This put up is the primary in a free collection exploring forecasting of spatially-determined knowledge over time. By spatially-determined I imply that regardless of the portions we’re making an attempt to foretell – be they univariate or multivariate time collection, of spatial dimensionality or not – the enter knowledge are given on a spatial grid.

For instance, the enter could possibly be atmospheric measurements, similar to sea floor temperature or strain, given at some set of latitudes and longitudes. The goal to be predicted may then span that very same (or one other) grid. Alternatively, it could possibly be a univariate time collection, like a meteorological index.

However wait a second, you could be pondering. For time-series prediction, we now have that time-honored set of recurrent architectures (e.g., LSTM, GRU), proper? Proper. We do; however, as soon as we feed spatial knowledge to an RNN, treating completely different areas as completely different enter options, we lose a vital structural relationship. Importantly, we have to function in each house and time. We would like each: recurrence relations and convolutional filters. Enter convolutional RNNs.

What to anticipate from this put up

Right this moment, we gained’t soar into real-world functions simply but. As an alternative, we’ll take our time to construct a convolutional LSTM (henceforth: convLSTM) in torch. For one, we now have to – there isn’t any official PyTorch implementation.

What’s extra, this put up can function an introduction to constructing your personal modules. That is one thing you could be aware of from Keras or not – relying on whether or not you’ve used customized fashions or somewhat, most well-liked the declarative outline -> compile -> match fashion. (Sure, I’m implying there’s some switch occurring if one involves torch from Keras customized coaching. Syntactic and semantic particulars could also be completely different, however each share the object-oriented fashion that permits for nice flexibility and management.)

Final however not least, we’ll additionally use this as a hands-on expertise with RNN architectures (the LSTM, particularly). Whereas the overall idea of recurrence could also be straightforward to understand, it’s not essentially self-evident how these architectures ought to, or may, be coded. Personally, I discover that impartial of the framework used, RNN-related documentation leaves me confused. What precisely is being returned from calling an LSTM, or a GRU? (In Keras this relies on the way you’ve outlined the layer in query.) I believe that when we’ve determined what we need to return, the precise code gained’t be that sophisticated. Consequently, we’ll take a detour clarifying what it’s that torch and Keras are giving us. Implementing our convLSTM might be much more easy thereafter.

A torch convLSTM

The code mentioned right here could also be discovered on GitHub. (Relying on whenever you’re studying this, the code in that repository could have advanced although.)

My place to begin was one of many PyTorch implementations discovered on the web, specifically, this one. Should you seek for “PyTorch convGRU” or “PyTorch convLSTM”, you’ll discover beautiful discrepancies in how these are realized – discrepancies not simply in syntax and/or engineering ambition, however on the semantic degree, proper on the heart of what the architectures could also be anticipated to do. As they are saying, let the customer beware. (Relating to the implementation I ended up porting, I’m assured that whereas quite a few optimizations might be potential, the essential mechanism matches my expectations.)

What do I count on? Let’s method this job in a top-down approach.

Enter and output

The convLSTM’s enter might be a time collection of spatial knowledge, every commentary being of measurement (time steps, channels, top, width).

Examine this with the standard RNN enter format, be it in torch or Keras. In each frameworks, RNNs count on tensors of measurement (timesteps, input_dim). input_dim is (1) for univariate time collection and better than (1) for multivariate ones. Conceptually, we could match this to convLSTM’s channels dimension: There could possibly be a single channel, for temperature, say – or there could possibly be a number of, similar to for strain, temperature, and humidity. The 2 extra dimensions present in convLSTM, top and width, are spatial indexes into the information.

In sum, we would like to have the ability to cross knowledge that:

• encompass a number of options,

• evolve in time, and

• are listed in two spatial dimensions.

How in regards to the output? We would like to have the ability to return forecasts for as many time steps as we now have within the enter sequence. That is one thing that torch RNNs do by default, whereas Keras equivalents don’t. (It’s a must to cross return_sequences = TRUE to acquire that impact.) If we’re curious about predictions for only a single time limit, we will all the time choose the final time step within the output tensor.

Nonetheless, with RNNs, it’s not all about outputs. RNN architectures additionally carry via hidden states.

What are hidden states? I fastidiously phrased that sentence to be as normal as potential – intentionally circling across the confusion that, for my part, typically arises at this level. We’ll try and clear up a few of that confusion in a second, however let’s first end our high-level necessities specification.

We would like our convLSTM to be usable in numerous contexts and functions. Numerous architectures exist that make use of hidden states, most prominently maybe, encoder-decoder architectures. Thus, we would like our convLSTM to return these as properly. Once more, that is one thing a torch LSTM does by default, whereas in Keras it’s achieved utilizing return_state = TRUE.

Now although, it truly is time for that interlude. We’ll type out the methods issues are known as by each torch and Keras, and examine what you get again from their respective GRUs and LSTMs.

Interlude: Outputs, states, hidden values … what’s what?

For this to stay an interlude, I summarize findings on a excessive degree. The code snippets within the appendix present easy methods to arrive at these outcomes. Closely commented, they probe return values from each Keras and torch GRUs and LSTMs. Operating these will make the upcoming summaries appear so much much less summary.

First, let’s take a look at the methods you create an LSTM in each frameworks. (I’ll usually use LSTM because the “prototypical RNN instance”, and simply point out GRUs when there are variations important within the context in query.)

In Keras, to create an LSTM you could write one thing like this:

lstm <- layer_lstm(models = 1)

The torch equal can be:

lstm <- nn_lstm(
  input_size = 2, # variety of enter options
  hidden_size = 1 # variety of hidden (and output!) options
)

Don’t concentrate on torch‘s input_size parameter for this dialogue. (It’s the variety of options within the enter tensor.) The parallel happens between Keras’ models and torch’s hidden_size. Should you’ve been utilizing Keras, you’re in all probability pondering of models because the factor that determines output measurement (equivalently, the variety of options within the output). So when torch lets us arrive on the identical consequence utilizing hidden_size, what does that imply? It signifies that by some means we’re specifying the identical factor, utilizing completely different terminology. And it does make sense, since at each time step present enter and former hidden state are added:

[
mathbf{h}_t = mathbf{W}_{x}mathbf{x}_t + mathbf{W}_{h}mathbf{h}_{t-1}
]

Now, about these hidden states.

When a Keras LSTM is outlined with return_state = TRUE, its return worth is a construction of three entities known as output, reminiscence state, and carry state. In torch, the identical entities are known as output, hidden state, and cell state. (In torch, we all the time get all of them.)

So are we coping with three various kinds of entities? We aren’t.

The cell, or carry state is that particular factor that units aside LSTMs from GRUs deemed liable for the “lengthy” in “lengthy short-term reminiscence”. Technically, it could possibly be reported to the person in any respect deadlines; as we’ll see shortly although, it’s not.

What about outputs and hidden, or reminiscence states? Confusingly, these actually are the identical factor. Recall that for every enter merchandise within the enter sequence, we’re combining it with the earlier state, leading to a brand new state, to be made used of within the subsequent step:

[
mathbf{h}_t = mathbf{W}_{x}mathbf{x}_t + mathbf{W}_{h}mathbf{h}_{t-1}
]

Now, say that we’re curious about taking a look at simply the ultimate time step – that’s, the default output of a Keras LSTM. From that standpoint, we will contemplate these intermediate computations as “hidden”. Seen like that, output and hidden states really feel completely different.

Nonetheless, we will additionally request to see the outputs for each time step. If we achieve this, there isn’t any distinction – the outputs (plural) equal the hidden states. This may be verified utilizing the code within the appendix.

Thus, of the three issues returned by an LSTM, two are actually the identical. How in regards to the GRU, then? As there isn’t any “cell state”, we actually have only one sort of factor left over – name it outputs or hidden states.

Let’s summarize this in a desk.

Now, about that public vs. non-public distinction. In each frameworks, we will receive outputs (hidden states) for each time step. The cell state, nevertheless, we will entry just for the final time step. That is purely an implementation choice. As we’ll see when constructing our personal recurrent module, there are not any obstacles inherent in protecting monitor of cell states and passing them again to the person.

Should you dislike the pragmatism of this distinction, you possibly can all the time go along with the maths. When a brand new cell state has been computed (primarily based on prior cell state, enter, overlook, and cell gates – the specifics of which we’re not going to get into right here), it’s remodeled to the hidden (a.ok.a. output) state making use of yet one more, specifically, the output gate:

[
h_t = o_t odot tanh(c_t)
]

Undoubtedly, then, hidden state (output, resp.) builds on cell state, including extra modeling energy.

Now it’s time to get again to our unique purpose and construct that convLSTM. First although, let’s summarize the return values obtainable from torch and Keras.

convLSTM, the plan

In each torch and Keras RNN architectures, single time steps are processed by corresponding Cell lessons: There’s an LSTM Cell matching the LSTM, a GRU Cell matching the GRU, and so forth. We do the identical for ConvLSTM. In convlstm_cell(), we first outline what ought to occur to a single commentary; then in convlstm(), we construct up the recurrence logic.

As soon as we’re finished, we create a dummy dataset, as reduced-to-the-essentials as may be. With extra complicated datasets, even synthetic ones, chances are high that if we don’t see any coaching progress, there are lots of of potential explanations. We would like a sanity verify that, if failed, leaves no excuses. Real looking functions are left to future posts.

A single step: convlstm_cell

Our convlstm_cell’s constructor takes arguments input_dim , hidden_dim, and bias, similar to a torch LSTM Cell.

However we’re processing two-dimensional enter knowledge. As an alternative of the standard affine mixture of recent enter and former state, we use a convolution of kernel measurement kernel_size. Inside convlstm_cell, it’s self$conv that takes care of this.

Be aware how the channels dimension, which within the unique enter knowledge would correspond to completely different variables, is creatively used to consolidate 4 convolutions into one: Every channel output might be handed to only one of many 4 cell gates. As soon as in possession of the convolution output, ahead() applies the gate logic, ensuing within the two sorts of states it must ship again to the caller.

library(torch)
library(zeallot)

convlstm_cell <- nn_module(
  
  initialize = operate(input_dim, hidden_dim, kernel_size, bias) {
    
    self$hidden_dim <- hidden_dim
    
    padding <- kernel_size %/% 2
    
    self$conv <- nn_conv2d(
      in_channels = input_dim + self$hidden_dim,
      # for every of enter, overlook, output, and cell gates
      out_channels = 4 * self$hidden_dim,
      kernel_size = kernel_size,
      padding = padding,
      bias = bias
    )
  },
  
  ahead = operate(x, prev_states) {

    c(h_prev, c_prev) %<-% prev_states
    
    mixed <- torch_cat(record(x, h_prev), dim = 2)  # concatenate alongside channel axis
    combined_conv <- self$conv(mixed)
    c(cc_i, cc_f, cc_o, cc_g) %<-% torch_split(combined_conv, self$hidden_dim, dim = 2)
    
    # enter, overlook, output, and cell gates (equivalent to torch's LSTM)
    i <- torch_sigmoid(cc_i)
    f <- torch_sigmoid(cc_f)
    o <- torch_sigmoid(cc_o)
    g <- torch_tanh(cc_g)
    
    # cell state
    c_next <- f * c_prev + i * g
    # hidden state
    h_next <- o * torch_tanh(c_next)
    
    record(h_next, c_next)
  },
  
  init_hidden = operate(batch_size, top, width) {
    
    record(
      torch_zeros(batch_size, self$hidden_dim, top, width, machine = self$conv$weight$machine),
      torch_zeros(batch_size, self$hidden_dim, top, width, machine = self$conv$weight$machine))
  }
)

Now convlstm_cell needs to be known as for each time step. That is finished by convlstm.

Iteration over time steps: convlstm

A convlstm could encompass a number of layers, similar to a torch LSTM. For every layer, we’re in a position to specify hidden and kernel sizes individually.

Throughout initialization, every layer will get its personal convlstm_cell. On name, convlstm executes two loops. The outer one iterates over layers. On the finish of every iteration, we retailer the ultimate pair (hidden state, cell state) for later reporting. The interior loop runs over enter sequences, calling convlstm_cell at every time step.

We additionally hold monitor of intermediate outputs, so we’ll have the ability to return the whole record of hidden_states seen through the course of. In contrast to a torch LSTM, we do that for each layer.

convlstm <- nn_module(
  
  # hidden_dims and kernel_sizes are vectors, with one ingredient for every layer in n_layers
  initialize = operate(input_dim, hidden_dims, kernel_sizes, n_layers, bias = TRUE) {
 
    self$n_layers <- n_layers
    
    self$cell_list <- nn_module_list()
    
    for (i in 1:n_layers) {
      cur_input_dim <- if (i == 1) input_dim else hidden_dims[i - 1]
      self$cell_list$append(convlstm_cell(cur_input_dim, hidden_dims[i], kernel_sizes[i], bias))
    }
  },
  
  # we all the time assume batch-first
  ahead = operate(x) {
    
    c(batch_size, seq_len, num_channels, top, width) %<-% x$measurement()
   
    # initialize hidden states
    init_hidden <- vector(mode = "record", size = self$n_layers)
    for (i in 1:self$n_layers) {
      init_hidden[[i]] <- self$cell_list[[i]]$init_hidden(batch_size, top, width)
    }
    
    # record containing the outputs, of size seq_len, for every layer
    # this is identical as h, at every step within the sequence
    layer_output_list <- vector(mode = "record", size = self$n_layers)
    
    # record containing the final states (h, c) for every layer
    layer_state_list <- vector(mode = "record", size = self$n_layers)

    cur_layer_input <- x
    hidden_states <- init_hidden
    
    # loop over layers
    for (i in 1:self$n_layers) {
      
      # each layer's hidden state begins from 0 (non-stateful)
      c(h, c) %<-% hidden_states[[i]]
      # outputs, of size seq_len, for this layer
      # equivalently, record of h states for every time step
      output_sequence <- vector(mode = "record", size = seq_len)
      
      # loop over time steps
      for (t in 1:seq_len) {
        c(h, c) %<-% self$cell_list[[i]](cur_layer_input[ , t, , , ], record(h, c))
        # hold monitor of output (h) for each time step
        # h has dim (batch_size, hidden_size, top, width)
        output_sequence[[t]] <- h
      }

      # stack hs forever steps over seq_len dimension
      # stacked_outputs has dim (batch_size, seq_len, hidden_size, top, width)
      # identical as enter to ahead (x)
      stacked_outputs <- torch_stack(output_sequence, dim = 2)
      
      # cross the record of outputs (hs) to subsequent layer
      cur_layer_input <- stacked_outputs
      
      # hold monitor of record of outputs or this layer
      layer_output_list[[i]] <- stacked_outputs
      # hold monitor of final state for this layer
      layer_state_list[[i]] <- record(h, c)
    }
 
    record(layer_output_list, layer_state_list)
  }
    
)

Calling the convlstm

Let’s see the enter format anticipated by convlstm, and easy methods to entry its completely different outputs.

Right here is an acceptable enter tensor.

# batch_size, seq_len, channels, top, width
x <- torch_rand(c(2, 4, 3, 16, 16))

First we make use of a single layer.

mannequin <- convlstm(input_dim = 3, hidden_dims = 5, kernel_sizes = 3, n_layers = 1)

c(layer_outputs, layer_last_states) %<-% mannequin(x)

We get again an inventory of size two, which we instantly break up up into the 2 sorts of output returned: intermediate outputs from all layers, and remaining states (of each varieties) for the final layer.

With only a single layer, layer_outputs[[1]]holds all the layer’s intermediate outputs, stacked on dimension two.

dim(layer_outputs[[1]])
# [1]  2  4  5 16 16

layer_last_states[[1]]is an inventory of tensors, the primary of which holds the one layer’s remaining hidden state, and the second, its remaining cell state.

dim(layer_last_states[[1]][[1]])
# [1]  2  5 16 16
dim(layer_last_states[[1]][[2]])
# [1]  2  5 16 16

For comparability, that is how return values search for a multi-layer structure.

mannequin <- convlstm(input_dim = 3, hidden_dims = c(5, 5, 1), kernel_sizes = rep(3, 3), n_layers = 3)
c(layer_outputs, layer_last_states) %<-% mannequin(x)

# for every layer, tensor of measurement (batch_size, seq_len, hidden_size, top, width)
dim(layer_outputs[[1]])
# 2  4  5 16 16
dim(layer_outputs[[3]])
# 2  4  1 16 16

# record of two tensors for every layer
str(layer_last_states)
# Checklist of three
#  $ :Checklist of two
#   ..$ :Float [1:2, 1:5, 1:16, 1:16]
#   ..$ :Float [1:2, 1:5, 1:16, 1:16]
#  $ :Checklist of two
#   ..$ :Float [1:2, 1:5, 1:16, 1:16]
#   ..$ :Float [1:2, 1:5, 1:16, 1:16]
#  $ :Checklist of two
#   ..$ :Float [1:2, 1:1, 1:16, 1:16]
#   ..$ :Float [1:2, 1:1, 1:16, 1:16]

# h, of measurement (batch_size, hidden_size, top, width)
dim(layer_last_states[[3]][[1]])
# 2  1 16 16

# c, of measurement (batch_size, hidden_size, top, width)
dim(layer_last_states[[3]][[2]])
# 2  1 16 16

Now we wish to sanity-check this module with the simplest-possible dummy knowledge.

Sanity-checking the convlstm

We generate black-and-white “motion pictures” of diagonal beams successively translated in house.

Every sequence consists of six time steps, and every beam of six pixels. Only a single sequence is created manually. To create that one sequence, we begin from a single beam:

library(torchvision)

beams <- vector(mode = "record", size = 6)
beam <- torch_eye(6) %>% nnf_pad(c(6, 12, 12, 6)) # left, proper, high, backside
beams[[1]] <- beam

Utilizing torch_roll() , we create a sample the place this beam strikes up diagonally, and stack the person tensors alongside the timesteps dimension.

for (i in 2:6) {
  beams[[i]] <- torch_roll(beam, c(-(i-1),i-1), c(1, 2))
}

init_sequence <- torch_stack(beams, dim = 1)

That’s a single sequence. Because of torchvision::transform_random_affine(), we nearly effortlessly produce a dataset of 100 sequences. Shifting beams begin at random factors within the spatial body, however all of them share that upward-diagonal movement.

sequences <- vector(mode = "record", size = 100)
sequences[[1]] <- init_sequence

for (i in 2:100) {
  sequences[[i]] <- transform_random_affine(init_sequence, levels = 0, translate = c(0.5, 0.5))
}

enter <- torch_stack(sequences, dim = 1)

# add channels dimension
enter <- enter$unsqueeze(3)
dim(enter)
# [1] 100   6  1  24  24

That’s it for the uncooked knowledge. Now we nonetheless want a dataset and a dataloader. Of the six time steps, we use the primary 5 as enter and attempt to predict the final one.

dummy_ds <- dataset(
  
  initialize = operate(knowledge) {
    self$knowledge <- knowledge
  },
  
  .getitem = operate(i) {
    record(x = self$knowledge[i, 1:5, ..], y = self$knowledge[i, 6, ..])
  },
  
  .size = operate() {
    nrow(self$knowledge)
  }
)

ds <- dummy_ds(enter)
dl <- dataloader(ds, batch_size = 100)

Here’s a tiny-ish convLSTM, skilled for movement prediction:

mannequin <- convlstm(input_dim = 1, hidden_dims = c(64, 1), kernel_sizes = c(3, 3), n_layers = 2)

optimizer <- optim_adam(mannequin$parameters)

num_epochs <- 100

for (epoch in 1:num_epochs) {
  
  mannequin$practice()
  batch_losses <- c()
  
  for (b in enumerate(dl)) {
    
    optimizer$zero_grad()
    
    # last-time-step output from final layer
    preds <- mannequin(b$x)[[2]][[2]][[1]]
  
    loss <- nnf_mse_loss(preds, b$y)
    batch_losses <- c(batch_losses, loss$merchandise())
    
    loss$backward()
    optimizer$step()
  }
  
  if (epoch %% 10 == 0)
    cat(sprintf("nEpoch %d, coaching loss:%3fn", epoch, imply(batch_losses)))
}

Epoch 10, coaching loss:0.008522

Epoch 20, coaching loss:0.008079

Epoch 30, coaching loss:0.006187

Epoch 40, coaching loss:0.003828

Epoch 50, coaching loss:0.002322

Epoch 60, coaching loss:0.001594

Epoch 70, coaching loss:0.001376

Epoch 80, coaching loss:0.001258

Epoch 90, coaching loss:0.001218

Epoch 100, coaching loss:0.001171

Loss decreases, however that in itself is just not a assure the mannequin has discovered something. Has it? Let’s examine its forecast for the very first sequence and see.

For printing, I’m zooming in on the related area within the 24×24-pixel body. Right here is the bottom fact for time step six:

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  1  0  0  0  0  0  0
0  0  0  0  1  0  0  0  0  0
0  0  0  0  0  1  0  0  0  0
0  0  0  0  0  0  1  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

And right here is the forecast. This doesn’t look unhealthy in any respect, given there was neither experimentation nor tuning concerned.

       [,1]  [,2]  [,3]  [,4]  [,5]  [,6]  [,7]  [,8]  [,9] [,10]
 [1,]  0.00  0.00  0.00  0.00  0.00  0.00  0.00  0.00  0.00     0
 [2,] -0.02  0.36  0.01  0.06  0.00  0.00  0.00  0.00  0.00     0
 [3,]  0.00 -0.01  0.71  0.01  0.06  0.00  0.00  0.00  0.00     0
 [4,] -0.01  0.04  0.00  0.75  0.01  0.06  0.00  0.00  0.00     0
 [5,]  0.00 -0.01 -0.01 -0.01  0.75  0.01  0.06  0.00  0.00     0
 [6,]  0.00  0.01  0.00 -0.07 -0.01  0.75  0.01  0.06  0.00     0
 [7,]  0.00  0.01 -0.01 -0.01 -0.07 -0.01  0.75  0.01  0.06     0
 [8,]  0.00  0.00  0.01  0.00  0.00 -0.01  0.00  0.71  0.00     0
 [9,]  0.00  0.00  0.00  0.01  0.01  0.00  0.03 -0.01  0.37     0
[10,]  0.00  0.00  0.00  0.00  0.00  0.00 -0.01 -0.01 -0.01     0

This could suffice for a sanity verify. Should you made it until the top, thanks to your endurance! In one of the best case, you’ll have the ability to apply this structure (or an analogous one) to your personal knowledge – however even when not, I hope you’ve loved studying about torch mannequin coding and/or RNN weirdness 😉

I, for one, am definitely wanting ahead to exploring convLSTMs on real-world issues within the close to future. Thanks for studying!

Appendix

This appendix incorporates the code used to create tables 1 and a couple of above.

Keras

LSTM

library(keras)

# batch of three, with 4 time steps every and a single function
enter <- k_random_normal(form = c(3L, 4L, 1L))
enter

# default args
# return form = (batch_size, models)
lstm <- layer_lstm(
  models = 1,
  kernel_initializer = initializer_constant(worth = 1),
  recurrent_initializer = initializer_constant(worth = 1)
)
lstm(enter)

# return_sequences = TRUE
# return form = (batch_size, time steps, models)
#
# notice how for every merchandise within the batch, the worth for time step 4 equals that obtained above
lstm <- layer_lstm(
  models = 1,
  return_sequences = TRUE,
  kernel_initializer = initializer_constant(worth = 1),
  recurrent_initializer = initializer_constant(worth = 1)
  # bias is by default initialized to 0
)
lstm(enter)

# return_state = TRUE
# return form = record of:
#                - outputs, of form: (batch_size, models)
#                - "reminiscence states" for the final time step, of form: (batch_size, models)
#                - "carry states" for the final time step, of form: (batch_size, models)
#
# notice how the primary and second record gadgets are similar!
lstm <- layer_lstm(
  models = 1,
  return_state = TRUE,
  kernel_initializer = initializer_constant(worth = 1),
  recurrent_initializer = initializer_constant(worth = 1)
)
lstm(enter)

# return_state = TRUE, return_sequences = TRUE
# return form = record of:
#                - outputs, of form: (batch_size, time steps, models)
#                - "reminiscence" states for the final time step, of form: (batch_size, models)
#                - "carry states" for the final time step, of form: (batch_size, models)
#
# notice how once more, the "reminiscence" state present in record merchandise 2 matches the final-time step outputs reported in merchandise 1
lstm <- layer_lstm(
  models = 1,
  return_sequences = TRUE,
  return_state = TRUE,
  kernel_initializer = initializer_constant(worth = 1),
  recurrent_initializer = initializer_constant(worth = 1)
)
lstm(enter)

GRU

# default args
# return form = (batch_size, models)
gru <- layer_gru(
  models = 1,
  kernel_initializer = initializer_constant(worth = 1),
  recurrent_initializer = initializer_constant(worth = 1)
)
gru(enter)

# return_sequences = TRUE
# return form = (batch_size, time steps, models)
#
# notice how for every merchandise within the batch, the worth for time step 4 equals that obtained above
gru <- layer_gru(
  models = 1,
  return_sequences = TRUE,
  kernel_initializer = initializer_constant(worth = 1),
  recurrent_initializer = initializer_constant(worth = 1)
)
gru(enter)

# return_state = TRUE
# return form = record of:
#    - outputs, of form: (batch_size, models)
#    - "reminiscence" states for the final time step, of form: (batch_size, models)
#
# notice how the record gadgets are similar!
gru <- layer_gru(
  models = 1,
  return_state = TRUE,
  kernel_initializer = initializer_constant(worth = 1),
  recurrent_initializer = initializer_constant(worth = 1)
)
gru(enter)

# return_state = TRUE, return_sequences = TRUE
# return form = record of:
#    - outputs, of form: (batch_size, time steps, models)
#    - "reminiscence states" for the final time step, of form: (batch_size, models)
#
# notice how once more, the "reminiscence state" present in record merchandise 2 matches the final-time-step outputs reported in merchandise 1
gru <- layer_gru(
  models = 1,
  return_sequences = TRUE,
  return_state = TRUE,
  kernel_initializer = initializer_constant(worth = 1),
  recurrent_initializer = initializer_constant(worth = 1)
)
gru(enter)

torch

LSTM (non-stacked structure)

library(torch)

# batch of three, with 4 time steps every and a single function
# we are going to specify batch_first = TRUE when creating the LSTM
enter <- torch_randn(c(3, 4, 1))
enter

# default args
# return form = (batch_size, models)
#
# notice: there's an extra argument num_layers that we may use to specify a stacked LSTM - successfully composing two LSTM modules
# default for num_layers is 1 although 
lstm <- nn_lstm(
  input_size = 1, # variety of enter options
  hidden_size = 1, # variety of hidden (and output!) options
  batch_first = TRUE # for simple comparability with Keras
)

nn_init_constant_(lstm$weight_ih_l1, 1)
nn_init_constant_(lstm$weight_hh_l1, 1)
nn_init_constant_(lstm$bias_ih_l1, 0)
nn_init_constant_(lstm$bias_hh_l1, 0)

# returns an inventory of size 2, specifically
#   - outputs, of form (batch_size, time steps, hidden_size) - given we specified batch_first
#       Be aware 1: If it is a stacked LSTM, these are the outputs from the final layer solely.
#               For our present goal, that is irrelevant, as we're proscribing ourselves to single-layer LSTMs.
#       Be aware 2: hidden_size right here is equal to models in Keras - each specify variety of options
#  - record of:
#    - hidden state for the final time step, of form (num_layers, batch_size, hidden_size)
#    - cell state for the final time step, of form (num_layers, batch_size, hidden_size)
#      Be aware 3: For a single-layer LSTM, the hidden states are already supplied within the first record merchandise.

lstm(enter)

GRU (non-stacked structure)

# default args
# return form = (batch_size, models)
#
# notice: there's an extra argument num_layers that we may use to specify a stacked GRU - successfully composing two GRU modules
# default for num_layers is 1 although 
gru <- nn_gru(
  input_size = 1, # variety of enter options
  hidden_size = 1, # variety of hidden (and output!) options
  batch_first = TRUE # for simple comparability with Keras
)

nn_init_constant_(gru$weight_ih_l1, 1)
nn_init_constant_(gru$weight_hh_l1, 1)
nn_init_constant_(gru$bias_ih_l1, 0)
nn_init_constant_(gru$bias_hh_l1, 0)

# returns an inventory of size 2, specifically
#   - outputs, of form (batch_size, time steps, hidden_size) - given we specified batch_first
#       Be aware 1: If it is a stacked GRU, these are the outputs from the final layer solely.
#               For our present goal, that is irrelevant, as we're proscribing ourselves to single-layer GRUs.
#       Be aware 2: hidden_size right here is equal to models in Keras - each specify variety of options
#  - record of:
#    - hidden state for the final time step, of form (num_layers, batch_size, hidden_size)
#    - cell state for the final time step, of form (num_layers, batch_size, hidden_size)
#       Be aware 3: For a single-layer GRU, these values are already supplied within the first record merchandise.
gru(enter)