def run():
test_images, test_labels = datasets.load_cifar10(is_train=False)
train_acc = []
train_cost = []
test_acc = []
test_cost = []
save_dir = "./models/network_01"
if not os.path.exists(save_dir):
os.mkdir(save_dir)
for epoch in xrange(100):
train_images, train_labels = datasets.load_cifar10(is_train=True)
num_epoch = 1
start_time = time.time()
network.fit(train_images, train_labels)
duration = time.time() - start_time
examples_per_sec = (num_epoch * len(train_images)) / duration
train_accuracy, train_loss = network.score(train_images, train_labels)
test_accuracy, test_loss = network.score(test_images, test_labels)
summary = {
"epoch": epoch,
"name": epoch,
"train_accuracy": train_accuracy,
"test_accuracy": test_accuracy,
"train_loss": train_loss,
"test_loss": test_loss,
"examples_per_sec": examples_per_sec,
}
print "[%(epoch)d][%(name)s]train-acc: %(train_accuracy).3f, train-loss: %(train_loss).3f," % summary,
print "test-acc: %(test_accuracy).3f, test-loss: %(test_loss).3f, %(examples_per_sec).1f examples/sec" % summary
train_acc.append(train_accuracy)
train_cost.append(train_loss)
test_acc.append(test_accuracy)
test_cost.append(test_loss)
network.save(save_dir + "/model_%d.ckpt" % epoch)
save("./json/network_01.json", train_acc, train_cost, test_acc, test_cost)
args = parser.parse_args()
embedding_dim = args.embedding_dim
ranking_loss = args.ranking_loss
snapshot = args.snapshot
net_data = np.load(args.initial_parameters).item()
image_placeholder = tf.placeholder(dtype=global_dtype,
shape=[batch_size, 227, 227, 3])
var_dict = nw.get_variable_dict(net_data)
SPP = args.spp
pooling = args.pooling
with tf.variable_scope("ranker") as scope:
feature_vec = nw.build_alexconvnet(image_placeholder,
var_dict,
embedding_dim,
SPP=SPP,
pooling=pooling)
score_func = nw.score(feature_vec)
# load pre-trained model
saver = tf.train.Saver(tf.global_variables())
sess = tf.Session(config=tf.ConfigProto())
sess.run(tf.global_variables_initializer())
saver.restore(sess, snapshot)
print "Snapshot: {}".format(snapshot)
start_time = time.time()
evaluate_FCDB()
print("--- %s seconds ---" % (time.time() - start_time))
parser.add_argument("--initial_parameters", help="Path to initial parameter file", type=str, default="alexnet.npy")
parser.add_argument("--ranking_loss", help="Type of ranking loss", type=str, choices=['ranknet', 'svm'], default='svm')
parser.add_argument("--snapshot", help="Name of the checkpoint files", type=str, default='./snapshots/model-spp-max')
parser.add_argument("--spp", help="Whether to use spatial pyramid pooling in the last layer or not", type=str2bool, default=True)
parser.add_argument("--pooling", help="Which pooling function to use", type=str, choices=['max', 'avg'], default='max')
args = parser.parse_args()
embedding_dim = args.embedding_dim
ranking_loss = args.ranking_loss
snapshot = args.snapshot
net_data = np.load(args.initial_parameters).item()
image_placeholder = tf.placeholder(dtype=global_dtype, shape=[batch_size,227,227,3])
var_dict = nw.get_variable_dict(net_data)
SPP = args.spp
pooling = args.pooling
with tf.variable_scope("ranker") as scope:
feature_vec = nw.build_alexconvnet(image_placeholder, var_dict, embedding_dim, SPP=SPP, pooling=pooling)
score_func = nw.score(feature_vec)
# load pre-trained model
saver = tf.train.Saver(tf.global_variables())
sess = tf.Session(config=tf.ConfigProto())
sess.run(tf.global_variables_initializer())
saver.restore(sess, snapshot)
print "Snapshot: {}".format(snapshot)
start_time = time.time()
evaluate_FCDB()
print("--- %s seconds ---" % (time.time() - start_time))