os.environ['CUDA_VISIBLE_DEVICES'] = ''.join(FLAGS.run_device)
config.gpu_options.allow_growth = FLAGS.allow_growth
config.gpu_options.per_process_gpu_memory_fraction = FLAGS.gpu_fraction
else:
print('Deploying Model on CPU')
# set up step
sess = tf.Session(config=config)
with tf.device(store_device):
global_step = tf.Variable(0, trainable=False, name='global_step', dtype=tf.int64)
# read data
reshape_size = [FLAGS.reshape_height, FLAGS.reshape_weight]
name_batch, image_batch, label_batch = get_next_batch(get_dataset(
dir=FLAGS.data_dir,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.epoch_num,
reshape_size=reshape_size,
normalize=False)
)
# inference
with arg_scope([get_variable], device=store_device):
with tf.device('/GPU:0'):
outputs_to_scales_to_logits, mean_loss = deeplab_v3_plus._build_deeplab(image_batch,
label_batch,
ignore_labels=[255],
FLAGS=FLAGS,
is_training=True)
score_map = outputs_to_scales_to_logits['semantic']['merged_logits']
print('Deploying Model on CPU')
# set up step
sess = tf.Session(config=config)
default_params.num_classes = FLAGS.num_classes
with tf.device(FLAGS.run_device):
global_step = tf.Variable(0,
trainable=False,
name='global_step',
dtype=tf.int64)
# read data
default_params.img_shape = [FLAGS.reshape_height, FLAGS.reshape_weight]
name_batch, image_batch, labels_batch, bboxes_batch = get_next_batch(
get_dataset(dir=FLAGS.data_dir,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.epoch_num,
reshape_size=default_params.img_shape))
# inference
with arg_scope([get_variable], device=store_device):
with tf.device('/CPU:0'):
with arg_scope(
ssd_arg_scope(weight_init=None,
weight_reg=weight_reg,
bias_init=tf.zeros_initializer,
bias_reg=bias_reg,
is_training=False)):
net, endpoints, prediction_gathers = ssd_vgg16(
image_batch, scope='ssd_vgg16_300')
# predictions of bboxes
print('Deploying Model on CPU')
run_device = '/CPU:0'
# set up step
sess = tf.Session(config=config)
with tf.device(run_device):
global_step = tf.Variable(0,
trainable=False,
name='global_step',
dtype=tf.int64)
# read data
reshape_size = [FLAGS.reshape_height, FLAGS.reshape_weight]
name_batch, image_batch, label_batch = get_next_batch(
get_dataset(dir=FLAGS.data_dir,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.epoch_num,
reshape_size=reshape_size))
# inference
with arg_scope([get_variable], device=store_device):
with tf.device('/GPU:0'):
net = get_net(FLAGS.net_name)
score_map, endpoints = net(image_batch,
num_classes=FLAGS.num_classes,
weight_init=None,
weight_reg=weight_reg,
bias_init=tf.zeros_initializer,
bias_reg=bias_reg)
# solve for mAP and loss
class_map = arg_max(score_map, axis=3, name='class_map')