To just quote its documentation: ‘Sacred is a tool to configure, organize, log and reproduce computational experiments’. Sacred's documentation itself is great. However, I figured some nice sample code was missing, that shows you how you run a full project.
In this repo you can find the code for a very simple feed forward neural network in Pytorch,
where we make use of Sacred. The code is based on Yunjey's code,
but quite heavily adapted for the current example. All the lines that are there for Sacred
are commented with #sacred.
python train_nn.py with num_epochs=10
sbatch train_nn.job
Sacred allows you to change your configuration in several ways. You can have al look at the documentation for a full overview of how to do this.
Here we use the update from the command line. In the code num_epochs=2, whereas we want
to update it to num_epochs=10. From the command line you can do this by using
with, followed by your update. In a job file you need to make sure to put your full
parameter update between quotation marks in case of integers: with 'num_epochs=10', as
otherwise your parameter is not recognized as an integer.
In some cases you may want to use a Sacred experiment in a new file. You can do this by importing it. An example for how that would go with the current setup:
import train_nn
ex = train_nn.ex
Then you can use your ex as you're used to.
You display all your results with Sacredboard. To run Sacredboard for the current setup:
sacredboard -m my-database
