def train_sample(feature_str, label, pos_train=0.5, neg_train=1000):
"""Perform training and testing using disjont samples from the full
set of label_values. This is equivalent to doing one round of cross
validation (see classipy.cross_validation) only it keeps the values around
for display.
Args:
"""
all_hashes = list(cass.get_image_hashes())
pos_hashes = [_[0] for _ in cass.buffered_get_row(cf_labels, label)]
neg_hashes = list(set(all_hashes) - set(pos_hashes))
if 0 < pos_train <= 1: pos_train = int(pos_train * len(pos_hashes))
if 0 < neg_train <= 1: neg_train = int(neg_train * len(neg_hashes))
# Choose a training sample and a testing sample
if len(pos_hashes) < pos_train:
raise ValueError('Not enough positive examples %s(%d)' % \
(label, len(pos_hashes)))
if len(neg_hashes) < neg_train:
raise ValueError('Not enough negative examples %s(%d)' % \
(label, len(neg_hashes)))
pos_sample = random.sample(pos_hashes, pos_train)
neg_sample = random.sample(neg_hashes, neg_train)
labels = [-1 for _ in neg_sample] + [1 for _ in pos_sample]
values = cass.get_feature_values(feature_str, neg_sample+pos_sample)
global label_values
label_values = zip(labels, values)
print 'Training classifier with sample %d' % len(label_values)
train_classifier(label_values)
def run_train_test(feature_str, label, split_opts={}):
"""
Args:
split_opts: passed to split_train_test
Returns:
(label, [(conf, gt, hash), ...]
where
label: string label, e.g. 'airplane'
conf: -inf to inf, prediction confidence
gt: -1 or 1 for negative or positive annotation
hash: image key for use with cass.get_image
"""
train, test = split_train_test(label, **split_opts)
# Train
labels = (L for L, k in train)
hashes = [k for L, k in train]
values = cass.get_feature_values(feature_str, hashes)
label_values = zip(labels, values)
print 'Training classifier with %d values' % len(label_values)
classifier = train_classifier(label_values)
# Test
hashes = [k for L, k in test]
values = list(cass.get_feature_values(feature_str, hashes))
print('Testing with %d(%d) values' % (len(hashes),len(values)))
conf_label = (classifier.predict(value) for value in values)
conf_gt_hash = [(pred*conf, L, k)
for ((pred, conf),), (L, k)
in zip(conf_label, test)]
conf_gt_hash = sorted(conf_gt_hash, key=lambda _: _[0])
return (label, conf_gt_hash)
