-
Notifications
You must be signed in to change notification settings - Fork 0
/
myoptimizer.py
213 lines (186 loc) · 12 KB
/
myoptimizer.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
import math
import random
import torch
from torch.optim.optimizer import Optimizer, required
class ALQ_optimizer(Optimizer):
"""Implement ALQ optimizer.
Arguments:
params (iterable): iterable of parameters to optimize or dicts defining parameter groups.
lr (float, optional): learning rate (default: 1e-3).
betas (Tuple[float, float], optional): coefficients used for computing running averages of gradient and its square (default: (0.9, 0.999)).
eps (float, optional): term added to the denominator to improve numerical stability (default: 1e-8).
weight_decay (float, optional): weight decay (L2 regularization) (default: 0).
Reference:
Adam optimizer by Pytorch:
https://pytorch.org/docs/stable/_modules/torch/optim/adam.html#Adam
On the Convergence of Adam and Beyond:
https://openreview.net/forum?id=ryQu7f-RZ
"""
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.):
# Check the validity
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
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))
defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)
super(ALQ_optimizer, self).__init__(params, defaults)
def __setstate__(self, state):
super(ALQ_optimizer, self).__setstate__(state)
def step(self, params_bin, mode, pruning_rate=None, closure=None):
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
# Check if this is a pruning step
if pruning_rate is not None:
importance_list = torch.tensor([])
for i, (p_bin, p) in enumerate(zip(params_bin, group['params'])):
if p.grad is None:
continue
# Compute the gradient in both w domain and alpha domain
grad = p.grad.data
grad_alpha = p_bin.construct_grad_alpha(grad)
state = self.state[p]
# Initialize the state parameters in both w domain and alpha domain
if len(state) == 0:
state['step_alpha'] = 0
state['exp_avg_alpha'] = torch.zeros_like(p_bin.alpha)
state['exp_avg_sq_alpha'] = torch.zeros_like(p_bin.alpha)
state['max_exp_avg_sq_alpha'] = torch.zeros_like(p_bin.alpha)
state['step'] = 0
state['exp_avg'] = torch.zeros_like(p.data)
state['exp_avg_sq'] = torch.zeros_like(p.data)
state['max_exp_avg_sq'] = torch.zeros_like(p.data)
if mode == 'coordinate':
# Update the state parameters in w domain
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
max_exp_avg_sq = state['max_exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(1 - beta1, grad)
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
# Maintain the maximum of all second moment running avg. till now
torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
# Update the state parameters in alpha domain
exp_avg_alpha, exp_avg_sq_alpha = state['exp_avg_alpha'], state['exp_avg_sq_alpha']
max_exp_avg_sq_alpha = state['max_exp_avg_sq_alpha']
state['step_alpha'] += 1
# L2 regularization on coordinates (in alpha domain)
if group['weight_decay'] != 0:
grad_alpha = grad_alpha.add(p_bin.alpha, alpha=group['weight_decay'])
# Decay the first and second moment running average coefficient
exp_avg_alpha.mul_(beta1).add_(1 - beta1, grad_alpha)
exp_avg_sq_alpha.mul_(beta2).addcmul_(1 - beta2, grad_alpha, grad_alpha)
# Maintain the maximum of all second moment running avg. till now
torch.max(max_exp_avg_sq_alpha, exp_avg_sq_alpha, out=max_exp_avg_sq_alpha)
# Use the max. for normalizing running avg. of gradient
denom_alpha = max_exp_avg_sq_alpha.sqrt().add_(group['eps'])
bias_correction1 = 1 - beta1 ** state['step_alpha']
bias_correction2 = 1 - beta2 ** state['step_alpha']
# Compute the pseudo gradient and the pseudo diagonal Hessian
pseudo_grad_alpha = (group['lr'] / bias_correction1) * exp_avg_alpha
pseudo_hessian_alpha = denom_alpha.div(math.sqrt(bias_correction2))
# Check if this is a pruning step
if pruning_rate is not None:
# Compute the integer used to determine the number of selected alpha's in this layer
float_tmp = p_bin.num_bin_filter.item()*pruning_rate[0]
int_tmp = int(float_tmp)
if random.random()<float_tmp-int_tmp:
int_tmp += 1
# Sort the importance of binary filters (alpha's) in this layer and select Top-k% (int_tmp) unimportant ones
p_bin_importance_list = p_bin.sort_importance_bin_filter(pseudo_grad_alpha, pseudo_hessian_alpha, int_tmp)
importance_list = torch.cat((importance_list,p_bin_importance_list), 0)
else:
# Take one optimization step on coordinates
p_bin.alpha.add_(-pseudo_grad_alpha/pseudo_hessian_alpha)
# Reconstruct the weight tensor from the current quantization
p_bin.update_w_FP()
tmp_p = p.detach()
tmp_p.zero_().add_(p_bin.w_FP.data)
elif mode == 'basis':
# Update the state parameters in w domain
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
max_exp_avg_sq = state['max_exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(1 - beta1, grad)
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
# Maintain the maximum of all second 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().add_(group['eps'])
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
# Compute the pseudo gradient and the pseudo diagonal Hessian
pseudo_grad = (group['lr'] / bias_correction1) * exp_avg
pseudo_hessian = denom.div(math.sqrt(bias_correction2))
# Take one optimization step on binary bases
p_bin.optimize_bin_basis(pseudo_grad, pseudo_hessian)
# Speed up with an optimization step on coordinates
p_bin.speedup(pseudo_grad, pseudo_hessian)
# Reconstruct the weight tensor from the current quantization
p_bin.update_w_FP()
tmp_p = p.detach()
tmp_p.zero_().add_(p_bin.w_FP.data)
# Update the state parameters in alpha domain (approximately)
state['step_alpha'] += 1
state['exp_avg_alpha'] = p_bin.construct_grad_alpha(exp_avg)
state['exp_avg_sq_alpha'] = p_bin.construct_hessian_alpha(exp_avg_sq)
state['max_exp_avg_sq_alpha'] = p_bin.construct_hessian_alpha(max_exp_avg_sq)
elif mode == 'ste':
# Update the state parameters in w domain
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
max_exp_avg_sq = state['max_exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(1 - beta1, grad)
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
# Maintain the maximum of all second 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().add_(group['eps'])
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
# Compute the pseudo gradient and the pseudo diagonal Hessian
pseudo_grad = (group['lr'] / bias_correction1) * exp_avg
pseudo_hessian = denom.div(math.sqrt(bias_correction2))
# Take one optimization step on binary bases
p_bin.optimize_bin_basis(pseudo_grad, pseudo_hessian)
# Speed up with an optimization step on coordinates
p_bin.speedup(pseudo_grad, pseudo_hessian)
# Update the maintained full precision weights
p_bin.update_w_FP(-pseudo_grad/pseudo_hessian)
# Reconstruct the weight tensor from the current quantization
tmp_p = p.detach()
tmp_p.zero_().add_(p_bin.reconstruct_w())
# Update the state parameters in alpha domain (approximately)
state['step_alpha'] += 1
state['exp_avg_alpha'] = p_bin.construct_grad_alpha(exp_avg)
state['exp_avg_sq_alpha'] = p_bin.construct_hessian_alpha(exp_avg_sq)
state['max_exp_avg_sq_alpha'] = p_bin.construct_hessian_alpha(max_exp_avg_sq)
# Check if this is a pruning step
if pruning_rate is not None:
# Resort the importance of selected binary filters (alpha's) over all layers
sorted_ind = torch.argsort(importance_list[:,-1])
# Compute the number of pruned alpha's in this iteration
# Note that unlike the paper, M_p varies over iterations here, but this does not influence the pruning schedule.
M_p = int(sorted_ind.nelement()*pruning_rate[1])
# Determine indexes of alpha's to be pruned
ind_prune = sorted_ind[:M_p]
list_prune = importance_list[ind_prune,:]
# Prune alpha's in each layer and reconstruct the weight tensor
for i, (p_bin, p) in enumerate(zip(params_bin, group['params'])):
p_bin.prune_alpha((torch.sort(list_prune[list_prune[:,0]==i,1])[0]).to(torch.int64))
p_bin.update_w_FP()
tmp_p = p.detach()
tmp_p.zero_().add_(p_bin.w_FP.data)
return loss