def _data(self):
file_train = '/home/dcs41/Documents/tf-Faster-RCNN/Data/data_clutter/clutter_mnist_train.tfrecords'
self.x, self.gt_boxes, self.im_dims = Data.batch_inputs(self.read_and_decode, file_train,
batch_size=self.flags['batch_size'])
file_valid = '/home/dcs41/Documents/tf-Faster-RCNN/Data/data_clutter/clutter_mnist_valid.tfrecords'
self.x_valid, self.gt_boxes_valid, self.im_dims_valid = Data.batch_inputs(self.read_and_decode, file_valid,
mode="eval", batch_size=1, num_threads=1, num_readers=1)
self.num_train_images = 55000
self.num_valid_images = 5000
self.num_test_images = 10000
def test_print_image(self):
""" Read data through self.sess and plot out """
threads, coord = Data.init_threads(self.sess) # Begin Queues
print("Running 100 iterations of simple data transfer from queue to np.array")
for i in range(100):
x, gt_boxes = self.sess.run([self.x, self.gt_boxes])
print(i)
# Plot an example
faster_rcnn_tests.plot_img(x[0], gt_boxes[0])
Data.exit_threads(threads, coord) # Exit Queues
def test_print_image(self):
""" Read data through self.sess and plot out """
threads, coord = Data.init_threads(self.sess) # Begin Queues
print("Running 100 iterations of simple data transfer from queue to np.array")
for i in range(100):
x, gt_boxes = self.sess.run([self.x['TRAIN'], self.gt_boxes['TRAIN']])
print(i)
# Plot an example
faster_rcnn_tests.plot_img(x[0], gt_boxes[0])
Data.exit_threads(threads, coord) # Exit Queues
def eval(self):
""" Loop Through Evaluation Dataset and calculate metrics """
info = dict()
for c in range(flags['num_classes']):
info[c] = list()
for i in range(self.num_valid_images):
bboxes, cls_score, gt_boxes, image = self.sess.run([self.roi_proposal_net_valid.get_rois(),
self.fast_rcnn_net_valid.get_cls_score(), self.gt_boxes_valid,
self.x_valid])
print(gt_boxes)
self.vis_detections(np.squeeze(image[0]), str(gt_boxes[0][4]), gt_boxes, bboxes)
Data.exit_threads(self.threads, self.coord) # Exit Queues
def print_test_image(self):
""" Takes in a .tfrecord file and plots the image batch with bounding box """
file = '/home/dcs41/Documents/tf-Faster-RCNN/Data/data_clutter/clutter_mnist_valid.tfrecords'
im_dims, gt_boxes, image = Data.batch_inputs(self.read_and_decode,
file,
batch_size=32)
self.sess.run(tf.local_variables_initializer())
self.sess.run(tf.global_variables_initializer())
threads, coord = Data.init_threads(self.sess)
_, gt_boxes, image_out = self.sess.run([im_dims, gt_boxes, image])
self.plot_img(image_out[0], gt_boxes[0])
Data.exit_threads(threads, coord)
def eval(self):
""" Loop Through Evaluation Dataset and calculate metrics """
info = dict()
for c in range(flags['num_classes']):
info[c] = list()
for i in range(self.num_valid_images):
bboxes, cls_score, gt_boxes, image = self.sess.run([
self.roi_proposal_net_valid.get_rois(),
self.fast_rcnn_net_valid.get_cls_score(), self.gt_boxes_valid,
self.x_valid
])
print(gt_boxes)
self.vis_detections(np.squeeze(image[0]), str(gt_boxes[0][4]),
gt_boxes, bboxes)
Data.exit_threads(self.threads, self.coord) # Exit Queues
def train(self):
""" Run training function. Save model upon completion """
epochs = 0
iterations = int(np.ceil(self.num_images['TRAIN']/self.flags['batch_size']) * self.flags['num_epochs'])
self.print_log('Training for %d iterations' % iterations)
for i in tqdm(range(iterations)):
summary = self._run_train_iter()
if self.step % self.flags['display_step'] == 0:
self._record_train_metrics()
bbox, cls = self.sess.run([self.fast_rcnn_net['TRAIN'].get_bbox_refinement(),
self.fast_rcnn_net['TRAIN'].get_cls_score()])
if self.step % (self.num_images['TRAIN']) == 0: # save model every epoch
self._save_model(section=epochs)
epochs += 1
self._record_training_step(summary)
Data.exit_threads(self.threads, self.coord) # Exit Queues
def train(self):
""" Run training function. Save model upon completion """
epochs = 0
iterations = int(np.ceil(self.num_train_images/self.flags['batch_size']) * self.flags['num_epochs'])
self.print_log('Training for %d iterations' % iterations)
for i in range(3):
summary = self._run_train_iter()
if self.step % self.flags['display_step'] == 0:
self._record_train_metrics()
bbox, cls = self.sess.run([self.fast_rcnn_net.get_bbox_refinement(), self.fast_rcnn_net.get_cls_score()])
self.print_log('Number of predictions: %d' % bbox.shape[0])
if (self.step % 2) == 0:
self._save_model(section=epochs)
epochs += 1
self._record_training_step(summary)
print(self.step)
Data.exit_threads(self.threads, self.coord) # Exit Queues
def _data(self):
file_train = '/home/dcs41/Documents/tf-Faster-RCNN/Data/data_clutter/clutter_mnist_train.tfrecords'
self.x, self.gt_boxes, self.im_dims = Data.batch_inputs(
self.read_and_decode,
file_train,
batch_size=self.flags['batch_size'])
file_valid = '/home/dcs41/Documents/tf-Faster-RCNN/Data/data_clutter/clutter_mnist_valid.tfrecords'
self.x_valid, self.gt_boxes_valid, self.im_dims_valid = Data.batch_inputs(
self.read_and_decode,
file_valid,
mode="eval",
batch_size=1,
num_threads=1,
num_readers=1)
self.num_train_images = 55000
self.num_valid_images = 5000
self.num_test_images = 10000
def train(self):
""" Run training function. Save model upon completion """
iterations = int(
np.ceil(self.num_train_images / self.flags['batch_size']) *
self.flags['num_epochs'])
threads, coord = Data.init_threads(self.sess) # Begin Queues
self.print_log('Training for %d iterations' % iterations)
for i in range(iterations):
if self.step % self.flags['display_step'] != 0:
summary = self._run_train_iter()
else:
summary = self._run_train_metrics_iter()
self._record_train_metrics()
self._record_training_step(summary)
print(self.step)
self._save_model(section=1)
Data.exit_threads(threads, coord) # Exit Queues
def __init__(self, flags_input, dictionary):
self.epoch = flags_input['file_epoch'] if flags_input['restore'] else 0
self.lr = flags['learning_rate']
super().__init__(flags_input, flags_input['run_num'], vram=cfg.VRAM, restore=flags_input['restore_num'])
self.print_log(dictionary)
self.print_log(flags_input)
self.threads, self.coord = Data.init_threads(self.sess)
def __init__(self, flags_input, dictionary):
self.epoch = flags_input['file_epoch'] if flags_input['restore'] else 0
self.lr = flags['learning_rate']
super().__init__(flags_input, flags_input['run_num'], vram=cfg.VRAM, restore=flags_input['restore_num'], restore_slim=flags_input['restore_slim_file'])
self.print_log(dictionary)
self.print_log(flags_input)
self.threads, self.coord = Data.init_threads(self.sess)
def _data(self):
# Initialize placeholder dicts
self.x = {}
self.gt_boxes = {}
self.im_dims = {}
# Train data
file_train = flags['data_directory'] + 'trans_mnist_train.tfrecords'
self.x['TRAIN'], self.gt_boxes['TRAIN'], self.im_dims['TRAIN'] = Data.batch_inputs(self.read_and_decode,
file_train, batch_size=self.flags['batch_size'])
# Validation data; ground truth boxes used for evaluation/visualization only
file_valid = flags['data_directory'] + 'trans_mnist_valid.tfrecords'
self.x['VALID'], self.gt_boxes['VALID'], self.im_dims['VALID'] = Data.batch_inputs(self.read_and_decode,
file_valid, mode="eval",
batch_size=self.flags['batch_size'],
num_threads=1, num_readers=1)
# Test data. No GT Boxes.
self.x['TEST'] = tf.placeholder(tf.float32, [None, 128, 128, 1])
self.im_dims['TEST'] = tf.placeholder(tf.int32, [None, 2])
self.num_images = {'TRAIN': 55000, 'VALID': 5000, 'TEST': 10000}
def __init__(self, flags_input, dictionary):
if flags_input['restore'] is True:
self.epochs = flags_input['file_epoch']
else: # not restore
self.epochs = 0
self.lr = flags['learning_rate']
super().__init__(flags_input,
flags_input['run_num'],
vram=0.2,
restore=flags_input['restore_num'])
self.print_log(dictionary)
self.print_log(flags_input)
self.threads, self.coord = Data.init_threads(self.sess)
def train(self):
""" Run training function. Save model upon completion """
epochs = 0
iterations = int(
np.ceil(self.num_train_images / self.flags['batch_size']) *
self.flags['num_epochs'])
self.print_log('Training for %d iterations' % iterations)
for i in range(3):
summary = self._run_train_iter()
if self.step % self.flags['display_step'] == 0:
self._record_train_metrics()
bbox, cls = self.sess.run([
self.fast_rcnn_net.get_bbox_refinement(),
self.fast_rcnn_net.get_cls_score()
])
self.print_log('Number of predictions: %d' % bbox.shape[0])
if (self.step % 2) == 0:
self._save_model(section=epochs)
epochs += 1
self._record_training_step(summary)
print(self.step)
Data.exit_threads(self.threads, self.coord) # Exit Queues
def main():
file_train = flags['data_directory'] + 'trans_mnist_train.tfrecords'
x, _, _ = Data.batch_inputs(read_and_decode,
file_train,
batch_size=flags['batch_size'])
x = tf.stack([x, x, x], 3)
with slim.arg_scope(resnet_arg_scope()):
_ = resnet50(x)
variables_to_restore = slim.get_model_variables()
saver = tf_saver.Saver(variables_to_restore)
with tf.Session() as sess:
saver.restore(
sess, "/home/kd/Documents/tf-Faster-RCNN/Models/resnet_v1_50.ckpt")
a = input('Now finished restoring...')
def _data(self):
# Initialize placeholder dicts
self.x = {}
self.gt_boxes = {}
self.im_dims = {}
# Train data
file_train = flags['data_directory'] + 'clutter_mnist_train.tfrecords'
self.x['TRAIN'], self.gt_boxes['TRAIN'], self.im_dims['TRAIN'] = Data.batch_inputs(self.read_and_decode,
file_train, batch_size=
self.flags['batch_size'])
# Validation data. No GT Boxes necessary.
file_valid = flags['data_directory'] + 'clutter_mnist_valid.tfrecords'
self.x['VALID'], _, self.im_dims['VALID'] = Data.batch_inputs(self.read_and_decode,
file_valid, mode="eval",
batch_size=
self.flags['batch_size'],
num_threads=1, num_readers=1)
# Test data. No GT Boxes.
self.x['TEST'] = tf.placeholder(tf.float32, [None, 128, 128, 1])
self.im_dims['TEST'] = tf.placeholder(tf.int32, [None, 2])
self.num_images = {'TRAIN': 55000, 'VALID': 5000, 'TEST': 10000}
def _data(self):
# Initialize placeholder dicts
self.x = {}
self.gt_boxes = {}
self.im_dims = {}
# Train data
file_train = flags['data_directory'] + 'clutter_mnist_train.tfrecords'
self.x['TRAIN'], self.gt_boxes['TRAIN'], self.im_dims[
'TRAIN'] = Data.batch_inputs(self.read_and_decode,
file_train,
batch_size=self.flags['batch_size'])
# Validation data. No GT Boxes.
self.x['EVAL'] = tf.placeholder(tf.float32, [None, 128, 128, 1])
self.im_dims['EVAL'] = tf.placeholder(tf.int32, [None, 2])
self.num_images = {'TRAIN': 55000, 'VALID': 5000, 'TEST': 10000}
def close(self):
Data.exit_threads(self.threads, self.coord)
def __init__(self, flags_input, run_num, restore):
super().__init__(flags_input, run_num, vram=0.3, restore=restore)
self.print_log("Seed: %d" % flags['seed'])
self.threads, self.coord = Data.init_threads(self.sess)
def __init__(self, flags_input):
super().__init__(flags_input, flags_input['run_num'], vram=0.2, restore=flags_input['restore_num'])
self.print_log("Seed: %d" % flags_input['seed'])
self.threads, self.coord = Data.init_threads(self.sess)
def close(self):
Data.exit_threads(self.threads, self.coord)
def __init__(self, flags_input, run_num, restore):
super().__init__(flags_input, run_num, vram=0.3, restore=restore)
self.print_log("Seed: %d" % flags['seed'])
self.threads, self.coord = Data.init_threads(self.sess)
def _data(self):
file = '/home/dcs41/Documents/tf-Faster-RCNN/Data/data_clutter/clutter_mnist_train.tfrecords'
self.x, self.gt_boxes, self.im_dims = Data.batch_inputs(
self.read_and_decode, file, batch_size=self.flags['batch_size'])
self.num_train_images = 55000
