Error when using a bundled Keras model on POSIT Workbench

[andrewbcooper]

Hello! Thank yo for this amazing packages! I trained a Keras model on my laptop, bundled it, uploaded it to our POSIT Workbench environment, and then tried to use it for predictions. Unfortunately, I got the error:

Error in if (tensorflow::tf_version() < "2.0.0") stop("TensorFlow version >= 2.0.0 is requires to load models in the SavedModel format.", :
argument is of length zero

My guess it's because we don't actually have TensorFlow installed on the POSIT Workbench servers (e.g., install_tensorflow(); install_keras()) even though we have both TensorFlow and Keras packages installed. tensorflow::tf_version() returns NULL.

Is there any way to use bundle to get predictions from a Keras model without installing TensorFlow / Keras? Our IT folks say it'll probably be 6-9 months before they can get around to that :/

Thanks!

Non-reproducable code:

On my laptop:

bundled_NNModel <-
bundle(NNModel)

save(bundled_NNModel,file="bundled_NNModel.RData")

After uploading to the server....

load("bundled_NNModel.RData")
library(keras)

r(
function(model_bundle, new_data) {
library(bundle)

model_object <- unbundle(model_bundle)

predict(model_object, new_data)

},
args = list(
model_bundle = bundled_NNModel,
new_data = as.matrix(testfeatures)
)
)

[juliasilge]

Unfortunately, bundling only solves the problem of how you serialize your model, not the software that you then need to use the model. For example, you likely will need bundle to store an xgboost model and use it in a new R session because of how xgboost works, but you also still need xgboost.

An option for making predictions that does not require the original R package is orbital, but currently it does not provide support for keras models.

[andrewbcooper]

Thank you for your lightning-fast response! That's kind of what I figured. Thankfully, it looks like I can get around it by using torch and luz which don't require IT to do anything :) And even more thankfully, your bundle package handles those beautifully! Thanks for this amazing package!

[juliasilge]

Let us know if you have further questions! 🙌