def plot_pca(X, y):
# apply PCA to data
pca = PCA(n_components=200)
pca.fit(X, y)
# plot variance explained by each component
plt.plot(np.cumsum(pca.explained_variance_ratio_))
plt.xlabel('Number of PCA components')
plt.ylabel('Total explained variance ratio')
plt.grid()
plt.savefig(get_file('output', 'graphs', 'pca_wing_explained.eps'),
bbox_inches='tight')
plt.clf()
# plot principal component vs angle
X_t = pca.transform(X)
plt.scatter(X_t[:, 0].flatten(), y.flatten())
plt.xlabel('PCA Component 1')
plt.ylabel('Wing angle (deg)')
plt.grid()
plt.savefig(get_file('output', 'graphs', 'pca_wing.eps'),
bbox_inches='tight')
plt.clf()
def main():
X = joblib.load(get_file('output', 'data', X_JOBLIB_NAME))
y = joblib.load(get_file('output', 'data', Y_JOBLIB_NAME))
for type in ['male', 'female']:
print('{} fly orientation detector...'.format(type.capitalize()))
train_once(X, y, type)
def plot_pca(X, y, type):
# apply PCA to data
pca = PCA(n_components=200)
pca.fit(X, y)
# plot variance explained by each component
plt.plot(np.cumsum(pca.explained_variance_ratio_))
plt.xlabel('Number of PCA components')
plt.ylabel('Total explained variance ratio')
plt.grid()
plt.savefig(get_file('output', 'graphs', 'pca_orient_{}_explained.eps'.format(type)), bbox_inches='tight')
plt.clf()
# show how data are separated by first two principal components
X_t = pca.transform(X)
for l, c, m in zip(range(2), ('blue', 'red'), ('o', 'x')):
plt.scatter(X_t[y == l, 0], X_t[y == l, 1], color=c, label=CATEGORIES[l], alpha=0.5, marker=m)
plt.legend(loc='upper right')
plt.xlabel('PCA 1')
plt.ylabel('PCA 2')
plt.grid()
plt.savefig(get_file('output', 'graphs', 'pca_orient_{}.eps'.format(type)), bbox_inches='tight')
plt.clf()
def train_once(X, y):
clf, X_train, X_test, y_train, y_test = train(X, y)
report_model_classification(clf, X_train, X_test, y_train, y_test)
tree.export_graphviz(clf,
out_file=get_file('output', 'diagrams',
'tree_is_fly.dot'),
feature_names=FEATURES,
class_names=CATEGORIES,
filled=True,
rounded=True,
special_characters=True)
joblib.dump(clf, get_file('output', 'models', CLF_JOBLIB_NAME))
def train_once(X, y):
clf, X_train, X_test, y_train, y_test = train(X, y, plot=True)
report_model_regression(clf,
X_train,
X_test,
y_train,
y_test,
units='radians')
joblib.dump(clf, get_file('output', 'models', CLF_JOBLIB_NAME))
def train_once(X, y, type):
clf, X_train, X_test, y_train, y_test = train(X, y, type, plot=True)
report_model_classification(clf, X_train, X_test, y_train, y_test)
joblib.dump(clf, get_file('output', 'models', clf_joblib_name(type)))
def __init__(self, type):
self.clf = joblib.load(get_file('output', 'models', clf_joblib_name(type)))
self.hog = make_hog()
def main():
X = joblib.load(get_file('output', 'data', X_JOBLIB_NAME))
y = joblib.load(get_file('output', 'data', Y_JOBLIB_NAME))
train_once(X, y)
def __init__(self):
self.clf = joblib.load(get_file('output', 'models', CLF_JOBLIB_NAME))
def main():
X, y = load_data()
joblib.dump(X, get_file('output', 'data', X_JOBLIB_NAME))
joblib.dump(y, get_file('output', 'data', Y_JOBLIB_NAME))
def load_data(tol_radians=0.1):
hog = make_hog()
X = []
y = []
img_count = 0
hand_labeled_count = 0
for anno in tqdm(get_annotations()):
img = cv2.imread(anno.image_path, 0)
mask = img_to_mask(img)
contours = find_contours(img, mask=mask, type='core')
for contour in contours:
type = contour_label(anno, contour)
if type == 'male':
break
else:
continue
if anno.count('mw') == 2 and anno.has('mh') and anno.has(
'ma') and anno.has('mp2'):
mh = anno.get('mh')[0]
ma = anno.get('ma')[0]
mp2 = anno.get('mp2')[0]
else:
continue
# make patch, which will compute the angle from the image
body_patch = crop_to_contour(img, contour)
# compute angle from labels
label_angle = np.arctan2(ma[1] - mh[1], mh[0] - ma[0])
# find out if the image is flipped or not
rotate_angle = body_patch.estimate_angle()
diff = abs(angle_diff(rotate_angle, label_angle))
if diff <= tol_radians:
pass
elif np.pi - tol_radians <= diff <= np.pi + tol_radians:
rotate_angle = rotate_angle + np.pi
else:
anno.warn(
'Could not properly determine whether image is flipped (diff={:0.1f} degrees)'
.format(degrees(diff)))
continue
# find center of fly
body_center = body_patch.estimate_center(absolute=True)
# create patch centered on fly
fly_patch = male_fly_patch(img, mask, body_center, rotate_angle)
if fly_patch is None:
continue
origin = bound_point((body_center[0], body_center[1]), img)
mp2_rel = [mp2[0] - origin[0], origin[1] - mp2[1]]
rot_mat = get_rotation_matrix(np.pi / 2 - rotate_angle)
mp2_rot = rot_mat.dot(mp2_rel)
wings = []
for mw in anno.get('mw'):
wing_rel = [mw[0] - origin[0], origin[1] - mw[1]]
wing_rot = rot_mat.dot(wing_rel) - mp2_rot
angle = np.arctan(abs(wing_rot[0]) / abs(wing_rot[1]))
wings.append({'x': wing_rot[0], 'angle': angle})
if len(wings) != 2:
anno.warn('Length of wings is not 2 for some reason, skipping...')
continue
if wings[0]['x'] > wings[1]['x']:
# wing 0 is right, wing 1 is left
right_wing_angle = wings[0]['angle']
left_wing_angle = wings[1]['angle']
else:
# wing 1 is right, wing 0 is left
right_wing_angle = wings[1]['angle']
left_wing_angle = wings[0]['angle']
# create a hog patches for both wings
data = []
data.append({
'hog_patch': make_hog_patch(fly_patch),
'angle': right_wing_angle
})
data.append({
'hog_patch': make_hog_patch(fly_patch.flip('horizontal')),
'angle': left_wing_angle
})
# add data to X and y
for datum in data:
X.append(patch_to_features(datum['hog_patch'], hog))
y.append(datum['angle'])
hand_labeled_count += 1
if DEBUG:
plt.imshow(datum['hog_patch'].img)
plt.show()
# increment img_count to indicate that this file was actually used
img_count += 1
# assemble features
X = np.array(X, dtype=float)
# assemble labels
y = np.array(y, dtype=float)
print('Used {} annotated images.'.format(img_count))
print('Used {} hand-labeled wings.'.format(hand_labeled_count))
report_labels_regression(np.degrees(y),
filename=get_file('output', 'graphs',
'labels_wing.eps'),
units='Wing Angle (degrees)')
return X, y