That is the primary publish in a collection introducing timeseries forecasting with torch
. It does assume some prior expertise with torch
and/or deep studying. However so far as time collection are involved, it begins proper from the start, utilizing recurrent neural networks (GRU or LSTM) to foretell how one thing develops in time.
On this publish, we construct a community that makes use of a sequence of observations to foretell a worth for the very subsequent time limit. What if we’d wish to forecast a sequence of values, similar to, say, per week or a month of measurements?
One factor we might do is feed again into the system the beforehand forecasted worth; that is one thing we’ll attempt on the finish of this publish. Subsequent posts will discover different choices, a few of them involving considerably extra complicated architectures. It is going to be fascinating to check their performances; however the important purpose is to introduce some torch
“recipes” that you could apply to your personal information.
We begin by inspecting the dataset used. It’s a lowdimensional, however fairly polyvalent and complicated one.
The vic_elec
dataset, out there by bundle tsibbledata
, supplies three years of halfhourly electrical energy demand for Victoria, Australia, augmented by sameresolution temperature data and a each day vacation indicator.
Rows: 52,608
Columns: 5
$ Time <dttm> 20120101 00:00:00, 20120101 00:30:00, 20120101 01:00:00,…
$ Demand <dbl> 4382.825, 4263.366, 4048.966, 3877.563, 4036.230, 3865.597, 369…
$ Temperature <dbl> 21.40, 21.05, 20.70, 20.55, 20.40, 20.25, 20.10, 19.60, 19.10, …
$ Date <date> 20120101, 20120101, 20120101, 20120101, 20120101, 20…
$ Vacation <lgl> TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRU…
Relying on what subset of variables is used, and whether or not and the way information is temporally aggregated, these information might serve as an instance quite a lot of totally different strategies. For instance, within the third version of Forecasting: Principles and Practice each day averages are used to show quadratic regression with ARMA errors. On this first introductory publish although, in addition to in most of its successors, we’ll try to forecast Demand
with out counting on extra data, and we maintain the unique decision.
To get an impression of how electrical energy demand varies over totally different timescales. Let’s examine information for 2 months that properly illustrate the Ushaped relationship between temperature and demand: January, 2014 and July, 2014.
First, right here is July.
vic_elec_2014 < vic_elec %>%
filter(year(Date) == 2014) %>%
choose(c(Date, Vacation)) %>%
mutate(Demand = scale(Demand), Temperature = scale(Temperature)) %>%
pivot_longer(Time, names_to = "variable") %>%
update_tsibble(key = variable)
vic_elec_2014 %>% filter(month(Time) == 7) %>%
autoplot() +
scale_colour_manual(values = c("#08c5d1", "#00353f")) +
theme_minimal()
It’s winter; temperature fluctuates under common, whereas electrical energy demand is above common (heating). There’s robust variation over the course of the day; we see troughs within the demand curve similar to ridges within the temperature graph, and vice versa. Whereas diurnal variation dominates, there is also variation over the times of the week. Between weeks although, we don’t see a lot distinction.
Evaluate this with the info for January:
We nonetheless see the robust circadian variation. We nonetheless see some dayofweek variation. However now it’s excessive temperatures that trigger elevated demand (cooling). Additionally, there are two durations of unusually excessive temperatures, accompanied by distinctive demand. We anticipate that in a univariate forecast, not making an allowance for temperature, this will probably be arduous – and even, not possible – to forecast.
Let’s see a concise portrait of how Demand
behaves utilizing feasts::STL()
. First, right here is the decomposition for July:
And right here, for January:
Each properly illustrate the robust circadian and weekly seasonalities (with diurnal variation considerably stronger in January). If we glance intently, we will even see how the pattern element is extra influential in January than in July. This once more hints at a lot stronger difficulties predicting the January than the July developments.
Now that now we have an concept what awaits us, let’s start by making a torch
dataset
.
Here’s what we intend to do. We need to begin our journey into forecasting through the use of a sequence of observations to foretell their rapid successor. In different phrases, the enter (x
) for every batch merchandise is a vector, whereas the goal (y
) is a single worth. The size of the enter sequence, x
, is parameterized as n_timesteps
, the variety of consecutive observations to extrapolate from.
The dataset
will replicate this in its .getitem()
technique. When requested for the observations at index i
, it is going to return tensors like so:
list(
x = self$x[start:end],
y = self$x[end+1]
)
the place begin:finish
is a vector of indices, of size n_timesteps
, and finish+1
is a single index.
Now, if the dataset
simply iterated over its enter so as, advancing the index one by one, these traces might merely learn
list(
x = self$x[i:(i + self$n_timesteps  1)],
y = self$x[self$n_timesteps + i]
)
Since many sequences within the information are comparable, we will scale back coaching time by making use of a fraction of the info in each epoch. This may be completed by (optionally) passing a sample_frac
smaller than 1. In initialize()
, a random set of begin indices is ready; .getitem()
then simply does what it usually does: search for the (x,y)
pair at a given index.
Right here is the entire dataset
code:
elec_dataset < dataset(
title = "elec_dataset",
initialize = perform(x, n_timesteps, sample_frac = 1) {
self$n_timesteps < n_timesteps
self$x < torch_tensor((x  train_mean) / train_sd)
n < length(self$x)  self$n_timesteps
self$begins < sort(sample.int(
n = n,
dimension = n * sample_frac
))
},
.getitem = perform(i) {
begin < self$begins[i]
finish < begin + self$n_timesteps  1
list(
x = self$x[start:end],
y = self$x[end + 1]
)
},
.size = perform() {
length(self$begins)
}
)
You will have observed that we normalize the info by globally outlined train_mean
and train_sd
. We but should calculate these.
The best way we cut up the info is simple. We use the entire of 2012 for coaching, and all of 2013 for validation. For testing, we take the “troublesome” month of January, 2014. You’re invited to check testing outcomes for July that very same yr, and evaluate performances.
vic_elec_get_year < perform(yr, month = NULL) {
vic_elec %>%
filter(year(Date) == yr, month(Date) == if (is.null(month)) month(Date) else month) %>%
as_tibble() %>%
choose(Demand)
}
elec_train < vic_elec_get_year(2012) %>% as.matrix()
elec_valid < vic_elec_get_year(2013) %>% as.matrix()
elec_test < vic_elec_get_year(2014, 1) %>% as.matrix() # or 2014, 7, alternatively
train_mean < mean(elec_train)
train_sd < sd(elec_train)
Now, to instantiate a dataset
, we nonetheless want to choose sequence size. From prior inspection, per week looks as if a good choice.
n_timesteps < 7 * 24 * 2 # days * hours * halfhours
Now we will go forward and create a dataset
for the coaching information. Let’s say we’ll make use of fifty% of the info in every epoch:
train_ds < elec_dataset(elec_train, n_timesteps, sample_frac = 0.5)
length(train_ds)
8615
Fast test: Are the shapes appropriate?
$x
torch_tensor
0.4141
0.5541
[...] ### traces eliminated by me
0.8204
0.9399
... [the output was truncated (use n=1 to disable)]
[ CPUFloatType{336,1} ]
$y
torch_tensor
0.6771
[ CPUFloatType{1} ]
Sure: That is what we needed to see. The enter sequence has n_timesteps
values within the first dimension, and a single one within the second, similar to the one function current, Demand
. As supposed, the prediction tensor holds a single worth, corresponding– as we all know – to n_timesteps+1
.
That takes care of a single inputoutput pair. As regular, batching is organized for by torch
’s dataloader
class. We instantiate one for the coaching information, and instantly once more confirm the result:
batch_size < 32
train_dl < train_ds %>% dataloader(batch_size = batch_size, shuffle = TRUE)
length(train_dl)
b < train_dl %>% dataloader_make_iter() %>% dataloader_next()
b
$x
torch_tensor
(1,.,.) =
0.4805
0.3125
[...] ### traces eliminated by me
1.1756
0.9981
... [the output was truncated (use n=1 to disable)]
[ CPUFloatType{32,336,1} ]
$y
torch_tensor
0.1890
0.5405
[...] ### traces eliminated by me
2.4015
0.7891
... [the output was truncated (use n=1 to disable)]
[ CPUFloatType{32,1} ]
We see the added batch dimension in entrance, leading to general form (batch_size, n_timesteps, num_features)
. That is the format anticipated by the mannequin, or extra exactly, by its preliminary RNN layer.
Earlier than we go on, let’s shortly create dataset
s and dataloader
s for validation and take a look at information, as properly.
valid_ds < elec_dataset(elec_valid, n_timesteps, sample_frac = 0.5)
valid_dl < valid_ds %>% dataloader(batch_size = batch_size)
test_ds < elec_dataset(elec_test, n_timesteps)
test_dl < test_ds %>% dataloader(batch_size = 1)
The mannequin consists of an RNN – of sort GRU or LSTM, as per the person’s selection – and an output layer. The RNN does a lot of the work; the singleneuron linear layer that outputs the prediction compresses its vector enter to a single worth.
Right here, first, is the mannequin definition.
mannequin < nn_module(
initialize = perform(sort, input_size, hidden_size, num_layers = 1, dropout = 0) {
self$sort < sort
self$num_layers < num_layers
self$rnn < if (self$sort == "gru") {
nn_gru(
input_size = input_size,
hidden_size = hidden_size,
num_layers = num_layers,
dropout = dropout,
batch_first = TRUE
)
} else {
nn_lstm(
input_size = input_size,
hidden_size = hidden_size,
num_layers = num_layers,
dropout = dropout,
batch_first = TRUE
)
}
self$output < nn_linear(hidden_size, 1)
},
ahead = perform(x) {
# listing of [output, hidden]
# we use the output, which is of dimension (batch_size, n_timesteps, hidden_size)
x < self$rnn(x)[[1]]
# from the output, we solely need the ultimate timestep
# form now could be (batch_size, hidden_size)
x < x[ , dim(x)[2], ]
# feed this to a single output neuron
# closing form then is (batch_size, 1)
x %>% self$output()
}
)
Most significantly, that is what occurs in ahead()
.

The RNN returns an inventory. The listing holds two tensors, an output, and a synopsis of hidden states. We discard the state tensor, and maintain the output solely. The excellence between state and output, or relatively, the way in which it’s mirrored in what a
torch
RNN returns, deserves to be inspected extra intently. We’ll do this in a second. 
Of the output tensor, we’re fascinated by solely the ultimate timestep, although.

Solely this one, thus, is handed to the output layer.

Lastly, the stated output layer’s output is returned.
Now, a bit extra on states vs. outputs. Think about Fig. 1, from Goodfellow, Bengio, and Courville (2016).
Let’s fake there are three time steps solely, similar to (t1), (t), and (t+1). The enter sequence, accordingly, consists of (x_{t1}), (x_{t}), and (x_{t+1}).
At every (t), a hidden state is generated, and so is an output. Usually, if our purpose is to foretell (y_{t+2}), that’s, the very subsequent commentary, we need to have in mind the entire enter sequence. Put otherwise, we need to have run by the entire equipment of state updates. The logical factor to do would thus be to decide on (o_{t+1}), for both direct return from ahead()
or for additional processing.
Certainly, return (o_{t+1}) is what a Keras LSTM or GRU would do by default. Not so its torch
counterparts. In torch
, the output tensor includes all of (o). This is the reason, in step two above, we choose the one time step we’re fascinated by – particularly, the final one.
In later posts, we’ll make use of greater than the final time step. Typically, we’ll use the sequence of hidden states (the (h)s) as an alternative of the outputs (the (o)s). So it’s possible you’ll really feel like asking, what if we used (h_{t+1}) right here as an alternative of (o_{t+1})? The reply is: With a GRU, this could not make a distinction, as these two are similar. With LSTM although, it could, as LSTM retains a second, particularly, the “cell,” state.
On to initialize()
. For ease of experimentation, we instantiate both a GRU or an LSTM based mostly on person enter. Two issues are price noting:

We cross
batch_first = TRUE
when creating the RNNs. That is required withtorch
RNNs after we need to persistently have batch gadgets stacked within the first dimension. And we do need that; it’s arguably much less complicated than a change of dimension semantics for one subtype of module. 
num_layers
can be utilized to construct a stacked RNN, similar to what you’d get in Keras when chaining two GRUs/LSTMs (the primary one created withreturn_sequences = TRUE
). This parameter, too, we’ve included for fast experimentation.
Let’s instantiate a mannequin for coaching. It is going to be a singlelayer GRU with thirtytwo models.
# coaching RNNs on the GPU at present prints a warning that will muddle
# the console
# see
# alternatively, use
# system < "cpu"
system < torch_device(if (cuda_is_available()) "cuda" else "cpu")
internet < mannequin("gru", 1, 32)
internet < internet$to(system = system)
In any case these RNN specifics, the coaching course of is totally normal.
optimizer < optim_adam(internet$parameters, lr = 0.001)
num_epochs < 30
train_batch < perform(b) {
optimizer$zero_grad()
output < internet(b$x$to(system = system))
goal < b$y$to(system = system)
loss < nnf_mse_loss(output, goal)
loss$backward()
optimizer$step()
loss$merchandise()
}
valid_batch < perform(b) {
output < internet(b$x$to(system = system))
goal < b$y$to(system = system)
loss < nnf_mse_loss(output, goal)
loss$merchandise()
}
for (epoch in 1:num_epochs) {
internet$practice()
train_loss < c()
coro::loop(for (b in train_dl) {
loss <train_batch(b)
train_loss < c(train_loss, loss)
})
cat(sprintf("nEpoch %d, coaching: loss: %3.5f n", epoch, mean(train_loss)))
internet$eval()
valid_loss < c()
coro::loop(for (b in valid_dl) {
loss < valid_batch(b)
valid_loss < c(valid_loss, loss)
})
cat(sprintf("nEpoch %d, validation: loss: %3.5f n", epoch, mean(valid_loss)))
}
Epoch 1, coaching: loss: 0.21908
Epoch 1, validation: loss: 0.05125
Epoch 2, coaching: loss: 0.03245
Epoch 2, validation: loss: 0.03391
Epoch 3, coaching: loss: 0.02346
Epoch 3, validation: loss: 0.02321
Epoch 4, coaching: loss: 0.01823
Epoch 4, validation: loss: 0.01838
Epoch 5, coaching: loss: 0.01522
Epoch 5, validation: loss: 0.01560
Epoch 6, coaching: loss: 0.01315
Epoch 6, validation: loss: 0.01374
Epoch 7, coaching: loss: 0.01205
Epoch 7, validation: loss: 0.01200
Epoch 8, coaching: loss: 0.01155
Epoch 8, validation: loss: 0.01157
Epoch 9, coaching: loss: 0.01118
Epoch 9, validation: loss: 0.01096
Epoch 10, coaching: loss: 0.01070
Epoch 10, validation: loss: 0.01132
Epoch 11, coaching: loss: 0.01003
Epoch 11, validation: loss: 0.01150
Epoch 12, coaching: loss: 0.00943
Epoch 12, validation: loss: 0.01106
Epoch 13, coaching: loss: 0.00922
Epoch 13, validation: loss: 0.01069
Epoch 14, coaching: loss: 0.00862
Epoch 14, validation: loss: 0.01125
Epoch 15, coaching: loss: 0.00842
Epoch 15, validation: loss: 0.01095
Epoch 16, coaching: loss: 0.00820
Epoch 16, validation: loss: 0.00975
Epoch 17, coaching: loss: 0.00802
Epoch 17, validation: loss: 0.01120
Epoch 18, coaching: loss: 0.00781
Epoch 18, validation: loss: 0.00990
Epoch 19, coaching: loss: 0.00757
Epoch 19, validation: loss: 0.01017
Epoch 20, coaching: loss: 0.00735
Epoch 20, validation: loss: 0.00932
Epoch 21, coaching: loss: 0.00723
Epoch 21, validation: loss: 0.00901
Epoch 22, coaching: loss: 0.00708
Epoch 22, validation: loss: 0.00890
Epoch 23, coaching: loss: 0.00676
Epoch 23, validation: loss: 0.00914
Epoch 24, coaching: loss: 0.00666
Epoch 24, validation: loss: 0.00922
Epoch 25, coaching: loss: 0.00644
Epoch 25, validation: loss: 0.00869
Epoch 26, coaching: loss: 0.00620
Epoch 26, validation: loss: 0.00902
Epoch 27, coaching: loss: 0.00588
Epoch 27, validation: loss: 0.00896
Epoch 28, coaching: loss: 0.00563
Epoch 28, validation: loss: 0.00886
Epoch 29, coaching: loss: 0.00547
Epoch 29, validation: loss: 0.00895
Epoch 30, coaching: loss: 0.00523
Epoch 30, validation: loss: 0.00935
Loss decreases shortly, and we don’t appear to be overfitting on the validation set.
Numbers are fairly summary, although. So, we’ll use the take a look at set to see how the forecast truly appears.
Right here is the forecast for January, 2014, thirty minutes at a time.
internet$eval()
preds < rep(NA, n_timesteps)
coro::loop(for (b in test_dl) {
output < internet(b$x$to(system = system))
preds < c(preds, output %>% as.numeric())
})
vic_elec_jan_2014 < vic_elec %>%
filter(year(Date) == 2014, month(Date) == 1) %>%
choose(Demand)
preds_ts < vic_elec_jan_2014 %>%
add_column(forecast = preds * train_sd + train_mean) %>%
pivot_longer(Time) %>%
update_tsibble(key = title)
preds_ts %>%
autoplot() +
scale_colour_manual(values = c("#08c5d1", "#00353f")) +
theme_minimal()
General, the forecast is superb, however it’s fascinating to see how the forecast “regularizes” essentially the most excessive peaks. This type of “regression to the imply” will probably be seen way more strongly in later setups, after we attempt to forecast additional into the long run.
Can we use our present structure for multistep prediction? We will.
One factor we will do is feed again the present prediction, that’s, append it to the enter sequence as quickly as it’s out there. Successfully thus, for every batch merchandise, we receive a sequence of predictions in a loop.
We’ll attempt to forecast 336 time steps, that’s, an entire week.
n_forecast < 2 * 24 * 7
test_preds < vector(mode = "listing", size = length(test_dl))
i < 1
coro::loop(for (b in test_dl) {
enter < b$x
output < internet(enter$to(system = system))
preds < as.numeric(output)
for(j in 2:n_forecast) {
enter < torch_cat(list(enter[ , 2:length(input), ], output$view(c(1, 1, 1))), dim = 2)
output < internet(enter$to(system = system))
preds < c(preds, as.numeric(output))
}
test_preds[[i]] < preds
i << i + 1
})
For visualization, let’s decide three nonoverlapping sequences.
test_pred1 < test_preds[[1]]
test_pred1 < c(rep(NA, n_timesteps), test_pred1, rep(NA, nrow(vic_elec_jan_2014)  n_timesteps  n_forecast))
test_pred2 < test_preds[[408]]
test_pred2 < c(rep(NA, n_timesteps + 407), test_pred2, rep(NA, nrow(vic_elec_jan_2014)  407  n_timesteps  n_forecast))
test_pred3 < test_preds[[817]]
test_pred3 < c(rep(NA, nrow(vic_elec_jan_2014)  n_forecast), test_pred3)
preds_ts < vic_elec %>%
filter(year(Date) == 2014, month(Date) == 1) %>%
choose(Demand) %>%
add_column(
iterative_ex_1 = test_pred1 * train_sd + train_mean,
iterative_ex_2 = test_pred2 * train_sd + train_mean,
iterative_ex_3 = test_pred3 * train_sd + train_mean) %>%
pivot_longer(Time) %>%
update_tsibble(key = title)
preds_ts %>%
autoplot() +
scale_colour_manual(values = c("#08c5d1", "#00353f", "#ffbf66", "#d46f4d")) +
theme_minimal()
Even with this very primary forecasting method, the diurnal rhythm is preserved, albeit in a strongly smoothed kind. There even is an obvious dayofweek periodicity within the forecast. We do see, nevertheless, very robust regression to the imply, even in loop situations the place the community was “primed” with the next enter sequence.
Hopefully this publish supplied a helpful introduction to time collection forecasting with torch
. Evidently, we picked a difficult time collection – difficult, that’s, for no less than two causes:

To appropriately issue within the pattern, exterior data is required: exterior data in type of a temperature forecast, which, “in actuality,” can be simply obtainable.

Along with the extremely essential pattern element, the info are characterised by a number of ranges of seasonality.
Of those, the latter is much less of an issue for the strategies we’re working with right here. If we discovered that some stage of seasonality went undetected, we might attempt to adapt the present configuration in numerous uncomplicated methods:

Use an LSTM as an alternative of a GRU. In principle, LSTM ought to higher have the ability to seize extra lowerfrequency parts on account of its secondary storage, the cell state.

Stack a number of layers of GRU/LSTM. In principle, this could permit for studying a hierarchy of temporal options, analogously to what we see in a convolutional neural community.
To deal with the previous impediment, greater adjustments to the structure can be wanted. We might try to try this in a later, “bonus,” publish. However within the upcoming installments, we’ll first dive into oftenused strategies for sequence prediction, additionally porting to numerical time collection issues which can be generally carried out in pure language processing.
Thanks for studying!
Picture by Nick Dunn on Unsplash
Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. 2016. Deep Studying. MIT Press.