def __init__(self):
signdata = SignData()
self.imageProcessClass = ImageProces()
self.X_train, self.y_train = signdata.getTrainFeatures()
self.X_test, self.y_test = signdata.getTestFeatures()
self.X_valid, self.y_valid = signdata.getValidFeatures()
self.X_train_aug, self.y_train_aug = signdata.getTrainAugmentsFeatures(
)
def readtfrecord():
tfRecordCls = tfRecordHandlerClass()
imgProcCls = ImageProces()
FILE = "./tfRecords/trafficSign_aug.tfRecords"
FILE = "./tfRecords/trafficSign_train.tfRecords"
filename_queue = tf.train.string_input_producer([ FILE ], num_epochs=4)
# get images labels from tf records
#images, labels = tfRecordCls.read_and_decode(filename_queue,BATCH_SIZE=100,NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN=124465)
images, labels = tfRecordCls.read_and_decode(filename_queue,BATCH_SIZE=100,NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN=34799)
#
#
#
init = tf.global_variables_initializer()
init2 = tf.local_variables_initializer()
with tf.Session() as sess:
sess.run( [init,init2] )
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
try:
#while True:
local_steps = 0
while not coord.should_stop():
img,label = sess.run([images,labels])
local_steps += 1
#
# if you want to apply gray scale from tfRecords...
#
img = list( map( lambda im : imgProcCls.getGrayScale(im) , img[:] ) )
img = np.array(img)
#print(img.shape)
except tf.errors.OutOfRangeError:
# This will be raised when you reach the end of an epoch (i.e. the
# iterator has no more elements).
print("REACHED. OutOfRange from EPOCH (string_input_producer)")
print(" local steps --> ",local_steps)
pass
# Perform any end-of-epoch computation here.
print('Done training, epoch reached')
coord.request_stop()
coord.join(threads)
def loadDownloadImage():
imageProcCls = ImageProces()
download_images = {}
downloadDir = "./DownloadsSign"
files = os.listdir(downloadDir)
for name in files:
filename = name.split(".")
if filename[-1] != "jpg" and filename[-1] != "png":
continue
print(name)
image = cv2.imread( os.path.join(downloadDir,name) )
image = cv2.resize(image,(32,32))
image = cv2.cvtColor(image,cv2.COLOR_BGR2RGB)
image = imageProcCls.getGrayScale(image) / 255.
download_images[filename[0]] = image
return download_images
def main(argv):
print("Preparing training data.")
#mnist = input_data.read_data_sets("./data/", validation_size=0, one_hot=True)
print("-" * 30)
print(" initialize resource")
imgProcCls = ImageProces()
sign_image = SignImageClass()
sign_image.imagePreprocessNormalize()
vgg19 = Vgg19()
print("-" * 40)
print(" defined hyper parameters")
print("dropout keep prob:", DEFINE.dropout_keep_prob)
print("epochs :", DEFINE.epochs)
print("BATCH_SIZE :", DEFINE.train_batch_size)
NUM_CLASS = 43
BATCH_SIZE = DEFINE.train_batch_size
x = tf.placeholder(tf.float32, [None, 32, 32, 1])
resized = tf.image.resize_images(x, (227, 227))
#y_ = tf.placeholder(tf.float32, [None, NUM_CLASS])
y_ = tf.placeholder(tf.int64, [None])
y_one_hot = tf.one_hot(y_, depth=NUM_CLASS, dtype=tf.float32)
# placeholder drop out
dropout_keep_prob = tf.placeholder(tf.float32)
logits = vgg19.model(resized, NUM_CLASS , dropout_keep_prob)
prob_img = tf.nn.softmax(logits)
cost = vgg19.cross_entropy(y_one_hot,logits)
train_op, global_step = vgg19.train(cost)
accuracy = vgg19.accuracy(logits, y_one_hot)
init = tf.global_variables_initializer()
init2 = tf.local_variables_initializer()
with tf.Session() as sess:
sess.run( [init,init2] )
saver = tf.train.Saver()
length_train_data = sign_image.train_aug_data_length()
length_valid_data = sign_image.valid_data_length()
g_step=tf.train.global_step(sess, global_step)
for it in range(DEFINE.epochs):
sign_image.shuffle_train_aug()
for offset in range(0,length_train_data,BATCH_SIZE):
features_batch,labels_batch = sign_image.batch_train_aug(offset,batch_size=BATCH_SIZE)
#features_batch = list( map( lambda im : imgProcCls.getGrayScale(im) , features_batch[:] ) )
#features_batch = np.array(features_batch) / 255.
feeds = {x:features_batch, y_:labels_batch , dropout_keep_prob:DEFINE.dropout_keep_prob }
train_op.run(feed_dict=feeds)
print(g_step)
#print('global_step: %s' % g_step)
total_acc = []
for offset in range(0,length_valid_data,BATCH_SIZE):
features_batch,labels_batch = sign_image.batch_valid(offset,batch_size=BATCH_SIZE)
#features_batch = list( map( lambda im : imgProcCls.getGrayScale(im) , features_batch[:] ) )
#features_batch = np.array(features_batch) / 255.
feeds = {x:features_batch, y_:labels_batch, dropout_keep_prob:DEFINE.dropout_keep_prob}
acc_, cost_ = sess.run([accuracy,cost],feed_dict=feeds)
total_acc.append( acc_ * BATCH_SIZE )
accuracy_ = np.sum( total_acc ) / np.float(length_valid_data)
print("EPOCH:%d validation - total accuracy : %.4f" % (it, accuracy_) )
def traing_testing():
tfRecordCls = tfRecordHandlerClass()
imgProcCls = ImageProces()
FILE_train = "./tfRecords/trafficSign_aug.tfRecords"
FILE_valid = "./tfRecords/trafficSign_test.tfRecords"
filename_queue_train = tf.train.string_input_producer([ FILE_train ], num_epochs=4)
filename_queue_test = tf.train.string_input_producer([ FILE_test ], num_epochs=1)
# get images labels from tf records
images_train, labels_train = tfRecordCls.read_and_decode(filename_queue_train,BATCH_SIZE=100,NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN=124465)
images_test, labels_test = tfRecordCls.read_and_decode(filename_queue_test,BATCH_SIZE=100,NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN=34799)
#
# define LeNet Model
#
x = tf.placeholder(tf.float32, (None, 32, 32, 1))
y_ = tf.placeholder(tf.int64, [None])
y_one_hot = tf.one_hot(y_, depth=43, dtype=tf.float32)
logits = LeNet(x,43)
with tf.variable_scope("cost") as scope:
#cost = tf.reduce_sum(tf.pow(pred_y - y_, 2))/(2*n_samples)
softmax = tf.nn.softmax_cross_entropy_with_logits(labels=y_one_hot, logits=logits)
cost = tf.reduce_mean(softmax)
with tf.variable_scope("train") as scope:
#train_op = tf.train.AdamOptimizer(1e-4).minimize(cost)
optimizer = tf.train.AdamOptimizer(learning_rate = 0.001)
train_op = optimizer.minimize( cost )
with tf.variable_scope("acc") as scope:
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(y_one_hot, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
init = tf.global_variables_initializer()
init2 = tf.local_variables_initializer()
with tf.Session() as sess:
sess.run( [init,init2] )
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
try:
local_steps = 0
while not coord.should_stop():
img,label = sess.run([images_train,labels_train])
local_steps += 1
#
# if you want to apply gray scale from tfRecords...
#
img = list( map( lambda im : imgProcCls.getGrayScale(im) , img[:] ) )
img = np.array(img)
feeds = {x:img, y_:label}
train_op.run(feed_dict=feeds)
if local_steps % 1244 == 0:
total_acc = []
for offset in range(0,length_valid_data,BATCH_SIZE):
features_batch,labels_batch = sign_image.batch_valid(offset,batch_size=BATCH_SIZE)
feeds = {x:features_batch, y_:labels_batch}
acc_, cost_ = sess.run([accuracy,cost],feed_dict=feeds)
total_acc.append( acc_ * BATCH_SIZE )
accuracy_ = np.sum( total_acc ) / np.float(length_valid_data)
print("EPOCH:%d total accuracy : %.4f" % (it, accuracy_) )
except tf.errors.OutOfRangeError:
# This will be raised when you reach the end of an epoch (i.e. the
# iterator has no more elements).
print("REACHED. OutOfRange from EPOCH (string_input_producer)")
print(" local steps --> ",local_steps)
pass
# Perform any end-of-epoch computation here.
print('Done training, epoch reached')
coord.request_stop()
coord.join(threads)
def traing_testing():
imgProcCls = ImageProces()
sign_image = SignImageClass()
sign_image.imagePreprocessNormalize()
BATCH_SIZE = 64
#
# define LeNet Model
#
x = tf.placeholder(tf.float32, (None, 32, 32, 1))
y_ = tf.placeholder(tf.int64, [None])
y_one_hot = tf.one_hot(y_, depth=43, dtype=tf.float32)
logits = LeNet(x,43)
with tf.variable_scope("cost") as scope:
#cost = tf.reduce_sum(tf.pow(pred_y - y_, 2))/(2*n_samples)
softmax = tf.nn.softmax_cross_entropy_with_logits(labels=y_one_hot, logits=logits)
cost = tf.reduce_mean(softmax)
with tf.variable_scope("train") as scope:
global_step = tf.Variable(0, name='global_step',trainable=False)
#train_op = tf.train.AdamOptimizer(1e-4).minimize(cost)
optimizer = tf.train.AdamOptimizer(learning_rate = 0.001)
train_op = optimizer.minimize( cost , global_step=global_step )
with tf.variable_scope("acc") as scope:
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(y_one_hot, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
print("Start Valuation of Augmentation Image train data ....")
init = tf.global_variables_initializer()
init2 = tf.local_variables_initializer()
with tf.Session() as sess:
sess.run( [init,init2] )
saver = tf.train.Saver()
EPOCH = 32
length_train_data = sign_image.train_aug_data_length()
length_valid_data = sign_image.valid_data_length()
for it in range(EPOCH):
#print("EPOCH", it)
#
# Augmentation image data shuffling
#
sign_image.shuffle_train_aug()
for offset in range(0,length_train_data,BATCH_SIZE):
features_batch,labels_batch = sign_image.batch_train_aug(offset,batch_size=BATCH_SIZE)
#features_batch = list( map( lambda im : imgProcCls.getGrayScale(im) , features_batch[:] ) )
#features_batch = np.array(features_batch) / 255.
feeds = {x:features_batch, y_:labels_batch}
train_op.run(feed_dict=feeds)
g_step=tf.train.global_step(sess, global_step)
#print('global_step: %s' % g_step)
total_acc = []
for offset in range(0,length_valid_data,BATCH_SIZE):
features_batch,labels_batch = sign_image.batch_valid(offset,batch_size=BATCH_SIZE)
#features_batch = list( map( lambda im : imgProcCls.getGrayScale(im) , features_batch[:] ) )
#features_batch = np.array(features_batch) / 255.
feeds = {x:features_batch, y_:labels_batch}
acc_, cost_ = sess.run([accuracy,cost],feed_dict=feeds)
total_acc.append( acc_ * BATCH_SIZE )
accuracy_ = np.sum( total_acc ) / np.float(length_valid_data)
print("EPOCH:%d validation - total accuracy : %.4f" % (it, accuracy_) )
save_path = saver.save(sess,"./lenet_aug_model/lenet2_aug.ckpt", global_step=g_step)
print("-" * 30 )
print("-- Model saved in file: ", save_path )
class SignImageClass():
def __init__(self):
signdata = SignData()
self.imageProcessClass = ImageProces()
self.X_train, self.y_train = signdata.getTrainFeatures()
self.X_test, self.y_test = signdata.getTestFeatures()
self.X_valid, self.y_valid = signdata.getValidFeatures()
self.X_train_aug, self.y_train_aug = signdata.getTrainAugmentsFeatures(
)
def imagePreprocessNormalize(self):
print("<SignImageClass> Image Preprocess...Normalize ")
print(" train test valid and augmentation train ...")
self.X_train = self.preprocessImages(self.X_train)
self.X_test = self.preprocessImages(self.X_test)
self.X_valid = self.preprocessImages(self.X_valid)
self.X_train_aug = self.preprocessImages(self.X_train_aug)
def train_aug_data_length(self):
return self.X_train_aug.shape[0]
def train_data_length(self):
return self.X_train.shape[0]
def test_data_length(self):
return self.X_test.shape[0]
def valid_data_length(self):
return self.X_valid.shape[0]
def getValidPreprocess(self):
return self.X_valid, self.y_valid
def label_one_hot(self):
pass
def shuffle_train_aug(self):
num_images = self.X_train_aug.shape[0]
r = np.random.permutation(num_images)
self.X_train_aug = self.X_train_aug[r]
self.y_train_aug = self.y_train_aug[r]
def shuffle_train(self):
num_images = self.X_prep_train.shape[0]
r = np.random.permutation(num_images)
self.X_prep_train = self.X_prep_train[r]
self.y_train = self.y_train[r]
def batch_train_aug(self, offset=0, batch_size=64):
features_batch = self.X_train_aug[offset:offset + batch_size]
labels_batch = self.y_train_aug[offset:offset + batch_size]
return features_batch, labels_batch
def batch_train(self, offset=0, batch_size=64):
features_batch = self.X_train[offset:offset + batch_size]
labels_batch = self.y_train[offset:offset + batch_size]
return features_batch, labels_batch
def batch_valid(self, offset=0, batch_size=64):
features_batch = self.X_valid[offset:offset + batch_size]
labels_batch = self.y_valid[offset:offset + batch_size]
return features_batch, labels_batch
def batch_test(self, offset=0, batch_size=64):
features_batch = self.X_test[offset:offset + batch_size]
labels_batch = self.y_test[offset:offset + batch_size]
return features_batch, labels_batch
def preprocessImages(self, images):
gray_images = list(
map(lambda im: self.imageProcessClass.getGrayScale(im), images[:]))
gray_images = np.array(gray_images) / 255.0
return gray_images
def preprocess_image(self, X):
# convert to gray scale
X = self.rgb2gray(X)
#X = getGrayScale(X)
# normalize with [0 1]
X_norm = (X / 255.).astype(np.float32)
X_norm = X_norm.reshape(X_norm.shape + (1, ))
return X_norm
def convert_to_full_records_prep(self):
tf_filenames = [
"signtraffic_train.tfrecords", "signtraffic_test.tfrecords",
"signtraffic_valid.tfrecords"
]
image_list = [self.X_prep_train, self.X_prep_test, self.X_prep_valid]
label_list = [self.y_train, self.y_test, self.y_valid]
tfRecordCls = tfRecordHandlerClass()
for idx, (images, labels) in enumerate(zip(image_list, label_list)):
tfRecordCls.convert_to_records(images, labels, tf_filenames[idx])
def convert_to_full_records(self):
tf_filenames = [
"trafficSign_train.tfRecords", "trafficSign_test.tfrecords",
"trafficSign_valid.tfrecords"
]
image_list = [self.X_train, self.X_test, self.X_valid]
label_list = [self.y_train, self.y_test, self.y_valid]
tfRecordCls = tfRecordHandlerClass()
for idx, (images, labels) in enumerate(zip(image_list, label_list)):
tfRecordCls.convert_to_records(images, labels, tf_filenames[idx])