Handling input and files in tensorflow-drake.Rmd

[wlandau]

Functions file_in(), file_out() and knitr_in() tell drake about file dependencies so make() can respond to changes. I suggest we use them in commands that deal with files.

plan <- drake_plan(
  customer_data = read_csv(file_in("customer_churn.csv")),
  save_rec_obj = save(rec_obj, file = file_out("rec_obj.RData")),
  # ...
)

I would also like to discuss the storage of rec_object and model_keras.

• What is the rationale behind creating rec_obj.RData? We can already readd(rec_obj), and in general, I do not think we should create special files too often. But then again, it is sometimes convenient to store certain results separately. In fact, in my day job, I just did this for a dashboard in order to avoid having to upload the entire .drake/ cache.
• If we want to allow other workflows to make use of the models we train, we might want to track the newmodel/ files using file_out(). Unfortunately, however, drake cannot hash/track entire directories. Is it possible to pick a subset of files that (1) have stable paths and (2) fully represent the important parts of the exported model?

[wlandau]

I take back what I said about newmodel above. Today, I enhanced drake so that file_out() and file_in() can handle entire directories (ref: ropensci/drake#795). In development drake, we can simply write drake_plan(save_model = keras:::export_savedmodel.keras.engine.training.Model(model_keras, file_out("newmodel"))). I plan to release this functionality to CRAN with drake version 7.1.0 on April 10.

[edgararuiz-zz]

Fixed by 7e0a686