with a value of 0 and a shape of (batch size, number of hidden units). Using tuples makes it easier to handle situations where the hidden state contains multiple variables (e.g. when combining multiple layers in an RNN where each layers requires initializing).

def init_rnn_state(batch_size, num_hiddens, ctx):
    return (nd.zeros(shape=(batch_size, num_hiddens), ctx=ctx), )

The following rnn function defines how to compute the hidden state and output in a time step. The activation function here uses the tanh function. As described in Section 4.1, the mean value of the \(\tanh\) function values is 0 when the elements are evenly distributed over the real numbers.

def rnn(inputs, state, params):
    # inputs shape: (num_steps, batch_size, vocab_size)
    W_xh, W_hh, b_h, W_hq, b_q = params
    H, = state
    outputs = []
    for X in inputs:
        H = nd.tanh(nd.dot(X, W_xh) + nd.dot(H, W_hh) + b_h)
        Y = nd.dot(H, W_hq) + b_q
    return nd.concat(*outputs, dim=0), (H,)

Now we have all functions defined, next we create a class to wrap these functions and store parameters.

# Save to the d2l package.
class RNNModelScratch(object):
    """A RNN Model based on scratch implementations"""
    def __init__(self, vocab_size, num_hiddens, ctx,
                 get_params, init_state, forward):
        self.vocab_size, self.num_hiddens = vocab_size, num_hiddens
        self.params = get_params(vocab_size, num_hiddens, ctx)
        self.init_state, self.forward_fn = init_state, forward

    def __call__(self, X, state):
        X = nd.one_hot(X.T, self.vocab_size)
        return self.forward_fn(X, state, self.params)

    def begin_state(self, batch_size, ctx):
        return self.init_state(batch_size, self.num_hiddens, ctx)

Let’s do a sanity check whether inputs and outputs have the correct dimensions, e.g. to ensure that the dimensionality of the hidden state hasn’t changed.

vocab_size, num_hiddens, ctx = len(vocab), 512, d2l.try_gpu()
model = RNNModelScratch(len(vocab), num_hiddens, ctx, get_params,
                        init_rnn_state, rnn)
state = model.begin_state(X.shape[0], ctx)
Y, new_state = model(X.as_in_context(ctx), state)
Y.shape, len(new_state), new_state[0].shape
((10, 28), 1, (2, 512))

We can see that the output shape is (number steps \(\times\) batch size, vocabulary size), while the state shape remains the same, i.e. (batch size, number of hidden units).

8.5.4. Prediction

We first explain the predicting function so we can regularly check the prediction during training. This function predicts the next num_predicts characters based on the prefix (a string containing several characters). For the beginning of the sequence, we only update the hidden state. After that we begin generating new characters and emitting them.

# Save to the d2l package.
def predict_ch8(prefix, num_predicts, model, vocab, ctx):
    state = model.begin_state(batch_size=1, ctx=ctx)
    outputs = [vocab[prefix[0]]]
    get_input = lambda: nd.array([outputs[-1]], ctx=ctx).reshape((1, 1))
    for y in prefix[1:]:  # Warmup state with prefix
        _, state = model(get_input(), state)
    for _ in range(num_predicts):  # Predict num_predicts steps
        Y, state = model(get_input(), state)
    return ''.join([vocab.idx_to_token[i] for i in outputs])

We test the predict_rnn function first. Given that we didn’t train the network it will generate nonsensical predictions. We initialize it with the sequence traveller and have it generate 10 additional characters.

predict_ch8('time traveller ', 10, model, vocab, ctx)
'time traveller  emmmmmmmm'

8.5.5. Gradient Clipping

For a sequence of length \(T\), we compute the gradients over these \(T\) time steps in an iteration, which results in a chain of matrix-products with length \(O(T)\) during backpropagating. As mentioned in Section 4.8, it might result in numerical instability, e.g. the gradients may either explode or vanish, when \(T\) is large. Therefore RNN models often need extra help to stabilize the training.

Recall that when solving an optimization problem, we take update steps for the weights \(\mathbf{w}\) in the general direction of the negative gradient \(\mathbf{g}_t\) on a minibatch, say \(\mathbf{w} - \eta \cdot \mathbf{g}_t\). Let’s further assume that the objective is well behaved, i.e. it is Lipschitz continuous with constant \(L\), i.e.

(8.5.1)\[|l(\mathbf{w}) - l(\mathbf{w}')| \leq L \|\mathbf{w} - \mathbf{w}'\|.\]

In this case we can safely assume that if we update the weight vector by \(\eta \cdot \mathbf{g}_t\) we will not observe a change by more than \(L \eta \|\mathbf{g}_t\|\). This is both a curse and a blessing. A curse since it limits the speed with which we can make progress, a blessing since it limits the extent to which things can go wrong if we move in the wrong direction.

Sometimes the gradients can be quite large and the optimization algorithm may fail to converge. We could address this by reducing the learning rate \(\eta\) or by some other higher order trick. But what if we only rarely get large gradients? In this case such an approach may appear entirely unwarranted. One alternative is to clip the gradients by projecting them back to a ball of a given radius, say \(\theta\) via

(8.5.2)\[\mathbf{g} \leftarrow \min\left(1, \frac{\theta}{\|\mathbf{g}\|}\right) \mathbf{g}.\]

By doing so we know that the gradient norm never exceeds \(\theta\) and that the updated gradient is entirely aligned with the original direction \(\mathbf{g}\). It also has the desirable side-effect of limiting the influence any given minibatch (and within it any given sample) can exert on the weight vectors. This bestows a certain degree of robustness to the model. Gradient clipping provides a quick fix to the gradient exploding. While it doesn’t entire solve the problem, it is one of the many techniques to alleviate it.

Below we define a function to clip the gradients of a model that is either a RNNModelScratch instance or a Gluon model. Also note that we compute the gradient norm over all parameters.

# Save to the d2l package.
def grad_clipping(model, theta):
    if isinstance(model, gluon.Block):
        params = [p.data() for p in model.collect_params().values()]
        params = model.params
    norm = math.sqrt(sum((p.grad ** 2).sum().asscalar() for p in params))
    if norm > theta:
        for param in params:
            param.grad[:] *= theta / norm

8.5.6. Training

Similar to Section 3.2, let’s first define the function to train the model on one data epoch. It differs to the models training from previous chapters in three places:

  1. Different sampling methods for sequential data (independent sampling and sequential partitioning) will result in differences in the initialization of hidden states.

  2. We clip the gradient before updating the model parameters. This ensures that the model doesn’t diverge even when gradients blow up at some point during the training process (effectively it reduces the stepsize automatically).

  3. We use perplexity to evaluate the model. This ensures that different tests are comparable.

When the consecutive sampling is used, we initialize the hidden state at the beginning of each epoch. Since the \(i\)-th example in the next mini-batch is adjacent to the current \(i\)-th example, so we next mini-batch can use the current hidden state directly, we only detach the gradient so that we only compute the gradients within a mini-batch. When using the random sampling, we need to re-initialize the hidden state for each iteration since each example is sampled with a random position. Same to the train_epoch_ch3 function (Section 3.2), we use generalized updater, which could be a Gluon trainer or a scratched implementation.

# Save to the d2l package.
def train_epoch_ch8(model, train_iter, loss, updater, ctx, use_random_iter):
    state, timer = None, d2l.Timer()
    metric = d2l.Accumulator(2)  # loss_sum, num_examples
    for X, Y in train_iter:
        if state is None or use_random_iter:
            # Initialize state when either it's the first iteration or
            # using random sampling.
            state = model.begin_state(batch_size=X.shape[0], ctx=ctx)
            for s in state: s.detach()
        y = Y.T.reshape((-1,))
        X, y = X.as_in_context(ctx), y.as_in_context(ctx)
        with autograd.record():
            py, state = model(X, state)
            l = loss(py, y).mean()
        grad_clipping(model, 1)
        updater(batch_size=1)  # Since used mean already.
        metric.add(l.asscalar() * y.size, y.size)
    return math.exp(metric[0]/metric[1]), metric[1]/timer.stop()

The training function again supports either we implement the model from scratch or using Gluon.

# Save to the d2l package.
def train_ch8(model, train_iter, vocab, lr, num_epochs, ctx,
    # Initialize
    loss = gluon.loss.SoftmaxCrossEntropyLoss()
    animator = d2l.Animator(xlabel='epoch', ylabel='perplexity',
                            legend=['train'], xlim=[1, num_epochs])
    if isinstance(model, gluon.Block):
        model.initialize(ctx=ctx, force_reinit=True, init=init.Normal(0.01))
        trainer = gluon.Trainer(model.collect_params(), 'sgd', {'learning_rate': lr})
        updater = lambda batch_size : trainer.step(batch_size)
        updater = lambda batch_size : d2l.sgd(model.params, lr, batch_size)

    predict = lambda prefix: predict_ch8(prefix, 50, model, vocab, ctx)
    # Train and check the progress.
    for epoch in range(num_epochs):
        ppl, speed = train_epoch_ch8(
            model, train_iter, loss, updater, ctx, use_random_iter)
        if epoch % 10 == 0:
            print(predict('time traveller'))
            animator.add(epoch+1, [ppl])
    print('Perplexity %.1f, %d tokens/sec on %s' % (ppl, speed, ctx))
    print(predict('time traveller'))

Finally we can train a model. Since we only use 10,000 tokens in the dataset, so here we need more data epochs to converge.

num_epochs, lr = 500, 1
train_ch8(model, train_iter, vocab, lr, num_epochs, ctx)
Perplexity 1.0, 45996 tokens/sec on gpu(0)
time traveller  it s against reason said filby  what reason said
traveller  it s against reason said filby  what reason said

Then let’s check the results to use a random sampling iterator.

train_ch8(model, train_iter, vocab, lr, num_epochs, ctx, use_random_iter=True)
Perplexity 1.3, 45498 tokens/sec on gpu(0)
time traveller  it s against reason said filby  what reason said
traveller smiled round at us then still smiling faintly and

In the following we will see how to improve significantly on the current model and how to make it faster and easier to implement.

8.5.7. Summary

  • Sequence models need state initialization for training.

  • Between sequential models you need to ensure to detach the gradient, to ensure that the automatic differentiation does not propagate effects beyond the current sample.

  • A simple RNN language model consists of an encoder, an RNN model and a decoder.

  • Gradient clipping prevents gradient explosion (but it cannot fix vanishing gradients).

  • Perplexity calibrates model performance across variable sequence length. It is the exponentiated average of the cross-entropy loss.

  • Sequential partitioning typically leads to better models.

8.5.8. Exercises

  1. Show that one-hot encoding is equivalent to picking a different embedding for each object.

  2. Adjust the hyperparameters to improve the perplexity.

    • How low can you go? Adjust embeddings, hidden units, learning rate, etc.

    • How well will it work on other books by H. G. Wells, e.g. The War of the Worlds.

  3. Modify the predict function such as to use sampling rather than picking the most likely next character.

    • What happens?

    • Bias the model towards more likely outputs, e.g. by sampling from \(q(w_t|w_{t-1}, \ldots w_1) \propto p^\alpha(w_t|w_{t-1}, \ldots w_1)\) for \(\alpha > 1\).

  4. Run the code in this section without clipping the gradient. What happens?

  5. Change adjacent sampling so that it does not separate hidden states from the computational graph. Does the running time change? How about the accuracy?

  6. Replace the activation function used in this section with ReLU and repeat the experiments in this section.

  7. Prove that the perplexity is the inverse of the harmonic mean of the conditional word probabilities.

8.5.9. Scan the QR Code to Discuss