def setUp(self):
self.samps = {
# Empty sample
"empty": ddm.Sample(aa([]), aa([]), 0),
# Simple sample
"simple": ddm.Sample(aa([1, 2]), aa([.5, .7]), 0),
# Sample with conditions
"conds": ddm.Sample(aa([1, 2, 3]), aa([]), 0,
cond1=(aa([1, 1, 2]), aa([]))),
# Sample with conditions and explicitly showing undecided
"condsexp": ddm.Sample(aa([1, 2, 3]), aa([]), 0,
cond1=(aa([1, 1, 2]), aa([]), aa([]))),
# Sample with undecided
"undec": ddm.Sample(aa([1, 2]), aa([.5, .7]), 2),
# Sample with undecided and conditions
"undeccond": ddm.Sample(aa([1, 2, 3]), aa([]), 3,
cond1=(aa([1, 1, 2]), aa([]), aa([2, 2, 1]))),
# For the adding test
"adda": ddm.Sample(aa([1]), aa([2, 4]), 3,
cond1=(aa(["a"]), aa(["a", "b"]), aa(["a", "b", "b"]))),
"addb": ddm.Sample(aa([1.5, 2, 1]), aa([]), 1,
cond1=(aa(["b", "b", "c"]), aa([]), aa(["d"]))),
# Two conditions
"two": ddm.Sample(aa([1]), aa([2]), 1,
conda=(aa(["a"]), aa(["b"]), aa(["a"])),
condb=(aa([1]), aa([2]), aa([2]))),
}
def test_LossLikelihood(self):
"""Likelihood loss function"""
# We can calculate likelihood for this simple case
m = self.FakePointModel(dt=.02)
sol = m.solve()
s = ddm.Sample(aa([.02]), aa([]))
expected = -np.log(np.sum(sol.corr)/m.dt)
assert np.isclose(expected, ddm.models.LossLikelihood(sample=s, dt=m.dt, T_dur=m.T_dur).loss(m))
# And for the uniform case we can assert equivalence
m = self.FakeUniformModel()
s1 = ddm.Sample(aa([.02, .05, .07, .12]), aa([.33, .21]))
s2 = ddm.Sample(aa([.13, .1, .02]), aa([.66, .15, .89]))
assert np.isclose(ddm.models.LossLikelihood(sample=s1, dt=m.dt, T_dur=m.T_dur).loss(m),
ddm.models.LossLikelihood(sample=s2, dt=m.dt, T_dur=m.T_dur).loss(m))
def test_BIC(self):
"""BIC loss function"""
# -2*Likelihood == BIC for a sample size of 1
m = self.FakePointModel(dt=.02)
sol = m.solve()
s = ddm.Sample(aa([.02]), aa([]))
expected = -np.log(np.sum(sol.corr)/m.dt)
assert np.isclose(ddm.models.LossBIC(sample=s, dt=m.dt, T_dur=m.T_dur, nparams=1, samplesize=1).loss(m),
2*ddm.models.LossLikelihood(sample=s, dt=m.dt, T_dur=m.T_dur).loss(m))
def test_LossSquaredError(self):
"""Squared error loss function"""
# Should be zero for empty sample when all undecided
m = self.FakeUndecidedModel()
s = ddm.Sample(aa([]), aa([]), undecided=1)
assert ddm.models.LossSquaredError(sample=s, dt=m.dt, T_dur=m.T_dur).loss(m) == 0
# Can also be determined precisely for the point model
m = self.FakePointModel()
sol = m.solve()
err = ddm.models.LossSquaredError(sample=s, dt=m.dt, T_dur=m.T_dur).loss(m)
assert np.isclose(err, np.sum(sol.corr)**2 + np.sum(sol.err)**2)
def prepare_sample_for_variable_drift(sample, resolution=RESOLUTION):
new_samples = []
for i in range(0, resolution):
corr = sample.corr.copy()
err = sample.err.copy()
undecided = sample.undecided
conditions = copy.deepcopy(sample.conditions)
conditions['driftnum'] = (np.asarray([i]*len(corr)),
np.asarray([i]*len(err)),
np.asarray([i]*undecided))
new_samples.append(ddm.Sample(corr, err, undecided, **conditions))
new_sample = new_samples.pop()
for s in new_samples:
new_sample += s
return new_sample
FakePointModel.solve_numerical = FakePointModel.solve
FakePointModel.solve_numerical_cn = FakePointModel.solve
FakePointModel.solve_numerical_implicit = FakePointModel.solve
FakePointModel.solve_numerical_explicit = FakePointModel.solve
m = FakePointModel()
s = m.solve()
so = OverlayNonDecisionUniform(nondectime=-.05, st=.1).apply(s)
so2 = OverlayNonDecisionUniform2(nondectime=-.05, st=.1).apply(s)
assert np.all(so.corr == so2.corr)
so = OverlayNonDecisionGamma(shape=4, scale=.01, nondectime=.1).apply(s)
print(list(so.corr))
print(np.sum(so.corr))
#m = FakePointModel(overlay=OverlayNonDecisionGamma(shape=ddm.Fittable(minval=1, maxval=3), scale=ddm.Fittable(minval=0, maxval=.2), nondectime=ddm.Fittable(minval=0, maxval=1)))
m = FakePointModel(overlay=OverlayNonDecisionGaussian(
ndsigma=ddm.Fittable(minval=0, maxval=1),
nondectime=ddm.Fittable(minval=0, maxval=1)))
#m = FakePointModel(overlay=OverlayNonDecisionUniform(st=ddm.Fittable(minval=0, maxval=1), nondectime=ddm.Fittable(minval=0, maxval=1)))
import ddm.plot
ddm.plot.model_gui(
m, ddm.Sample(np.asarray([1, 1.1, 1.2, 1.3, 1.4, 1.5]), np.asarray([1])))
from ddm import ICGaussian
r = ICPointRew(x0=.99)
#print(r.get_IC(np.linspace(-1, 1, 201), dx=.01, conditions={"highreward": 1}))
def setUp(self):
# You need some gymnastics to get numpy to accept an array of
# same-length tuples
_tuple_same_length = np.empty(3, dtype=object)
_tuple_same_length[:] = [(3, 1, 2), (3, 3, 3), (5, 4, 3)]
self.samps = {
# Empty sample
"empty":
ddm.Sample(aa([]), aa([]), 0),
# Simple sample
"simple":
ddm.Sample(aa([1, 2]), aa([.5, .7]), 0),
# Sample with conditions
"conds":
ddm.Sample(aa([1, 2, 3]), aa([]), 0, cond1=(aa([1, 1,
2]), aa([]))),
# Sample with conditions as strings
"condsstr":
ddm.Sample(aa([1, 2, 3]),
aa([]),
0,
cond1=(aa(["x", "yy", "z z z"]), aa([]))),
# Sample with conditions as tuples
"condstuple":
ddm.Sample(aa([1, 2, 3]),
aa([]),
0,
cond1=(aa([(3, 1, 2), (5, 5), (1, 1, 1, 1, 1)],
dtype=object), aa([]))),
# Sample with conditions as tuples which are the same length
"condstuplesame":
ddm.Sample(aa([1, 2, 3]),
aa([]),
0,
cond1=(_tuple_same_length, aa([]))),
# Sample with conditions and explicitly showing undecided
"condsexp":
ddm.Sample(aa([1, 2, 3]),
aa([]),
0,
cond1=(aa([1, 1, 2]), aa([]), aa([]))),
# Sample with undecided
"undec":
ddm.Sample(aa([1, 2]), aa([.5, .7]), 2),
# Sample with undecided and conditions
"undeccond":
ddm.Sample(aa([1, 2, 3]),
aa([]),
3,
cond1=(aa([1, 1, 2]), aa([]), aa([2, 2, 1]))),
# For the adding test
"adda":
ddm.Sample(aa([1]),
aa([2, 4]),
3,
cond1=(aa(["a"]), aa(["a", "b"]), aa(["a", "b", "b"]))),
"addb":
ddm.Sample(aa([1.5, 2, 1]),
aa([]),
1,
cond1=(aa(["b", "b", "c"]), aa([]), aa(["d"]))),
# Two conditions
"two":
ddm.Sample(aa([1]),
aa([2]),
1,
conda=(aa(["a"]), aa(["b"]), aa(["a"])),
condb=(aa([1]), aa([2]), aa([2]))),
}
