k1lib.schedule module

This module allows you to make and combine a bunch of schedules, and setup the optimizer so that it changes hyperparameter values based on the schedule. Highly recommend you check out the tutorials section on this. This is exposed automatically with:

from k1lib.imports import *
schedule.Fn # exposed

class k1lib.schedule.Fn(f: Callable[[float], float], param: Optional[str] = None)

Bases: object

__init__(f: Callable[[float], float], param: Optional[str] = None)

Creates a new schedule based on some custom function. Example:

s = schedule.Fn(lambda x: x**2)
s(0.2) # returns 0.04

# you can also use this as a decorator
@schedule.Fn
def s(x):
    return x**2

Parameters

• f – function (domain should always in [0, 1]), can be op

• param – (optional) Parameter to schedule (e.g “lr”) if using ParamScheduler

property value

iter(n: int)

Returns an n-step iterator evenly divided in range [0, 1]. Example:

s = schedule.Fn(lambda x: x+2)
list(s.iter(6)) # returns [2.0, 2.2, 2.4, 2.6, 2.8, 3.0]

modifyOutput(f: Callable[[float], float]) → Fn

Returns a new Fn that has its output modified. Example:

s = Fn(lambda x: x+2)
s.modifyOutput(lambda x: x**2) # now s's function is (x+2)**2

k1lib.schedule.linear(low, high, param: Optional[str] = None)

Sharply goes from low to high

k1lib.schedule.smooth(low, high, param: Optional[str] = None)

Smoothly goes from low to high

k1lib.schedule.hump(low, high, param: Optional[str] = None)

Smoothly rises up (30%), then down (70%)

k1lib.schedule.exp(low, high, param: Optional[str] = None)

Rises/drops quickly, then rate of change gets smaller and smaller

class k1lib.schedule.ParamScheduler(css: str, *schedules: List[Fn])

Bases: Callback

Schedules a param in parts of the network.

Parameters

• css – the selected parts of the network to schedule

• schedules – (obvious)

startBatch()

startRun()

Schedules a param in parts of the network.

Parameters

• css – the selected parts of the network to schedule

• schedules – (obvious)