禁止R中Keras的控制台输出

Question

I am trying to use an R-script in a docker environment to do some classification on the input that I am piping to it. The script should only print the output of the classification, however I always get Using TensorFlow backend. as output of the first line where a function of Keras is called.

I made a small working example.

library(keras)

v <- (c(1,2,3))
print(v)
vCat<-to_categorical(v)

I want to display this as output:

[1] 1 2 3

I get however

[1] 1 2 3
Using TensorFlow backend.

So I searched around on google and stackoverflow about sink and suppress... and other stuff, from for example Suppress automatic output to console in R , Suppress one command's output in R and Suppress automatic output to console in R . I then tried the following piece of code:

library(keras, quietly = T)

v <- (c(1,2,3))
print(v) 
sink("/dev/null")
capture.output(suppressMessages(suppressWarnings(
suppressPackageStartupMessages(
vCat<-to_categorical(v) ))), file = "/dev/null")
sink()

This still doesn't suppress the Using TensorFlow backend. message. The script is called in a docker environment using littler, from the command line: r test.R . Note: when just running the script in Rstudio I don't get the message.

Answer 1

This is a Keras issue (ie not an R or Tensorflow one) and, as I have already commented, there have been several people complaining about this , but a solution has not been proposed by Keras maintainers so far.

A hack , as already suggested in the issue thread above , is to locate the keras/backend/__init__.py file and comment out the following line :

sys.stderr.write('Using TensorFlow backend.\n')

Last time I checked, Keras in R only uses the Tensorflow backend, so you shouldn't need to repeat this for the other backend options in __init__.py .