def cgtostring(vec):
tmpstr = []
for c in vec:
if np.absolute(c) < self.prec:
tmpstr.append(" ")
continue
tmpstr.append("$%s$" % (lu.latexify(c)))
tmpstr = " & ".join(tmpstr)
return tmpstr
def tostring(vec, j):
tmpstr = []
for m, c in enumerate(vec):
if np.absolute(c) < self.prec:
continue
tmpstr.append("%s |%d,%+d\\rangle" %
(lu.latexify(c), j, m - j))
tmpstr = " + ".join(tmpstr)
tmpstr = " ".join(["$", tmpstr, "$"])
return tmpstr
K = 4 # Number of trajectories
n = 20 # dim of solution space
m = 2 # dim of objective space
rs = circle_points(K) # preference
pmtl_refs = 1. / rs
pf = create_pf()
methods = ["EPO", "PMTL"]
colors = list(mcolors.TABLEAU_COLORS.values())
print(colors)
markers = {"EPO": '*', "PMTL": '^'}
msz = {"EPO": 90, "PMTL": 35}
for method in methods:
latexify(fig_width=2.25, fig_height=1.8)
fig, ax = plt.subplots()
fig.subplots_adjust(left=.12, bottom=.12, right=.975, top=.975)
label = 'Pareto Front' if init_type == "easy" and method == "EPO" else ''
ax.plot(pf[:, 0], pf[:, 1], lw=2, c='k', label=label)
for k, r in enumerate(rs):
if init_type == "hard":
x0 = np.random.uniform(0.15 * (-1)**(k), .5 * (-1)**(k), n)
else:
x0 = np.random.uniform(-0.4, 0.4, n)
l0, _ = concave_fun_eval(x0)
label = r'$l(\theta^0)$' if init_type == "hard" and k == 0 and method == "PMTL" else ''
plt.scatter([l0[0]], [l0[1]],
c=colors[k],
from problems.toy_biobjective import circle_points, concave_fun_eval, create_pf
from solvers import epo_search, pareto_mtl_search, linscalar, moo_mtl_search, meta_search
import matplotlib.pyplot as plt
from latex_utils import latexify
if __name__ == '__main__':
K = 4 # Number of trajectories
n = 20 # dim of solution space
m = 2 # dim of objective space
rs = circle_points(K) # preference
pmtl_K = 5
pmtl_refs = circle_points(pmtl_K, 0, np.pi / 2)
methods = ['EPO', 'PMTL', 'MOOMTL', 'LinScalar', "Meta"]
latexify(fig_width=2., fig_height=1.55)
ss, mi = 0.1, 100
pf = create_pf()
for method in methods:
fig, ax = plt.subplots()
fig.subplots_adjust(left=.12, bottom=.12, right=.97, top=.97)
ax.plot(pf[:, 0], pf[:, 1], lw=2, c='k', label='Pareto Front')
last_ls = []
for k, r in enumerate(rs):
r_inv = 1. / r
ep_ray = 1.1 * r_inv / np.linalg.norm(r_inv)
ep_ray_line = np.stack([np.zeros(m), ep_ray])
label = r'$r^{-1}$ ray' if k == 0 else ''
ax.plot(ep_ray_line[:, 0],
ep_ray_line[:, 1],
color='k',
def case1_satisfyingMCF():
n = args.n # dim of solution space
m = args.m # dim of objective space
##construct x0
x0 = np.zeros(n)
x0[range(0, n, 2)] = -0.2
x0[range(1, n, 2)] = -0.2
eps_set = [0.8, 0.6, 0.4, 0.2]
color_set = ["c", "g", "orange", "b"]
latexify(fig_width=2.2, fig_height=1.8)
l0, _ = concave_fun_eval(x0)
max_iters = args.max_iters
relax = True
pf = create_pf()
fig = plt.figure()
fig.subplots_adjust(left=.12, bottom=.12, right=.9, top=.9)
label = 'Pareto\nFront' if relax else ''
plt.plot(pf[:, 0], pf[:, 1], lw=2.0, c='k', label=label)
label = r'$l(\theta^0)$'
plt.scatter([l0[0]], [l0[1]], c='r', s=40)
plt.annotate(label, xy=(l0[0] + 0.03, l0[1]), xytext=(l0[0] + 0.03, l0[1]))
for idx, eps0 in enumerate(eps_set):
if eps0 == 0.2:
eps_plot = np.array([[i * 0.1 * 0.903, 0.2] for i in range(11)])
plt.plot(eps_plot[:, 0],
eps_plot[:, 1],
color="gray",
label=r'$\epsilon$',
lw=1,
ls='--')
elif eps0 == 0.4:
eps_plot = np.array([[i * 0.1 * 0.807, 0.4] for i in range(11)])
plt.plot(eps_plot[:, 0],
eps_plot[:, 1],
color="gray",
lw=1,
ls='--')
elif eps0 == 0.6:
eps_plot = np.array([[i * 0.1 * 0.652, 0.6] for i in range(11)])
plt.plot(eps_plot[:, 0],
eps_plot[:, 1],
color="gray",
lw=1,
ls='--')
elif eps0 == 0.8:
eps_plot = np.array([[i * 0.1 * 0.412, 0.8] for i in range(11)])
plt.plot(eps_plot[:, 0],
eps_plot[:, 1],
color="gray",
lw=1,
ls='--')
else:
print("error eps0")
exit()
c = color_set[idx]
_, res = hco_search(concave_fun_eval,
x=x0,
deltas=[0.5, 0.5],
para_delta=0.5,
lr_delta=args.lr_delta,
relax=False,
eps=[eps0, args.eps1, args.eps2],
max_iters=max_iters,
n_dim=args.n,
step_size=args.step_size,
grad_tol=args.grad_tol,
store_xs=args.store_xs)
ls = res['ls']
alpha = 1.0
zorder = 1
# plt.plot(ls[:, 0], ls[:, 1], c=c, lw=2.0, alpha=alpha, zorder=zorder)
plt.plot(ls[:, 0], ls[:, 1], c=c, lw=2.0)
print(ls[-1])
plt.scatter(ls[[-1], 0], ls[[-1], 1], c=c, s=40)
plt.xlabel(r'$l_1$')
plt.ylabel(r'$l_2$', rotation="horizontal")
plt.legend(loc='lower left', handletextpad=0.3, framealpha=0.9)
ax = plt.gca()
ax.xaxis.set_label_coords(1.05, -0.02)
ax.yaxis.set_label_coords(-0.02, 1.02)
ax.spines['right'].set_color('none')
ax.spines['top'].set_color('none')
plt.savefig('figures/satifying' + '.pdf')
plt.close()
from latex_utils import latexify
import os
baseline = "indiv"
methods = ["EPO", "PMTL", "LinScalar"]
markers = {"EPO": "*", "PMTL": "^", "LinScalar": "s", "indiv": "o"}
marker_szs = {"EPO": 5, "PMTL": 3.5, "LinScalar": 2}
colors = {"EPO": "g", "PMTL": "b", "LinScalar": "r"}
dataset = "rf1"
model = 'fnn'
niters, nprefs = 100, 10
folder = f"results"
m = 8
latexify(fig_width=3.3, fig_height=1.8)
# Baseline stem plot
file = f"{baseline}_{dataset}_{model}_{niters}.pkl"
results = pkl.load(open(os.path.join(folder, file), 'rb'))
baseline_l = []
for j in results:
res = results[j]["res"]
lj = res["training_losses"][-1][j] / (m * 1000)
baseline_l.append(lj)
markerline, stemlines, baseline = plt.stem(baseline_l, label='Baseline',
use_line_collection=True,
basefmt=' ')
plt.setp(stemlines, 'linewidth', 8, 'alpha', 0.5,
'color', 'gray', 'zorder', -5)
plt.setp(markerline, 'ms', 8, 'color', 'gray', 'zorder', -5, 'marker', '_')
from latex_utils import latexify
import os
baseline = "indiv"
methods = ["EPO", "PMTL", "LinScalar"]
markers = {"EPO": "*", "PMTL": "^", "LinScalar": "s", "indiv": "o"}
marker_szs = {"EPO": 5, "PMTL": 3.5, "LinScalar": 2}
colors = {"EPO": "g", "PMTL": "b", "LinScalar": "r"}
dataset = "emotion"
model = 'fnn'
niters, nprefs = 200, 10
folder = f"results"
m = 6
latexify(fig_width=4, fig_height=1.8)
fig = plt.figure()
fig.subplots_adjust(left=0.12, bottom=0.1, right=.78, top=1)
# Baseline stem plot
file = f"{baseline}_{dataset}_{model}_{niters}.pkl"
results = pkl.load(open(os.path.join(folder, file), 'rb'))
baseline_l = []
for j in results:
res = results[j]["res"]
lj = res["training_losses"][-1][j] / m
baseline_l.append(lj)
markerline, stemlines, baseline = plt.stem(baseline_l, label='Baseline',
use_line_collection=True,
basefmt=' ')
plt.setp(stemlines, 'linewidth', 8, 'alpha', 0.5, 'color', 'gray', 'zorder', -5)
xs, ys, zs = proj3d.proj_transform(xs3d, ys3d, zs3d, renderer.M)
self.set_positions((xs[0], ys[0]), (xs[1], ys[1]))
FancyArrowPatch.draw(self, renderer)
if __name__ == '__main__':
K = 3 # Number of trajectories
n = 20 # dim of solution space
m = 3 # dim of objective space
rs = sphere_points(K)
max_iters = 500
ss = [0.05, 0.2, 0.05, 0.05, 0.02, 0.05]
colors = list(mcolors.TABLEAU_COLORS.values())
pf, pf_tri, ls, tri = create_pf()
latexify(fig_width=4., fig_height=3.6)
params = {
'axes.labelsize': 16,
'axes.titlesize': 16,
'font.size': 15,
'legend.fontsize': 18,
'xtick.labelsize': 12,
'ytick.labelsize': 12
}
mpl.rcParams.update(params)
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
ax.plot_trisurf(pf[:, 0],
pf[:, 1],
pf[:, 2],
a.append(a0)
if side in ["right", "both"]:
a1 = ax.annotate(caps[1],
xy=(xdata[1], ydata[1]),
zorder=2,
**anno_args)
a.append(a1)
return (line, tuple(a))
if __name__ == '__main__':
theta = np.linspace(-4, 4, 150)
l1 = [f1(np.array([x])) for x in theta]
l2 = [f2(np.array([x])) for x in theta]
latexify(fig_width=3, fig_height=3.0)
fig, ax = plt.subplots()
fig.subplots_adjust(left=0.025, bottom=.15, right=.975, top=.975)
ax.spines['left'].set_position('zero')
ax.spines['right'].set_color('none') # turn off the right spine/ticks
ax.spines['top'].set_color('none') # turn off the top spine/ticks
ax.xaxis.tick_bottom()
ax.plot(theta, l1, color="black", label=r'$l_1(\theta)$', lw=2)
ax.plot(theta, l2, color="red", label=r'$l_2(\theta)$', lw=2)
# Comment out the twin axis code below for ppt
axt = ax.twinx()
axt.axis("off")
# add_interval(axt, (-3, -1.05), (0.02, 0.02), "r", "()",
# r"$\theta^0 \preccurlyeq -\mathbf{1}/\sqrt{n}$ or " +
if k != baseline)
data[dataset]["baseline_loss"] = [
([0, max_task1_loss + 0.05], [base_task2_loss, base_task2_loss]),
([base_task1_loss, base_task1_loss], [0, max_task2_loss + 0.05]),
]
data[dataset]["baseline_acc"] = [
([0, 1], [base_task2_acc, base_task2_acc]),
([base_task1_acc, base_task1_acc], [0, 1]),
]
# setup preferences
data[dataset]["rs"] = data[dataset][methods[0]]["rs"]
latexify(fig_width=3, fig_height=2)
for dataset in datasets:
fig, ax = plt.subplots(1, 1)
x_min, x_max = data[dataset]["baseline_loss"][0][0]
y_min, y_max = data[dataset]["baseline_loss"][1][1]
ax.set_xlim(left=x_min, right=x_max + 0.05)
ax.set_ylim(bottom=y_min, top=y_max + 0.05)
ax.set_xlabel(r"Task 1 Loss")
ax.set_ylabel(r"Task 2 Loss")
for i, (xs, ys) in enumerate(data[dataset]["baseline_loss"]):
ax.plot(xs, ys, lw=1, alpha=0.3, c="k")
colors = []
for i, r in enumerate(data[dataset]["rs"]):
label = r"Preference" if i == 0 else ""
