def get_batch(dataset_dir, num_readers, batch_size, out_shape, net, anchors, num_preprocessing_threads, file_pattern = '*.tfrecord', is_training = True): dataset = sythtextprovider.get_datasets(dataset_dir,file_pattern = file_pattern) provider = slim.dataset_data_provider.DatasetDataProvider( dataset, num_readers=num_readers, common_queue_capacity=20 * batch_size, common_queue_min=10 * batch_size, shuffle=True) [image, shape, glabels, gbboxes] = provider.get(['image', 'shape', 'object/label', 'object/bbox']) image, glabels, gbboxes,num = \ ssd_vgg_preprocessing.preprocess_image(image, glabels,gbboxes, out_shape,is_training=is_training) gclasses, glocalisations, gscores = \ net.bboxes_encode( glabels, gbboxes, anchors, num) batch_shape = [1] + [len(anchors)] * 3 r = tf.train.batch( tf_utils.reshape_list([image, gclasses, glocalisations, gscores]), batch_size=batch_size, num_threads=num_preprocessing_threads, capacity=5 * batch_size) b_image, b_gclasses, b_glocalisations, b_gscores= \ tf_utils.reshape_list(r, batch_shape) return [b_image, b_gclasses, b_glocalisations, b_gscores]
def get_batch(dataset_dir,
num_readers,
batch_size,
out_shape,
net,
anchors,
FLAGS,
file_pattern='*.tfrecord',
is_training=True,
shuffe=False):
dataset = sythtextprovider.get_datasets(dataset_dir,
file_pattern=file_pattern)
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
num_readers=num_readers,
common_queue_capacity=512 * 16 + 20 * batch_size,
common_queue_min=512 * 16,
shuffle=shuffe)
[image, shape, glabels, gbboxes, height, width] = provider.get(
['image', 'shape', 'object/label', 'object/bbox', 'height', 'width'])
if is_training:
image, glabels, gbboxes,num = \
txt_preprocessing.preprocess_image(image, glabels, gbboxes, height, width,
out_shape,use_whiten=FLAGS.use_whiten,is_training=is_training)
glocalisations, gscores = \
net.bboxes_encode( gbboxes, anchors, num)
batch_shape = [1] + [len(anchors)] * 2
r = tf.train.shuffle_batch(tf_utils.reshape_list(
[image, glocalisations, gscores]),
batch_size=batch_size,
num_threads=FLAGS.num_preprocessing_threads,
capacity=100 * batch_size,
min_after_dequeue=50 * batch_size)
b_image, b_glocalisations, b_gscores= \
tf_utils.reshape_list(r, batch_shape)
return b_image, b_glocalisations, b_gscores
else:
image, glabels, gbboxes,bbox_img, num = \
txt_preprocessing.preprocess_image(image, glabels,gbboxes, height,width,
out_shape,use_whiten=FLAGS.use_whiten,is_training=is_training)
glocalisations, gscores = \
net.bboxes_encode( gbboxes, anchors, num)
batch_shape = [1] * 4 + [len(anchors)] * 2
r = tf.train.batch(tf_utils.reshape_list(
[image, glabels, gbboxes, bbox_img, glocalisations, gscores]),
batch_size=batch_size,
num_threads=FLAGS.num_preprocessing_threads,
capacity=50 * batch_size,
dynamic_pad=True)
image, glabels, gbboxes,g_bbox_img,glocalisations, gscores = \
tf_utils.reshape_list(r, batch_shape)
return image, glabels, gbboxes, g_bbox_img, glocalisations, gscores
def main(_):
if not FLAGS.dataset_dir:
raise ValueError('You must supply the dataset directory with --dataset_dir')
tf.logging.set_verbosity(tf.logging.DEBUG)
with tf.Graph().as_default():
# Config model_deploy. Keep TF Slim Models structure.
# Useful if want to need multiple GPUs and/or servers in the future.
deploy_config = model_deploy.DeploymentConfig(
num_clones=FLAGS.num_clones,
clone_on_cpu=FLAGS.clone_on_cpu,
replica_id=0,
num_replicas=1,
num_ps_tasks=0)
# Create global_step.
with tf.device(deploy_config.variables_device()):
global_step = slim.create_global_step()
# Select the dataset.
#dataset = dataset_factory.get_dataset(
# FLAGS.dataset_name, FLAGS.dataset_split_name, FLAGS.dataset_dir)
dataset = sythtextprovider.get_datasets(FLAGS.dataset_dir)
# Get the SSD network and its anchors.
#ssd_class = nets_factory.get_network(FLAGS.model_name)
#ssd_params = ssd_class.default_params._replace(num_classes=FLAGS.num_classes)
text_net = txtbox_300.TextboxNet()
text_shape = text_net.params.img_shape
print 'text_shape '+ str(text_shape)
text_anchors = text_net.anchors(text_shape)
print len(text_anchors)
# Select the preprocessing function.
'''
preprocessing_name = FLAGS.preprocessing_name or FLAGS.model_name
image_preprocessing_fn = preprocessing_factory.get_preprocessing(
preprocessing_name, is_training=True)
'''
#tf_utils.print_configuration(FLAGS.__flags, ssd_params,
# dataset.data_sources, FLAGS.train_dir)
# =================================================================== #
# Create a dataset provider and batches.
# =================================================================== #
with tf.device(deploy_config.inputs_device()):
with tf.name_scope(FLAGS.dataset_name + '_data_provider'):
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
num_readers=FLAGS.num_readers,
common_queue_capacity=20 * FLAGS.batch_size,
common_queue_min=10 * FLAGS.batch_size,
shuffle=True)
# Get for SSD network: image, labels, bboxes.
[image, shape, glabels, gbboxes] = provider.get(['image', 'shape',
'object/label',
'object/bbox'])
init_op = tf.global_variables_initializer()
# Pre-processing image, labels and bboxes.
image, glabels, gbboxes,num = \
ssd_vgg_preprocessing.preprocess_image(image, glabels,gbboxes,
text_shape,is_training=True,
data_format='NHWC')
# Encode groundtruth labels and bboxes.
print 'bboxes num' + str(gbboxes.get_shape())
print 'glabes' + str(tf.shape(glabels))
glocalisations, gscores = \
text_net.bboxes_encode( gbboxes, text_anchors,num)
batch_shape = [1] + [len(text_anchors)] * 2
# Training batches and queue.
r = tf.train.batch(
tf_utils.reshape_list([image, glocalisations, gscores]),
batch_size=FLAGS.batch_size,
num_threads=FLAGS.num_preprocessing_threads,
capacity=5 * FLAGS.batch_size)
print 'r shape' + str(r[0]) + str(r[1]) + str(r[10])
b_image, b_glocalisations, b_gscores= \
tf_utils.reshape_list(r, batch_shape)
# Intermediate queueing: unique batch computation pipeline for all
# GPUs running the training.
batch_queue = slim.prefetch_queue.prefetch_queue(
tf_utils.reshape_list([b_image, b_glocalisations,b_gscores]),
capacity=2 * deploy_config.num_clones)
# =================================================================== #
# Define the model running on every GPU.
# =================================================================== #
def clone_fn(batch_queue):
#Allows data parallelism by creating multiple
#clones of network_fn.
# Dequeue batch.
b_image, b_glocalisations, b_gscores = \
tf_utils.reshape_list(batch_queue.dequeue(), batch_shape)
# Construct SSD network.
arg_scope = text_net.arg_scope(weight_decay=FLAGS.weight_decay)
with slim.arg_scope(arg_scope):
localisations, logits, end_points = \
text_net.net(b_image, is_training=True)
# Add loss function.
text_net.losses(logits, localisations,
b_glocalisations, b_gscores,
match_threshold=FLAGS.match_threshold,
negative_ratio=FLAGS.negative_ratio,
alpha=FLAGS.loss_alpha,
label_smoothing=FLAGS.label_smoothing)
return end_points
# Gather initial summaries.
summaries = set(tf.get_collection(tf.GraphKeys.SUMMARIES))
# =================================================================== #
# Add summaries from first clone.
# =================================================================== #
clones = model_deploy.create_clones(deploy_config, clone_fn, [batch_queue])
first_clone_scope = deploy_config.clone_scope(0)
# Gather update_ops from the first clone. These contain, for example,
# the updates for the batch_norm variables created by network_fn.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, first_clone_scope)
# Add summaries for end_points.
end_points = clones[0].outputs
for end_point in end_points:
x = end_points[end_point]
summaries.add(tf.summary.histogram('activations/' + end_point, x))
summaries.add(tf.summary.scalar('sparsity/' + end_point,
tf.nn.zero_fraction(x)))
# Add summaries for losses and extra losses.
for loss in tf.get_collection(tf.GraphKeys.LOSSES, first_clone_scope):
summaries.add(tf.summary.scalar(loss.op.name, loss))
for loss in tf.get_collection('EXTRA_LOSSES', first_clone_scope):
summaries.add(tf.summary.scalar(loss.op.name, loss))
# Add summaries for variables.
for variable in slim.get_model_variables():
summaries.add(tf.summary.histogram(variable.op.name, variable))
# =================================================================== #
# Configure the moving averages.
# =================================================================== #
if FLAGS.moving_average_decay:
moving_average_variables = slim.get_model_variables()
variable_averages = tf.train.ExponentialMovingAverage(
FLAGS.moving_average_decay, global_step)
else:
moving_average_variables, variable_averages = None, None
# =================================================================== #
# Configure the optimization procedure.
# =================================================================== #
with tf.device(deploy_config.optimizer_device()):
learning_rate = tf_utils.configure_learning_rate(FLAGS,
dataset.num_samples,
global_step)
optimizer = tf_utils.configure_optimizer(FLAGS, learning_rate)
summaries.add(tf.summary.scalar('learning_rate', learning_rate))
if FLAGS.moving_average_decay:
# Update ops executed locally by trainer.
update_ops.append(variable_averages.apply(moving_average_variables))
# Variables to train.
variables_to_train = tf_utils.get_variables_to_train(FLAGS)
# and returns a train_tensor and summary_op
total_loss, clones_gradients = model_deploy.optimize_clones(
clones,
optimizer,
var_list=variables_to_train)
# Add total_loss to summary.
summaries.add(tf.summary.scalar('total_loss', total_loss))
# Create gradient updates.
grad_updates = optimizer.apply_gradients(clones_gradients,
global_step=global_step)
update_ops.append(grad_updates)
update_op = tf.group(*update_ops)
train_tensor = control_flow_ops.with_dependencies([update_op], total_loss,
name='train_op')
# Add the summaries from the first clone. These contain the summaries
summaries |= set(tf.get_collection(tf.GraphKeys.SUMMARIES,
first_clone_scope))
# Merge all summaries together.
summary_op = tf.summary.merge(list(summaries), name='summary_op')
# =================================================================== #
# Kicks off the training.
# =================================================================== #
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=FLAGS.gpu_memory_fraction)
config = tf.ConfigProto(log_device_placement=False,
gpu_options=gpu_options,
allow_soft_placement = True)
saver = tf.train.Saver(max_to_keep=5,
keep_checkpoint_every_n_hours=1.0,
write_version=2,
pad_step_number=False)
slim.learning.train(
train_tensor,
logdir=FLAGS.train_dir,
master='',
is_chief=True,
init_fn=tf_utils.get_init_fn(FLAGS),
summary_op=summary_op,
number_of_steps=FLAGS.max_number_of_steps,
log_every_n_steps=FLAGS.log_every_n_steps,
save_summaries_secs=FLAGS.save_summaries_secs,
saver=saver,
save_interval_secs=FLAGS.save_interval_secs,
session_config=config,
sync_optimizer=None)
def get_batch(dataset_dir,
num_readers,
batch_size,
out_shape,
net,
anchors,
FLAGS,
file_pattern='*.tfrecord',
is_training=True,
shuffe=False):
dataset = sythtextprovider.get_datasets(dataset_dir,
file_pattern=file_pattern)
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
num_readers=num_readers,
common_queue_capacity=20 * batch_size,
common_queue_min=10 * batch_size,
shuffle=shuffe)
[image, shape, glabels, gbboxes, corx, cory] = provider.get([
'image', 'shape', 'object/label', 'object/bbox', 'object/corx',
'object/cory'
])
corx = tf.expand_dims(corx, -1)
cory = tf.expand_dims(cory, -1)
cord = tf.concat([corx, cory], -1)
if is_training:
image, glabels, gbboxes, cord, num = \
txt_preprocessing.preprocess_image(image, glabels,gbboxes, cord,
out_shape,is_training=is_training)
glocalisations, glabels, glinks = \
net.bboxes_encode(cord, anchors ,num)
batch_shape = [1] + [len(anchors)] * 3
r = tf.train.batch(
tf_utils.reshape_list([image, glocalisations, glabels, glinks]),
batch_size=batch_size,
num_threads=FLAGS.num_preprocessing_threads,
capacity=5 * batch_size,
)
b_image, b_glocalisations, b_glabels, b_glinks= \
tf_utils.reshape_list(r, batch_shape)
return b_image, b_glocalisations, b_glabels, b_glinks
else:
image, labels, bboxes, cord, num = \
txt_preprocessing.preprocess_image(image, glabels,gbboxes, cord,
out_shape,is_training=is_training)
glocalisations, glabels, glinks = \
net.bboxes_encode(cord, anchors ,num)
batch_shape = [1] * 3 + [len(anchors)] * 3
r = tf.train.batch(
tf_utils.reshape_list(
[image, labels, cord, glocalisations, glabels, glinks]),
batch_size=batch_size,
num_threads=FLAGS.num_preprocessing_threads,
capacity=5 * batch_size,
)
b_image, b_labels, b_cord, b_glocalisations, b_glabels, b_glinks= \
tf_utils.reshape_list(r, batch_shape)
return b_image, b_labels, b_cord, b_glocalisations, b_glabels, b_glinks
def run():
if not FLAGS.dataset_dir:
raise ValueError(
'You must supply the dataset directory with --dataset_dir')
print('-----start test-------')
if not os.path.exists(save_dir):
os.makedirs(save_dir)
with tf.device('/GPU:0'):
dataset = sythtextprovider.get_datasets(FLAGS.dataset_dir)
print(dataset)
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
num_readers=FLAGS.num_readers,
common_queue_capacity=20 * FLAGS.batch_size,
common_queue_min=10 * FLAGS.batch_size,
shuffle=True)
print('provider:', provider)
[image, shape, glabels, gbboxes, x1, x2, x3, x4, y1, y2, y3,
y4] = provider.get([
'image', 'shape', 'object/label', 'object/bbox',
'object/oriented_bbox/x1', 'object/oriented_bbox/x2',
'object/oriented_bbox/x3', 'object/oriented_bbox/x4',
'object/oriented_bbox/y1', 'object/oriented_bbox/y2',
'object/oriented_bbox/y3', 'object/oriented_bbox/y4'
])
print('image:', image)
print('shape:', shape)
print('glabel:', glabels)
print('gboxes:', gbboxes)
gxs = tf.transpose(tf.stack([x1, x2, x3, x4])) #shape = (N,4)
gys = tf.transpose(tf.stack([y1, y2, y3, y4]))
image = tf.identity(image, 'input_image')
text_shape = (384, 384)
image, glabels, gbboxes, gxs, gys = ssd_vgg_preprocessing.preprocess_image(
image,
glabels,
gbboxes,
gxs,
gys,
text_shape,
is_training=True,
data_format='NHWC')
x1, x2, x3, x4 = tf.unstack(gxs, axis=1)
y1, y2, y3, y4 = tf.unstack(gys, axis=1)
text_net = txtbox_384.TextboxNet()
text_anchors = text_net.anchors(text_shape)
e_localisations, e_scores, e_labels = text_net.bboxes_encode(
glabels, gbboxes, text_anchors, gxs, gys)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.7)
config = tf.ConfigProto(log_device_placement=False,
gpu_options=gpu_options,
allow_soft_placement=True)
with tf.Session(config=config) as sess:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess, coord)
j = 0
all_time = 0
try:
while not coord.should_stop() and j < show_pic_sum:
start_time = time.time()
image_sess, label_sess, gbbox_sess, x1_sess, x2_sess, x3_sess, x4_sess, y1_sess, y2_sess, y3_sess, y4_sess, p_localisations, p_scores, p_labels = sess.run(
[
image, glabels, gbboxes, x1, x2, x3, x4, y1, y2, y3,
y4, e_localisations, e_scores, e_labels
])
end_time = time.time() - start_time
all_time += end_time
image_np = image_sess
# print(image_np)
# print('label_sess:',label_sess)
p_labels_concat = np.concatenate(p_labels)
p_scores_concat = np.concatenate(p_scores)
debug = False
if debug is True:
print(p_labels)
print('l_labels:',
len(p_labels_concat[p_labels_concat.nonzero()]),
p_labels_concat[p_labels_concat.nonzero()])
print('p_socres:',
len(p_scores_concat[p_scores_concat.nonzero()]),
p_scores_concat[p_scores_concat.nonzero()])
# print(img_np.shape)
print('label_sess:',
np.array(list(label_sess)).shape, list(label_sess))
img_np = np.array(image_np)
cv2.imwrite('{}/{}.png'.format(save_dir, j), img_np)
img_np = cv2.imread('{}/{}.png'.format(save_dir, j))
h, w, d = img_np.shape
label_sess = list(label_sess)
# for i , label in enumerate(label_sess):
i = 0
num_correct = 0
for label in label_sess:
# print(int(label) == 1)
if int(label) == 1:
num_correct += 1
img_np = draw_polygon(img_np, x1_sess[i] * w,
y1_sess[i] * h, x2_sess[i] * w,
y2_sess[i] * h, x3_sess[i] * w,
y3_sess[i] * h, x4_sess[i] * w,
y4_sess[i] * h)
if int(label) == 0:
img_np = draw_polygon(img_np,
x1_sess[i] * w,
y1_sess[i] * h,
x2_sess[i] * w,
y2_sess[i] * h,
x3_sess[i] * w,
y3_sess[i] * h,
x4_sess[i] * w,
y4_sess[i] * h,
color=(0, 0, 255))
i += 1
img_np = cv2.cvtColor(img_np, cv2.COLOR_BGR2RGB)
cv2.imwrite(
'{}'.format(os.path.join(save_dir,
str(j) + '.png')), img_np)
j += 1
print('correct:', num_correct)
except tf.errors.OutOfRangeError:
print('done')
finally:
print('done')
coord.request_stop()
print('all time:', all_time, 'average:', all_time / show_pic_sum)
coord.join(threads=threads)
'image/object/bbox/label': tf.VarLenFeature(dtype=tf.int64),
#'image/object/bbox/label_text' : tf.VarLenFeature(dtype=tf.string),
'image/format': tf.FixedLenFeature((), tf.string, default_value='jpeg'),
'image/encoded': tf.FixedLenFeature((), tf.string, default_value=''),
}, name='features')
# image was saved as uint8, so we have to decode as uint8.
image = tf.decode_raw(tfrecord_features['image/encoded'], tf.uint8)
shape = tf.cast(tfrecord_features['image/shape'], tf.int64)
#image = tf.reshape(image, shape)
height = tf.cast(tfrecord_features['image/height'],tf.int64)
width = tf.cast(tfrecord_features['image/width'],tf.int64)
"""
dataset_dir = '/Users/xiaodiu/Documents/github/projecttextbox/TextBoxes-TensorFlow/data/sythtext/'
dataset = get_datasets(dataset_dir)
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
num_readers=1,
common_queue_capacity=20 * 32,
common_queue_min=10 * 32,
shuffle=True)
# Get for SSD network: image, labels, bboxes.
[
image,
shape,
height,
width,
glabels,
gbboxes,
for min_s in min_scala:
for max_s in max_scala:
scales = [min_s + i * (max_s - min_s) / 6 for i in range(7)]
anchor_sizes = [(512 * scales[i], 512 * scales[i] + 50)
for i in range(7)]
with tf.Graph().as_default():
# build a net
params = txtbox512.TextboxNet.default_params
params = params._replace(anchor_sizes=anchor_sizes)
text_net = txtbox512.TextboxNet(params)
text_shape = text_net.params.img_shape
print 'text_shape ' + str(text_shape)
text_anchors = text_net.anchors(text_shape)
## dataset provider
dataset = sythtextprovider.get_datasets('../data/ICDAR2013/',
file_pattern='*.tfrecord')
data_provider = slim.dataset_data_provider.DatasetDataProvider(
dataset, common_queue_capacity=32, common_queue_min=2)
[image, shape, glabels, gbboxes] = \
data_provider.get(['image', 'shape',
'object/label',
'object/bbox'])
dst_image, glabels, gbboxes,num = \
txt_preprocessing.preprocess_image(image, glabels,gbboxes,
text_shape,is_training=True)
glocalisations, gscores = \
text_net.bboxes_encode( gbboxes, text_anchors, num)