# model load & save configs
flags.DEFINE_string('summaries_dir',
'volume/TF_Logs/TensorflowModelZoo/resnet50_deform/',
'where to store summary log')
flags.DEFINE_string(
'pretrained_ckpts',
'/home/chenyifeng/TF_Models/ptrain/ILSVRC/ResNet_50_V1_GN/model.ckpt',
'where to load pretrained model')
flags.DEFINE_string('last_ckpt', None, 'where to load last saved model')
FLAGS = flags.FLAGS
# config devices
store_device = parse_device_name(FLAGS.store_device)
run_device = parse_device_name(FLAGS.run_device)
config = tf.ConfigProto(log_device_placement=False, allow_soft_placement=True)
if not 'CPU' in FLAGS.run_device:
GPU_NUMS = len(FLAGS.run_device.split(','))
print(
'====================================Deploying Model on {} GPUs===================================='
.format(GPU_NUMS))
os.environ['CUDA_VISIBLE_DEVICES'] = ''.join(FLAGS.run_device)
config.gpu_options.allow_growth = FLAGS.allow_growth
else:
GPU_NUMS = 1
print(
'====================================Deploying Model on CPU===================================='
)
'/home/chenyifeng/TF_Models/atrain/SEGS/fcn/mgpu',
'where to load last saved model')
tf.app.flags.DEFINE_string('next_ckpt',
'/home/chenyifeng/TF_Models/atrain/SEGS/fcn/mgpu',
'where to store current model')
tf.app.flags.DEFINE_integer('save_per_step', 1000, 'save model per xxx steps')
FLAGS = tf.app.flags.FLAGS
if (FLAGS.reshape_height is None
or FLAGS.reshape_weight is None) and FLAGS.batch_size != 1:
assert 0, 'Can' 't Stack Images Of Different Shapes, Please Speicify Reshape Size!'
store_device = parse_device_name(FLAGS.store_device)
run_device = parse_device_name(FLAGS.run_device)
config = tf.ConfigProto(log_device_placement=False, allow_soft_placement=True)
if FLAGS.run_device in '01234567':
print('Deploying Model on {} GPU Card'.format(''.join(FLAGS.run_device)))
# 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')
weight_reg = regularizer(mode=FLAGS.weight_reg_func,
scale=FLAGS.weight_reg_scale)
bias_reg = regularizer(mode=FLAGS.bias_reg_func, scale=FLAGS.bias_reg_scale)
net = get_net(FLAGS.net_name)