def analyze_features(best_features='best_weak_learners_200.pkl', features_file='all_predict_features.pkl',
num_pos=4916, num_neg=7960, top=None):
"""
:param best_features: file containing a list of the best classifiers selected by adaboost
:param features_file: file containing a matrix of predicted features
:param num_pos: number of positive images in the training set
:param num_neg: number of negative images in the training set
:return:
"""
best_features = pickle_load(best_features)
predict_features = pickle_load(features_file)
sum_of_alphas = 0
n_estimators = len(best_features)
labels = np.zeros(num_pos+num_neg, dtype=np.int8)
labels[:num_pos] = 1
if top is None:
top = n_estimators
accuracy_rates = np.empty(top, dtype=np.float32)
false_positive_rates = np.empty(top, dtype=np.float32)
for i, estimator in enumerate(best_features[:top]):
min_feature_num, min_error, error_rate, threshold, parity , alpha = estimator
sum_of_alphas += alpha
pred_features = np.ravel(predict_features[min_feature_num].todense())
if i == 0:
combined_threshold = pred_features * alpha
else:
combined_threshold += pred_features * alpha
pred = combined_threshold > sum_of_alphas / 2
cmat = confusion_matrix(labels, pred)
# compute the false positive of the current final classifier
# (using all features selected by Adaboost up to i+1)
false_positive_rates[i] = cmat[1, 0] / (cmat[0, 0] + cmat[1, 0])
# compute the accuracy of the final classifier
# (using all features selected by Adaboost up to i+1)
accuracy_rates[i] = (cmat[0, 0] + cmat[1, 1]) / len(pred)
return accuracy_rates, false_positive_rates
def analyze_features(best_features='adaboost300.pkl', features_file='pred_features.pkl',
num_pos=4916, num_neg=7960):
from sklearn.metrics import confusion_matrix
best_features = pickle_load(best_features)
features = pickle_load(features_file)
sum_of_alphas = 0
n_estimators = len(best_features)
labels = np.zeros(num_pos+num_neg, dtype=np.int8)
labels[:num_pos] = 1
accuracy_rates = np.empty(n_estimators, dtype=np.float32)
false_positive_rates = np.empty(n_estimators, dtype=np.float32)
for i, estimator in enumerate(best_features):
min_feature_num, min_error, error_rate, threshold, parity, alpha = estimator
sum_of_alphas += alpha
if i == 0:
combined_threshold = features[min_feature_num] * alpha
else:
combined_threshold += features[min_feature_num] * alpha
pred = combined_threshold > sum_of_alphas / 2
cmat = confusion_matrix(labels, pred)
# compute the false positive of the current final classifier
# (using all features selected by Adaboost up to i+1)
false_positive_rates[i] = cmat[1, 0] / (cmat[0, 0] + cmat[1, 0])
# compute the accuracy of the final classifier
# (using all features selected by Adaboost up to i+1)
accuracy_rates[i] = (cmat[0, 0] + cmat[1, 1]) / len(pred)
return accuracy_rates, false_positive_rates
def feature_train(data_path, weak_learner_file='all_weak_learners.pkl',
features_file='predict_error.pkl', num_pos=4916, num_neg=7960, n_estimators=200):
start = time.time()
print 'Starting Adaboost training ...'
parent_dir = os.getcwd()
os.chdir(data_path)
weak_learners = pickle_load(weak_learner_file)
features = pickle_load(features_file)
os.chdir(parent_dir)
# setup the initial weights
data_weights = np.empty(num_pos+num_neg, dtype=np.float32)
data_weights[:num_pos] = 1 / 2 / num_pos
data_weights[num_pos:] = 1 / 2 / num_neg
labels = np.zeros(num_pos+num_neg, dtype=np.int8)
labels[:num_pos] = 1
best_weak_learners = []
for i in xrange(n_estimators):
# find the best weak learner by computing the minimum error
round_start = time.time()
error = np.dot(features, data_weights)
# select the best weak learner
min_feature_num = np.argmin(error)
min_error = error[min_feature_num]
pred_errs = features[min_feature_num]
error_rate, threshold, parity = weak_learners[min_feature_num]
# update the data weights
beta = min_error / (1 - min_error)
alpha = -np.log(beta)
Beta = np.copy(data_weights)
Beta.fill(beta)
data_weights = data_weights * np.power(Beta, np.logical_not(pred_errs))
data_weights = data_weights / np.sum(data_weights) # normalize the weights
hist, bin_edges = np.histogram(data_weights, bins=10, range=(0, 1), density=True)
print hist
# ensure the feature selected will NOT be selected again
# this ensures that the previously selected feature have the worst error rates
features[min_feature_num] = 2
print 'Iter #{:3d}'.format(i+1)
print '-' * 30
print 'Feature = {: 06d}'.format(min_feature_num)
print 'W. Error = {:0.6f}'.format(min_error)
print 'Error = {:0.6f}'.format(error_rate)
print 'Threshold = {:0.6f}'.format(threshold)
print 'Beta = {:0.6f}'.format(beta)
print 'Alpha = {:0.6f}'.format(alpha)
print 'This round took %5.2f secs.' % (time.time() - round_start)
print '-' * 30
print
# save this weak classifier
best_weak_learners.append(TrainedFeatures(min_feature_num, min_error, error_rate,
threshold, parity, alpha))
print 'Finished Adaboost training in %5.2f secs.' % (time.time() - start)
return best_weak_learners
if key == 27:
stop_inspection = True
break
if key == ord('n'):
break
cv2.imshow(win_name, img)
if stop_inspection:
break
cv2.destroyAllWindows()
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Feature Inspection')
parser.add_argument('-features', nargs='+', type=int, default=[66214],
help='inspect features by their id number')
parser.add_argument('-all', action='store_true',
help='inspect all features chosen by Adaboost')
args = parser.parse_args()
if not args.all and len(args.features):
inspect_features(args.features) # inspect individual or a set of features given the feature id
if args.all:
best_features = pickle_load('adaboost10.pkl')
# best_features = pickle_load('best_weak_learners_10.pkl')
inspect_best_features(best_features)