3

I'm building a R package for binary classification and I'm using opencpu to host it. Currently I've saved the h5 file as .RData file(serialized), which is then loaded in the environment using the .onLoad() function in R. This enables the R script to use the environment variable to load keras model using keras::unserialized_model().

I've tried directly using keras::load_model_hdf5() in the code, but after building and deploying on opencpu, when I try to hit the prediction API, I get error

ioerror: unable to open file (unable to open file: name = '/home/modelfile_26feb.h5', errno = 13, error message = 'permission denied', flags = 0, o_flags = 0)

I have changed permission for the file(777) and even the groups but still getting the error.

I even tried putting the file in inst/extdata folder so that it gets in the package but still same error.

Can anyone help on this, or suggest some alternative to load the h5 model directly?

Swapnesh Khare
  • 109
  • 1
  • 10

1 Answers1

0

Which OS does OpenCPU run on? Why does it try to write in /home/, this is very unusual? The best solution is to adapt your code to write in getwd() or tempdir(). Even better is to store data in a local database or redis server and let R read it from there, so you don't need disk access at all.

If you run on Ubuntu Server, reading from /home/ is not permitted by default. If you want to allow this, you need to add apparmor rules, see section 3.5 of the server manual.

Some relevant topics from the opencpu mailing list:

Jeroen Ooms
  • 31,998
  • 35
  • 134
  • 207
  • Thanks, the 'keras in opencpu' linked solved it. It was due to apparmor rules. For others doing the same thing as reading the model file from a different location, remember to put `keras::k_clear_session()` at the end to avoid file reading errors due to multithreading ([link](https://github.com/tensorflow/tensorflow/issues/14356)) – Swapnesh Khare Mar 15 '19 at 07:14