Closed luigiferiani closed 5 years ago
Please double-check I haven't messed up anything before merging!
Hey, this looks good but I remember having a problems with the conda-forge packages and Windows. I think conda-forge works better for macos/linux but some packages maybe not up to day for Windows, while I think anaconda might have the opposite.
I suggest to rather than enforce the channel in requirements.txt
to state it in the installation instructions like:
For OSX:
conda install --file requirements.txt -c conda-forge
For Windows
conda install --file requirements.txt -c anaconda
Additionally, somebody might want to use pip
so that opens the possibility. Although i am not sure if all the packages are there...
Yep, that makes sense. At this moment we are without a Windows 10 machine, but we're planning to get one exactly for tierpsy testing purposes.
Tensorflow and conda-forge
The pinned
keras=2.1.5
in the general channel comes withtensorflow=1.10.0=eigen...
, with abysmal performances (at least on MacOS).keras=2.1.5
via conda-forge comes withtensorflow=1.0.0=py36_0
, with better performance. Solution Updated the installation instructions so thatconda-forge
will be the natural choice for users. Also updated therequirements.txt
file so installing from source also usesconda-forge
.Multithreading
Enforced
OMP_NUM_THREADS='1'
when launching a subprocess. This will improve performances when analysing in parallel a number of videos close to the number of cores available. However performance will be degraded if only a few files are analysed in parallel, so we should think about using a parameter for this.Other installation issues
Updated the installation instruction with steps to deal with an error I've encountered while installing tierpsy recently. Also suggesting a version of
opencv
viaconda
that works on MacOS.