That is the ultimate publish in a fourpart introduction to timeseries forecasting with torch
. These posts have been the story of a quest for multiplestep prediction, and by now, we’ve seen three completely different approaches: forecasting in a loop, incorporating a multilayer perceptron (MLP), and sequencetosequence fashions. Right here’s a fast recap.

As one ought to when one units out for an adventurous journey, we began with an indepth research of the instruments at our disposal: recurrent neural networks (RNNs). We educated a mannequin to foretell the very subsequent commentary in line, after which, considered a intelligent hack: How about we use this for multistep prediction, feeding again particular person predictions in a loop? The end result , it turned out, was fairly acceptable.

Then, the journey actually began. We constructed our first mannequin “natively” for multistep prediction, relieving the RNN a little bit of its workload and involving a second participant, a tinyish MLP. Now, it was the MLP’s activity to challenge RNN output to a number of time factors sooner or later. Though outcomes had been fairly passable, we didn’t cease there.

As a substitute, we utilized to numerical time sequence a method generally utilized in pure language processing (NLP): sequencetosequence (seq2seq) prediction. Whereas forecast efficiency was not a lot completely different from the earlier case, we discovered the method to be extra intuitively interesting, because it displays the causal relationship between successive forecasts.
At this time we’ll enrich the seq2seq strategy by including a brand new element: the consideration module. Initially launched round 2014, consideration mechanisms have gained huge traction, a lot so {that a} latest paper title begins out “Consideration is Not All You Want”.
The thought is the next.
Within the basic encoderdecoder setup, the decoder will get “primed” with an encoder abstract only a single time: the time it begins its forecasting loop. From then on, it’s by itself. With consideration, nonetheless, it will get to see the entire sequence of encoder outputs once more each time it forecasts a brand new worth. What’s extra, each time, it will get to zoom in on these outputs that appear related for the present prediction step.
This can be a significantly helpful technique in translation: In producing the subsequent phrase, a mannequin might want to know what a part of the supply sentence to concentrate on. How a lot the method helps with numerical sequences, in distinction, will possible depend upon the options of the sequence in query.
As earlier than, we work with vic_elec
, however this time, we partly deviate from the best way we used to make use of it. With the unique, bihourly dataset, coaching the present mannequin takes a very long time, longer than readers will wish to wait when experimenting. So as a substitute, we mixture observations by day. With a view to have sufficient knowledge, we prepare on years 2012 and 2013, reserving 2014 for validation in addition to posttraining inspection.
We’ll try and forecast demand as much as fourteen days forward. How lengthy, then, needs to be the enter sequences? This can be a matter of experimentation; all of the extra so now that we’re including within the consideration mechanism. (I believe that it won’t deal with very lengthy sequences so effectively).
Beneath, we go together with fourteen days for enter size, too, however that will not essentially be the very best selection for this sequence.
n_timesteps < 7 * 2
n_forecast < 7 * 2
elec_dataset < dataset(
title = "elec_dataset",
initialize = operate(x, n_timesteps, sample_frac = 1) {
self$n_timesteps < n_timesteps
self$x < torch_tensor((x  train_mean) / train_sd)
n < size(self$x)  self$n_timesteps  1
self$begins < type(pattern.int(
n = n,
dimension = n * sample_frac
))
},
.getitem = operate(i) {
begin < self$begins[i]
finish < begin + self$n_timesteps  1
lag < 1
listing(
x = self$x[start:end],
y = self$x[(start+lag):(end+lag)]$squeeze(2)
)
},
.size = operate() {
size(self$begins)
}
)
batch_size < 32
train_ds < elec_dataset(elec_train, n_timesteps)
train_dl < train_ds %>% dataloader(batch_size = batch_size, shuffle = TRUE)
valid_ds < elec_dataset(elec_valid, n_timesteps)
valid_dl < valid_ds %>% dataloader(batch_size = batch_size)
test_ds < elec_dataset(elec_test, n_timesteps)
test_dl < test_ds %>% dataloader(batch_size = 1)
Mannequinwise, we once more encounter the three modules acquainted from the earlier publish: encoder, decoder, and toplevel seq2seq module. Nonetheless, there may be a further element: the consideration module, utilized by the decoder to acquire consideration weights.
Encoder
The encoder nonetheless works the identical manner. It wraps an RNN, and returns the ultimate state.
encoder_module < nn_module(
initialize = operate(kind, input_size, hidden_size, num_layers = 1, dropout = 0) {
self$kind < kind
self$rnn < if (self$kind == "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
)
}
},
ahead = operate(x) {
# return outputs for all timesteps, in addition to lasttimestep states for all layers
x %>% self$rnn()
}
)
Consideration module
In primary seq2seq, each time it needed to generate a brand new worth, the decoder took into consideration two issues: its prior state, and the earlier output generated. In an attentionenriched setup, the decoder moreover receives the entire output from the encoder. In deciding what subset of that output ought to matter, it will get assist from a brand new agent, the eye module.
This, then, is the eye module’s raison d’être: Given present decoder state and effectively as full encoder outputs, get hold of a weighting of these outputs indicative of how related they’re to what the decoder is at present as much as. This process leads to the socalled consideration weights: a normalized rating, for every time step within the encoding, that quantify their respective significance.
Consideration could also be applied in a variety of other ways. Right here, we present two implementation choices, one additive, and one multiplicative.
Additive consideration
In additive consideration, encoder outputs and decoder state are generally both added or concatenated (we select to do the latter, under). The ensuing tensor is run by way of a linear layer, and a softmax is utilized for normalization.
attention_module_additive < nn_module(
initialize = operate(hidden_dim, attention_size) {
self$consideration < nn_linear(2 * hidden_dim, attention_size)
},
ahead = operate(state, encoder_outputs) {
# operate argument shapes
# encoder_outputs: (bs, timesteps, hidden_dim)
# state: (1, bs, hidden_dim)
# multiplex state to permit for concatenation (dimensions 1 and a pair of should agree)
seq_len < dim(encoder_outputs)[2]
# ensuing form: (bs, timesteps, hidden_dim)
state_rep < state$permute(c(2, 1, 3))$repeat_interleave(seq_len, 2)
# concatenate alongside characteristic dimension
concat < torch_cat(listing(state_rep, encoder_outputs), dim = 3)
# run by way of linear layer with tanh
# ensuing form: (bs, timesteps, attention_size)
scores < self$consideration(concat) %>%
torch_tanh()
# sum over consideration dimension and normalize
# ensuing form: (bs, timesteps)
attention_weights < scores %>%
torch_sum(dim = 3) %>%
nnf_softmax(dim = 2)
# a normalized rating for each supply token
attention_weights
}
)
Multiplicative consideration
In multiplicative consideration, scores are obtained by computing dot merchandise between decoder state and all the encoder outputs. Right here too, a softmax is then used for normalization.
attention_module_multiplicative < nn_module(
initialize = operate() {
NULL
},
ahead = operate(state, encoder_outputs) {
# operate argument shapes
# encoder_outputs: (bs, timesteps, hidden_dim)
# state: (1, bs, hidden_dim)
# permit for matrix multiplication with encoder_outputs
state < state$permute(c(2, 3, 1))
# put together for scaling by variety of options
d < torch_tensor(dim(encoder_outputs)[3], dtype = torch_float())
# scaled dot merchandise between state and outputs
# ensuing form: (bs, timesteps, 1)
scores < torch_bmm(encoder_outputs, state) %>%
torch_div(torch_sqrt(d))
# normalize
# ensuing form: (bs, timesteps)
attention_weights < scores$squeeze(3) %>%
nnf_softmax(dim = 2)
# a normalized rating for each supply token
attention_weights
}
)
Decoder
As soon as consideration weights have been computed, their precise utility is dealt with by the decoder. Concretely, the tactic in query, weighted_encoder_outputs()
, computes a product of weights and encoder outputs, ensuring that every output could have applicable affect.
The remainder of the motion then occurs in ahead()
. A concatenation of weighted encoder outputs (typically referred to as “context”) and present enter is run by way of an RNN. Then, an ensemble of RNN output, context, and enter is handed to an MLP. Lastly, each RNN state and present prediction are returned.
decoder_module < nn_module(
initialize = operate(kind, input_size, hidden_size, attention_type, attention_size = 8, num_layers = 1) {
self$kind < kind
self$rnn < if (self$kind == "gru") {
nn_gru(
input_size = input_size,
hidden_size = hidden_size,
num_layers = num_layers,
batch_first = TRUE
)
} else {
nn_lstm(
input_size = input_size,
hidden_size = hidden_size,
num_layers = num_layers,
batch_first = TRUE
)
}
self$linear < nn_linear(2 * hidden_size + 1, 1)
self$consideration < if (attention_type == "multiplicative") attention_module_multiplicative()
else attention_module_additive(hidden_size, attention_size)
},
weighted_encoder_outputs = operate(state, encoder_outputs) {
# encoder_outputs is (bs, timesteps, hidden_dim)
# state is (1, bs, hidden_dim)
# ensuing form: (bs * timesteps)
attention_weights < self$consideration(state, encoder_outputs)
# ensuing form: (bs, 1, seq_len)
attention_weights < attention_weights$unsqueeze(2)
# ensuing form: (bs, 1, hidden_size)
weighted_encoder_outputs < torch_bmm(attention_weights, encoder_outputs)
weighted_encoder_outputs
},
ahead = operate(x, state, encoder_outputs) {
# encoder_outputs is (bs, timesteps, hidden_dim)
# state is (1, bs, hidden_dim)
# ensuing form: (bs, 1, hidden_size)
context < self$weighted_encoder_outputs(state, encoder_outputs)
# concatenate enter and context
# NOTE: this repeating is finished to compensate for the absence of an embedding module
# that, in NLP, would give x the next proportion within the concatenation
x_rep < x$repeat_interleave(dim(context)[3], 3)
rnn_input < torch_cat(listing(x_rep, context), dim = 3)
# ensuing shapes: (bs, 1, hidden_size) and (1, bs, hidden_size)
rnn_out < self$rnn(rnn_input, state)
rnn_output < rnn_out[[1]]
next_hidden < rnn_out[[2]]
mlp_input < torch_cat(listing(rnn_output$squeeze(2), context$squeeze(2), x$squeeze(2)), dim = 2)
output < self$linear(mlp_input)
# shapes: (bs, 1) and (1, bs, hidden_size)
listing(output, next_hidden)
}
)
seq2seq
module
The seq2seq
module is mainly unchanged (other than the truth that now, it permits for consideration module configuration). For an in depth clarification of what occurs right here, please seek the advice of the earlier publish.
seq2seq_module < nn_module(
initialize = operate(kind, input_size, hidden_size, attention_type, attention_size, n_forecast,
num_layers = 1, encoder_dropout = 0) {
self$encoder < encoder_module(kind = kind, input_size = input_size, hidden_size = hidden_size,
num_layers, encoder_dropout)
self$decoder < decoder_module(kind = kind, input_size = 2 * hidden_size, hidden_size = hidden_size,
attention_type = attention_type, attention_size = attention_size, num_layers)
self$n_forecast < n_forecast
},
ahead = operate(x, y, teacher_forcing_ratio) {
outputs < torch_zeros(dim(x)[1], self$n_forecast)
encoded < self$encoder(x)
encoder_outputs < encoded[[1]]
hidden < encoded[[2]]
# listing of (batch_size, 1), (1, batch_size, hidden_size)
out < self$decoder(x[ , n_timesteps, , drop = FALSE], hidden, encoder_outputs)
# (batch_size, 1)
pred < out[[1]]
# (1, batch_size, hidden_size)
state < out[[2]]
outputs[ , 1] < pred$squeeze(2)
for (t in 2:self$n_forecast) {
teacher_forcing < runif(1) < teacher_forcing_ratio
enter < if (teacher_forcing == TRUE) y[ , t  1, drop = FALSE] else pred
enter < enter$unsqueeze(3)
out < self$decoder(enter, state, encoder_outputs)
pred < out[[1]]
state < out[[2]]
outputs[ , t] < pred$squeeze(2)
}
outputs
}
)
When instantiating the toplevel mannequin, we now have a further selection: that between additive and multiplicative consideration. Within the “accuracy” sense of efficiency, my checks didn’t present any variations. Nonetheless, the multiplicative variant is loads sooner.
internet < seq2seq_module("gru", input_size = 1, hidden_size = 32, attention_type = "multiplicative",
attention_size = 8, n_forecast = n_forecast)
Identical to final time, in mannequin coaching, we get to decide on the diploma of trainer forcing. Beneath, we go together with a fraction of 0.0, that’s, no forcing in any respect.
optimizer < optim_adam(internet$parameters, lr = 0.001)
num_epochs < 1000
train_batch < operate(b, teacher_forcing_ratio) {
optimizer$zero_grad()
output < internet(b$x, b$y, teacher_forcing_ratio)
goal < b$y
loss < nnf_mse_loss(output, goal[ , 1:(dim(output)[2])])
loss$backward()
optimizer$step()
loss$merchandise()
}
valid_batch < operate(b, teacher_forcing_ratio = 0) {
output < internet(b$x, b$y, teacher_forcing_ratio)
goal < b$y
loss < nnf_mse_loss(output, goal[ , 1:(dim(output)[2])])
loss$merchandise()
}
for (epoch in 1:num_epochs) {
internet$prepare()
train_loss < c()
coro::loop(for (b in train_dl) {
loss <train_batch(b, teacher_forcing_ratio = 0.0)
train_loss < c(train_loss, loss)
})
cat(sprintf("nEpoch %d, coaching: loss: %3.5f n", epoch, imply(train_loss)))
internet$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, imply(valid_loss)))
}
# Epoch 1, coaching: loss: 0.83752
# Epoch 1, validation: loss: 0.83167
# Epoch 2, coaching: loss: 0.72803
# Epoch 2, validation: loss: 0.80804
# ...
# ...
# Epoch 99, coaching: loss: 0.10385
# Epoch 99, validation: loss: 0.21259
# Epoch 100, coaching: loss: 0.10396
# Epoch 100, validation: loss: 0.20975
For visible inspection, we choose a couple of forecasts from the check set.
internet$eval()
test_preds < vector(mode = "listing", size = size(test_dl))
i < 1
vic_elec_test < vic_elec_daily %>%
filter(yr(Date) == 2014, month(Date) %in% 1:4)
coro::loop(for (b in test_dl) {
output < internet(b$x, b$y, teacher_forcing_ratio = 0)
preds < as.numeric(output)
test_preds[[i]] < preds
i << i + 1
})
test_pred1 < test_preds[[1]]
test_pred1 < c(rep(NA, n_timesteps), test_pred1, rep(NA, nrow(vic_elec_test)  n_timesteps  n_forecast))
test_pred2 < test_preds[[21]]
test_pred2 < c(rep(NA, n_timesteps + 20), test_pred2, rep(NA, nrow(vic_elec_test)  20  n_timesteps  n_forecast))
test_pred3 < test_preds[[41]]
test_pred3 < c(rep(NA, n_timesteps + 40), test_pred3, rep(NA, nrow(vic_elec_test)  40  n_timesteps  n_forecast))
test_pred4 < test_preds[[61]]
test_pred4 < c(rep(NA, n_timesteps + 60), test_pred4, rep(NA, nrow(vic_elec_test)  60  n_timesteps  n_forecast))
test_pred5 < test_preds[[81]]
test_pred5 < c(rep(NA, n_timesteps + 80), test_pred5, rep(NA, nrow(vic_elec_test)  80  n_timesteps  n_forecast))
preds_ts < vic_elec_test %>%
choose(Demand, Date) %>%
add_column(
ex_1 = test_pred1 * train_sd + train_mean,
ex_2 = test_pred2 * train_sd + train_mean,
ex_3 = test_pred3 * train_sd + train_mean,
ex_4 = test_pred4 * train_sd + train_mean,
ex_5 = test_pred5 * train_sd + train_mean) %>%
pivot_longer(Date) %>%
update_tsibble(key = title)
preds_ts %>%
autoplot() +
scale_color_hue(h = c(80, 300), l = 70) +
theme_minimal()
We are able to’t instantly evaluate efficiency right here to that of earlier fashions in our sequence, as we’ve pragmatically redefined the duty. The principle aim, nonetheless, has been to introduce the idea of consideration. Particularly, methods to manually implement the method – one thing that, when you’ve understood the idea, it’s possible you’ll by no means should do in apply. As a substitute, you’d possible make use of present instruments that include torch
(multihead consideration and transformer modules), instruments we might introduce in a future “season” of this sequence.
Thanks for studying!
Photograph by David Clode on Unsplash