fret 0.2.1

REProducible Experimental environment.

fret

Framework for Reproducible ExperimenTs

Installation

From pip:

pip install fret

From source: clone the repository and then run: python setup.py install.

Tutorial

Basic Usage

Create a file named app.py with content:

import fret

@fret.command
def run(ws):
    model = ws.build()
    print(model)

@fret.configurable
class Model:
    def __init__(self, x=3, y=4):
        ...

Then under the same directory, you can run:

$ fret config Model
[ws/_default] configured "main" as "Model" with: x=3, y=4
$ fret run
Model(x=3, y=4)
$ fret config Model -x 5 -y 10
[ws/_default] configured "main" as "Model" with: x=5, y=10
$ fret run
Model(x=5, y=10)

Using Workspace

You can specify different configuration in different workspace:

$ fret -w ws/model1 config Model
[ws/model1] configured "main" as "Model" with: x=3, y=4
$ fret -w ws/model2 config Model -x 5 -y 10
[ws/model2] configured "main" as "Model" with: x=5, y=10
$ fret -w ws/model1 run
Model(x=3, y=4)
$ fret -w ws/model2 run
Model(x=5, y=10)

Save/Load

import fret

@fret.command
def train(ws):
    model = ws.build()
    model.train()
    ws.save(model, 'trained')

@fret.command
def test(ws):
    model = ws.load('ws/best/snapshot/main.trained.pt')
    print(model.weight)

@fret.configurable
@fret.stateful('weight')
class Model:
    def __init__(self):
        self.weight = 0
    def train(self):
        self.weight = 23

$ fret -w ws/best config Model
[ws/_default] configured "main" as "Model"
$ fret -w ws/best train
$ fret test
23

An Advanced Workflow

In app.py:

import time
import fret

@fret.configurable
@fret.stateful('value')
class Model:
    def __init__(self):
        self.value = 0

@fret.command
def resumable(ws):
    model = ws.build()
    with ws.run('exp-1') as run:
        run.register(model)
        cnt = run.acc()
        for e in fret.nonbreak(run.range(5)):
            # with `nonbreak`, the program always finish this loop before exit
            model.value += e
            time.sleep(0.2)
            cnt += 1
            print('current epoch: %d, sum: %d, cnt: %d' %
                  (e, model.value, cnt))

Then you can stop and restart this program anytime, with consistent results:

$ fret resumable
current epoch: 0, sum: 0, cnt: 1
current epoch: 1, sum: 1, cnt: 2
^CW SIGINT received. Delaying KeyboardInterrupt.
current epoch: 2, sum: 3, cnt: 3
Traceback (most recent call last):
    ...
KeyboardInterrupt
W cancelled by user
$ fret resumable
current epoch: 3, sum: 6, cnt: 4
current epoch: 4, sum: 10, cnt: 5

Submodule

@fret.configurable
class A:
    def __init__(self, foo):
        ...

@fret.configurable(submodules=['sub'])
class B:
    def __init__(self, sub, bar=3):
        self.sub = sub(foo='bar')   # call sub to build submodule

$ fret config sub A
[ws/_default] configured "sub" as "A"
$ fret config B
[ws/_default] configured "main" as "B" with: sub='sub', bar=3
$ fret run
B(sub=A(), bar=3)

Inheritance

@fret.configurable
class A:
    def __init__(self, foo='bar', sth=3):
        ...

@fret.configurable
class B(A):
    def __init__(self, bar=3, **others):
        super().__init__(**others)
        ...

$ fret config B -foo baz -bar 0
[ws/_default] configured "main" as "B" with: bar=0, foo='baz', sth=3
$ fret run
B(bar=0, foo='baz', sth=3)

Internals

>>> config = fret.Configuration({'foo': 'bar'})
>>> config
foo='bar'

TODO

• fret.Configuration: high-level class for configuration
• fret.Workspace: module build/save/load (by tag or by path), persistency
• ws.run() context manager, run.value(), run.acc(), run.range(), run.register()
• fret.App, global app object
• CLI: entry point logic, testing
• @fret.configurable and @fret.command clean-up
• fret.toml: global configuration, configuration by class name, named workspace
• Java/GNU style command line args, shorthands, better logic for boolean default value
• Full docstring
• Documents and examples
• Utility commands like show log, check module, etc.