All 3 entries tagged Sc-Rtrp
No other Warwick Blogs use the tag Sc-Rtrp on entries | View entries tagged Sc-Rtrp at Technorati | There are no images tagged Sc-Rtrp on this blog
February 02, 2021
I wanted to upgrade my FastAI installation to V2, using my existing ubuntu installation of anaconda. After a few wrong tipps on google, I converged to the following.
First, I made a new conda environment, i.e.
$ conda create -n FastAI2
Then I installed all that was needed into the new environment:
$ conda update -n base -c defaults conda
$ conda install -c fastai -c pytorch fastai
$ conda install jupyter
$ conda install -c fastai nbdev
OK, that seems to work, i.e. the example jupyter notebooks from the FastAI V2 course now run.
October 23, 2020
This set of commands allows to load a working Keras/TF/scikit ML environment which will run on a GPU if present
$ module load GCC/8.3.0 CUDA/10.1.243 OpenMPI/3.1.4 $ module load IPython/7.9.0-Python-3.7.4 $ module load TensorFlow/2.1.0-Python-3.7.4 $ module load Keras
$ module load scikit-learn
$ jupyter notebook &
using the versions
tensorflow: 2.1.0 , keras: 2.3.1 sklearn: 0.21.3
October 16, 2020
Follow-up to Installing your own TensorFlow/Kears + FastAI on Ubuntu from Rudo's blog
We (the DisQS team) have now moved back from Cergy to the University of Warwick where the Scientific Computing RTP provides a centrally managed Linux desktop setup as well as bespoke HPC solutions, including some GPU-enabled machines. This means that we need to switch from our self-managed Ubuntu/Anaconda solution to the managed SC-RTP desktop. The following lines, executed in an SC-RTP terminal, allow us to use PyTorch, or preferred ML package at the moment, as follows
module load GCC/8.3.0 OpenMPI/3.1.4
module load IPython/7.9.0-Python-3.7.4
module load PyTorch
module load scikit-learn
pip install --user torchvision==0.4.0
module save ML (or some such name)
and later one can retrieve these settings via
module restore ML
Please not that we need to install torchvision locally via pip as it's not available on the SC-RTP system as a module. The pip mechanism is what replaces the "conda install ..." used under anaconda for missing libraries. Also, we found that the newer torchvision version 0.7.0 gave some error with our codes, so we are using an older version that seems to work. This may not be needed for your code. Obviously, the use of "module save/restore" is optional, but very convenient.
jupyter notebook &
then works as before/desired.