Deep Learning Training Experience on Windows 10 and Anaconda Need Improve

harryzhu · September 14, 2019, 6:42am

I am trying to train deep learning model using keras R package and related Anaconda, Rstudio toolchains on Windows 10. However, I found it is not production-ready at all.

Hope the rstudio-keras team can solve these issues as soon as possible. see the more issue information:

github.com/rstudio/rstudio

Keras crashes R session with fatal error

opened 02:46PM - 13 Sep 19 UTC

closed 03:30PM - 16 Sep 19 UTC

ghost

external

### System details RStudio Edition : desktop RStudio Version : 1.1.456 OS Version : windows 7 R Version : 3.6.1 ### Steps to reproduce the problem I'm using rstudio in anaconda. I installed keras with ``` install.packages("keras") library(keras) install_keras() ``` Then i tried to run the example code from https://keras.rstudio.com/ ``` library(keras) mnist <- dataset_mnist() x_train <- mnist$train$x y_train <- mnist$train$y x_test <- mnist$test$x y_test <- mnist$test$y # reshape x_train <- array_reshape(x_train, c(nrow(x_train), 784)) x_test <- array_reshape(x_test, c(nrow(x_test), 784)) # rescale x_train <- x_train / 255 x_test <- x_test / 255 y_train <- to_categorical(y_train, 10) y_test <- to_categorical(y_test, 10) model <- keras_model_sequential() model %>% layer_dense(units = 256, activation = 'relu', input_shape = c(784)) %>% layer_dropout(rate = 0.4) %>% layer_dense(units = 128, activation = 'relu') %>% layer_dropout(rate = 0.3) %>% layer_dense(units = 10, activation = 'softmax') ``` ### Describe the problem in detail When it gets to the `model <- keras_model_sequential()` line, it runs for a few seconds before R crashes with the "r session aborted. R encountered a fatal error". One thing i noticed that was strange in anaconda, there are two "r-reticulate" environments after installing keras. I tried deleting one but both disappeared. I'm new to anaconda so i'm not sure if that's weird behaviour or not. ### Describe the behavior you expected I expect it to run without issue.

Due to these bug issues, I have to switch to JupyterLab and Mxnet with GPU, which is more productive.

kevinushey · September 19, 2019, 6:18pm

The associated issue you linked to appears to have been solved. Can you provide more details as to what problem you're seeing, exactly?

harryzhu · October 1, 2019, 12:15am

As the solution says, The Anaconda version RStudio will lead a wrong way, in that, the most of deep learning tutorials are linked to anaconda which will introduce more people get confusing set-up experience.

So, to update the Anaconda version RStudio or write an official introduction blog seems a better option.

harryzhu · October 6, 2019, 2:55am

After Tensorflow 2.0 published, the most of API conflicts were fixed. In my experience, installing tensorflow by virtualenv which works on my laptop without any error anymore.

In conclusion, training your deep learning model in RStudio without anaconda is the best option right now.

kevinushey · October 11, 2019, 6:14pm

You should also (hopefully) have success using Anaconda environments, with tensorflow installed through pip (from PyPI) rather than the Conda repositories. This is the workflow generally recommended by the RStudio TensorFlow team.

system · November 1, 2019, 6:14pm

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.