def get_data_examples(source):
#while True:
indices = range(len(source[0]))
if settings.shuffle:
random.shuffle(indices)
for i in indices:
image = source[0][i]
image_left = source[0][max(0, i - 1)]
image_right = source[0][min(i + 1, len(source[0]) - 1)]
pose_x = source[1][0][i]
pose_q = source[1][1][i]
pose_x_left = source[1][0][max(0, i - 1)]
pose_q_left = source[1][1][max(0, i - 1)]
pose_x_right = source[1][0][min(i + 1, len(source[0]) - 1)]
pose_q_right = source[1][1][min(i + 1, len(source[0]) - 1)]
# print type(pose_x)
# print pose_x
# print pose_q
# print pose_x_left
m1_2, a1_2 = getError(pose_x, pose_q, pose_x_left, pose_q_left)
m2_3, a2_3 = getError(pose_x, pose_q, pose_x_right, pose_q_right)
m1_3, a1_3 = getError(pose_x_left, pose_q_left, pose_x_right,
pose_q_right)
if m1_2 > settings.distanceThreshold or m2_3 > settings.distanceThreshold:
continue
if a1_2 > settings.angleThreshold or a2_3 > settings.angleThreshold:
continue
#global samplesCounter
# samplesCounter+=1
# print samplesCounter
yield np.asarray([image_left, image, image_right]), np.asarray(
[pose_x, pose_x_left,
pose_x_right]), np.asarray([pose_q, pose_q_left, pose_q_right])
def validate(model):
directory = settings.directory
datasource = utilities.get_data(settings.testsetpath)
datagen = utilities.limited_gen_data(datasource)
settings.saveMean = False
#model = cnn_lstm.create_cnn_lstm(weightsfile)
#sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
#model.compile(optimizer=sgd, loss='categorical_crossentropy')
print('\nin validate method!!!!!!')
print(type(model))
posxs = []
posqs = []
howmanyaccepted = 0
counter = 0
print('looping on test set!')
for ims, xs, qs in datagen:
print(len(ims))
howmanyaccepted += 1
print howmanyaccepted
inputs = np.zeros([1, 3, 3, 224, 224])
inputs[0, :] = ims
out = model.predict(inputs)
posx = out[0][0][1] #.mean(0)#xyz
posq = out[1][0][1] #.mean(0)#wpqr
actualx = xs[1] #.mean(0)
actualq = qs[1] #.mean(0)
errx, theta = getError(posx, posq, actualx, actualq)
posxs.append(errx)
posqs.append(theta)
print('error should report here!')
print 'errx ', errx, ' m and ', 'errq ', theta, ' degrees'
return np.median(posxs), np.median(posqs), howmanyaccepted
p6 = float(p6)
posexyzwpqr = np.asarray((p0, p1, p2, p3, p4, p5, p6))
img = utilities.ResizeCropImage(imread(directory + fname)).astype(
np.float32)
img = img.transpose((2, 0, 1))
img[0, :, :] -= meanImage[0, :, :].mean()
img[1, :, :] -= meanImage[1, :, :].mean()
img[2, :, :] -= meanImage[2, :, :].mean()
inputs[0, counter % settings.stepSize, :, :, :] = img
labels[counter % settings.
stepSize, :] = posexyzwpqr #np.asarray((p0,p1,p2,p3,p4,p5,p6))
counter += 1
if counter % 3 == 0:
m1_2, a1_2 = getError(labels[0][:3], labels[0][3:7], labels[1][:3],
labels[1][3:7])
m2_3, a2_3 = getError(labels[1][:3], labels[1][3:7], labels[2][:3],
labels[2][3:7])
m1_3, a1_3 = getError(labels[0][:3], labels[0][3:7], labels[2][:3],
labels[2][3:7])
if m1_2 > settings.distanceThreshold or m2_3 > settings.distanceThreshold:
print m1_2, m2_3
continue
if a1_2 > settings.angleThreshold or a2_3 > settings.angleThreshold:
print a1_2, a2_3
continue
howmanyaccepted += 1
out = model.predict(inputs)
# print out
# print out[0].shape #(1,3,3)
#labels = np.zeros([3, 7])
for ims,xs,qs in datagen:
howmanyaccepted+=1
print howmanyaccepted
inputs = np.zeros([1, 3, 3, 224, 224])
inputs[0,:]=ims
out = model.predict(inputs)
# print out
# print out[0].shape #(1,3,3)
#print out
posx = out[0][0][1]#.mean(0)#xyz
posq = out[1][0][1]#.mean(0)#wpqr
#print "actual:"
actualx = xs[1]#.mean(0)
actualq = qs[1]#.mean(0)
errx, theta = getError(posx,posq,actualx,actualq)
posxs.append(errx)
posqs.append(theta)
print 'errx ', errx, ' m and ', 'errq ', theta, ' degrees'
#inputs = np.zeros([1, 3, 3, 224, 224])
#labels = np.zeros([ 3, 7])
print 'median error', np.median(posxs), ' m and ', np.median(posqs), ' degrees'
print 'accepted test sequences: ',howmanyaccepted
# poses.append((p0,p1,p2,p3,p4,p5,p6))
# images.append(directory+fname)
# imgs.append(img)
# print len(imgs)
# print np.asarray(imgs).shape