File size: 9,251 Bytes
2042bc3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
245
246
247
248
249
250
251
252
253
254
255
import numpy as np
import torch
from torch.autograd import Function
import cvxpy as cp

def _np(x): return x.detach().cpu().numpy()
def _sym(P): return 0.5 * (P + P.T)

CACHE = True
_QP_CACHE = {}
_EQ_CACHE = {}

def _qp_cvx_osqp(Pn, qn, Gn, hn, An, bn, eps_abs=1e-5, eps_rel=1e-5, max_iter=10000):
    Pn = _sym(Pn).astype(np.float64)
    qn = qn.reshape(-1).astype(np.float64)
    Gn = Gn.astype(np.float64); hn = hn.reshape(-1).astype(np.float64)
    An = An.astype(np.float64); bn = bn.reshape(-1).astype(np.float64)

    n = Pn.shape[0]
    m = Gn.shape[0]
    p = An.shape[0]
    key = (n, m, p)

    if (not CACHE) or (key not in _QP_CACHE):
        x = cp.Variable(n)
        Pp = cp.Parameter((n, n), symmetric=True)
        qp = cp.Parameter(n)
        cons = []
        ineq = eq = None
        Gp = hp = Ap = bp = None
        if m > 0:
            Gp = cp.Parameter((m, n))
            hp = cp.Parameter(m)
            ineq = (Gp @ x <= hp)
            cons.append(ineq)
        if p > 0:
            Ap = cp.Parameter((p, n))
            bp = cp.Parameter(p)
            eq = (Ap @ x == bp)
            cons.append(eq)
        obj = cp.Minimize(0.5 * cp.quad_form(x, cp.psd_wrap(Pp)) + qp @ x)
        prob = cp.Problem(obj, cons)
        bundle = {"x": x, "Pp": Pp, "qp": qp, "Gp": Gp, "hp": hp, "Ap": Ap, "bp": bp, "ineq": ineq, "eq": eq, "prob": prob}
        if CACHE:
            _QP_CACHE[key] = bundle
    else:
        bundle = _QP_CACHE[key]

    bundle["Pp"].value = Pn
    bundle["qp"].value = qn
    if m > 0:
        bundle["Gp"].value = Gn
        bundle["hp"].value = hn
    if p > 0:
        bundle["Ap"].value = An
        bundle["bp"].value = bn

    bundle["prob"].solve(
        solver=cp.OSQP,
        warm_start=True,
        verbose=False,
        eps_abs=eps_abs,
        eps_rel=eps_rel,
        max_iter=max_iter,
        polish=True,
    )
    st = bundle["prob"].status
    if st not in (cp.OPTIMAL, cp.OPTIMAL_INACCURATE):
        raise RuntimeError(f"CVXPY/OSQP status: {st}")

    xval = np.asarray(bundle["x"].value, dtype=np.float64).reshape(-1)
    lam = np.asarray(bundle["ineq"].dual_value, dtype=np.float64).reshape(-1) if m > 0 else np.zeros((0,), dtype=np.float64)
    nu = np.asarray(bundle["eq"].dual_value, dtype=np.float64).reshape(-1) if p > 0 else np.zeros((0,), dtype=np.float64)
    y = np.concatenate([lam, nu], axis=0)
    return xval, y

def _eq_qp_cvx_osqp(Pn, gnp, Ab, eps_abs=1e-5, eps_rel=1e-5, max_iter=10000):
    Pn = _sym(Pn).astype(np.float64)
    gnp = gnp.reshape(-1).astype(np.float64)
    Ab = Ab.astype(np.float64)

    k, n = Ab.shape
    key = (n, k)

    if (not CACHE) or (key not in _EQ_CACHE):
        z = cp.Variable(n)
        Pp = cp.Parameter((n, n), symmetric=True)
        gp = cp.Parameter(n)
        Ap = cp.Parameter((k, n))
        con = (Ap @ z == 0)
        obj = cp.Minimize(0.5 * cp.quad_form(z, cp.psd_wrap(Pp)) + gp @ z)
        prob = cp.Problem(obj, [con])
        bundle = {"z": z, "Pp": Pp, "gp": gp, "Ap": Ap, "con": con, "prob": prob}
        if CACHE:
            _EQ_CACHE[key] = bundle
    else:
        bundle = _EQ_CACHE[key]

    bundle["Pp"].value = Pn
    bundle["gp"].value = gnp
    bundle["Ap"].value = Ab

    bundle["prob"].solve(
        solver=cp.OSQP,
        warm_start=True,
        verbose=False,
        eps_abs=eps_abs,
        eps_rel=eps_rel,
        max_iter=max_iter,
        polish=True,
    )
    st = bundle["prob"].status
    if st not in (cp.OPTIMAL, cp.OPTIMAL_INACCURATE):
        raise RuntimeError(f"CVXPY/OSQP status: {st}")

    zval = np.asarray(bundle["z"].value, dtype=np.float64).reshape(-1)
    yb = np.asarray(bundle["con"].dual_value, dtype=np.float64).reshape(-1)
    return zval, yb

def bpqp_backward(xn, yn, Pn, Gn, hn, An, gnp, m, p, act_tol):
    lam = yn[:m] if m > 0 else np.zeros((0,), dtype=np.float64)
    if m > 0:
        resid = Gn @ xn - hn
        active = np.where((resid > -act_tol) | (lam > act_tol))[0].astype(np.int64)
    else:
        active = np.zeros((0,), dtype=np.int64)

    rows = []
    if active.size > 0:
        rows.append(Gn[active, :])
    if p > 0:
        rows.append(An)

    n = xn.size
    if len(rows) == 0:
        try:
            z = -np.linalg.solve(Pn, gnp)
        except np.linalg.LinAlgError:
            z = -np.linalg.solve(Pn + 1e-8 * np.eye(n), gnp)
        return z, np.zeros((0,), dtype=np.float64), active

    Ab = np.vstack(rows)
    z, yb = _eq_qp_cvx_osqp(Pn, gnp.astype(np.float64), Ab)
    return z, yb, active

def BPQPLayer(sign=1, act_tol=1e-6):
    class _Layer(Function):
        @staticmethod
        def forward(ctx, P, q, G, h, A, b):
            batched = (P.dim() == 3)
            B = P.shape[0] if batched else 1
            xs, ys, ms, ps = [], [], [], []

            for i in range(B):
                Pi = P[i] if batched else P
                qi = q[i] if q.dim() == 2 else q
                Gi = G[i] if G.dim() == 3 else G
                hi = h[i] if h.dim() == 2 else h
                Ai = A[i] if A.dim() == 3 else A
                bi = b[i] if b.dim() == 2 else b

                Pn = _sym(_np(Pi)).astype(np.float64)
                qn = (sign * _np(qi)).reshape(-1).astype(np.float64)
                Gn = _np(Gi).astype(np.float64)
                hn = _np(hi).reshape(-1).astype(np.float64)
                An = _np(Ai).astype(np.float64)
                bn = _np(bi).reshape(-1).astype(np.float64)

                m, p = Gn.shape[0], An.shape[0]
                x_np, y_np = _qp_cvx_osqp(Pn, qn, Gn, hn, An, bn)

                xs.append(torch.from_numpy(x_np).to(device=Pi.device, dtype=Pi.dtype))
                ys.append(torch.from_numpy(y_np).to(device=Pi.device, dtype=Pi.dtype))
                ms.append(m); ps.append(p)

            x_out = torch.stack(xs, 0) if batched else xs[0]
            y_out = torch.stack(ys, 0) if batched else ys[0]

            ctx.save_for_backward(P, q, G, h, A, b, x_out, y_out)
            ctx.meta = (batched, B, sign, act_tol, ms, ps)
            return x_out

        @staticmethod
        def backward(ctx, grad_output):
            P, q, G, h, A, b, x, y = ctx.saved_tensors
            batched, B, sign, act_tol, ms, ps = ctx.meta

            gP = torch.zeros_like(P); gq = torch.zeros_like(q); gG = torch.zeros_like(G)
            gh = torch.zeros_like(h); gA = torch.zeros_like(A); gb = torch.zeros_like(b)

            for i in range(B):
                Pi = P[i] if batched else P
                qi = q[i] if q.dim() == 2 else q
                hi = h[i] if h.dim() == 2 else h
                Gi = G[i] if G.dim() == 3 else G
                Ai = A[i] if A.dim() == 3 else A
                bi = b[i] if b.dim() == 2 else b

                xi = x[i] if batched else x
                yi = y[i] if batched else y
                gi = grad_output[i] if batched else grad_output

                Pn = _sym(_np(Pi)).astype(np.float64)
                Gn = _np(Gi).astype(np.float64)
                hn = _np(hi).reshape(-1).astype(np.float64)
                An = _np(Ai).astype(np.float64)
                m, p = Gn.shape[0], An.shape[0]

                z, yb, active = bpqp_backward(
                    _np(xi), _np(yi), Pn, Gn, hn, An, _np(gi), m, p, act_tol
                )

                zt = torch.from_numpy(z).to(device=Pi.device, dtype=Pi.dtype)
                gq_i = sign * zt
                gP_i = 0.5 * (torch.outer(zt, xi) + torch.outer(xi, zt))

                lam = yi[:m] if m > 0 else torch.empty((0,), device=Pi.device, dtype=Pi.dtype)
                nu = yi[m:m+p] if p > 0 else torch.empty((0,), device=Pi.device, dtype=Pi.dtype)

                k = int(active.size)
                mu = torch.from_numpy(yb[:k]).to(device=Pi.device, dtype=Pi.dtype) if k > 0 else torch.empty((0,), device=Pi.device, dtype=Pi.dtype)
                eta = torch.from_numpy(yb[k:k+p]).to(device=Pi.device, dtype=Pi.dtype) if p > 0 else torch.empty((0,), device=Pi.device, dtype=Pi.dtype)

                gG_i = torch.zeros_like(Gi); gh_i = torch.zeros_like(hi)
                if m > 0 and k > 0:
                    at = torch.tensor(active, device=Pi.device, dtype=torch.long)
                    lam_act = lam.index_select(0, at)
                    block = mu[:, None] * xi[None, :] + lam_act[:, None] * zt[None, :]
                    gG_i.index_copy_(0, at, block)
                    gh_i.index_copy_(0, at, -mu)

                if p > 0:
                    gb_i = -eta
                    gA_i = eta[:, None] * xi[None, :] + nu[:, None] * zt[None, :]
                else:
                    gb_i = torch.zeros_like(bi)
                    gA_i = torch.zeros_like(Ai)

                if batched: gP[i] = gP_i
                else: gP = gP + gP_i
                if q.dim() == 2: gq[i] = gq_i
                else: gq = gq + gq_i
                if G.dim() == 3: gG[i] = gG_i
                else: gG = gG + gG_i
                if h.dim() == 2: gh[i] = gh_i
                else: gh = gh + gh_i
                if A.dim() == 3: gA[i] = gA_i
                else: gA = gA + gA_i
                if b.dim() == 2: gb[i] = gb_i
                else: gb = gb + gb_i

            return gP, gq, gG, gh, gA, gb

    return _Layer.apply