def pivot_selection_timed(*args, **kwargs):
print('Pivot selection')
tinit = time()
s_pivots, t_pivots = pivot_selection(*args, **kwargs)
pivot_time = time() - tinit
print(f'\t[pivot selection took {pivot_time:.3f} seconds]')
return s_pivots, t_pivots, pivot_time
mds_home= '../datasets/MDS'
dataset_home='../datasets/Webis-CLS-10'
nfolds=5
outfile = './DCI.varpivot.dat'
if exists(outfile):
rperf = Result.load(outfile, False)
else:
rperf = Result(['dataset', 'task', 'method', 'fold', 'npivots', 'acc', 'dci_time', 'svm_time'])
pivot_range = [10,25,50,100,250,500,1000,1500,2000,2500,5000]
for source, target, fold, taskname in MDS_task_generator(abspath(mds_home), nfolds=nfolds):
s_pivots, t_pivots = pivot_selection(max(pivot_range), source.X, source.y, source.U, target.U,
source.V, target.V,
phi=1, cross=True)
for npivots in pivot_range:
for dcf in ['cosine','linear']:
dci = DCI(dcf=dcf, unify=False, post='normal')
acc, dci_time, svm_time, _ = DCIclassify(source, target, s_pivots[:npivots], t_pivots[:npivots], dci, optimize=True)
rperf.add(dataset='MDS', task=taskname, method=str(dci), fold=fold, npivots=npivots, acc=acc, dci_time=dci_time, svm_time=svm_time)
rperf.pivot(index=['dataset', 'task','npivots'], values=['acc', 'dci_time', 'svm_time'])
rperf.dump(outfile)
for source, target, oracle, taskname in WebisCLS10_task_generator(abspath(dataset_home)):
s_pivots, t_pivots = pivot_selection(max(pivot_range), source.X, source.y, source.U, target.U,
source.V, target.V,
oracle=oracle, phi=30, cross=False)
for npivots in pivot_range:
dataset_home = '../datasets/Webis-CLS-10'
rperf = Result([
'dataset', 'task', 'method', 'acc', 'pivot_t', 'dci_t', 'svm_t', 'test_t'
])
for source, target, oracle, taskname in WebisCLS10_crossdomain_crosslingual_task_generator(
os.path.abspath(dataset_home)):
# pivot selection
tinit = time()
s_pivots, t_pivots = pivot_selection(npivots,
source.X,
source.y,
source.U,
target.U,
source.V,
target.V,
oracle=oracle,
phi=30,
show=min(10, npivots),
cross=True)
pivot_time = time() - tinit
print('pivot selection took {:.3f} seconds'.format(pivot_time))
dci = DCI(dcf=dcf, unify=True, post='normal')
acc, dci_time, svm_time, test_time = DCIclassify(source,
target,
s_pivots,
t_pivots,
dci,
optimize=optimize)