specifying a GPU to train on (to allow multi-tasking)

import os
os.environ["CUDA_VISIBLE_DEVICES"]="1"

specify the GPU number at the end. same ID’s as gpu/0, gpu/1 etc. this will keep the other GPUs empty, allowing you to run other models on them.

using pip in conda virtual environments

for the primary guide to using conda environments, see the official guide

however, i noticed that my system has the problem that even when in my environment, pip is run from the base evironment, as mentioned in this issue

in order to maintain different environments with different versions of e.g. tensorflow, keras and keras-contrib, i needed to use the solution given:

1. install pip to the environment as detailed in the official guide: conda install -n myenv pip
2. get environments list with conda info --envs
3. use the full path to pip to install:

(oldkeras) derek@mymachine:~$ miniconda3/envs/oldkeras/bin/pip install --upgrade Keras==2.0.6

…this seemed to help me maintain two versions of Keras for backwards compatibility

managing conda virtual environments with jupyter lab

this site details the instructions i used

1. create a conda env
2. pip install ipykernel (using the full pip path if needed, as described above)
3. activate the environment with source activate myenv
4. ipython kernel install --user --name=<kernelname>

…this will require a restart of the jupyter lab server

sklearn multithreading e.g. in GridSearchCV

using multithreaded objects in sklearn can cause the training to freeze, e.g. using GridSearchCV with n_jobs>1. the following seems to work as an alternative:

from sklearn.externals.joblib import parallel_backend
grid = GridSearchCV(...)
with parallel_backend('threading'):
    grid.fit(x_train, y_train)

source & discussion: StackOverflow