import math
import helper
import posenet
import numpy as np
from keras.optimizers import Adam
if __name__ == "__main__":
# Test model
model = posenet.create_posenet()
model.load_weights('trained_weights.h5')
adam = Adam(lr=0.0001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0,
clipvalue=2.0)
model.compile(optimizer=adam,
loss={
'cls1_fc_pose_xyz': posenet.euc_loss1x,
'cls1_fc_pose_wpqr': posenet.euc_loss1q,
'cls2_fc_pose_xyz': posenet.euc_loss2x,
'cls2_fc_pose_wpqr': posenet.euc_loss2q,
'cls3_fc_pose_xyz': posenet.euc_loss3x,
'cls3_fc_pose_wpqr': posenet.euc_loss3q
})
dataset_train, dataset_test = helper.getKings()
X_test = np.squeeze(np.array(dataset_test.images))
y_test = np.squeeze(np.array(dataset_test.poses))
import helper
import posenet
import numpy as np
from keras.optimizers import Adam
from keras.callbacks import ModelCheckpoint
if __name__ == "__main__":
# Variables
batch_size = 75
# Train model
model = posenet.create_posenet('posenet.npy',
True) # GoogLeNet (Trained on Places)
adam = Adam(lr=0.001, clipvalue=1.5)
model.compile(optimizer=adam,
loss={
'cls1_fc_pose_xyz': posenet.euc_loss1x,
'cls1_fc_pose_wpqr': posenet.euc_loss1q,
'cls2_fc_pose_xyz': posenet.euc_loss2x,
'cls2_fc_pose_wpqr': posenet.euc_loss2q,
'cls3_fc_pose_xyz': posenet.euc_loss3x,
'cls3_fc_pose_wpqr': posenet.euc_loss3q
})
dataset_train, dataset_test = helper.getKings()
X_train = np.squeeze(np.array(dataset_train.images))
y_train = np.squeeze(np.array(dataset_train.poses))
y_train_x = y_train[:, 0:3]
y_train_q = y_train[:, 3:7]
def test(TEST_ALL_IMAGES):
# root_path = '/ssd/data/fxpal_dataset/posenetDataset/posenet_brendan/posenet_data/'
# model_name = '2017_09_04_22_37_train_kin_no_drift'
# model_path = root_path + 'model/' + model_name + '_posenet.h5'
### NEW MODELS ###
root_path = '/ssd/posenet_training_output/'
model_name = '2018_11_09_00_00_train_train_data_training_largedataset_w_halloweendecor_rotated_BN_more_oct2018'
model_path = root_path + 'weights/checkpoint_weights/' + model_name + '_posenet.h5'
data_dict = readDataFromFile(
root_path +
'training_data_info/{}_posenet_data.h5py'.format(model_name))
# Test model
model = posenet.create_posenet()
if 0:
model_json = model.to_json()
with open(model_name + '_model_json_file.json', 'w') as f:
f.write(model_json)
graph = tf.get_default_graph()
# load model weight
model.load_weights(model_path)
# get image mean value
image_mean = data_dict['image_mean']
# print('image mean values before loading: {}'.format(image_mean) )
image_mean = image_mean.flatten()
image_mean = image_mean.tolist()
with open(model_name + '_image_mean.json', 'w') as outfile:
json.dump(image_mean, outfile)
image_mean = load_image_mean(model_name)
# read file to get image file name and ground truth pose
# test data path
if TEST_ALL_IMAGES:
test_data_directory = '/home/mitesh/Documents/keras_to_tensorflow/'
fread = open(test_data_directory + 'test_images_new/groundTruth.csv',
'r')
else:
test_image_name = '/home/mitesh/Documents/keras_to_tensorflow/test_images_new/2300_image_save.png'
## save results to file
if TEST_ALL_IMAGES:
fwrite = open('results' + model_name + '.csv', 'w')
for string in fread:
groundtruth = []
temp = re.split("[ ' \n ,]+", string)
# temp = string.split(',')
img = cv2.imread(test_data_directory + temp[0])
fwrite.write(temp[0] + ',')
for idx in range(1, len(temp) - 1):
groundtruth.append(float(temp[idx]))
fwrite.write(temp[idx] + ',')
########### predict
# Downsize the image and crop
image = subtractMean(img, image_mean)
images = np.empty((1, 224, 224, 3), dtype=float)
images[0] = image
# This 'with' statement is to address a bug with keras/tensorflow. It
# doesn't have any functional effect.
with graph.as_default():
robot_pose = model.predict(images, batch_size=1, verbose=0)
groundtruth = np.asarray(groundtruth)
predictedPose = getPoseFromTensor(robot_pose)
for pose in predictedPose:
value = "%.5f" % pose
fwrite.write(value + ',')
error_xyz, theta_error = error(predictedPose, groundtruth)
fwrite.write(str(error_xyz) + ',')
fwrite.write(str(theta_error) + '\n')
print(
'img: {}, predicted pose is: {}, ground truth is: {}, position error: {} meters ,orientation error: {} degree'
.format(temp[0], predictedPose, np.asarray(groundtruth),
error_xyz, theta_error))
fwrite.close()
else:
img = cv2.imread(test_image_name)
image = subtractMean(img, image_mean)
images = np.empty((1, 224, 224, 3), dtype=float)
images[0] = image
with graph.as_default():
robot_pose = model.predict(images, batch_size=1, verbose=0)
predictedPose = getPoseFromTensor(robot_pose)
print('img: {}, predicted pose is: {}'.format(test_image_name,
predictedPose))