m = 40
n = 2
A = np.zeros((m, n))
b = np.zeros(m)
for i in range(m):
a = np.random.randn(n)
a = a / np.linalg.norm(a)
bi = 1 + np.random.rand(1)
A[i, :] = a
b[i] = bi
q = QCML()
q.parse('''
dimensions m n
variables x(n) r
parameters A(m,n) b(m)
maximize r
A*x + r <= b
''')
q.canonicalize()
q.dims = {'m': m, 'n': n}
q.codegen("python")
socp_data = q.prob2socp(locals())
# stuffed variable size
n = socp_data['G'].shape[1]
parser.add_argument('-n',
type=int,
help='Number of templates in dictionary')
parser.add_argument(
'-c',
'--codegen',
help='Codegen type to use (python, matlab, or C; default python)',
default='python')
args = parser.parse_args()
m, n = (args.m, args.n)
print "Running CBP example...."
# TODO: takeaways from this example: "diag" constructor?
p = QCML(debug=True)
p.parse("""
dimensions m n
variable c(n)
variable u(n)
variable v(n)
parameter noise positive
parameter lambda(n)
parameter data(m)
parameter dictc(m,n)
parameter dictu(m,n)
parameter dictv(m,n)
parameter radii(n,n) # diagonal matrix
parameter rctheta(n,n) # diagonal matrix
minimize noise*norm(data - (dictc*c + dictu*u + dictv*v)) + lambda'*c
subject to
s_right = np.sign(diff_right)
diff_right = np.abs(diff_right)
diff_right = s_right * np.log(diff_right + 1)
s_right = np.abs(diff_right)
s_right = 1.0 / (s_right + 1)
#s_down is the scaling we want to associate with each difference. Large differences
#in the original image are given small weight in the objective
s_down = np.diag(s_down)
# <codecell>
from qcml import QCML
import cvxopt
p = QCML()
rows = (m - 1) * n
cols = m * n
s = '''
dimension cols
dimension rows
variable y(cols)
parameter top(rows,cols)
parameter bottom(rows,cols)
parameter left(rows,cols)
parameter right(rows,cols)
parameter diff_down(rows)
parameter diff_right(rows)
