keras-tcn for R

[Ujjwal4CULS]

Would you please le me know how to interface this keras-tcn module in R or RStudio?

[philipperemy]

@Ujjwal4CULS maybe try this:

• https://rdrr.io/cran/kerasR/
• https://github.com/statsmaths/kerasR (same but more documentation).

Keras-TCN is a 100% Keras layer. So if you can make it work with a simple Keras model, it should work with this one.

Let me know.

[Steviey]

+1 Has the layer to be hardcoded registered anywhere?

suppressPackageStartupMessages(library(reticulate))
    use_condaenv("keras_tensorflow")

    suppressPackageStartupMessages(library(tensorflow))
    suppressPackageStartupMessages(library(keras))
    #suppressPackageStartupMessages(library(keras3))
    #suppressPackageStartupMessages(library(kerasR))
    suppressPackageStartupMessages(library(abind))
    

Caused by error in `layer_tcn()`:
! could not find function "layer_tcn"

[philipperemy]

Not that I am aware of.

[philipperemy]

Here’s a fully working snippet in R that demonstrates using Keras-TCN via reticulate:

Prerequisites (Run once)
Ensure you have the required packages and Python dependencies installed:

install.packages("keras")
install.packages("tensorflow")
install.packages("reticulate")

library(keras)
library(tensorflow)
library(reticulate)

# Install keras-tcn in the Python environment used by reticulate
reticulate::py_install("keras-tcn")

TCN Model in R

The task to learn is pretty simple. Given any random input, always yield a 1.

library(keras)
library(reticulate)

# Import keras-tcn
tcn <- import("tcn")

# Generate some dummy data (time series)
set.seed(42)
x_train <- array(runif(1000), dim = c(100, 10, 1))  # 100 samples, 10 time steps, 1 feature
y_train <- array(rep(1,100))       # 100 target values

# Define Input Layer (Explicit Integer Shape)
inputs <- keras$Input(shape = c(as.integer(10), as.integer(1)))

# Apply TCN Layer
tcn_output <- tcn$TCN(
  nb_filters = as.integer(64),
  kernel_size = as.integer(6),
  dilations = list(as.integer(1), as.integer(2), as.integer(4), as.integer(8)),
  return_sequences = FALSE
)(inputs)

# Dense Output Layer
outputs <- keras$layers$Dense(units = as.integer(1), activation = "linear")(tcn_output)

# Create Model
model <- keras$Model(inputs = inputs, outputs = outputs)

# Compile Model
model$compile(
  optimizer = keras$optimizers$Adam(0.001),
  loss = "mse"
)

# Train Model
model$fit(
  x_train, y_train,
  epochs = as.integer(100),
  batch_size = as.integer(16)
)

# Predict
preds <- model$predict(x_train[1:5,,])
print(preds)

Expected output (preds should be around 1):

[philipperemy]

I'll close this question. Feel free to re-open it if I missed something!