k1lib.schedule module

k1lib.schedule.combine(lambdas: List[Callable[[float], float]], ratios: Optional[List[float]] = None) → Callable[[float], float]

Combine multiple different schedules.

Parameters

• ratios – weighting diferent functions. Does not have to add up to 1.

• lambdas – functions with 1 float input in [0, 1]

k1lib.schedule.decorate(f: Callable[[float, float, float], float]) → Callable[[float], float]

Decorator, transforms f(low, high, x) to (low, high) -> f(x).

k1lib.schedule.linear(low, high)

k1lib.schedule.cosine(low, high)

k1lib.schedule.oneCycle(low, high)

k1lib.schedule.decay(low, high)

class k1lib.schedule.Schedule(param: str, scheduleF: Callable[[float], float])

Bases: object

startBatch(paramGroup: dict, progress: float)

property value

static linear(param: str, low: float, high: float) → Callable[[float], float]

Sharply goes from low to high

static cosine(param: str, low: float, high: float) → Callable[[float], float]

Smoothly goes from low to high

static oneCycle(param: str, low: float, high: float, upPortion: float = 0.3) → Callable[[float], float]

Goes from middle-low (0.8*low + 0.2*high) to high to low

Parameters

upPortion – the percent of portion that goes up

static decay(param: str, low: float, high: float) → Callable[[float], float]

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

class k1lib.schedule.ParamScheduler(css: str, schedules: Union[List[k1lib.schedule.Schedule], k1lib.schedule.Schedule])

Bases: k1lib.callbacks.callbacks.Callback

Schedules a param in parts of the network.

Parameters

• css – the selected parts of the network to schedule

• schedules – (obvious)

startRun()

meta:private