-
Notifications
You must be signed in to change notification settings - Fork 15
/
Copy pathoptim_n2n.py
170 lines (149 loc) · 6.41 KB
/
optim_n2n.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
#!/usr/bin/env python3
import sys
import os
import argparse
import json
import random
import shutil
import torch
from torch import cuda
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
class OptimN2N:
def __init__(self, loss_fn, model, model_update_params,
lr=[1],
iters=20,
acc_param_grads=True,
max_grad_norm = 0,
eps = 0.00001,
momentum=0.5):
self.iters = iters
self.lr = lr
self.loss_fn = loss_fn
self.eps = eps
self.max_grad_norm = max_grad_norm
self.model = model
self.momentum = momentum
self.acc_param_grads = acc_param_grads
if self.acc_param_grads:
self.params = model_update_params
self.param_grads = [torch.zeros([self.iters] + list(p.size())).type_as(p.data)
for p in self.params]
def forward(self, input, y, verbose=False):
self.seeds = np.random.randint(3435, size=self.iters)
return self.forward_mom(input, y, verbose)
def backward(self, grad_output, verbose=False):
grads = self.backward_mom(grad_output, verbose)
return grads
def grad_norm(self, g_list, g_norm_list):
for g, g_norm in zip(g_list, g_norm_list):
g_norm2 = (g**2).sum(1)**0.5
g.div_(g_norm2.unsqueeze(1).expand_as(g)).mul_(g_norm.unsqueeze(1).expand_as(g))
def clip_grad_norm(self, parameters, max_norm, norm_type=2):
if len(parameters) > 0:
max_norm = float(max_norm)
norm_type = float(norm_type)
if norm_type == float('inf'):
total_norm = max(p.abs().max() for p in parameters)
else:
total_norm = 0
for p in parameters:
param_norm = p.norm(norm_type)
total_norm += param_norm ** norm_type
total_norm = total_norm ** (1. / norm_type)
clip_coef = max_norm / (total_norm + 1e-6)
if clip_coef < 1:
for p in parameters:
p.mul_(clip_coef)
def forward_mom(self, input, y, verbose=False):
self.y = y
self.input_grads = [torch.zeros([self.iters] + list(x.size())).type_as(x.data) for x in input]
self.mom_params = [torch.zeros(x.size()).type_as(x) for x in self.input_grads]
self.input_cache = [torch.zeros(x.size()).type_as(x) for x in self.input_grads]
self.all_z = []
if self.acc_param_grads:
for p in self.param_grads:
p.zero_()
for k in range(self.iters):
self.all_z.append(Variable(torch.cuda.FloatTensor(input[0].size()).normal_(0, 1)))
torch.manual_seed(int(self.seeds[k]))
loss = self.loss_fn(input, self.y, self.model, self.all_z[k])
if self.acc_param_grads:
all_input_params = input + self.params
else:
all_input_params = input
all_grads_k = torch.autograd.grad(loss, all_input_params, retain_graph = True)
input_grad_k = all_grads_k[:len(input)]
param_grads_k = all_grads_k[len(input):]
if self.max_grad_norm > 0:
self.clip_grad_norm([input_grad_k[0].data], self.max_grad_norm)
self.clip_grad_norm([input_grad_k[1].data], self.max_grad_norm)
if self.acc_param_grads:
for i, p in enumerate(param_grads_k):
self.param_grads[i][k].copy_(p.data)
for i, x_grad_k in enumerate(input_grad_k):
self.input_cache[i][k].copy_(input[i].data)
self.input_grads[i][k].copy_(x_grad_k.data)
for i in range(len(self.mom_params)):
if k == 0:
self.mom_params[i][k] = -input_grad_k[i].data
else:
self.mom_params[i][k] = self.mom_params[i][k-1]*self.momentum -input_grad_k[i].data
lr_k_list = [lr for lr in self.lr]
input = [Variable(x.data + lr_k * p[k], requires_grad=True) for x, p, lr_k in
zip(input, self.mom_params, lr_k_list)]
if verbose:
print('mom', k, loss.data[0])
return input
def backward_mom(self, grad_output, verbose=False):
input_kp1_grad = [g.data for g in grad_output]
p_kp1_grad = [torch.zeros(x.size()).type_as(x) for x in input_kp1_grad]
rev_iters = self.iters
for k in reversed(range(rev_iters)):
lr_k_list = [lr for lr in self.lr]
input_k_grad = input_kp1_grad
p_kp1_grad = [p + lr_k*x for p, x, lr_k in zip(p_kp1_grad, input_kp1_grad, lr_k_list)]
input_k_rv = []
input_H_xx_v = []
r = self.eps
for i in range(len(p_kp1_grad)):
v = p_kp1_grad[i]
x_k = self.input_cache[i][k]
x_k_rv = Variable((x_k + r*v).type_as(x_k), requires_grad = True)
input_k_rv.append(x_k_rv)
if self.acc_param_grads:
all_input_params = input_k_rv + self.params
else:
all_input_params = input_k_rv
torch.manual_seed(int(self.seeds[k]))
loss = self.loss_fn(input_k_rv, self.y, self.model, self.all_z[k])
all_grads_rv_k = torch.autograd.grad(loss, all_input_params, retain_graph=True)
if self.max_grad_norm > 0:
self.clip_grad_norm([g.data for g in all_grads_rv_k], self.max_grad_norm)
input_grads_rv_k = all_grads_rv_k[:len(input_k_rv)]
param_grads_rv_k = all_grads_rv_k[len(input_k_rv):]
if self.acc_param_grads:
H_wx_v_list = []
for i, p_grad_rv_k in enumerate(param_grads_rv_k):
H_wx_v = (p_grad_rv_k.data - self.param_grads[i][k]) / r
H_wx_v_list.append(H_wx_v)
if self.params[i].grad is None:
self.params[i].grad = Variable(torch.zeros(self.params[i].size()).type_as(
self.params[i].data))
if self.max_grad_norm > 0:
self.clip_grad_norm(H_wx_v_list, self.max_grad_norm)
for i in range(len(self.params)):
self.params[i].grad.data += -H_wx_v_list[i]
for i, x_k_rv_grad in enumerate(input_grads_rv_k):
H_xx_v = (x_k_rv_grad.data - self.input_grads[i][k])/r
input_H_xx_v.append(H_xx_v)
input_kp1_grad = [x_kp1_grad - H_xx_v
for (x_kp1_grad, H_xx_v) in zip(input_kp1_grad, input_H_xx_v)]
if self.max_grad_norm > 0:
self.clip_grad_norm(input_kp1_grad, self.max_grad_norm)
p_kp1_grad = [p.mul_(self.momentum) for p in p_kp1_grad]
if verbose:
print('mom', k, input_kp1_grad[0][0].norm())
return input_kp1_grad