cloudml
cloudml copied to clipboard
Any jobs that uses `{keras}` functions fail
Jobs of any simplest scripts fail with the following error.json
:
{
"message": "could not find a Python environment for /usr/local/bin/python3",
"traceback": ["stop(sprintf(fmt, ...), call. = call.)", "stopf(\"could not find a Python environment for %s\", python)", "python_info(python)", "reticulate::py_install(packages = c(package, extra_packages), \n envname = envname, method = method, conda = conda, python_version = conda_python_version, \n pip = TRUE, ...)", "install_tensorflow()", "eval(ei, envir)", "eval(ei, envir)", "withVisible(eval(ei, envir))", "source(file = file, local = envir, echo = echo, encoding = encoding)", "withCallingHandlers({\n source(file = file, local = envir, echo = echo, encoding = encoding)\n write_run_property(\"completed\", TRUE)\n}, error = function(e) {\n write_run_metadata(\"error\", list(message = e$message, traceback = capture_stacktrace(sys.calls())))\n stop(e)\n})", "force(expr)", "tryCatch({\n tfruns::training_run(file = deploy$entrypoint, context = deploy$context, \n config = \"cloudml\", flags = deploy$overlay, encoding = \"UTF-8\", \n echo = TRUE, view = FALSE, run_dir = run_dir)\n}, error = function(e) {\n training_error <<- e\n})"]
}
The logs look like that:
"Error: Installation of TensorFlow not found."
"Python environments searched for 'tensorflow' package:"
" /usr/bin/python3.7"
" /usr/bin/python3.5"
" /usr/bin/python2.7"
"You can install TensorFlow using the install_tensorflow() function."
"Execution halted"
I have a feeling that this is related to the updated keras::install_keras()
function.
I also have tried:
- Previous versions of
runtimeVersion
(1.15) and Python (2.7) - Use
install_tensorflow()
inside the script