def _knockout_pass(f_id, classifier, train_data, folds, to_censor):
macro_scores = []
for fold_num, fold in enumerate(folds, start=1):
train_set = train_data - fold
test_set = fold
assert len(train_set) + len(test_set) == len(train_data)
train_vecs = [d for d in _censor_sparse_vectors_gen(
(v for _, v in train_set), to_censor)]
train_lbls = [l for l, _ in train_set]
classifier._liblinear_train(train_lbls, train_vecs)
test_vecs = [d for d in _censor_sparse_vectors_gen(
(v for _, v in test_set), to_censor)]
test_lbls = (l for l, _ in test_set)
res_tup =score_classifier_by_tup(classifier, (test_lbls, test_vecs))
macro_scores.append(res_tup[0])
mean = mean(macro_scores)
return f_id, mean
def _lexical_descent(classifiers, datasets, outdir, verbose=False,
worker_pool=None, no_simstring_cache=False, use_test_set=False):
# Check that we can in fact do a lexical descent for the classifier
for classifier_name in classifiers:
assert ('SIMSTRING' in classifier_name
or 'TSURUOKA' in classifier_name
or 'GAZETTER' in classifier_name)
for classifier_name, classifier_class in classifiers.iteritems():
print 'Classifier:', classifier_name
classifier = classifier_class()
for dataset_name, dataset_getter in datasets.iteritems():
print 'Dataset:', dataset_name
if verbose:
print >> stderr, 'Reading data...',
train_set, dev_set, test_set = dataset_getter()
if use_test_set:
train, test = list(chain(train_set, dev_set)), list(test_set)
else:
train, test = list(train_set), list(dev_set)
del train_set, dev_set, test_set
if verbose:
print >> stderr, 'Done!'
if not no_simstring_cache:
simstring_caching((classifier_name, ),
(train, test, ), verbose=verbose)
train_lbls, train_vecs = classifier._gen_lbls_vecs(train)
test_lbls, test_vecs = classifier._gen_lbls_vecs(test)
train_vecs = [hashabledict(d) for d in train_vecs]
test_vecs = [hashabledict(d) for d in test_vecs]
train_uncensored_vecs = deepcopy(train_vecs)
# Generate the folds for all iterations
folds = [f for f in _k_folds(5,
set(izip(train_lbls, train_vecs)))] #XXX: Constant
# XXX: This is an ugly hack and bound to break:
# Locate which vector ID;s that are used by SimString features and
# by which feature
from classifier.simstring.features import SIMSTRING_FEATURES
sf_ids = [f().get_id() for f in SIMSTRING_FEATURES]
vec_idxs_by_feat_id = defaultdict(set)
for sf_id in sf_ids:
for f_id in classifier.vec_index_by_feature_id:
# NOTE: Not 100% safe check, could match by accident
if sf_id in f_id:
vec_idxs_by_feat_id[sf_id].add(
classifier.vec_index_by_feature_id[f_id])
# Which ones never fired?
i = 0
for i, sf_id in enumerate((id for id in sf_ids
if id not in vec_idxs_by_feat_id), start=1):
print sf_id, 'never fired'
else:
print '{} SimString feature(s) never fired'.format(i)
res_dic = defaultdict(lambda : defaultdict(lambda : '-'))
# Iteratively find the best candidate
to_evaluate = set((f_id for f_id in vec_idxs_by_feat_id))
removed = set()
iteration = 1
last_macro_score = None
while to_evaluate:
print 'Iteration:', iteration
print 'Censoring vectors...',
# Censor everything we have removed so far
idxs_to_censor = set(i for i in chain(
*(vec_idxs_by_feat_id[f_id] for f_id in removed)))
train_vecs = [d for d in _censor_sparse_vectors_gen(
train_vecs, idxs_to_censor)]
train_data = set(izip(train_lbls, train_vecs))
train_folds = []
for fold in folds:
f_lbls = (l for l, _ in fold)
f_vecs = (d for d in _censor_sparse_vectors_gen(
(v for _, v in fold), idxs_to_censor))
train_folds.append(set(izip(f_lbls, f_vecs)))
print 'Done!'
print 'Training and evaluating a model of our current state...',
classifier._liblinear_train(train_lbls, train_vecs)
print 'Done!'
test_censored_vecs = [d for d in _censor_sparse_vectors_gen(
test_vecs, idxs_to_censor)]
curr_macro_score = score_classifier_by_tup(classifier,
(test_lbls, test_censored_vecs))[0]
print 'Current state on test is: {}'.format(curr_macro_score)
if last_macro_score is not None:
print 'Last state was: {} (diff: {})'.format(last_macro_score,
curr_macro_score - last_macro_score)
last_macro_score = curr_macro_score
# Prepare to go parallel
f_args = ((f_id, classifier, train_data, train_folds,
to_censor) for f_id, to_censor
in vec_idxs_by_feat_id.iteritems() if f_id in to_evaluate)
# Also cram in our non-censored one in there
f_args = chain(((None, classifier, train_data, train_folds,
set()), ), f_args)
score_by_knockout = {}
print 'Evaluating knockouts ({} in total)'.format(
len(to_evaluate) + 1)
# TODO: A bit reduntant, prettify!
if worker_pool is not None:
i = 1
for f_id, mean in worker_pool.imap_unordered(
__knockout_pass, f_args):
score_by_knockout[f_id] = mean
print 'it: {} k: {} res: {} {}'.format(
iteration, i, f_id, mean)
i += 1
else:
for i, args in enumerate(f_args, start=1):
f_id, mean = _knockout_pass(*args)
score_by_knockout[f_id] = mean
print 'it: {} k: {} res: {} {}'.format(
iteration, i, f_id, mean)
# Set the result dictionary
for f_id, mean in score_by_knockout.iteritems():
res_dic[str(iteration)][f_id] = mean
# And write the results incrementally for each round
with open(join_path(outdir, 'descent_{}_{}.md'.format(
classifier_name, dataset_name)), 'w') as md_file:
from md import dict_to_table
md_file.write(dict_to_table(res_dic, total=False, perc=False))
md_file.write('\n')
# Find the best scoring one...
scores = [(s, f_id)
for f_id, s in score_by_knockout.iteritems()]
scores.sort()
scores.reverse()
best_score, best_f_id = scores[0]
print 'Round winner: {} with {}'.format(best_f_id, best_score)
if best_f_id is None:
# We are done, no removal gave a better score
break
removed.add(best_f_id)
to_evaluate.remove(best_f_id)
iteration += 1
if removed:
# TODO: Could do more metrics here?
print 'Training and evaluating a model of our previous state...',
classifier._liblinear_train(train_lbls, train_uncensored_vecs)
before_macro_score = score_classifier_by_tup(classifier,
(test_lbls, test_vecs))[0]
print 'Done!'
print 'Training and evaluating a model of our current state...',
train_censored_vecs = [d for d in _censor_sparse_vectors_gen(train_vecs,
set(i for i in chain(*(vec_idxs_by_feat_id[f_id] for f_id in removed))))]
classifier._liblinear_train(train_lbls, train_censored_vecs)
print 'Done!'
test_censored_vecs = [d for d in _censor_sparse_vectors_gen(test_vecs,
set(i for i in chain(*(vec_idxs_by_feat_id[f_id] for f_id in removed))))]
after_macro_score = score_classifier_by_tup(classifier,
(test_lbls, test_censored_vecs))[0]
res_str = 'Before: {} After: {}'.format(before_macro_score,
after_macro_score)
print res_str
print 'Happy?'
else:
res_str = 'Unable to remove any lexical resource to make improvements...'
print res_str
# Ugly but saves the final result safely
with open(join_path(outdir, 'descent_{}_{}.txt'.format(
classifier_name, dataset_name)), 'w') as res_file:
res_file.write(res_str)
