Skip to content

Instantly share code, notes, and snippets.

@redknightlois
Last active August 9, 2023 20:50
Show Gist options
  • Save redknightlois/c4023d393eb8f92bb44b2ab582d7ec20 to your computer and use it in GitHub Desktop.
Save redknightlois/c4023d393eb8f92bb44b2ab582d7ec20 to your computer and use it in GitHub Desktop.
Ralamb optimizer (RAdam + LARS trick)
class Ralamb(Optimizer):
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0):
defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)
self.buffer = [[None, None, None] for ind in range(10)]
super(Ralamb, self).__init__(params, defaults)
def __setstate__(self, state):
super(Ralamb, self).__setstate__(state)
def step(self, closure=None):
loss = None
if closure is not None:
loss = closure()
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('Ralamb does not support sparse gradients')
p_data_fp32 = p.data.float()
state = self.state[p]
if len(state) == 0:
state['step'] = 0
state['exp_avg'] = torch.zeros_like(p_data_fp32)
state['exp_avg_sq'] = torch.zeros_like(p_data_fp32)
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)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
beta1, beta2 = group['betas']
# Decay the first and second moment running average coefficient
# m_t
exp_avg.mul_(beta1).add_(1 - beta1, grad)
# v_t
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
state['step'] += 1
buffered = self.buffer[int(state['step'] % 10)]
if state['step'] == buffered[0]:
N_sma, radam_step_size = buffered[1], buffered[2]
else:
buffered[0] = state['step']
beta2_t = beta2 ** state['step']
N_sma_max = 2 / (1 - beta2) - 1
N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t)
buffered[1] = N_sma
# more conservative since it's an approximated value
if N_sma >= 5:
radam_step_size = group['lr'] * math.sqrt((1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / (N_sma_max - 2)) / (1 - beta1 ** state['step'])
else:
radam_step_size = group['lr'] / (1 - beta1 ** state['step'])
buffered[2] = radam_step_size
if group['weight_decay'] != 0:
p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32)
# more conservative since it's an approximated value
radam_step = p_data_fp32.clone()
if N_sma >= 5:
denom = exp_avg_sq.sqrt().add_(group['eps'])
radam_step.addcdiv_(-radam_step_size, exp_avg, denom)
else:
radam_step.add_(-radam_step_size, exp_avg)
radam_norm = radam_step.pow(2).sum().sqrt()
weight_norm = p.data.pow(2).sum().sqrt().clamp(0, 10)
if weight_norm == 0 or radam_norm == 0:
trust_ratio = 1
else:
trust_ratio = weight_norm / radam_norm
state['weight_norm'] = weight_norm
state['adam_norm'] = radam_norm
state['trust_ratio'] = trust_ratio
if N_sma >= 5:
p_data_fp32.addcdiv_(-radam_step_size * trust_ratio, exp_avg, denom)
else:
p_data_fp32.add_(-radam_step_size * trust_ratio, exp_avg)
p.data.copy_(p_data_fp32)
return loss
@redknightlois
Copy link
Author

No, this was a prototype that I knocked up in a few hours time. Feel free to add those and I will update it.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment