Long Short-Term Networks or LSTMs are a popular and powerful type of Recurrent Neural Network, or RNN.

They can be quite difficult to configure and apply to arbitrary sequence prediction problems, even with well defined and “easy to use” interfaces like those provided in the Keras deep learning library in Python.

One reason for this difficulty in Keras is the use of the TimeDistributed wrapper layer and the need for some LSTM layers to return sequences rather than single values.

In this tutorial, you will discover different ways to configure LSTM networks for sequence prediction, the role that the TimeDistributed layer plays, and exactly how to use it.

After completing this tutorial, you will know:

- How to design a one-to-one LSTM for sequence prediction.
- How to design a many-to-one LSTM for sequence prediction without the TimeDistributed Layer.
- How to design a many-to-many LSTM for sequence prediction with the TimeDistributed Layer.

Let’s get started.

## Tutorial Overview

This tutorial is divided into 5 parts; they are:

- TimeDistributed Layer
- Sequence Learning Problem
- One-to-One LSTM for Sequence Prediction
- Many-to-One LSTM for Sequence Prediction (without TimeDistributed)
- Many-to-Many LSTM for Sequence Prediction (with TimeDistributed)

### Environment

This tutorial assumes a Python 2 or Python 3 development environment with SciPy, NumPy, and Pandas installed.

The tutorial also assumes scikit-learn and Keras v2.0+ are installed with either the Theano or TensorFlow backend.

For help setting up your Python environment, see the post:

### Need help with LSTMs for Sequence Prediction?

Take my free 7-day email course and discover 6 different LSTM architectures (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

## TimeDistributed Layer

LSTMs are powerful, but hard to use and hard to configure, especially for beginners.

An added complication is the TimeDistributed Layer (and the former *TimeDistributedDense* layer) that is cryptically described as a layer wrapper:

This wrapper allows us to apply a layer to every temporal slice of an input.

How and when are you supposed to use this wrapper with LSTMs?

The confusion is compounded when you search through discussions about the wrapper layer on the Keras GitHub issues and StackOverflow.

For example, in the issue “When and How to use TimeDistributedDense,” fchollet (Keras’ author) explains:

TimeDistributedDense applies a same Dense (fully-connected) operation to every timestep of a 3D tensor.

This makes perfect sense if you already understand what the TimeDistributed layer is for and when to use it, but is no help at all to a beginner.

This tutorial aims to clear up confusion around using the TimeDistributed wrapper with LSTMs with worked examples that you can inspect, run, and play with to help your concrete understanding.

## Sequence Learning Problem

We will use a simple sequence learning problem to demonstrate the TimeDistributed layer.

In this problem, the sequence [0.0, 0.2, 0.4, 0.6, 0.8] will be given as input one item at a time and must be in turn returned as output, one item at a time.

Think of it as learning a simple echo program. We give 0.0 as input, we expect to see 0.0 as output, repeated for each item in the sequence.

We can generate this sequence directly as follows:

1 2 3 4 |
from numpy import array length = 5 seq = array([i/float(length) for i in range(length)]) print(seq) |

Running this example prints the generated sequence:

1 |
[ 0. 0.2 0.4 0.6 0.8] |

The example is configurable and you can play with longer/shorter sequences yourself later if you like. Let me know about your results in the comments.

## One-to-One LSTM for Sequence Prediction

Before we dive in, it is important to show that this sequence learning problem can be learned piecewise.

That is, we can reframe the problem into a dataset of input-output pairs for each item in the sequence. Given 0, the network should output 0, given 0.2, the network must output 0.2, and so on.

This is the simplest formulation of the problem and requires the sequence to be split into input-output pairs and for the sequence to be predicted one step at a time and gathered outside of the network.

The input-output pairs are as follows:

1 2 3 4 5 6 |
X, y 0.0, 0.0 0.2, 0.2 0.4, 0.4 0.6, 0.6 0.8, 0.8 |

The input for LSTMs must be three dimensional. We can reshape the 2D sequence into a 3D sequence with 5 samples, 1 time step, and 1 feature. We will define the output as 5 samples with 1 feature.

1 2 |
X = seq.reshape(5, 1, 1) y = seq.reshape(5, 1) |

We will define the network model as having 1 input with 1 time step. The first hidden layer will be an LSTM with 5 units. The output layer with be a fully-connected layer with 1 output.

The model will be fit with efficient ADAM optimization algorithm and the mean squared error loss function.

The batch size was set to the number of samples in the epoch to avoid having to make the LSTM stateful and manage state resets manually, although this could just as easily be done in order to update weights after each sample is shown to the network.

The complete code listing is provided below:

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 |
from numpy import array from keras.models import Sequential from keras.layers import Dense from keras.layers import LSTM # prepare sequence length = 5 seq = array([i/float(length) for i in range(length)]) X = seq.reshape(len(seq), 1, 1) y = seq.reshape(len(seq), 1) # define LSTM configuration n_neurons = length n_batch = length n_epoch = 1000 # create LSTM model = Sequential() model.add(LSTM(n_neurons, input_shape=(1, 1))) model.add(Dense(1)) model.compile(loss='mean_squared_error', optimizer='adam') print(model.summary()) # train LSTM model.fit(X, y, epochs=n_epoch, batch_size=n_batch, verbose=2) # evaluate result = model.predict(X, batch_size=n_batch, verbose=0) for value in result: print('%.1f' % value) |

Running the example first prints the structure of the configured network.

We can see that the LSTM layer has 140 parameters. This is calculated based on the number of inputs (1) and the number of outputs (5 for the 5 units in the hidden layer), as follows:

1 2 3 4 |
n = 4 * ((inputs + 1) * outputs + outputs^2) n = 4 * ((1 + 1) * 5 + 5^2) n = 4 * 35 n = 140 |

We can also see that the fully connected layer only has 6 parameters for the number of inputs (5 for the 5 inputs from the previous layer), number of outputs (1 for the 1 neuron in the layer), and the bias.

1 2 3 |
n = inputs * outputs + outputs n = 5 * 1 + 1 n = 6 |

1 2 3 4 5 6 7 8 9 10 11 |
_________________________________________________________________ Layer (type) Output Shape Param # ================================================================= lstm_1 (LSTM) (None, 1, 5) 140 _________________________________________________________________ dense_1 (Dense) (None, 1, 1) 6 ================================================================= Total params: 146.0 Trainable params: 146 Non-trainable params: 0.0 _________________________________________________________________ |

The network correctly learns the prediction problem.

1 2 3 4 5 |
0.0 0.2 0.4 0.6 0.8 |

## Many-to-One LSTM for Sequence Prediction (without Â TimeDistributed)

In this section, we develop an LSTM to output the sequence all at once, although without the TimeDistributed wrapper layer.

The input for LSTMs must be three dimensional. We can reshape the 2D sequence into a 3D sequence with 1 sample, 5 time steps, and 1 feature. We will define the output as 1 sample with 5 features.

1 2 |
X = seq.reshape(1, 5, 1) y = seq.reshape(1, 5) |

Immediately, you can see that the problem definition must be slightly adjusted to support a network for sequence prediction without a TimeDistributed wrapper. Specifically, output one vector rather build out an output sequence one step at a time. The difference may sound subtle, but it is important to understanding the role of the TimeDistributed wrapper.

We will define the model as having one input with 5 time steps. The first hidden layer will be an LSTM with 5 units. The output layer is afully-connectedd layer with 5 neurons.

1 2 3 4 5 6 |
# create LSTM model = Sequential() model.add(LSTM(5, input_shape=(5, 1))) model.add(Dense(length)) model.compile(loss='mean_squared_error', optimizer='adam') print(model.summary()) |

Next, we fit the model for only 500 epochs and a batch size of 1 for the single sample in the training dataset.

1 2 |
# train LSTM model.fit(X, y, epochs=500, batch_size=1, verbose=2) |

Putting this all together, the complete code listing is provided below.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 |
from numpy import array from keras.models import Sequential from keras.layers import Dense from keras.layers import LSTM # prepare sequence length = 5 seq = array([i/float(length) for i in range(length)]) X = seq.reshape(1, length, 1) y = seq.reshape(1, length) # define LSTM configuration n_neurons = length n_batch = 1 n_epoch = 500 # create LSTM model = Sequential() model.add(LSTM(n_neurons, input_shape=(length, 1))) model.add(Dense(length)) model.compile(loss='mean_squared_error', optimizer='adam') print(model.summary()) # train LSTM model.fit(X, y, epochs=n_epoch, batch_size=n_batch, verbose=2) # evaluate result = model.predict(X, batch_size=n_batch, verbose=0) for value in result[0,:]: print('%.1f' % value) |

Running the example first prints a summary of the configured network.

We can see that the LSTM layer has 140 parameters as in the previous section.

The LSTM units have been crippled and will each output a single value, providing a vector of 5 values as inputs to the fully connected layer. The time dimension or sequence information has been thrown away and collapsed into a vector of 5 values.

We can see that the fully connected output layer has 5 inputs and is expected to output 5 values. We can account for the 30 weights to be learned as follows:

1 2 3 |
n = inputs * outputs + outputs n = 5 * 5 + 5 n = 30 |

The summary of the network is reported as follows:

1 2 3 4 5 6 7 8 9 10 11 |
_________________________________________________________________ Layer (type) Output Shape Param # ================================================================= lstm_1 (LSTM) (None, 5) 140 _________________________________________________________________ dense_1 (Dense) (None, 5) 30 ================================================================= Total params: 170.0 Trainable params: 170 Non-trainable params: 0.0 _________________________________________________________________ |

The model is fit, printing loss information before finalizing and printing the predicted sequence.

The sequence is reproduced correctly, but as a single piece rather than stepwise through the input data. We may have used a Dense layer as the first hidden layer instead of LSTMs as this usage of LSTMs does not take much advantage of their full capability for sequence learning and processing.

1 2 3 4 5 |
0.0 0.2 0.4 0.6 0.8 |

## Many-to-Many LSTM for Sequence Prediction (with TimeDistributed)

In this section, we will use the TimeDistributed layer to process the output from the LSTM hidden layer.

There are two key points to remember when using the TimeDistributed wrapper layer:

**The input must be (at least) 3D**. This often means that you will need to configure your last LSTM layer prior to your TimeDistributed wrapped Dense layer to return sequences (e.g. set the “return_sequences” argument to “True”).**The output will be 3D**. This means that if your TimeDistributed wrapped Dense layer is your output layer and you are predicting a sequence, you will need to resize your y array into a 3D vector.

We can define the shape of the output as having 1 sample, 5 time steps, and 1 feature, just like the input sequence, as follows:

1 |
y = seq.reshape(1, length, 1) |

We can define the LSTM hidden layer to return sequences rather than single values by setting the “*return_sequences*” argument to true.

1 |
model.add(LSTM(n_neurons, input_shape=(length, 1), return_sequences=True)) |

This has the effect of each LSTM unit returning a sequence of 5 outputs, one for each time step in the input data, instead of single output value as in the previous example.

We also can use the TimeDistributed on the output layer to wrap a fully connected Dense layer with a single output.

1 |
model.add(TimeDistributed(Dense(1))) |

The single output value in the output layer is key. It highlights that we intend to output one time step from the sequence for each time step in the input. It just so happens that we will process 5 time steps of the input sequence at a time.

The TimeDistributed achieves this trick by applying the same Dense layer (same weights) to the LSTMs outputs for one time step at a time. In this way, the output layer only needs one connection to each LSTM unit (plus one bias).

For this reason, the number of training epochs needs to be increased to account for the smaller network capacity. I doubled it from 500 to 1000 to match the first one-to-one example.

Putting this together, the full code listing is provided below.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 |
from numpy import array from keras.models import Sequential from keras.layers import Dense from keras.layers import TimeDistributed from keras.layers import LSTM # prepare sequence length = 5 seq = array([i/float(length) for i in range(length)]) X = seq.reshape(1, length, 1) y = seq.reshape(1, length, 1) # define LSTM configuration n_neurons = length n_batch = 1 n_epoch = 1000 # create LSTM model = Sequential() model.add(LSTM(n_neurons, input_shape=(length, 1), return_sequences=True)) model.add(TimeDistributed(Dense(1))) model.compile(loss='mean_squared_error', optimizer='adam') print(model.summary()) # train LSTM model.fit(X, y, epochs=n_epoch, batch_size=n_batch, verbose=2) # evaluate result = model.predict(X, batch_size=n_batch, verbose=0) for value in result[0,:,0]: print('%.1f' % value) |

Running the example, we can see the structure of the configured network.

We can see that as in the previous example, we have 140 parameters in the LSTM hidden layer.

The fully connected output layer is a very different story. In fact, it matches the one-to-one example exactly. One neuron that has one weight for each LSTM unit in the previous layer, plus one for the bias input.

This does two important things:

- Allows the problem to be framed and learned as it was defined, that is one input to one output, keeping the internal process for each time step separate.
- Simplifies the network by requiring far fewer weights such that only one time step is processed at a time.

The one simpler fully connected layer is applied to each time step in the sequence provided from the previous layer to build up the output sequence.

1 2 3 4 5 6 7 8 9 10 11 |
_________________________________________________________________ Layer (type) Output Shape Param # ================================================================= lstm_1 (LSTM) (None, 5, 5) 140 _________________________________________________________________ time_distributed_1 (TimeDist (None, 5, 1) 6 ================================================================= Total params: 146.0 Trainable params: 146 Non-trainable params: 0.0 _________________________________________________________________ |

Again, the network learns the sequence.

1 2 3 4 5 |
0.0 0.2 0.4 0.6 0.8 |

We can think of the framing of the problem with time steps and a TimeDistributed layer as a more compact way of implementing the one-to-one network in the first example. It may even be more efficient (space or time wise) at aÂ larger scale.

## Further Reading

Below are some resources and discussions on the TimeDistributed layer you may like to dive in into.

- TimeDistributed Layer in the Keras API
- TimeDistributed code on GitHub
- The difference between ‘Dense’ and ‘TimeDistributedDense’ of ‘Keras’ on StackExchange
- When and How to use TimeDistributedDenseÂ onÂ GitHub

## Summary

In this tutorial, you discovered how to develop LSTM networks for sequence prediction and the role of the TimeDistributed layer.

Specifically, you learned:

- How to design a one-to-one LSTM for sequence prediction.
- How to design a many-to-one LSTM for sequence prediction without the TimeDistributed Layer.
- How to design a many-to-many LSTM for sequence prediction with the TimeDistributed Layer.

Do you have any questions?

Ask your questions in the comments below and I will do my best to answer them.

Hi, Jason, nice article on TimeDistributed layer!

Basically, there’re three configurations for X (and thus y):

1. (5,1,1) – 5 batchs, 1 time step, 1 feature/step – result shape (5,1)

2. (1,5,1) – 1 batch, 5 time steps, 1 feature/step – result shape (1,5)

3. (1,1,5) – 1 batch, 1 time step, 5 features/step

in article, you discussed previous 2 configures.

I did experiment of config 3, result same shape (1, 5) as 2 does, ’cause X input only 1 batch (which contains 1 sample, which has 5 features.) this config surely lost time information.

3 differ from 2 in two ways:

1) how we/model frame the problem: sequence should be framed as multi time steps as 2

2) different number of LSTM params: config 2 has 140, while config 3 has 220! (big input vector)

Q:

in section ‘many to one without TimeDistributed’, with config 2, you said “The time dimension or sequence information has been thrown away and collapsed into a vector of 5 values.” — that surprise me a little bit.

– does that mean, for seq-to-seq problem, we should always use TimeDistributed?

– what situation suites config 2 (samples, multi-time-steps, features)?

I guess for sequence-to-vector problem (predict one target one time step), config 2 is fine. But for sequence-to-sequence problem discussed here, config 2 is not the right choice, go TimeDistributed.

Very nice, yes I agree.

Generally, we must model sequences as time steps. BPTT will use the sequence data to estimate the gradient. LSTMs have memory, but we cannot rely on them to remember everything (e.g. sequence length of 1).

We can configure an MLP or LSTM to output a vector. For an LSTM, if we output a vector of n values for one time step, each output is considered by the LSTM as a feature, not a time step. Thus it is a many-to-one architecture. The vector may contain timesteps, but the LSTM is not outputting time steps, it is outputting features.

This is no more or less valid, it may require more weights and may give better or worse performance.

Does that make sense?

Any recommendations when facing a one-to-many problem?

They often need more training than you think and consider using bidirectional inputs and regularization on input connections.

This post is great! Thanks for being about the only person to actually explain simply what the TimeDistributed wrapper is doing.

I tried it out with audio vocal data to attempt generation of new speech. I’d previously got basic results with a plain Dense layer on the output.

With the TimeDistributed the network of lstms learned fast. But the result was just to return a rough version of the seed data inputted during generation. This appears to be modelling the equality function, when what I expected was something resembling the sequence following the seed.

My X input is an array of batches, timesteps, and vocal properties. Just a longer version of your example. My y output for measuring error is effectively the same data, just one timestamp later for each batch (time sequence).

Since your examples are for equality modelling, it’s hard to tell if I’ve missed a concept. Any thoughts on why this seems to generate equality rather than next timesteps, from my basic description?

By the way, my original project without TimeDistributed is found at http://babble-rnn.consected.com in case you’re interested in extra context.

Perhaps you need to fit for longer or require more training data?

I wondered about that. I think my mistake may be simple…

Imagine the sequence I was trying to learn was 1,2,3,4,5,6,7,8 (which I’d normalise in the range 0:1). In the standard Keras LSTM example without TimeDistributed I’d have:

input X[0] = [0,1,2]

output y[0] = [3]

X[1] = [1,2,3]

y[1] = [4]

…

So in the TimeDistributed setup I reported above, I tried:

X[0] = [0,1,2]

y[0] = [1,2,3]

X[1] = [1,2,3]

y[1] = [2,3,4]

…

In other words, I was offsetting the intended output by just a single timestep for each batch to be learned.

But I’m guessing that I should really offset the output to be learned by the full number of timesteps in each batch:

X[0] = [0,1,2]

y[0] = [3,4,5]

X[1] = [1,2,3]

y[1] = [4,5,6]

Is the latter example what I should be doing? Intuitively, this would explain why I was learning something close to equality in my first run. But from multiple readings of your code in the post it is not clear to me that this is the case.

I’m not sure I follow. There are indeed many ways to frame a sequence prediction problem.

The simplest framing is sequence in => sequence out where either in or out could be one or more time steps.

Keep one sequence as one sample if possible.

Hmm, I think I have missed something big here. Please humour and let me try once more.

In the standard LSTM examples on Keras, if I was to learn a long time sequence (for example integers incrementing in the range 1..100000), I would pick a shorter segment of the total sequence to pass to the LSTM (I split my corpus into sub-batches that represent the number of LSTM timesteps), then the output to learn would be just the next item in the sequence. There is no TimeDistributed output, so I get one result to calculate error against.

input set: 1,2,3

desired output: 4

then repeat with other sub-batches in the same way (and Keras scrambles the order), so the next one may be…

input set: 473, 474, 475

desired output: 476

If that makes sense, then allow me to ask simply what the input and output should be for the TimeDistributed setup. Would it be

(option A)

input set: 1, 2, 3

desired output: 2, 3, 4

(option B)

input set: 1, 2, 3

desired output: 4, 5, 6

(option C)

something else entirely.

Am I making more sense now?

Your example shows input set and desired output being the same, which says to me that the net will just learn the equality function. Again, am I missing something?

Thanks again for your help.

Yes, good question.

Option B.

In your first example you have a many-to-one time step predictive model. In option B you have a many-to-many time step predictive model. The TimeDistributed wrapper would allow you to use the same Dense layer to output each time step in the output sequence, in this case one output time step per input time step.

I hope that helps.

Hi Jason, thanks for the great article!

Having looked through this article and forums online, is it correct to say that if we were to do many-to-one prediction (an input vector with an output value), it will be straightforward and faster to just use the Dense layer?

In the case where we want to do many-to-many predictions (multiple input vectors/matrices with an output vector/matrix), TimeDistributed layer should be used instead?

Yes, but in the latter case the dense is wrapped in the timedistributed.

That does, thank you! For some reason I couldn’t get that from your post, so thanks for taking the time to explain in more detail.

Thanks, that cleared up return_sequences for me, but I still don’t fully understand what TimeDistributed does.

In the last example (Many-to-Many): If I change TimeDistributed(Dense(1)) to just Dense(1), neither the output shape nor the number of parameters changes and it works just as well. What is the difference between these two options in this case?

Note the number of weights in the network.

Without the TimeDistributed wrapper, the Dense is connected to the output from each time step. With the wrapper, the same Dense is applied to each time step.

It’s a question of how you want to model the problem. Let the Dense combine the time steps and output a vector or process each time step one at a time.

Does that help?

Hey,

Thanks for amazing tutorial.

This shows simple echo program implementation right ?

I want something like –

Input(For time period 2012-2013) – 1,2,5,3,6,4,7,8,9,5

output(For Time Period 2014) – 1,3,4

The output sequence should be generated based on the input sequence, kindly guide me on that.

Sounds great.

What is the issue exactly? You can use code from the blog directly and adapt it for your problem. Where are you having trouble?

How to create a pyramid of LSTMs. i.e. the input to the first node of 2nd layer LSTM will be output at t1 and t2 of first layer LSTM, similrly 2nd node of the 2nd layer will use t3 and t4 from first layer, and so on..

You mean a stacked LSTM?

Hello! Is the a way to have DIFFERENT length of input and output-timesteps?

Like, I have series with 100 timesteps in the past and will learn next 10 in the feature?

TimeDistributed requires equal length.

If I output return_sequence=false in the last LSTM and Dense with 10 neurones, would it be the same?

Thanks You!

Sorry, I’m not sure I follow, can you restate your question?

Generally, different numbers of times steps on the input and output are referred to as seq2seq problems and are perhaps best addressed with an encoder-decoder network.