FAAL/FAAL.py at main · TrustAI/FAAL · GitHub

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
# +
import cvxpy as cp
import torch
import numpy as np
from torch.autograd import Variable
import torchattacks
import warnings
import mosek
import os
import torch.nn as nn
warnings.filterwarnings("ignore")
os.environ['MOSEKLM_LICENSE_FILE'] = "mosek.lic"

class DAW:

    def __init__(self,
                 train_batch_size,
                 r_choice,
                 learning_approach = "kl",
                 output_return = "weight"
                 ):

        self.train_batch_size = train_batch_size
        self.numerical_eps = 0.000001
        self.learning_approach = learning_approach
        self.output_return = output_return
        self.r_choice = r_choice


    def _initialise_FAAL_problem(self,y):

        # The primal - inner maximisation problem.
        N = self.train_batch_size
        N2= len(y)

        Pemp = 1/N * np.ones(N)  # Change for a diffrent Pemp

        # # Parameter controlling robustness to misspecification
        # α = cp.Constant(self.a_choice)
        # Parameter controlling robustness to statistical error
        r = cp.Constant(self.r_choice)

        # Primal variables and constraints, indep o_initialise_HD_problem2f problem
        self.p = cp.Variable(shape=N, nonneg=True)


        self.nn_loss = cp.Parameter(shape=N)
        self.nn_loss.value = [1/N]*N  # Initialising

        self.worst = cp.Parameter()
        self.worst.value = 0.1  # Initialising

        y = y.cpu()

        # Objective function
        aa = cp.multiply(self.p[0:N][y], self.nn_loss[y])

        objective = cp.Maximize(cp.sum(aa))

        # Simplex constraints
        simplex_constraints = [cp.sum(self.p) == 1]

        kl_constraint = [cp.sum(cp.kl_div(Pemp, self.p)) <= r]
        # t = cp.Variable(name="t", shape=N)
        # kl_constraint = [cp.sum(t) <= r, cp.constraints.exponential.ExpCone(-1*t, Pemp, self.p)]


        complete_constraints = simplex_constraints + kl_constraint


        # Problem definition
        self.model = cp.Problem(
            objective=objective,
            constraints=complete_constraints)


    def solve_weight(self,  y,inf_loss = None, device='cuda'):

        '''Solving the primal problem.
           Returning the weighted loss as a tensor Pytorch can autodiff'''


        if self.learning_approach == "kl":

            self._initialise_FAAL_problem(y)

        else:
            assert 0


        if self.r_choice > 0:

            if self.output_return == 'weights':
                self.nn_loss.value = np.array(inf_loss.cpu().detach().numpy())

            self.worst.value = np.max(self.nn_loss.value) # DPP step


            try:
                self.model.solve(solver=cp.ECOS)
                # ECOS is normally faster than MOSEK for conic problems (it is built for this purpose),
                # but generally also more unstable.
                # We will revert to MOSEK incase of solving issues.
                # This should happen very infrequently (<1/1000 calls or so, depending on α, r)

            except:

                try:
                    # self.nn_loss.value += self.numerical_eps # Small amt of noise incase its a numerical issue
                    self.worst.value = np.max(self.nn_loss.value) # Must also re-instate worst-case for DPP
                    self.model.solve(solver=cp.MOSEK)
                    # MOSEK is the second fastest,
                    # But also occasionally fails when α and r are too large.

                except:
                    self.model.solve(solver=cp.SCS)
                    # Last resort. Rarely needed.


            weights = Variable(torch.from_numpy(self.p.value),
                               requires_grad=True).to(device).float() # Converting primal weights to tensors


            if self.output_return == 'weights':

                return weights.detach()

            else:
                raise Exception("Not a valid choice of output, please pass pytorch_loss_function if using Pytorch or weights if using another framework")

        else: # If we use only epsilon (could be zero or not)

            if self.output_return == 'weights':
                return torch.ones((self.train_batch_size)).cuda()*1/(self.train_batch_size)