-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
executable file
·244 lines (206 loc) · 11.7 KB
/
trainer.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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
"""
Code for Non-reversible Parallel Tempering for Deep Posterior Approximation
(c) Wei Deng
Nov 24, 2022
Note that in Bayesian settings, the lr 2e-6 and weight decay 25 are equivalent to lr 0.1 and weight decay 5e-4 in standard setups.
"""
#!/usr/bin/python
import math
import copy
import sys
import os
import time
import csv
import argparse
import random
import collections
from random import shuffle
import pickle
#from tqdm import tqdm ## better progressbar
from math import exp
from sys import getsizeof
import numpy as np
## import pytorch modules
import torch
import torch.optim as optim
from torch.autograd import Variable
import torch.nn as nn
from torchvision import datasets, transforms
from torchvision.datasets import ImageFolder
from torchvision.transforms import ToTensor
from torch.utils.data import DataLoader
import torch.utils.data as data
import torchvision.datasets as datasets
## Import helper functions
from tools import BayesEval, StochasticWeightAvg
from SGHMC import SGHMC
CUDA_EXISTS = torch.cuda.is_available()
""" For standard baseline ensemble """
def trainer_vanilla_ensemble(nets, train_loader, test_loader, pars):
criterion = nn.CrossEntropyLoss()
samplers = {}
for idx in range(pars.chains):
samplers[idx] = optim.SGD(nets[idx].parameters(), lr=pars.lr_min, momentum=pars.momentum, weight_decay=pars.wdecay)
BMA = BayesEval(pars.data)
for epoch in range(pars.sn):
for idx in range(pars.chains):
nets[idx].train()
for i, (images, labels) in enumerate(train_loader):
images = Variable(images).cuda() if CUDA_EXISTS else Variable(images)
labels = Variable(labels).cuda() if CUDA_EXISTS else Variable(labels)
for idx in range(pars.chains):
current_loss = criterion(nets[idx](images), labels)
samplers[idx].zero_grad()
current_loss.backward()
samplers[idx].step()
""" Report results """
if epoch % pars.chains == 0:
for idx in range(pars.chains-1, -1, -1):
BMA.eval(pars.data, nets[idx], test_loader, criterion, bma=True, uq=True)
print('Epoch {} lr: {:.4f} Acc: {:0.2f} BMA: {:0.2f} Best Acc: {:0.2f} Best BMA: {:0.2f}'.format(\
epoch, pars.lr_min, BMA.cur_acc, BMA.bma_acc, BMA.best_cur_acc, BMA.best_bma_acc))
print('Epoch {} NLL: {:.1f} Best NLL: {:.1f} BMA NLL: {:.1f} Best BMA NLL: {:.1f}'.format(epoch, BMA.nll, BMA.best_nll, BMA.bma_nll, BMA.best_bma_nll))
print('')
""" Train baselines cyclic SGHMC + SWAG """
def trainer_cyc_ensemble(nets, train_loader, test_loader, pars):
criterion = nn.CrossEntropyLoss()
sampler, BMA, SWAG = SGHMC(nets[0]), BayesEval(pars.data), StochasticWeightAvg(pars.data, cycle=100)
print('Ensemble the initial modes first for fair comparison')
for idx in range(pars.chains):
nets[idx].eval()
BMA.eval(pars.data, nets[idx], test_loader, criterion, weight=10, bma=True, uq=True, print_tag=False)
print('Initial ensemble Acc: {:0.2f} BMA: {:0.2f}'.format(BMA.cur_acc, BMA.bma_acc))
""" the 0.1 learning rate in frequentist is equivalent to 0.1 / N in Bayesian,
where N is the number of total data points """
N, warm = 50000, 0.85
for epoch in range(pars.sn):
""" cosine learning rate """
sub_sn = 100
base_lr = 0.1 / N
cur_beta = (epoch % sub_sn) * 1.0 / sub_sn
sampler.lr = base_lr / 2 * (np.cos(np.pi * min(cur_beta, warm)) + 1)
if (epoch % sub_sn) * 1.0 / sub_sn == 0:
print('Cooling down for optimization')
sampler.T /= 1e10
elif epoch % sub_sn == int(warm * sub_sn):
print('Heating up for sampling')
sampler.T *= 1e10
nets[0].train()
for i, (images, labels) in enumerate(train_loader):
images = Variable(images).cuda() if CUDA_EXISTS else Variable(images)
labels = Variable(labels).cuda() if CUDA_EXISTS else Variable(labels)
sampler.net.zero_grad()
loss = criterion(sampler.net(images), labels) * N
loss.backward()
sampler.step(images, labels)
if cur_beta >= warm:
SWAG.update(epoch // sub_sn, nets[0], cur_beta, warm)
if (epoch + 1) % sub_sn == 0 and epoch >= sub_sn - 1:
SWAG.inference(pars.data, epoch // sub_sn, train_loader, test_loader, criterion, pars.lr_min/N, repeats=10)
BMA.eval(pars.data, nets[0], test_loader, criterion, bma=True, uq=True)
print('Epoch {} lr: {:.2e} T: {:.2e} Acc: {:0.2f} BMA: {:0.2f} Best Acc: {:0.2f} Best BMA: {:0.2f}'.format(\
epoch, sampler.lr, sampler.T, BMA.cur_acc, BMA.bma_acc, BMA.best_cur_acc, BMA.best_bma_acc))
print('Epoch {} NLL: {:.1f} Best NLL: {:.1f} BMA NLL: {:.1f} Best BMA NLL: {:.1f}'.format(epoch, BMA.nll, BMA.best_nll, BMA.bma_nll, BMA.best_bma_nll))
else:
if epoch % 10 == 0:
BMA.eval(pars.data, nets[0], test_loader, criterion, bma=False, uq=False)
print('Epoch {} lr: {:.2e} T: {:.2e} Acc: {:0.2f} BMA: {:0.2f} Best Acc: {:0.2f} Best BMA: {:0.2f}'.format(\
epoch, sampler.lr, sampler.T, BMA.cur_acc, BMA.bma_acc, BMA.best_cur_acc, BMA.best_bma_acc))
else:
print('Epoch {} lr: {:.2e} T: {:.2e}'.format(epoch, sampler.lr, sampler.T))
""" For our proposed parallel tempering methods """
def trainer(nets, train_loader, test_loader, pars):
criterion = nn.CrossEntropyLoss()
ratio = (pars.lr_max / pars.lr_min) ** (1. / (pars.chains-1 + 1e-10))
current_lr = np.array([round(pars.lr_min * ratio**power, 4) for power in range(pars.chains)])
print('Initial learning rate {} in ensemble period'.format(' '.join(map(str, current_lr))))
samplers, BMAS = {}, []
for idx in range(pars.chains):
samplers[idx] = optim.SGD(nets[idx].parameters(), lr=current_lr[idx], momentum=pars.momentum, weight_decay=pars.wdecay)
BMAS.append(BayesEval(pars.data))
gates = [1] * pars.chains
cumulative_swap, tentative_swap = [0.] * pars.chains, [0.] * pars.chains
correction, iters = pars.correction, 0
if pars.window_custom == 0:
window = int(np.ceil((np.log(pars.chains) + np.log(np.log(pars.chains))) / -np.log(1 - pars.swap_rate)))
else:
window = pars.window_custom
print('The optimal window size is {}'.format(window))
""" Calculate the number of round trips """
index_set = np.arange(pars.chains)
last_visited = []
for _ in range(pars.chains):
last_visited.append(set())
round_trip, warm_epoch = 0, 5
start = time.time()
for idx in [0, 1, 2]:
print('Ensemble the initial modes first')
for inner_idx in range(pars.chains):
nets[inner_idx].eval()
BMAS[idx].eval(pars.data, nets[inner_idx], test_loader, criterion, weight=10, bma=True, uq=True, print_tag=False)
for epoch in range(pars.sn):
for idx in range(pars.chains):
nets[idx].train()
for i, (images, labels) in enumerate(train_loader):
loss_chains, swap_indicator = [0] * pars.chains, [0] * (pars.chains - 1)
images = Variable(images).cuda() if CUDA_EXISTS else Variable(images)
labels = Variable(labels).cuda() if CUDA_EXISTS else Variable(labels)
""" Open the gate to allow a swap in the current """
if iters % window == 0:
gates = [1] * pars.chains
stepsize = min(0.5, pars.gamma / (iters**0.8 + 1000.))
iters = iters + 1
""" reset the last learning rate """
last_lr = copy.deepcopy(current_lr)
for idx in range(pars.chains):
current_loss = criterion(nets[idx](images), labels)
samplers[idx].zero_grad()
current_loss.backward()
samplers[idx].step()
loss_chains[idx] = current_loss.item()
""" Swap module """
if idx > 0 and pars.swap_rate > 0:
swap_indicator[idx-1] = 1. if loss_chains[idx] + correction < loss_chains[idx-1] else 0.
tentative_swap[idx] += swap_indicator[idx-1]
if (idx % 2 == ((iters // window) % 2)) and gates[idx] == 1 and swap_indicator[idx-1] == 1:
""" swap models """
temporary = pickle.loads(pickle.dumps(nets[idx-1]))
nets[idx-1].load_state_dict(nets[idx].state_dict())
nets[idx].load_state_dict(temporary.state_dict())
""" swap losses """
loss_chains[idx-1], loss_chains[idx] = loss_chains[idx], loss_chains[idx-1]
""" swap index to record changes of round trips """
index_set[idx-1], index_set[idx] = index_set[idx], index_set[idx-1]
""" Close the gate in the current window """
gates[idx] = 0
cumulative_swap[idx] += 1.
""" avoid consecutive swaps """
print('Epoch {} Swap chain {} (loss {:0.3f}) with chain {} (loss {:0.3f}) Cumulative swaps {} iters {}'.format(\
epoch, idx-1, loss_chains[idx-1], idx, loss_chains[idx], int(cumulative_swap[idx]), iters))
""" reset sampler """
samplers[idx] = optim.SGD(nets[idx].parameters(), lr=current_lr[idx], momentum=pars.momentum, weight_decay=pars.wdecay)
samplers[idx-1] = optim.SGD(nets[idx-1].parameters(), lr=current_lr[idx-1], momentum=pars.momentum, weight_decay=pars.wdecay)
""" count the number of round trips: visited two ends """
last_visited[index_set[0]].add(0)
last_visited[index_set[pars.chains-1]].add(pars.chains-1)
""" count the number of round trips: returned to the original position """
if len(last_visited[idx]) == 2 and idx == index_set[idx]:
last_visited[idx] = set()
round_trip += 1
""" Adaptive learning rates for equi-acceptance rates """
for idx in range(1, pars.chains-1):
current_lr[idx] = (last_lr[idx-1] + max(0, last_lr[idx] - last_lr[idx-1]) * np.exp((swap_indicator[idx-1] - pars.swap_rate) * stepsize) \
+ last_lr[idx+1] - max(0, last_lr[idx+1] - last_lr[idx]) * np.exp((swap_indicator[idx] - pars.swap_rate) * stepsize)) / 2
for group in samplers[idx].param_groups:
group['lr'] = current_lr[idx]
correction = correction + (np.mean(swap_indicator) - pars.swap_rate) * stepsize / pars.scale
""" Report results """
for idx in range(pars.chains-1, -1, -1):
BMAS[idx].eval(pars.data, nets[idx], test_loader, criterion, bma=True, uq=False)
print('Epoch {} Chain {} lr: {:.4f} Acc: {:0.2f} BMA: {:0.2f} Best Acc: {:0.2f} Best BMA: {:0.2f} Loss: {:0.3f} Window {} Trips {} Swaps/tentative/rate/target {}/ {:.1e}/ {:.1e}/ {:.1e} Corrections: {:0.3f}'.format(\
epoch, idx, current_lr[idx], BMAS[idx].cur_acc, BMAS[idx].bma_acc, BMAS[idx].best_cur_acc, BMAS[idx].best_bma_acc, \
np.array(loss_chains[idx]).sum(), window, round_trip, int(cumulative_swap[idx]), tentative_swap[idx] / iters, cumulative_swap[idx] / iters, pars.swap_rate, correction))
print('Epoch {} Chain {} NLL: {:.1f} Best NLL: {:.1f} BMA NLL: {:.1f} Best BMA NLL: {:.1f}'.format(epoch, 0, BMAS[0].nll, BMAS[0].best_nll, BMAS[0].bma_nll, BMAS[0].best_bma_nll))
print('')
end = time.time()
print('Time used {:.2f}s'.format(end - start))