def run_single_test(data_dir, output_dir):
from detection import train_detector, detect
from keras import backend as K
from keras.models import load_model
from os import environ
from os.path import abspath, dirname, join
train_dir = join(data_dir, 'train')
test_dir = join(data_dir, 'test')
train_gt = read_csv(join(train_dir, 'gt.csv'))
train_img_dir = join(train_dir, 'images')
print('training')
train_detector(train_gt, train_img_dir, fast_train=True)
code_dir = dirname(abspath(__file__))
model = load_model(join(code_dir, 'model_e300.hdf5'))
test_img_dir = join(test_dir, 'images')
print('detecting')
detected_points = detect(model, test_img_dir)
save_csv(detected_points, join(output_dir, 'output.csv'))
if environ.get('KERAS_BACKEND') == 'tensorflow':
K.clear_session()
def run_single_test(data_dir, output_dir):
from detection import train_detector, detect
from keras import backend as K
from keras.models import load_model
from os.path import abspath, dirname, join
train_dir = join(data_dir, 'train')
test_dir = join(data_dir, 'test')
train_gt = read_csv(join(train_dir, 'gt.csv'))
train_img_dir = join(train_dir, 'images')
model = train_detector(train_gt, train_img_dir, fast_train=True)
code_dir = dirname(abspath(__file__))
model = load_model(join(code_dir, 'facepoints_model.hdf5'))
test_img_dir = join(test_dir, 'images')
detected_points = detect(model, test_img_dir)
save_csv(detected_points, join(output_dir, 'output.csv'))
K.clear_session()
return img_shapes
def compute_metric(detected, gt, img_shapes):
res = 0.0
for filename, coords in detected.items():
n_rows, n_cols = img_shapes[filename]
diff = (coords - gt[filename])
diff[::2] /= n_cols
diff[1::2] /= n_rows
diff *= 100
res += (diff**2).mean()
return res / len(detected.keys())
train_gt = read_csv(join(train_dir, 'gt.csv'))
train_img_dir = join(train_dir, 'images')
model = train_detector(train_gt, train_img_dir, fast_train=False)
model.save('facepoints_model.hdf5')
#model = train_detector(train_gt, train_img_dir, fast_train=True)
model = load_model('facepoints_model.hdf5')
test_img_dir = join(test_dir, 'images')
detected_points = detect(model, test_img_dir)
test_gt = read_csv(join(test_dir, 'gt.csv'))
img_shapes = read_img_shapes(test_dir)
error = compute_metric(detected_points, test_gt, img_shapes)
print('Error: ', error)
pil_draw.ellipse(
(pt2[0] - radius, pt2[1] - radius, pt2[0] + radius, pt2[1] + radius), fill=gt_color)
pil_img.save(res_dir + '/' + impaths[i])
#if (len(argv) != 2) and (len(argv) != 4):
# stdout.write('Usage: %s train_dir test_dir [-v results_dir]\n' % argv[0])
# exit(1)
start_time = time.time()
train_dir = argv[1]
#test_dir = argv[2]
visualisation_needed = (len(argv) > 2) and (argv[2] == '-v')
if visualisation_needed:
res_dir = argv[3]
if visualisation_needed:
train_imgs, train_gt,test_paths = load_data(train_dir,True)
else:
train_imgs, train_gt = load_data(train_dir)
X_train,X_test,y_train,y_test = cross_validation.train_test_split(train_imgs,train_gt,
test_size = 0.1, train_size = 0.5,random_state = 2 )
model = train_detector(X_train, y_train)
del X_train,y_train
detection_results = np.array(detect(model, X_test))
print("Result: %.4f" % compute_metrics(X_test, detection_results, y_test))
if visualisation_needed:
visualise(X_test, detection_results, y_test, res_dir, test_paths)
end_time = time.time()
print("Running time:", round(end_time - start_time, 2),
's (' + str(round((end_time - start_time) / 60, 2)) + " minutes)")
(pt2[0] - radius, pt2[1] - radius, pt2[0] + radius, pt2[1] + radius), fill=gt_color)
pil_img.save(res_dir + '/' + impaths[i])
if (len(argv) != 3) and (len(argv) != 5):
stdout.write('Usage: %s train_dir test_dir [-v results_dir]\n' % argv[0])
exit(1)
start_time = time.time()
train_dir = argv[1]
test_dir = argv[2]
visualisation_needed = (len(argv) > 3) and (argv[3] == '-v')
if visualisation_needed:
res_dir = argv[4]
train_imgs = load_imgs(train_dir)
train_gt = load_gt(train_dir)
model = train_detector(train_imgs, train_gt)
del train_imgs, train_gt
if visualisation_needed:
test_imgs = list(load_imgs(test_dir))
test_gt = list(load_gt(test_dir))
test_paths = list(load_paths(test_dir))
else:
test_imgs = list(load_imgs(test_dir))
test_gt = list(load_gt(test_dir))
detection_results = np.array(detect(model, test_imgs, test_gt))
print("Result: %.4f" % compute_metrics(test_imgs, detection_results, test_gt))
if visualisation_needed:
visualise(test_imgs, detection_results, test_gt, res_dir, test_paths)
end_time = time.time()
print("Running time:", round(end_time - start_time, 2),
's (' + str(round((end_time - start_time) / 60, 2)) + " minutes)")
train_imgs, train_gt = load_data(train_dir)
print("Data has been loaded")
train_num = len(train_imgs)
test_num = TEST_SIZE
mask = range(test_num)
test_imgs = train_imgs[mask]
test_gt = train_gt[mask]
mask = range(test_num, 6 * test_num)
train_imgs = train_imgs[mask]
train_gt = train_gt[mask]
try:
barnet = train_detector(train_imgs, train_gt)
barnet.model.save_weights(model_dir, overwrite=True)
layer = barnet.model.layers[0]
weights = np.array(layer.get_weights()[0])
grid = visualize_grid(weights.transpose(0, 2, 3, 1))
grid = resize(grid, (512, 512, 3))
grid_img = Image.fromarray(grid.astype('uint8'))
grid_img.save(res_dir + '/layer_weights.jpg')
finally:
del train_imgs, train_gt
print("CNN has been trained")
detection_results = np.array(detect(barnet, test_imgs))