def pre(cfg):
name = cfg["method"]
sigma = cfg["sigma"]
scale = cfg["scale"]
K = cfg["K"]
kernel = RBFKernel(sigma=sigma,scale=scale)
fastLogDet = FastIVM(K, kernel, 1.0)
if name == "Greedy":
opt = Greedy(K, fastLogDet)
if name == "IndependentSetImprovement":
opt = IndependentSetImprovement(K, fastLogDet)
elif name == "Random":
opt = Random(K, fastLogDet, cfg["run_id"])
elif name == "SieveStreaming":
e = cfg["epsilon"]
opt = SieveStreaming(K, fastLogDet, 1.0, e)
elif name == "SieveStreaming++":
e = cfg["epsilon"]
opt = SieveStreamingPP(K, fastLogDet, 1.0, e)
elif name == "Salsa":
e = cfg["epsilon"]
opt = Salsa(K, fastLogDet, 1.0, e)
elif name == "ThreeSieves":
e = cfg["epsilon"]
T = cfg["T"]
opt = ThreeSieves(K, fastLogDet, 1.0, e, "sieve", T)
return opt
# Target solutions depending on the kernel
target_rbf = np.array(sorted([
[0.0,0.0],
[1.0,1.0],
[0.0,1.0]
]))
target_poly = np.array(sorted([
[0.0,1.0],
[1.0,1.0],
[1.0,1.0]
]))
# Define all the kernel / submodular function combinations
kernel = RBFKernel(sigma=1,scale=1)
ivm_rbf = FastIVM(K, kernel = kernel, sigma = 1.0)
kernel = PolyKernel()
ivm_custom_kernel_class = FastIVM(K, kernel = kernel, sigma = 1.0)
ivm_custom_kernel_function = FastIVM(K, kernel = poly_kernel, sigma = 1.0)
ivm_custom_class = FastLogdet(K)
ivm_custom_function = ivm
optimizers = {}
### GREEDY ###
optimizers["Greedy with IVM + RBF"] = Greedy(K, ivm_rbf)
optimizers["Greedy with IVM + poly kernel class"] = Greedy(K, ivm_custom_kernel_class)
optimizers["Greedy with IVM + poly kernel function"] = Greedy(K, ivm_custom_kernel_function)
optimizers["Greedy with custom IVM class"] = Greedy(K, ivm_custom_class)