Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Built-in cosine sim and inplace ops #7

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Aug 27, 2020
Merged

Built-in cosine sim and inplace ops #7

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 5 additions & 8 deletions adamp/adamp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim.optimizer import Optimizer, required
import math

Expand All @@ -26,11 +27,7 @@ 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
return F.cosine_similarity(x, y, dim=1, eps=eps).abs_()

def _projection(self, p, grad, perturb, delta, wd_ratio, eps):
wd = 1
Expand Down Expand Up @@ -77,8 +74,8 @@ def step(self, closure=None):
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']

exp_avg.mul_(beta1).add_(1 - beta1, grad)
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)

denom = (exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(group['eps'])
step_size = group['lr'] / bias_correction1
Expand All @@ -98,6 +95,6 @@ def step(self, closure=None):
p.data.mul_(1 - group['lr'] * group['weight_decay'] * wd_ratio)

# Step
p.data.add_(-step_size, perturb)
p.data.add_(perturb, alpha=-step_size)

return loss
14 changes: 5 additions & 9 deletions adamp/sgdp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim.optimizer import Optimizer, required
import math

Expand All @@ -26,11 +27,7 @@ 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
return F.cosine_similarity(x, y, dim=1, eps=eps).abs_()

def _projection(self, p, grad, perturb, delta, wd_ratio, eps):
wd = 1
Expand All @@ -54,7 +51,6 @@ def step(self, closure=None):
loss = closure()

for group in self.param_groups:
weight_decay = group['weight_decay']
momentum = group['momentum']
dampening = group['dampening']
nesterov = group['nesterov']
Expand All @@ -71,7 +67,7 @@ def step(self, closure=None):

# SGD
buf = state['momentum']
buf.mul_(momentum).add_(1 - dampening, grad)
buf.mul_(momentum).add_(grad, alpha=1 - dampening)
if nesterov:
d_p = grad + momentum * buf
else:
Expand All @@ -83,10 +79,10 @@ def step(self, closure=None):
d_p, wd_ratio = self._projection(p, grad, d_p, group['delta'], group['wd_ratio'], group['eps'])

# Weight decay
if weight_decay != 0:
if group['weight_decay'] > 0:
p.data.mul_(1 - group['lr'] * group['weight_decay'] * wd_ratio / (1-momentum))

# Step
p.data.add_(-group['lr'], d_p)
p.data.add_(d_p, alpha=-group['lr'])

return loss