My data are not timeseries, but it has sequential properties.
Consider one sample:
data1 = matrix(rnorm(10, 0, 1), nrow = 1)
label1 = rnorm(1, 0, 1)
label1 is a function of the data1, but the data matrix is not a timeseries. I suppose that label is a function of not just one data sample, but more older samples, which are naturally ordered in time (not sampled randomly), in other words, data samples are dependent with one another.
I have a batch of examples, say, 16.
With that I want to understand how I can design an RNN/LSTM model which will memorize all 16 examples from the batch to construct the internal state. I am especially confused with the seq_len parameter, which as I understand is specifically about the length of the timeseries used as an input to a network, which is not case.
Now this piece of code (taken from a timeseries example) only confuses me because I don't see how my task fits in.
rm(symbol)
symbol <- rnn.graph.unroll(seq_len = 5,
num_rnn_layer = 1,
num_hidden = 50,
input_size = NULL,
num_embed = NULL,
num_decode = 1,
masking = F,
loss_output = "linear",
dropout = 0.2,
ignore_label = -1,
cell_type = "lstm",
output_last_state = F,
config = "seq-to-one")
graph.viz(symbol, type = "graph", direction = "LR",
graph.height.px = 600, graph.width.px = 800)
train.data <- mx.io.arrayiter(
data = matrix(rnorm(100, 0, 1), ncol = 20)
, label = rnorm(20, 0, 1)
, batch.size = 20
, shuffle = F
)
Sure, you can treat them as time steps, and apply LSTM. Also check out this example: https://github.com/apache/incubator-mxnet/tree/master/example/multivariate_time_series as it might be relevant for your case.