# Ranger deep learning optimizer - RAdam + Lookahead + calibrated adaptive LR combined. # https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer # Ranger has now been used to capture 12 records on the FastAI leaderboard. #This version = 9.13.19A #Credits: #RAdam --> https://github.com/LiyuanLucasLiu/RAdam #Lookahead --> rewritten by lessw2020, but big thanks to Github @LonePatient and @RWightman for ideas from their code. #Lookahead paper --> MZhang,G Hinton https://arxiv.org/abs/1907.08610 # Calibrated anisotropic adaptive learning rates - https://arxiv.org/abs/1908.00700v2 #summary of changes: #full code integration with all updates at param level instead of group, moves slow weights into state dict (from generic weights), #supports group learning rates (thanks @SHolderbach), fixes sporadic load from saved model issues. #changes 8/31/19 - fix references to *self*.N_sma_threshold; #changed eps to 1e-5 as better default than 1e-8. import math import torch from torch.optim.optimizer import Optimizer, required class RangerVA(Optimizer): def __init__(self, params, lr=1e-3, alpha=0.5, k=6, n_sma_threshhold=5, betas=(.95, 0.999), eps=1e-5, weight_decay=0, amsgrad=True, transformer='softplus', smooth=50, grad_transformer='square'): #parameter checks if not 0.0 <= alpha <= 1.0: raise ValueError('Invalid slow update rate: {}'.format(alpha)) if not 1 <= k: raise ValueError('Invalid lookahead steps: {}'.format(k)) if not lr > 0: raise ValueError('Invalid Learning Rate: {}'.format(lr)) if not eps > 0: raise ValueError('Invalid eps: {}'.format(eps)) #parameter comments: # beta1 (momentum) of .95 seems to work better than .90... #N_sma_threshold of 5 seems better in testing than 4. #In both cases, worth testing on your dataset (.90 vs .95, 4 vs 5) to make sure which works best for you. #prep defaults and init torch.optim base defaults = dict(lr=lr, alpha=alpha, k=k, step_counter=0, betas=betas, n_sma_threshhold=n_sma_threshhold, eps=eps, weight_decay=weight_decay, smooth=smooth, transformer=transformer, grad_transformer=grad_transformer, amsgrad=amsgrad) super().__init__(params,defaults) #adjustable threshold self.n_sma_threshhold = n_sma_threshhold #look ahead params self.alpha = alpha self.k = k #radam buffer for state self.radam_buffer = [[None,None,None] for ind in range(10)] #self.first_run_check=0 #lookahead weights #9/2/19 - lookahead param tensors have been moved to state storage. #This should resolve issues with load/save where weights were left in GPU memory from first load, slowing down future runs. #self.slow_weights = [[p.clone().detach() for p in group['params']] # for group in self.param_groups] #don't use grad for lookahead weights #for w in it.chain(*self.slow_weights): # w.requires_grad = False def __setstate__(self, state): super().__setstate__(state) def step(self, closure=None): loss = None #note - below is commented out b/c I have other work that passes back the loss as a float, and thus not a callable closure. #Uncomment if you need to use the actual closure... #if closure is not None: #loss = closure() #Evaluate averages and grad, update param tensors for group in self.param_groups: for p in group['params']: if p.grad is None: continue grad = p.grad.data.float() if grad.is_sparse: raise RuntimeError('Ranger optimizer does not support sparse gradients') amsgrad = group['amsgrad'] smooth = group['smooth'] grad_transformer = group['grad_transformer'] p_data_fp32 = p.data.float() state = self.state[p] #get state dict for this param if len(state) == 0: #if first time to run...init dictionary with our desired entries #if self.first_run_check==0: #self.first_run_check=1 #print("Initializing slow buffer...should not see this at load from saved model!") state['step'] = 0 state['exp_avg'] = torch.zeros_like(p_data_fp32) state['exp_avg_sq'] = torch.zeros_like(p_data_fp32) if amsgrad: # Maintains max of all exp. moving avg. of sq. grad. values state['max_exp_avg_sq'] = torch.zeros_like(p.data) #look ahead weight storage now in state dict state['slow_buffer'] = torch.empty_like(p.data) state['slow_buffer'].copy_(p.data) else: state['exp_avg'] = state['exp_avg'].type_as(p_data_fp32) state['exp_avg_sq'] = state['exp_avg_sq'].type_as(p_data_fp32) #begin computations exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq'] beta1, beta2 = group['betas'] if amsgrad: max_exp_avg_sq = state['max_exp_avg_sq'] #compute variance mov avg exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad) #compute mean moving avg exp_avg.mul_(beta1).add_(1 - beta1, grad) ##transformer if grad_transformer == 'square': grad_tmp = grad**2 elif grad_transformer == 'abs': grad_tmp = grad.abs() exp_avg_sq.mul_(beta2).add_((1 - beta2)*grad_tmp) if amsgrad: # Maintains the maximum of all 2nd moment running avg. till now torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq) # Use the max. for normalizing running avg. of gradient denomc = max_exp_avg_sq.clone() else: denomc = exp_avg_sq.clone() if grad_transformer == 'square': #pdb.set_trace() denomc.sqrt_() state['step'] += 1 if group['weight_decay'] != 0: p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32) bias_correction1 = 1 - beta1 ** state['step'] bias_correction2 = 1 - beta2 ** state['step'] step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1 # ...let's use calibrated alr if group['transformer'] =='softplus': sp = torch.nn.Softplus( smooth) denomf = sp( denomc) p_data_fp32.addcdiv_(-step_size, exp_avg, denomf ) else: denom = exp_avg_sq.sqrt().add_(group['eps']) p_data_fp32.addcdiv_(-step_size * group['lr'], exp_avg, denom) p.data.copy_(p_data_fp32) #integrated look ahead... #we do it at the param level instead of group level if state['step'] % group['k'] == 0: slow_p = state['slow_buffer'] #get access to slow param tensor slow_p.add_(self.alpha, p.data - slow_p) #(fast weights - slow weights) * alpha p.data.copy_(slow_p) #copy interpolated weights to RAdam param tensor return loss