csc - Tools for non standard Python execution

Install with

pip install csc

Patterns

Load a script as a module

train_script = csc.load("train.py") 
train_script.train_func()

Extracting local variable from functions

def add(x, y):
    z = x + y
    return z

res = csc.call(add, 1, 2)
assert res.__return__ == 3
assert res.x == 1
assert res.y == 2
assert res.z == 3

Execute scripts with code cells

Consider a script to define and train a model

#%% [setup] Parameters
...

#%% [setup] Setup
...

#%% Train
...

#%% Save
...

To run the parameters cell, then overwrite the parameters, and finally run set, use:

script = csc.Script("external_script.py")
script["Parameters"].run()
script.ns.hidden_units = 64
script["Setup":].run()

Splicing scripts

Same effect as in the previous example:

script = csc.Script("external_script.py")
with csc.splice(script, "Parameters"):
    script.ns.hidden_units = 64

Using cell tags

To only define the model without training or saving the results, a subset of the script can be selected via tags:

# execute any cell tagged with "setup"
script[lambda: tags & {"setup"}].run()

# can also be combined with splicing to modify the parameters
with csc.splice(script[lambda: tags & {"setup"}], "Parameters"):
    script.ns.hidden_units = 64

Creating temporary modules

In a Jupyter notebook, first register the magic function

import csc
csc.autoconfig()

Afterwards a module can be defined via

%%csc.module my_cool_module

def add(x, y):
    return x + y

It can be used as any other module. For example:

import my_cool_module

assert my_cool_module.add(1, 2) == 3

License

This package is licensed under the MIT License. See LICENSE for details.

The function csc._utils.capture_frame is adapted from this stackoverflow post by Niklas Rosenstein licensed under CC-BY-SA 4.0.