def run_single_test(data_dir, output_dir):
from fit_and_classify import fit_and_classify, extract_hog
from glob import glob
from numpy import zeros
from os.path import basename, join
from skimage.io import imread
train_dir = join(data_dir, 'train')
test_dir = join(data_dir, 'test')
def read_gt(gt_dir):
fgt = open(join(gt_dir, 'gt.csv'))
next(fgt)
lines = fgt.readlines()
filenames = []
labels = zeros(len(lines))
for i, line in enumerate(lines):
filename, label = line.rstrip('\n').split(',')
filenames.append(filename)
labels[i] = int(label)
return filenames, labels
def extract_features(path, filenames):
hog_length = len(extract_hog(imread(join(path, filenames[0]))))
data = zeros((len(filenames), hog_length))
for i in range(0, len(filenames)):
filename = join(path, filenames[i])
data[i, :] = extract_hog(imread(filename))
return data
train_filenames, train_labels = read_gt(train_dir)
test_filenames = []
for path in sorted(glob(join(test_dir, '*png'))):
test_filenames.append(basename(path))
print('started train features extraction:', time())
train_features = extract_features(train_dir, train_filenames)
print('started test features extraction:', time())
test_features = extract_features(test_dir, test_filenames)
print('Started svm training:', time())
y = fit_and_classify(train_features, train_labels, test_features)
with open(join(output_dir, 'output.csv'), 'w') as fout:
for i, filename in enumerate(test_filenames):
print('%s,%d' % (filename, y[i]), file=fout)
labels[i] = int(label)
return filenames, labels
def extract_features(path, filenames):
hog_length = len(
extract_hog(imread(join(path, filenames[0]), plugin='matplotlib')))
data = zeros((len(filenames), hog_length))
for i in range(0, len(filenames)):
filename = join(path, filenames[i])
data[i, :] = extract_hog(imread(filename, plugin='matplotlib'))
return data
if len(argv) != 3:
print('Usage: %s train_data_path test_data_path' % argv[0])
exit(0)
train_data_path = argv[1]
test_data_path = argv[2]
train_filenames, train_labels = read_gt(train_data_path)
test_filenames, test_labels = read_gt(test_data_path)
train_features = extract_features(train_data_path, train_filenames)
test_features = extract_features(test_data_path, test_filenames)
y = fit_and_classify(train_features, train_labels, test_features)
print('Accuracy: %.4f' % (sum(test_labels == y) / float(test_labels.shape[0])))
#!/usr/bin/python
from sys import argv, stdout, exit
from numpy import array, loadtxt
import pandas as pd
from fit_and_classify import fit_and_classify
if len(argv) != 3:
stdout.write('Usage: %s train_file test_file\n' % argv[0])
exit(1)
train_file = argv[1]
test_file = argv[2]
train = loadtxt(train_file, delimiter=',', skiprows=1)
train_features = train[:, :-1]
train_labels = train[:, -1]
test = loadtxt(test_file, delimiter=',', skiprows=1)
test_features = test[:, :-1]
test_labels = test[:, -1]
stdout.write('%.4f\n' % (sum(test_labels == fit_and_classify(train_features, train_labels, test_features)) / float(test_labels.shape[0])))
return (filenames, rois, labels)
def extract_features(path, filenames, rois):
hog_length = len(extract_hog(imread(path + '/' + filenames[0], plugin='matplotlib'), rois[0, :]))
print(rois[0, :])
data = zeros((len(filenames), hog_length))
for i in range(0, len(filenames)):
filename = path + '/' + filenames[i]
data[i, :] = extract_hog(imread(filename, plugin='matplotlib'), rois[i, :])
return data
if len(argv) != 3:
stdout.write('Usage: %s train_data_path test_data_path\n' % argv[0])
exit(1)
train_data_path = argv[1]
test_data_path = argv[2]
(train_filenames, train_rois, train_labels) = read_gt(train_data_path)
(test_filenames, test_rois, test_labels) = read_gt(test_data_path)
train_features = extract_features(train_data_path, train_filenames, train_rois)
test_features = extract_features(test_data_path, test_filenames, test_rois)
y = fit_and_classify(train_features, train_labels, test_features)
stdout.write('%.4f\n' % (sum(test_labels == y) / float(test_labels.shape[0])))
def run_single_test(data_dir, output_dir, seed, first):
from fit_and_classify import fit_and_classify, extract_hog, extract_data
from glob import glob
from numpy import zeros
from os.path import basename, join
from skimage.io import imread
train_dir = join(data_dir, 'train')
test_dir = join(data_dir, 'test')
def read_gt(gt_dir):
fgt = open(join(gt_dir, 'gt.csv'))
next(fgt)
lines = fgt.readlines()
filenames = []
labels = zeros(len(lines))
for i, line in enumerate(lines):
filename, label = line.rstrip('\n').split(',')[:2]
filenames.append(filename)
labels[i] = int(label)
return filenames, labels
HOG_FILENAME = 'train_hog_file_size80_block4_442fixed_10bins.csv'
def dump_features(path, filenames):
hog_length = len(extract_hog(imread(join(path, filenames[0]))))
data = zeros((len(filenames), hog_length))
with open(HOG_FILENAME, mode='w') as hog_file:
hog_writer = csv.writer(hog_file, delimiter=',', quotechar='"', quoting=csv.QUOTE_MINIMAL)
hog_writer.writerow(['filename', 'hog_vector'])
for i in range(0, len(filenames)):
filename = join(path, filenames[i])
data[i, :] = extract_hog(imread(filename))
hog_writer.writerow([filename, ','.join(np.asarray(np.round(data[i], 3), dtype=str))])
if i % 100 == 0:
print('{} done'.format(i))
def extract_features(path, filenames):
hog_length = len(extract_hog(imread(join(path, filenames[0]))))
data = zeros((len(filenames), hog_length))
hog_data = pd.read_csv(HOG_FILENAME)
hog_data = hog_data.set_index('filename')
for i in range(0, len(filenames)):
filename = join(path, filenames[i])
data[i, :] = np.asarray(hog_data.loc[filename].hog_vector.split(','), dtype=float)
if i % 5000 == 0:
print('{} done'.format(i))
train_data, test_data = extract_data(seed)
train_data['filenames'] = 'public_tests/00_test_img_input/train/' + train_data['filenames']
test_data['filenames'] = 'public_tests/00_test_img_input/train/' + test_data['filenames']
train_data = train_data.merge(hog_data, how='inner', left_on='filenames', right_on='filename')
test_data = test_data.merge(hog_data, how='inner', left_on='filenames', right_on='filename')
return train_data, test_data
train_filenames, train_labels = read_gt(train_dir)
test_filenames = []
for path in sorted(glob(join(test_dir, '*png'))):
test_filenames.append(basename(path))
# train_features = extract_features(train_dir, train_filenames, 'train_hog_file.csv')
# test_features = extract_features(test_dir, test_filenames, 'test_hog_file.csv')
if seed == first:
dump_features(train_dir, train_filenames)
train_data, test_data = extract_features(train_dir, train_filenames)
train_features = np.stack(train_data['hog_vector'].apply(lambda x: np.asarray(x.split(','), dtype=float)))
test_features = np.stack(test_data['hog_vector'].apply(lambda x: np.asarray(x.split(','), dtype=float)))
train_labels = np.array(train_data['class_id'])
test_labels = np.array(test_data['class_id'])
y = fit_and_classify(train_features, train_labels, test_features)
print("Length of test: {}".format(len(y)))
correct = 0
incorrect = 0
conf_matrix = np.zeros((43, 43), dtype=int)
miss = np.zeros((43), dtype=int)
total = np.zeros((43), dtype=int)
for i in range(len(y)):
conf_matrix[y[i], test_labels[i]] += 1
total[test_labels[i]]+=1
if y[i] == test_labels[i]:
correct += 1
else:
incorrect += 1
miss[test_labels[i]]+=1
# print("Name: {}, expected: {}, actual: {}".format(test_data.filenames[i], test_labels[i], y[i]))
print(HOG_FILENAME)
print("Correct: {}; Incorrect: {}; Accuracy: {}; Seed: {}".format(correct, incorrect, correct / len(y) * 100, seed))
print("Per class stats:")
for i in range(43):
conf_matrix[i][i] //= 100
print("Total: {}, missed: {}, miss percentage: {}".format(total[i], miss[i], miss[i] / total[i]))
df_cm = pd.DataFrame(conf_matrix, index=[i for i in range(43)], columns=[i for i in range(43)])
plt.figure(figsize=(15, 11))
sn.heatmap(df_cm, annot=True)
plt.show()
