You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
97 lines
3.2 KiB
97 lines
3.2 KiB
4 years ago
|
"""
|
||
|
|
SGDP Optimizer Implementation copied from https://github.com/clovaai/AdamP/blob/master/adamp/sgdp.py
|
||
|
|
|
||
|
|
Paper: `Slowing Down the Weight Norm Increase in Momentum-based Optimizers` - https://arxiv.org/abs/2006.08217
|
||
|
|
Code: https://github.com/clovaai/AdamP
|
||
|
|
|
||
|
|
Copyright (c) 2020-present NAVER Corp.
|
||
|
|
MIT license
|
||
|
|
"""
|
||
|
|
|
||
|
|
import torch
|
||
|
|
import torch.nn as nn
|
||
|
|
from torch.optim.optimizer import Optimizer, required
|
||
|
|
import math
|
||
|
|
|
||
|
|
class SGDP(Optimizer):
|
||
|
|
def __init__(self, params, lr=required, momentum=0, dampening=0,
|
||
|
|
weight_decay=0, nesterov=False, eps=1e-8, delta=0.1, wd_ratio=0.1):
|
||
|
|
defaults = dict(lr=lr, momentum=momentum, dampening=dampening, weight_decay=weight_decay,
|
||
|
|
nesterov=nesterov, eps=eps, delta=delta, wd_ratio=wd_ratio)
|
||
|
|
super(SGDP, self).__init__(params, defaults)
|
||
|
|
|
||
|
|
def _channel_view(self, x):
|
||
|
|
return x.view(x.size(0), -1)
|
||
|
|
|
||
|
|
def _layer_view(self, x):
|
||
|
|
return x.view(1, -1)
|
||
|
|
|
||
|
|
def _cosine_similarity(self, x, y, eps, view_func):
|
||
|
|
x = view_func(x)
|
||
|
|
y = view_func(y)
|
||
|
|
|
||
|
|
x_norm = x.norm(dim=1).add_(eps)
|
||
|
|
y_norm = y.norm(dim=1).add_(eps)
|
||
|
|
dot = (x * y).sum(dim=1)
|
||
|
|
|
||
|
|
return dot.abs() / x_norm / y_norm
|
||
|
|
|
||
|
|
def _projection(self, p, grad, perturb, delta, wd_ratio, eps):
|
||
|
|
wd = 1
|
||
|
|
expand_size = [-1] + [1] * (len(p.shape) - 1)
|
||
|
|
for view_func in [self._channel_view, self._layer_view]:
|
||
|
|
|
||
|
|
cosine_sim = self._cosine_similarity(grad, p.data, eps, view_func)
|
||
|
|
|
||
|
|
if cosine_sim.max() < delta / math.sqrt(view_func(p.data).size(1)):
|
||
|
|
p_n = p.data / view_func(p.data).norm(dim=1).view(expand_size).add_(eps)
|
||
|
|
perturb -= p_n * view_func(p_n * perturb).sum(dim=1).view(expand_size)
|
||
|
|
wd = wd_ratio
|
||
|
|
|
||
|
|
return perturb, wd
|
||
|
|
|
||
|
|
return perturb, wd
|
||
|
|
|
||
|
|
def step(self, closure=None):
|
||
|
|
loss = None
|
||
|
|
if closure is not None:
|
||
|
|
loss = closure()
|
||
|
|
|
||
|
|
for group in self.param_groups:
|
||
|
|
weight_decay = group['weight_decay']
|
||
|
|
momentum = group['momentum']
|
||
|
|
dampening = group['dampening']
|
||
|
|
nesterov = group['nesterov']
|
||
|
|
|
||
|
|
for p in group['params']:
|
||
|
|
if p.grad is None:
|
||
|
|
continue
|
||
|
|
grad = p.grad.data
|
||
|
|
state = self.state[p]
|
||
|
|
|
||
|
|
# State initialization
|
||
|
|
if len(state) == 0:
|
||
|
|
state['momentum'] = torch.zeros_like(p.data)
|
||
|
|
|
||
|
|
# SGD
|
||
|
|
buf = state['momentum']
|
||
|
|
buf.mul_(momentum).add_(1 - dampening, grad)
|
||
|
|
if nesterov:
|
||
|
|
d_p = grad + momentum * buf
|
||
|
|
else:
|
||
|
|
d_p = buf
|
||
|
|
|
||
|
|
# Projection
|
||
|
|
wd_ratio = 1
|
||
|
|
if len(p.shape) > 1:
|
||
|
|
d_p, wd_ratio = self._projection(p, grad, d_p, group['delta'], group['wd_ratio'], group['eps'])
|
||
|
|
|
||
|
|
# Weight decay
|
||
|
|
if weight_decay != 0:
|
||
|
|
p.data.mul_(1 - group['lr'] * group['weight_decay'] * wd_ratio / (1-momentum))
|
||
|
|
|
||
|
|
# Step
|
||
|
|
p.data.add_(-group['lr'], d_p)
|
||
|
|
|
||
|
|
return loss
|