import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from torch.optim.optimizer import Optimizer, required
[docs]
class SONG(torch.optim.Optimizer):
r"""
Stochastic Optimization for NDCG (SONG) and its top-K variant (K-SONG) is used for optimizing :obj:`~libauc.losses.NDCGLoss`. The key update steps are summarized as follows:
.. math::
1. & \mathbf{u}_{q,i}^{t+1} = (1-\gamma_0)\mathbf{u}_{q,i}^{t} + \gamma_0 \frac{1}{|B_q|} \sum_{x^{\prime}\in B_q} \exp(h_q(x^{\prime};\mathbf{w})-h_q(x;\mathbf{w}))
2. & G(\mathbf{w}_t) = \frac{1}{|Q_t|} \frac{1}{|B_q^+|} \frac{1}{|B_q|} \sum_{q\in Q_t} \sum_{x_i^q\in B_q^+} \sum_{x_j^q\in B_q} \frac{1}{\mathbf{u}_{q,i}^{t+1}} \nabla_{\mathbf{w}} (h_q(x_j^q;\mathbf{w}_t)-h_q(x_i^q;\mathbf{w}_t))
3. & m_{t+1} = \beta_1 m_{t} + (1-\beta_1) G(\mathbf{w}_t)
4. & \mathbf{w}_{t+1} = \mathbf{w}_t - \eta_1 m_{t+1} (or Adam style)
Args:
params (iterable): iterable of parameters to optimize
lr (float): learning rate
mode (str): optimization mode, 'sgd' or 'adam' (default: ``'sgd'``)
clip_value (float, optional): gradient clipping value (default: ``1.0``)
weight_decay (float, optional): weight decay (L2 penalty) (default: ``1e-5``)
epoch_decay (float, optional): epoch decay (epoch-wise l2 penalty) (default: ``0.0``)
momentum (float, optional): momentum factor for 'sgd' mode (default: ``0.9``)
betas (Tuple[float, float], optional): coefficients used for computing
running averages of gradient and its square for 'adam' mode (default: ``(0.9, 0.999)``)
eps (float, optional): term added to the denominator to improve
numerical stability for 'adam' mode (default: ``1e-8``)
amsgrad (bool, optional): whether to use the AMSGrad variant of 'adam' mode
from the paper `On the Convergence of Adam and Beyond` (default: ``False``)
verbose (bool, optional): whether to print optimization progress (default: ``True``)
device (torch.device, optional): the device used for optimization, e.g., 'cpu' or 'cuda' (default: ``None``)
Example:
>>> optimizer = libauc.optimizers.SONG(model.parameters(), lr=0.1, momentum=0.9)
>>> optimizer.zero_grad()
>>> loss_fn(predictions, batch).backward() # loss_fn can be ListwiseCE_Loss or NDCG_Loss
>>> optimizer.step()
Reference:
.. [1] Qiu, Zi-Hao, Hu, Quanqi, Zhong, Yongjian, Zhang, Lijun, and Yang, Tianbao.
"Large-scale Stochastic Optimization of NDCG Surrogates for Deep Learning with Provable Convergence."
Proceedings of the 39th International Conference on Machine Learning. 2022.
https://arxiv.org/abs/2202.12183
"""
def __init__(self,
params,
lr=required,
clip_value=1.0,
weight_decay=0,
epoch_decay=0,
mode='sgd',
momentum=0.9,
dampening=0,
nesterov=False,
betas=(0.9, 0.999),
eps=1e-8,
amsgrad=False,
verbose=False,
device=None,
**kwargs):
if not device:
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
else:
self.device = device
if lr is not required and lr < 0.0:
raise ValueError("Invalid learning rate: {}".format(lr))
if momentum < 0.0:
raise ValueError("Invalid momentum value: {}".format(momentum))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {}".format(eps))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
if not 0.0 <= weight_decay:
raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
if not isinstance(mode, str):
raise ValueError("Invalid mode type: {}".format(mode))
self.params = list(params)
self.lr = lr
self.mode = mode.lower()
self.model_ref = self.__init_model_ref__(self.params) if epoch_decay > 0 else None
self.model_acc = self.__init_model_acc__(self.params) if epoch_decay > 0 else None
self.T = 0 # for epoch_decay
self.steps = 0 # total optimization steps
self.verbose = verbose # print updates for lr/regularizer
self.epoch_decay = epoch_decay
assert self.mode in ['adam', 'sgd'], "Keyword is not found in [`adam`, `sgd`]!"
defaults = dict(lr=lr, weight_decay=weight_decay, epoch_decay=epoch_decay,
momentum=momentum, dampening=dampening, nesterov=nesterov,
betas=betas, eps=eps, amsgrad=amsgrad,
clip_value=clip_value, model_ref=self.model_ref, model_acc=self.model_acc)
if nesterov and (momentum <= 0 or dampening != 0): # sgd
raise ValueError("Nesterov momentum requires a momentum and zero dampening")
super(SONG, self).__init__(self.params, defaults)
def __setstate__(self, state):
r"""
# Set default options for sgd mode and adam mode
"""
super(SONG, self).__setstate__(state)
for group in self.param_groups:
if self.mode == 'sgd':
group.setdefault('nesterov', False)
elif self.mode == 'adam':
group.setdefault('amsgrad', False)
else:
NotImplementedError
def __init_model_ref__(self, params):
model_ref = []
if not isinstance(params, list):
params = list(params)
for var in params:
if var is not None:
model_ref.append(torch.empty(var.shape).normal_(mean=0, std=0.01).to(self.device))
return model_ref
def __init_model_acc__(self, params):
model_acc = []
if not isinstance(params, list):
params = list(params)
for var in params:
if var is not None:
model_acc.append(torch.zeros(var.shape, dtype=torch.float32, device=self.device, requires_grad=False).to(self.device))
return model_acc
[docs]
@torch.no_grad()
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
for group in self.param_groups:
self.lr = group['lr']
weight_decay = group['weight_decay']
momentum = group['momentum']
dampening = group['dampening']
nesterov = group['nesterov']
clip_value = group['clip_value']
epoch_decay = group['epoch_decay']
model_ref = group['model_ref']
model_acc = group['model_acc']
if self.mode == 'sgd':
for i, p in enumerate(group['params']):
if p.grad is None:
continue
if epoch_decay > 0:
d_p = torch.clamp(p.grad.data , -clip_value, clip_value) + epoch_decay*(p.data - model_ref[i].data) + weight_decay*p.data
else:
d_p = torch.clamp(p.grad.data , -clip_value, clip_value) + weight_decay*p.data
if momentum != 0:
param_state = self.state[p]
if 'momentum_buffer' not in param_state:
buf = param_state['momentum_buffer'] = torch.clone(d_p).detach()
else:
buf = param_state['momentum_buffer']
buf.mul_(momentum).add_(d_p, alpha=1 - dampening)
if nesterov:
d_p = d_p.add(buf, alpha=momentum)
else:
d_p = buf
p.add_(d_p, alpha=-group['lr'])
if epoch_decay > 0:
model_acc[i].data = model_acc[i].data + p.data
elif self.mode == 'adam':
for i, p in enumerate(group['params']):
if p.grad is None:
continue
#grad = p.grad
if epoch_decay > 0:
grad = torch.clamp(p.grad.data , -clip_value, clip_value) + epoch_decay*(p.data - model_ref[i].data) + weight_decay*p.data
else:
grad = torch.clamp(p.grad.data , -clip_value, clip_value) + weight_decay*p.data
if grad.is_sparse:
raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')
amsgrad = group['amsgrad']
state = self.state[p]
# State initialization
if len(state) == 0:
state['step'] = 0
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
if amsgrad:
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
if amsgrad:
max_exp_avg_sq = state['max_exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
#if group['weight_decay'] != 0:
# grad = grad.add(p, alpha=group['weight_decay'])
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
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
denom = (max_exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(group['eps'])
else:
denom = (exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(group['eps'])
step_size = group['lr'] / bias_correction1
p.addcdiv_(exp_avg, denom, value=-step_size)
if epoch_decay > 0:
model_acc[i].data = model_acc[i].data + p.data
self.steps += 1
self.T += 1
return loss
[docs]
def update_lr(self, decay_factor=None):
if decay_factor != None:
self.param_groups[0]['lr'] = self.param_groups[0]['lr']/decay_factor
print ('Reducing learning rate to %.5f !'%(self.param_groups[0]['lr']))
[docs]
def update_regularizer(self, decay_factor=None):
if decay_factor != None:
self.param_groups[0]['lr'] = self.param_groups[0]['lr']/decay_factor
if self.verbose:
print ('Reducing learning rate to %.5f @ T=%s!'%(self.param_groups[0]['lr'], self.steps))
if self.verbose:
print ('Updating regularizer @ T=%s!'%(self.steps))
for i, param in enumerate(self.model_ref):
self.model_ref[i].data = self.model_acc[i].data/self.T
for i, param in enumerate(self.model_acc):
self.model_acc[i].data = torch.zeros(param.shape, dtype=torch.float32, device=self.device, requires_grad=False).to(self.device)
self.T = 0