import torch import torch.optim as optim import numpy as np import copy from ... import sync from ...cfg_holder import cfg_unique_holder as cfguh def singleton(class_): instances = {} def getinstance(*args, **kwargs): if class_ not in instances: instances[class_] = class_(*args, **kwargs) return instances[class_] return getinstance @singleton class get_scheduler(object): def __init__(self): self.lr_scheduler = {} def register(self, lrsf, name): self.lr_scheduler[name] = lrsf def __call__(self, cfg): if cfg is None: return None if isinstance(cfg, list): schedulers = [] for ci in cfg: t = ci.type schedulers.append( self.lr_scheduler[t](**ci.args)) if len(schedulers) == 0: raise ValueError else: return compose_scheduler(schedulers) t = cfg.type return self.lr_scheduler[t](**cfg.args) def register(name): def wrapper(class_): get_scheduler().register(class_, name) return class_ return wrapper class template_scheduler(object): def __init__(self, step): self.step = step def __getitem__(self, idx): raise ValueError def set_lr(self, optim, new_lr, pg_lrscale=None): """ Set Each parameter_groups in optim with new_lr New_lr can be find according to the idx. pg_lrscale tells how to scale each pg. """ # new_lr = self.__getitem__(idx) pg_lrscale = copy.deepcopy(pg_lrscale) for pg in optim.param_groups: if pg_lrscale is None: pg['lr'] = new_lr else: pg['lr'] = new_lr * pg_lrscale.pop(pg['name']) assert (pg_lrscale is None) or (len(pg_lrscale)==0), \ "pg_lrscale doesn't match pg" @register('constant') class constant_scheduler(template_scheduler): def __init__(self, lr, step): super().__init__(step) self.lr = lr def __getitem__(self, idx): if idx >= self.step: raise ValueError return self.lr @register('poly') class poly_scheduler(template_scheduler): def __init__(self, start_lr, end_lr, power, step): super().__init__(step) self.start_lr = start_lr self.end_lr = end_lr self.power = power def __getitem__(self, idx): if idx >= self.step: raise ValueError a, b = self.start_lr, self.end_lr p, n = self.power, self.step return b + (a-b)*((1-idx/n)**p) @register('linear') class linear_scheduler(template_scheduler): def __init__(self, start_lr, end_lr, step): super().__init__(step) self.start_lr = start_lr self.end_lr = end_lr def __getitem__(self, idx): if idx >= self.step: raise ValueError a, b, n = self.start_lr, self.end_lr, self.step return b + (a-b)*(1-idx/n) @register('multistage') class constant_scheduler(template_scheduler): def __init__(self, start_lr, milestones, gamma, step): super().__init__(step) self.start_lr = start_lr m = [0] + milestones + [step] lr_iter = start_lr self.lr = [] for ms, me in zip(m[0:-1], m[1:]): for _ in range(ms, me): self.lr.append(lr_iter) lr_iter *= gamma def __getitem__(self, idx): if idx >= self.step: raise ValueError return self.lr[idx] class compose_scheduler(template_scheduler): def __init__(self, schedulers): self.schedulers = schedulers self.step = [si.step for si in schedulers] self.step_milestone = [] acc = 0 for i in self.step: acc += i self.step_milestone.append(acc) self.step = sum(self.step) def __getitem__(self, idx): if idx >= self.step: raise ValueError ms = self.step_milestone for idx, (mi, mj) in enumerate(zip(ms[:-1], ms[1:])): if mi <= idx < mj: return self.schedulers[idx-mi] raise ValueError #################### # lambda schedular # #################### class LambdaWarmUpCosineScheduler(template_scheduler): """ note: use with a base_lr of 1.0 """ def __init__(self, base_lr, warm_up_steps, lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0): cfgt = cfguh().cfg.train bs = cfgt.batch_size if 'gradacc_every' not in cfgt: print('Warning, gradacc_every is not found in xml, use 1 as default.') acc = cfgt.get('gradacc_every', 1) self.lr_multi = base_lr * bs * acc self.lr_warm_up_steps = warm_up_steps self.lr_start = lr_start self.lr_min = lr_min self.lr_max = lr_max self.lr_max_decay_steps = max_decay_steps self.last_lr = 0. self.verbosity_interval = verbosity_interval def schedule(self, n): if self.verbosity_interval > 0: if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_lr}") if n < self.lr_warm_up_steps: lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start self.last_lr = lr return lr else: t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps) t = min(t, 1.0) lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * ( 1 + np.cos(t * np.pi)) self.last_lr = lr return lr def __getitem__(self, idx): return self.schedule(idx) * self.lr_multi class LambdaWarmUpCosineScheduler2(template_scheduler): """ supports repeated iterations, configurable via lists note: use with a base_lr of 1.0. """ def __init__(self, base_lr, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0): cfgt = cfguh().cfg.train # bs = cfgt.batch_size # if 'gradacc_every' not in cfgt: # print('Warning, gradacc_every is not found in xml, use 1 as default.') # acc = cfgt.get('gradacc_every', 1) # self.lr_multi = base_lr * bs * acc self.lr_multi = base_lr assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths) self.lr_warm_up_steps = warm_up_steps self.f_start = f_start self.f_min = f_min self.f_max = f_max self.cycle_lengths = cycle_lengths self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths)) self.last_f = 0. self.verbosity_interval = verbosity_interval def find_in_interval(self, n): interval = 0 for cl in self.cum_cycles[1:]: if n <= cl: return interval interval += 1 def schedule(self, n): cycle = self.find_in_interval(n) n = n - self.cum_cycles[cycle] if self.verbosity_interval > 0: if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, " f"current cycle {cycle}") if n < self.lr_warm_up_steps[cycle]: f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle] self.last_f = f return f else: t = (n - self.lr_warm_up_steps[cycle]) / (self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle]) t = min(t, 1.0) f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * ( 1 + np.cos(t * np.pi)) self.last_f = f return f def __getitem__(self, idx): return self.schedule(idx) * self.lr_multi @register('stable_diffusion_linear') class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2): def schedule(self, n): cycle = self.find_in_interval(n) n = n - self.cum_cycles[cycle] if self.verbosity_interval > 0: if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, " f"current cycle {cycle}") if n < self.lr_warm_up_steps[cycle]: f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle] self.last_f = f return f else: f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle]) self.last_f = f return f