def create_tr_vl_ts_nc(path, n):
"""nested circles"""
if not os.path.exists(path):
os.makedirs(path)
x, y = generate_nested_circles(n)
nbr = x.shape[0]
l1 = int(nbr*2/3.)
l2 = int(nbr * ((2/3.) + 1/6.))
trainx, trainy = x[:l1, :], y[:l1]
validx, validy = x[l1:l2, :], y[l1:l2]
testx, testy = x[l2:, :], y[l2:]
trfig = plot_classes(
trainy, trainx, "", 0., "g.tr 2D: nested circles." + str(l1))
vlfig = plot_classes(
validy, validx, "", 0., "g.vl 2D: nested circles." + str(l2-l1))
tsfig = plot_classes(
testy, testx, "", 0., "g.tst 2D: nested circles." + str(y.size - l2))
trfig.savefig(path + "/traingfig.png", bbox_inches='tight')
vlfig.savefig(path + "/validfig.png", bbox_inches='tight')
tsfig.savefig(path + "/testfig.png", bbox_inches='tight')
# dump
with open(path+"/nc.pkl", "w") as f:
stuff = {"trainx": trainx, "trainy": trainy,
"validx": validx, "validy": validy,
"testx": testx, "testy": testy}
pkl.dump(stuff, f, protocol=pkl.HIGHEST_PROTOCOL)
def generate_2d_checkboard(x_born, y_born, s, ss):
"""x_born: [-1, 1], y_born:[-1, 1], s=10, ss=20
"""
linex = np.linspace(x_born[0], x_born[1], s, endpoint=False)
liney = np.linspace(y_born[0], y_born[1], s, endpoint=False)
x, y = [], []
start_y = True
for ix in range(linex.size - 1):
lx, lxnext = linex[ix], linex[ix+1]
for iy in range(liney.size - 1):
ly, lynext = liney[iy], liney[iy+1]
linexx = np.linspace(lx, lxnext, ss, endpoint=False)
lineyy = np.linspace(ly, lynext, ss, endpoint=False)
xv, yv = np.meshgrid(linexx, lineyy)
for i in range(xv.shape[0]):
for j in range(yv.shape[0]):
x.append([xv[i, j], yv[i, j]])
y.append(start_y)
start_y = not start_y
y = np.array(y) * 1.
x = np.array(x)
mega = np.hstack((x, y.reshape(y.size, 1)))
for i in range(500):
np.random.shuffle(mega)
print i
x = mega[:, :-1]
y = mega[:, -1]
print x.shape, y.shape
fig = plot_classes(y, x, "", 0., "generated 2D: checkboard.")
fig.savefig("data/2d/cb2d_generated.png", bbox_inches='tight')
return x, y
def create_tr_vl_ts_cb(path):
if not os.path.exists(path):
os.makedirs(path)
x, y = generate_2d_checkboard([-1, 1], [-1, 1], 10, 20)
nbr = x.shape[0]
l1 = int(nbr*2/3.)
l2 = int(nbr * ((2/3.) + 1/6.))
trainx, trainy = x[:l1, :], y[:l1]
validx, validy = x[l1:l2, :], y[l1:l2]
testx, testy = x[l2:, :], y[l2:]
trfig = plot_classes(trainy, trainx, "", 0., "g.tr 2D: checkboard.")
vlfig = plot_classes(validy, validx, "", 0., "g.vl 2D: checkboard.")
tsfig = plot_classes(testy, testx, "", 0., "g.tst 2D: checkboard.")
trfig.savefig(path + "/traingfig.png", bbox_inches='tight')
vlfig.savefig(path + "/validfig.png", bbox_inches='tight')
tsfig.savefig(path + "/testfig.png", bbox_inches='tight')
# dump
with open(path+"/cb.pkl", "w") as f:
stuff = {"trainx": trainx, "trainy": trainy,
"validx": validx, "validy": validy,
"testx": testx, "testy": testy}
pkl.dump(stuff, f, protocol=pkl.HIGHEST_PROTOCOL)
def generate_all_2d_data(path):
if not os.path.exists(path):
os.makedirs(path)
mn1, cov1 = [1, 0], [[1, -0.5], [-0.5, 1]]
mn2, cov2 = [4, 0], [[1, 0], [0, 1]]
trainx, trainy, trainfig = generate_2d_data_bin(25000, mn1,
cov1, mn2, cov2)
minx = np.min(trainx, axis=0)
maxx = np.max(trainx, axis=0)
trainx = (trainx - minx)/(maxx - minx)
trainfig = plot_classes(trainy, trainx, "", 0.,
"generated 2D: multivariate normal.")
validx, validy, validfig = generate_2d_data_bin(5000, mn1,
cov1, mn2, cov2)
validx = (validx - minx)/(maxx - minx)
validfig = plot_classes(trainy, trainx, "", 0.,
"generated 2D: multivariate normal.")
testx, testy, testfig = generate_2d_data_bin(5000, mn1,
cov1, mn2, cov2)
trainfig.savefig(path + "/traingfig.png", bbox_inches='tight')
validfig.savefig(path + "/validfig.png", bbox_inches='tight')
testfig.savefig(path + "/testfig.png", bbox_inches='tight')
def generate_2d_data_bin(nbr, mn1, cov1, mn2, cov2):
"""Generate 2D points using multivariate normal distribution.
nbr: number of samples per class."""
x1 = np.random.multivariate_normal(mn1, cov1, nbr)
x2 = np.random.multivariate_normal(mn2, cov2, nbr)
y1 = np.zeros((nbr, 1), dtype=np.float32)
y2 = np.ones((nbr, 1), dtype=np.float32)
x = np.vstack((x1, x2))
y = np.vstack((y1, y2))
print x.shape, y.shape
mega = np.hstack((x, y.reshape(y.size, 1)))
for i in range(100):
np.random.shuffle(mega)
x = mega[:, :-1]
y = mega[:, -1]
fig = plot_classes(y, x, "", 0., "generated 2D: multivariate normal.")
return x, y, fig
def generate_nested_circles(n):
limits = [0, 1./3, 2./3, 1, 2]
np.random.seed(0)
X = np.random.rand(n, 2)*2-1
Xd = np.sqrt((X**2).sum(axis=1))
Y = np.zeros((n, ), dtype='bool')
classe = True
for b1, b2 in zip(limits[:-1], limits[1:]):
(idx, ) = np.nonzero(np.logical_and(b1 < Xd, Xd <= b2))
Y[idx] = classe
classe = not classe
Y = Y.astype(np.float32)
mega = np.hstack((X, Y.reshape(Y.size, 1)))
for i in range(500):
np.random.shuffle(mega)
print i
x = mega[:, :-1]
y = mega[:, -1]
print x.shape, y.shape
fig = plot_classes(y, x, "", 0., "generated 2D: nested circles.")
fig.savefig("data/nestedcircle/nc_generated.png", bbox_inches='tight')
return x, y