Excessive (?) memory occupation during hyperparameter tuning with Tensorflow & Keras

During the holidays, I've been happily playing with keras & tensorflow in RStudio Cloud, on the UCI wine quality dataset. I defined a worfklow, where I download & preprocess the data, then I fit multiple models in order to tune the numerous hyperparameters: a README.md file has been provided so that you can easily rerun the analysis.

https://rstudio.cloud/project/161413

The dataset is very small ( ~500 kB), the task is relatively simple (classify the wine quality to one of 4 classes , according to the 11 features), and each model is very small, by modern standards (1 hidden layer, max ~5000 parameters). I try to fit ~400 models, however, I've not been able to fit all of them, not even once. Discussing on the RStudio Cloud channel, it seems to be a memory issue:

However, RStudio Cloud provides 1 GB of RAM, each model is very small, and AFAIU, Tensorflow is not growing the computational graph abnormally (each model is fit in a new graph, so the graph size doesn't keep growing). So, why do Tensorflow & Keras grow their memory occupation so much? Am I doing something wrong? Can you help me fix this? Thanks in advance

FYI, the project doesn't seem to have public access, so no one else can look at it.

3 Likes

Whoops! Thanks @mara! I made the original project public, but I forgot to make this "copy" public too. It should be fixed now! BTW, you just made me realize that RStudio Cloud is by far the best way to run R on mobile :grin:

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