def main(_):
gpu_options = tf.GPUOptions(allow_growth=True)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
network = crossnet.CrossNet(sess)
network.load_data(data_list, data_dir)
misc.pprint(network.config.__flags)
network.train_test()
def main(_):
gpu_options = tf.GPUOptions(allow_growth=True)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
# gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.7)
# config = tf.ConfigProto(gpu_options=gpu_options)
# with tf.Session(config=config) as sess:
network = nasnet.CrossNet(sess)
network.load_data(data_list, data_dir)
misc.pprint(network.config.__flags)
network.train_test()
def load_data(self, image_list, image_dir=""):
"""Create a queue that outputs batches of images and labels
label 0~3: [sky, bldg, road, tree]
"""
self.data_names = []
with open(image_list, 'r') as fid:
for line in fid.readlines():
names = [
os.path.join(image_dir, name.strip())
for name in line.split(',')
]
keys = ['im_a', 'im_g', 'lb_g']
self.data_names.append(dict(zip(keys, names)))
shuffle(self.data_names)
self.num_samples = len(self.data_names)
misc.pprint('[*] load %d samples from "%s"' %
(self.num_samples, image_list))
def build_model(self, data, is_training=True):
raw_aerial, raw_ground, label_ground = data
self.image_aerial = misc.preprocess_image(raw_aerial,
self.szs.image_aerial)
self.image_ground = misc.preprocess_image(raw_ground,
self.szs.image_ground)
self.prob_ground = misc.preprocess_label(label_ground,
self.num_classes,
self.szs.after_transf)
self.im_aerial = misc.proprocess_image(self.image_aerial)
self.im_ground = misc.proprocess_image(self.image_ground)
self.feat_aerial = models.pixelnet(self.image_aerial,
self.num_classes,
is_training=is_training,
batch_norm=self.batch_norm)
misc.pprint(
self.feat_aerial.get_shape().as_list()) # print the feature size
if is_training:
feat_aerial_small = self.feat_aerial
else:
feat_aerial_small = tf.image.resize_bilinear(
self.feat_aerial, self.szs.before_transf)
weights = models.compute_transfweights(self.szs.before_transf,
self.szs.after_transf,
self.conditioned,
is_training=is_training,
batch_norm=self.batch_norm)
self.feat_aerial2ground = models.transfnet(feat_aerial_small, weights,
self.szs.after_transf)
if is_training:
self.merged = tf.summary.merge_all()
self.summarizer = tf.summary.FileWriter(self.log_dir,
self.sess.graph)
with tf.name_scope("Loss"):
self.loss_class = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(
None, self.prob_ground, self.feat_aerial2ground))
self.loss_reg = tf.add_n(
tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))
self.loss = self.loss_class + self.loss_reg
with tf.name_scope("Optimizer"):
with tf.control_dependencies(
tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
self.step = tf.Variable(0,
name='global_step',
trainable=False)
self.optim = tf.train.AdamOptimizer(
tf.train.exponential_decay(
0.001, self.step, 5000, .7, staircase=True
) # not sure if this is necessary for Adam optimizer
).minimize(self.loss, global_step=self.step)
self.saver = tf.train.Saver(max_to_keep=10,
write_version=tf.train.SaverDef.V2)
misc.pprint("[*] build model.")
self.transfweights, self.transfbiases = tf.get_collection(
'transformer_weights')
self.prob_aerial = tf.nn.softmax(self.feat_aerial)
self.prob_aerial2ground = tf.nn.softmax(self.feat_aerial2ground)
with tf.name_scope('Vis'):
self.visual = [ \
self.image_aerial, self.image_ground,
tf.cast(self.prob_aerial, tf.float32)/self.num_classes,
tf.cast(self.prob_ground, tf.float32)/self.num_classes,
tf.cast(self.prob_aerial2ground, tf.float32)/self.num_classes,
self.transfweights]
def restore(self):
ckpt = tf.train.get_checkpoint_state(self.ckpt_dir)
self.saver.restore(self.sess, ckpt.model_checkpoint_path)
misc.pprint("[*] restore checkpoint from '%s'." % self.ckpt_dir)
parser.add_argument('--weight_decay', type=float, default=1e-4)
parser.add_argument('--meta_lr', type=float, default=0.01)
parser.add_argument('--base_lr', type=float,
default=0.01) # Learning rate for the inner loop
parser.add_argument(
'--update_step', type=int,
default=100) # The number of updates for the inner loop
parser.add_argument(
'--eval_weights', type=str,
default=None) # The meta-trained weights for meta-eval phase
parser.add_argument('--meta_label', type=str,
default='exp1') # Additional label for meta-train
# Set and print the parameters
args = parser.parse_args()
pprint(vars(args))
# Set manual seed for PyTorch
if args.seed == 0:
print('Using random seed.')
torch.backends.cudnn.benchmark = True
else:
print('Using manual seed:', args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Start trainer for pre-train, meta-train or meta-eval
if args.phase == 'meta_train':
trainer = MetaTrainer(args)