def __init__(self, detection_threshold=5):
self.classifier = baxter.BaxterClassifier()
self.top_results = 3
# IOU threshold to combine two overlapping bounding box
self.overlap_threshold = 0.75
# logit threshold to be considered as detection
self.threshold = detection_threshold
# Coefficient applied when two boxes are merged
self.overlap_bonus = 1.1
[self.image_mean, self.image_std] = inputProcessor.getNormalizationData(
self.classifier.training_file)
self.classifier.saver = tf.train.Saver()
self.classifier.saver.restore(
self.classifier.sess, self.classifier.weights_file)
def main(argvs):
baxterClassifier = baxter.BaxterClassifier()
[meanImage, std] = inputProcessor.getNormalizationData(
baxterClassifier.training_file)
# Start Tensorflow Session
with baxterClassifier.sess as sess:
baxterClassifier.saver = tf.train.Saver()
print("weight file to restore ... : ", baxterClassifier.weights_file)
baxterClassifier.saver.restore(
baxterClassifier.sess, baxterClassifier.weights_file)
cv2.waitKey(1000)
print("starting session... ")
batch = inputProcessor.get_custom_dataset_batch(
50, "data/spoon_test_data.csv", meanImage, std)
image_batch = batch[0]
label_batch = batch[1]
batch_size = len(label_batch)
prediction = tf.argmax(baxterClassifier.logits, 1)
trueLabel = np.argmax(label_batch, 1)
result = sess.run(prediction, feed_dict={
baxterClassifier.x: image_batch,
baxterClassifier.batch_size: batch_size,
baxterClassifier.dropout_rate: 1})
print("=============")
print(result)
print(trueLabel)
print("=============\n\n")
train_accuracy = baxterClassifier.accuracy.eval(feed_dict={baxterClassifier.x: image_batch,
baxterClassifier.y: label_batch,
baxterClassifier.batch_size: batch_size,
baxterClassifier.dropout_rate: 1})
print("\nTest Accuracy %.2f \n\n" % (train_accuracy))
def main():
baxterClassifier = BaxterClassifier()
[meanImage, std] = inputProcessor.getNormalizationData(
baxterClassifier.training_file)
# Start Tensorflow Session
with baxterClassifier.sess as sess:
baxterClassifier.saver = tf.train.Saver()
cv2.waitKey(1000)
print("starting session... ")
# INITIALIZE VARIABLES
# merged = tf.summary.merge(["accuracy", "cross_entropy"])
# train_writer = tf.summary.FileWriter("model/summary/")
sess.run(tf.initialize_all_variables())
# START TRAINING
batch_index = 0
i = 0
while batch_index < 30000:
print("starting " + str(i) + "th with batch index : " +
str(batch_index) + " training iteration..")
i += 1
###################################################
# GET BATCH (FOR CIFAR DATA SET)
# batch_size = 50
# batch = inputProcessor.get_next_cifar(batch_size, batch_index)
# image_batch = batch[0]
# label_batch = batch[1]
# batch_index = batch_index + batch[2]
# batch_size = len(label_batch)
###################################################
# GET BATCH (FOR IMAGENET DATASET)
# batch = inputProcessor.get_imagenet_batch(
# "data/train_data.csv", 10)
# image_batch = batch[0]
# label_batch = batch[1]
# batch_index = batch_index + 100
# batch_size = len(label_batch)
###################################################
# GET BATCH FOR CUSTOM DATASET AND (FOR CALTECH DATASET)
batch = inputProcessor.get_custom_dataset_batch(
32, baxterClassifier.training_file, meanImage, std)
image_batch = batch[0]
label_batch = batch[1]
batch_index = batch_index + 64
batch_size = len(label_batch)
###################################################
# PERIODIC PRINT-OUT FOR CHECKING
if i % 20 == 0:
prediction = tf.argmax(baxterClassifier.logits, 1)
trueLabel = np.argmax(label_batch, 1)
result = sess.run(prediction, feed_dict={
baxterClassifier.x: image_batch,
baxterClassifier.batch_size: batch_size,
baxterClassifier.dropout_rate: 1})
print("=============")
print(result)
print(trueLabel)
print("=============\n\n")
# summary, train_accuracy = sess.run(
# [merged, baxterClassifier.accuracy], feed_dict={baxterClassifier.x: image_batch,
# baxterClassifier.y: label_batch,
# baxterClassifier.batch_size: batch_size,
# baxterClassifier.dropout_rate: 1})
# train_writer.add_summary(summary, i)
train_accuracy = baxterClassifier.accuracy.eval(feed_dict={baxterClassifier.x: image_batch,
baxterClassifier.y: label_batch,
baxterClassifier.batch_size: batch_size,
baxterClassifier.dropout_rate: 1})
print("\nStep %d, Training Accuracy %.2f \n\n" % (i,
train_accuracy))
# ACTUAL TRAINING PROCESS
baxterClassifier.train_op.run(feed_dict={baxterClassifier.x: image_batch,
baxterClassifier.y: label_batch,
baxterClassifier.batch_size: batch_size,
baxterClassifier.dropout_rate: 0.5})
# DONE.. SAVE MODEL
save_path = baxterClassifier.saver.save(
sess, baxterClassifier.weights_file)
print("saving model to ", save_path)
def callback(self, ros_data):
'''Callback function of subscribed topic.
Here images get converted and features detected'''
if VERBOSE :
print 'received image of type: "%s"' % ros_data.format
#### direct conversion to IMAGE NP ARRAY ####
np_arr = np.fromstring(ros_data.data, np.uint8)
image_np = cv2.imdecode(np_arr, cv2.CV_LOAD_IMAGE_COLOR)
top_results = 2 # number of crops to show for detection
predicting_class = 1
predictions = []
# TODO : tensorflow image detection starts here
[meanImage, std] = inputProcessor.getNormalizationData("src/beginner_tutorials/scripts/data/custom_train_data.csv")
baxterClassifier = baxter.BaxterClassifier()
# Start Tensorflow Session
with baxterClassifier.sess as sess:
baxterClassifier.saver = tf.train.Saver()
print("weight file to restore ... : ", baxterClassifier.weights_file)
baxterClassifier.saver.restore(
baxterClassifier.sess, "src/beginner_tutorials/scripts/"+baxterClassifier.weights_file)
cv2.waitKey(1000)
print("starting session... ")
# GET IMAGE FROM USER INPUT
batch = inputProcessor.regionProposal(image_np)
if batch is None:
print("something went wrong when getting images crops ... ")
return
original_img = batch[0]
image_batch = batch[1]
boundingBoxInfo = batch[2]
batch_size = len(image_batch)
# CREATE INPUT IMAGE BATCH
input_image = np.zeros(
[len(image_batch), baxterClassifier.img_size, baxterClassifier.img_size, 3])
for x in range(batch_size):
input_image[x] = (image_batch[x] - meanImage) / std
# RUN CASCADING DETECTOR
print("batch size : ", batch_size)
print("input tensor size : ", input_image.shape)
prediction = sess.run(baxterClassifier.logits, feed_dict={
baxterClassifier.x: input_image,
baxterClassifier.batch_size: batch_size,
baxterClassifier.dropout_rate: 1})
# filter correctly detected crops
for y in range(batch_size):
prob = prediction[y][predicting_class]
boundingBox = boundingBoxInfo[y]
predictions.append([prob, boundingBox])
# sort crops by logit values
predictions.sort(reverse=True)
for i in range(top_results):
boundingBoxData = predictions[i]
print(boundingBoxData)
x = boundingBoxData[1][0]
y = boundingBoxData[1][1]
winW = boundingBoxData[1][2]
winH = boundingBoxData[1][3]
# if boundingBoxData[0] > threshold:
cv2.rectangle(original_img, (x, y),
(x + winW, y + winH), (0, 255, 0), 2)
cv2.imshow("Window", original_img)
cv2.waitKey(1000)
time.sleep(1)
def main(argvs):
[meanImage,
std] = inputProcessor.getNormalizationData("data/custom_train_data.csv")
baxterClassifier = baxter.BaxterClassifier(argvs)
top_results = 2 # number of crops to show for detection
# Start Tensorflow Session
with baxterClassifier.sess as sess:
baxterClassifier.saver = tf.train.Saver()
print("weight file to restore ... : ", baxterClassifier.weights_file)
baxterClassifier.saver.restore(baxterClassifier.sess,
baxterClassifier.weights_file)
cv2.waitKey(1000)
print("starting session... ")
while True:
# GET USER INPUT
predictions = []
img_filename = raw_input('image location: ')
predictingClass = int(raw_input('class value: '))
# GET IMAGE FROM USER INPUT
batch = inputProcessor.get_sliding_window_img_crops(img_filename)
if batch is None:
print("wrong user input regarding image or labels ")
original_img = batch[0]
image_batch = batch[1]
boundingBoxInfo = batch[2]
batch_size = len(image_batch)
# CREATE INPUT IMAGE BATCH
input_image = np.zeros([
len(image_batch), baxterClassifier.img_size,
baxterClassifier.img_size, 3
])
for x in range(batch_size):
input_image[x] = (image_batch[x] - meanImage) / std
# RUN CASCADING DETECTOR
print("batch size : ", batch_size)
print("input tensor size : ", input_image.shape)
prediction = sess.run(baxterClassifier.logits,
feed_dict={
baxterClassifier.x: input_image,
baxterClassifier.batch_size: batch_size,
baxterClassifier.dropout_rate: 1
})
# filter correctly detected crops
for y in range(batch_size):
prob = prediction[y][predictingClass]
boundingBox = boundingBoxInfo[y]
predictions.append([prob, boundingBox])
# sort crops by logit values
predictions.sort(reverse=True)
for i in range(top_results):
boundingBoxData = predictions[i]
print(boundingBoxData)
x = boundingBoxData[1][0]
y = boundingBoxData[1][1]
winW = boundingBoxData[1][2]
winH = boundingBoxData[1][3]
# if boundingBoxData[0] > threshold:
cv2.rectangle(original_img, (x, y), (x + winW, y + winH),
(0, 255, 0), 2)
cv2.imshow("Window", original_img)
cv2.waitKey(1000)
time.sleep(1)