def test_last_axis_precision_recall_curve():
dims = [2, 3, 4, 5]
for ndim in [1, 2, 3, 4]:
if ndim == 1:
d = 10
else:
d = dims[:ndim]
a = np.random.binomial(n=1, p=.5, size=d).astype('float32')
p = np.random.random(d)
prec, recall, _ = tmetrics.last_axis_precision_recall_curve(a, p)
print 'd: {}'.format(d)
print prec
print recall
print _
def test_precision_recall_curves_all_dims(n_iter=1):
dims = [(50,), (50, 100), (50, 30, 40), (5, 10, 5, 3)]
int_types = [T.fvector, T.fmatrix, T.ftensor3, T.ftensor4]
float_types = [T.fvector, T.fmatrix, T.ftensor3, T.ftensor4]
for d, int_type, float_type in zip(dims[1:], int_types[1:], float_types[1:]):
yt = int_type('yt')
yp = float_type('yp')
a = float_type('a')
b = float_type('b')
gpu_auc = tmetrics.auc(a, b)
get_auc = theano.function([a, b], gpu_auc)
p_expr, r_expr, t_expr = tmetrics.precision_recall_curves(yt, yp)
pr_auc = tmetrics.auc(r_expr, p_expr)
f = theano.function([yt, yp], [p_expr, r_expr, t_expr, pr_auc])
for epoch in xrange(n_iter):
true = np.random.binomial(n=1, p=.5, size=d).astype('float32')
predicted = np.random.random((d)).astype('float32')
precision, recall, thresh, avg = f(true, predicted)
refp, refr, reft = tmetrics.last_axis_precision_recall_curve(true, predicted)
try:
assert np.allclose(precision, refp, equal_nan=True)
assert np.allclose(recall, refr, equal_nan=True)
assert np.allclose(thresh, reft, equal_nan=True)
assert np.allclose(avg, get_auc(refr.astype('float32'), refp.astype('float32')), equal_nan=True)
except:
print true
print predicted
print 'precision'
print precision
print 'ref precision'
print refp
print 'recall'
print recall
print recall.shape
print 'ref recall'
print refr
print refr.shape
print thresh
print reft
print avg
print get_auc(refr.astype('float32'), refp.astype('float32'))
raise
