-
Notifications
You must be signed in to change notification settings - Fork 1
/
algs.py
156 lines (129 loc) · 4.52 KB
/
algs.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
import numpy as np
import cvxpy as cp
import mosek
import torch
import torch.nn
from torch import optim
from torch.nn.modules.module import Module
from torch.utils.data import DataLoader
import torch.nn.functional as F
###########################################
# Algorithms for selecting sample weights
def uniform(sample_losses_history):
num_samples = sample_losses_history.shape[1]
a = np.ones((num_samples,), dtype=np.float) / num_samples
return a
def adalpboost(sample_accuracy_history, sample_weights_history, eta):
acc = sample_accuracy_history[-1]
weight = sample_weights_history[-1]
weight *= np.exp(-eta * acc)
weight /= weight.sum()
return weight
# (Regularized) LPBoost
# Dual problem of LPBoost
def lpboost(sample_accuracy_history, obj_value, alpha,
beta=None, verbose=False, solve_for_lbd=False):
# Input shape: each history - (epoch, num_samples)
# Output: group_weights size: (num_samples,)
num_epochs, n = sample_accuracy_history.shape
m = alpha * n
w = cp.Variable(n)
g = cp.Variable()
objective = cp.Minimize(g) if beta is None else cp.Minimize(g - cp.sum(cp.entr(w)) / beta)
constraints = [sample_accuracy_history[i, :] @ w <= g for i in range(num_epochs)]
constraints.append(cp.sum(w) == 1)
constraints.append(0 <= w)
constraints.append(w <= 1 / m)
prob = cp.Problem(objective, constraints)
result = lpsolver(prob, verbose)
if obj_value is not None:
obj_value.append(result)
if solve_for_lbd:
lbd = [constraints[i].dual_value for i in range(num_epochs)]
lbd = np.array(lbd)
lbd[lbd < 0] = 0
lbd /= lbd.sum()
return lbd
ans = w.value
ans[ans < 0] = 0
ans /= ans.sum()
return ans
############################
# Find optimal lambda (model weights)
# Solve the dual problem of LPBoost and use the values of primal variables
def find_opt_lbd(acc_history, alpha, verbose=False):
print('==> Computing optimal model weights...')
lbd = lpboost(acc_history, None, alpha, None, verbose, True)
print('Optimal model weights found.')
print('Model weights: {}'.format(lbd))
return lbd
###########################################
# Other functions
def test(model: Module, loader: DataLoader, criterion, device: str, label_id):
"""Test the avg and group acc of the model"""
model.eval()
total_correct = 0
total_loss = 0
total_num = 0
l_rec = []
c_rec = []
with torch.no_grad():
for _, (inputs, targets) in enumerate(loader):
inputs, targets = inputs.to(device), targets.to(device)
labels = targets if label_id is None else label_id(targets)
outputs = model(inputs)
predictions = torch.argmax(outputs, dim=1)
c = (predictions == labels)
c_rec.append(c.detach().cpu().numpy())
correct = c.sum().item()
l = criterion(outputs, labels).view(-1)
l_rec.append(l.detach().cpu().numpy())
loss = l.sum().item()
total_correct += correct
total_loss += loss
total_num += len(inputs)
print('Acc: {} ({} of {})'.format(total_correct / total_num, total_correct, total_num))
print('Avg Loss: {}'.format(total_loss / total_num))
l_vec = np.concatenate(l_rec)
c_vec = np.concatenate(c_rec)
return total_correct / total_num, total_loss / total_num, \
c_vec, l_vec
def erm(model: Module, loader: DataLoader, optimizer: optim.Optimizer,
criterion, scheduler, device: str, iters=0):
"""Empirical Risk Minimization (ERM)"""
model.train()
iteri = 0
for _, (inputs, targets) in enumerate(loader):
inputs, targets = inputs.to(device), targets.to(device)
outputs = model(inputs)
loss = criterion(outputs, targets).mean()
optimizer.zero_grad()
loss.backward()
optimizer.step()
iteri += 1
if iteri == iters:
break
if scheduler is not None:
scheduler.step()
############################
# LP Solver
# Try multiple solvers because single solver might fail
def lpsolver(prob, verbose=False):
print('=== LP Solver ===')
solvers = [cp.MOSEK, cp.ECOS_BB]
for s in solvers:
print('==> Invoking {}...'.format(s))
try:
result = prob.solve(solver=s, verbose=verbose)
return result
except cp.error.SolverError as e:
print('==> Solver Error')
print('==> Invoking MOSEK simplex method...')
try:
result = prob.solve(solver=cp.MOSEK,
mosek_params={mosek.iparam.optimizer: mosek.optimizertype.free_simplex},
bfs=True, verbose=verbose)
return result
except cp.error.SolverError as e:
print('==> Solver Error')
raise cp.error.SolverError('All solvers failed.')