def test(self):
TXT = 'train/testImageList.txt'
template = '''################## Summary #####################
Test Number: %d
Time Consume: %.03f s
FPS: %.03f
LEVEL - %d
Mean Error:
Left Eye = %f
Right Eye = %f
Nose = %f
Left Mouth = %f
Right Mouth = %f
Failure:
Left Eye = %f
Right Eye = %f
Nose = %f
Left Mouth = %f
Right Mouth = %f
'''
t = time.clock()
# Restore variables from disk.
try:
saver = tf.train.Saver()
ckpt = os.path.join(self.logdir, 'model.ckpt')
saver.restore(self.sess, ckpt)
print("Model restored.")
except ValueError:
print("Model not in model.ckpt")
return
data = getDataFromTxt(TXT)
error = np.zeros((len(data), 5))
for i in range(len(data)):
imgPath, bbox, landmarkGt, _, _, _, _ = data[i]
print imgPath
img = cv2.imread(imgPath, cv2.IMREAD_GRAYSCALE)
assert (img is not None)
logger("process %s" % imgPath)
# Crop
f_bbox = bbox.subBBox(-0.05, 1.05, -0.05, 1.05)
f_face = img[int(f_bbox.top):int(f_bbox.bottom) + 1,
int(f_bbox.left):int(f_bbox.right) + 1]
# Resize
f_face = cv2.resize(f_face, (39, 39))
f_face = f_face.reshape((39, 39, 1))
f_face = f_face / 255.0
landmarkP = self.sess.run(self.yhat_1,
feed_dict={self.x: [f_face]})
landmarkP = landmarkP.reshape((5, 2))
# real landmark
landmarkP = bbox.reprojectLandmark(landmarkP)
landmarkGt = bbox.reprojectLandmark(landmarkGt)
error[i] = evaluateError(landmarkGt, landmarkP, bbox)
t = time.clock() - t
N = len(error)
fps = N / t
errorMean = error.mean(0)
# failure
failure = np.zeros(5)
threshold = 0.05
for i in range(5):
failure[i] = float(sum(error[:, i] > threshold)) / N
# log string
s = template % (N, t, fps, 0, errorMean[0], errorMean[1], errorMean[2], \
errorMean[3], errorMean[4], failure[0], failure[1], failure[2], \
failure[3], failure[4])
print s
name = 'fl_' + str(self.attribute) + '_crossstitch'
logfile = 'log/' + name + '.log'
with open(logfile, 'w') as fd:
fd.write(s)
fd.write('\n')
alphas = [
var for var in tf.all_variables()
if 'alpha' in var.name and 'Adam' not in var.name
]
a_ws = [
var for var in tf.all_variables()
if 'a_w' in var.name and 'Adam' not in var.name
]
b_ws = [
var for var in tf.all_variables()
if 'b_w' in var.name and 'Adam' not in var.name
]
for alpha in alphas:
fd.write(alpha.name + ": " +
str(self.sess.run(tf.reduce_mean(alpha))) + '\n')
for a_w, b_w in zip(a_ws, b_ws):
fd.write(a_w.name + ": " +
str(self.sess.run(tf.reduce_mean(a_w))) + '\n')
fd.write(b_w.name + ": " +
str(self.sess.run(tf.reduce_mean(b_w))) + '\n')
# plot error hist
plotError(error, name)
def test(self):
template = '''################## Summary #####################
Test Number: %d
Time Consume: %.03f s
FPS: %.03f
LEVEL - %d
Mean Error:
Left Eye = %f
Right Eye = %f
Nose = %f
Left Mouth = %f
Right Mouth = %f
Failure:
Left Eye = %f
Right Eye = %f
Nose = %f
Left Mouth = %f
Right Mouth = %f
'''
t = time.clock()
# Restore variables from disk.
try:
saver = tf.train.Saver()
ckpt = os.path.join(self.logdir, 'model.ckpt')
saver.restore(self.sess, ckpt)
print("Model restored.")
except ValueError:
print("Model not in model.ckpt")
return
TXT = 'train/testImageList.txt'
data = getDataFromTxt(TXT)
error = np.zeros((len(data), 5))
for i in range(len(data)):
imgPath, bbox, landmarkGt, _, _, _, _ = data[i]
img = cv2.imread(imgPath, cv2.IMREAD_GRAYSCALE)
assert(img is not None)
logger("process %s" % imgPath)
#landmarkP = P(img, bbox)
# Crop
f_bbox = bbox.subBBox(-0.05, 1.05, -0.05, 1.05)
f_face = img[int(f_bbox.top):int(f_bbox.bottom)+1,int(f_bbox.left):int(f_bbox.right)+1]
# Resize
f_face = cv2.resize(f_face, (39, 39))
f_face = f_face.reshape((39, 39, 1))
f_face = f_face / 255.0
landmarkP = self.sess.run(self.yhat, feed_dict={self.x: [f_face]})
landmarkP = landmarkP.reshape((5,2))
# real landmark
landmarkP = bbox.reprojectLandmark(landmarkP)
landmarkGt = bbox.reprojectLandmark(landmarkGt)
error[i] = evaluateError(landmarkGt, landmarkP, bbox)
t = time.clock() - t
N = len(error)
fps = N / t
errorMean = error.mean(0)
print(error)
print(errorMean)
# failure
failure = np.zeros(5)
threshold = 0.05
for i in range(5):
failure[i] = float(sum(error[:, i] > threshold)) / N
# log string
s = template % (N, t, fps, 0, errorMean[0], errorMean[1], errorMean[2], \
errorMean[3], errorMean[4], failure[0], failure[1], failure[2], \
failure[3], failure[4])
print s
logfile = 'log/regression.log'
with open(logfile, 'w') as fd:
fd.write(s)
# plot error hist
plotError(error, 'regression')