Executing keras aborts RStudio session - keras

Question

I am trying to install keras in a laptop with R and RStudio previously installed.
I first installed Anaconda3 following the instructions here https://docs.anaconda.com/anaconda/install/linux/.
Second, I followed these instructions to install keras and tensorflow in RStudio https://github.com/FrancisArgnR/Guide-Keras-R (cpu instructions), but specifying conda method:
install.packages('devtools')
devtools::install_github("rstudio/keras")
library(keras)
install_keras(method = c("conda"))
However, whenever I tried to run keras functions (data <- dataset_mnist()), RStudio session aborts. When I use R in the terminal I get the error:
> library(keras)
> data<-dataset_mnist()
*** caught illegal operation ***
address 0x7fb3e50fe820, cause 'illegal operand'
Traceback:
1: py_module_import(module, convert = convert)
2: import(module)
3: doTryCatch(return(expr), name, parentenv, handler)
4: tryCatchOne(expr, names, parentenv, handlers[[1L]])
5: tryCatchList(expr, classes, parentenv, handlers)
6: tryCatch(import(module), error = clear_error_handler())
7: py_resolve_module_proxy(x)
8: `$.python.builtin.module`(keras, "datasets")
9: keras$datasets
10: dataset_mnist()
Possible actions:
1: abort (with core dump, if enabled)
2: normal R exit
3: exit R without saving workspace
4: exit R saving workspace
Selection:
I found a similar error previously reported here https://github.com/rstudio/tensorflow/issues/228, although the "Traceback" is not exactly the same. I tried specifying an earlier version of Tensorflow as suggested, but the error perstists.
install_keras(method = c("conda"),tensorflow = "1.5")
I tired serveral tensorflow versions, from 1.9 to 1.0, but I obtain the same behaviour.
I also tried installing keras from CRAN with install.packages("keras") and repeating all the procedure, but nothing changed. If I do not specify the conda method, the same happens, but in addition, I obtain the following errors when running install_keras():
ERROR: spyder 3.3.6 requires pyqt5<5.13; python_version >= "3", which is not installed.
ERROR: spyder 3.3.6 requires pyqtwebengine<5.13; python_version >= "3", which is not installed.
ERROR: astroid 2.3.1 requires typed-ast<1.5,>=1.4.0; implementation_name == "cpython" and python_version < "3.8", which is not installed.
ERROR: astroid 2.3.1 has requirement six==1.12, but you'll have six 1.13.0 which is incompatible.
Some details in case they can be useful:
R version 3.6.1, Platform: x86_64-pc-linux-gnu (64-bit)
RStudio: Version 1.2.5019
OS: Ubuntu 19.10
Processor: Intel® Celeron(R) CPU N3450 # 1.10GHz × 4
5,6 GiB RAM

Answer 1

It seems that your problem is caused by the lack of AVX instructions support of your CPU.
First you should know that when you run install_keras(method = "conda") a new environment (usually named r-tensorflow) is created. All the Python libraries required to run Tensorflow within R will be installed in this environment. When you run Keras code in R what R does is call the Tensorflow library of this environment.
Most of the pre-built binaries of Tensorflow are compiled on a CPU supporting the AVX instructions set given that this instructions allow a drastic speed up in certain floating point operations. Intel started to support these instructions in 2011. Although your CPU seems to be produced in 2016 it is a Celeron and as we can see here:
Not all CPUs from the listed families support AVX. Generally, CPUs with the commercial denomination "Core i3/i5/i7" support them, whereas "Pentium" and "Celeron" CPUs don't.
You can check this by running on your linux terminal lscpu | grep avx. If nothing is shown your CPU does not support these instructions.
Under these circumstances you have two options:
Build Tensorflow from source in order to compile the code given your CPU features. You can find more info here
Find a pre-built binary of Tensorflow built for CPU without AVX support. Something like this