def train_graph():
input_hyper = init_hyper()
input_pls = {
'feature':
tf.placeholder(dtype=tf.float32,
shape=[None, input_hyper['hidden_dim']])
}
vnet = VTranse()
vnet.create_graph(input_hyper['N_each_batch'], input_hyper['index_sp'],
input_hyper['index_cls'], input_hyper['N_cls'],
input_hyper['N_rela'])
graph_visual = Graph(input_hyper['layer_num'], input_hyper['hidden_dim'],
input_hyper['num_cls'], input_pls,
input_hyper['knowledge'])
graph_text = Graph(input_hyper['layer_num'], input_hyper['hidden_dim'],
input_hyper['num_cls'], input_pls,
input_hyper['knowledge'])
text_layer_out = graph_text.get_layer_out()
optimizer = tf.train.AdamOptimizer(learning_rate=input_hyper['lr_rate'])
train_var = tf.trainable_variables()
restore_var = [
var for var in train_var if 'vgg_16' in var.name or 'RD' in var.name
]
saver_res = tf.train.Saver(restore_var)
with tf.Session() as sess:
# init
init = tf.global_variables_initializer()
sess.run(init)
saver_res.restore(sess, input_hyper['model_path'])
roidb_read = read_roidb(input_hyper['roidb_path'])
train_roidb = roidb_read['train_roidb']
test_roidb = roidb_read['test_roidb']
N_train = len(train_roidb)
N_test = len(test_roidb)
for epoch in range(input_hyper['num_epoch']):
for roidb_id in range(N_train):
roidb_use = train_roidb[roidb_id]
if len(roidb_use['rela_gt']) == 0:
continue
rd_loss_temp, acc_temp, diff = vnet.train_predicate(
sess, roidb_use, None)
diff = np.array(diff)
print(diff.shape)
vf = []
print(np.array(input_hyper['rel_emb']).shape)
num_batch = diff.shape[0]
# for i in range(num_batch):
# num_nodes = diff[i][0]
#
# vf.append(visual_feature)
feed_dict = {}
feed_dict.update({input_pls['feature']: input_hyper['rel_emb']})
text_out = sess.run(text_layer_out, feed_dict=feed_dict)
print(text_out)
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
init = tf.global_variables_initializer()
sess.run(init)
saver_res.restore(sess, res_path)
t = 0.0
rd_loss = 0.0
acc = 0.0
for r in range(N_round):
for roidb_id in range(N_train):
roidb_use = train_roidb[roidb_id]
if len(roidb_use['rela_gt']) == 0:
continue
rd_loss_temp, acc_temp = vnet.train_predicate(
sess, roidb_use, RD_train)
rd_loss = rd_loss + rd_loss_temp
acc = acc + acc_temp
t = t + 1.0
if t % N_show == 0:
print("t: {0}, rd_loss: {1}, acc: {2}".format(
t, rd_loss / N_show, acc / N_show))
rd_loss = 0.0
acc = 0.0
if t % N_save == 0:
save_path = cfg.DIR + 'vtranse/pred_para/vg_vgg/vg_vgg' + format(
int(t / N_save), '04') + '.ckpt'
print("saving model to {0}".format(save_path))
saver.save(sess, save_path)
rd_loss_val = 0.0
acc_val = 0.0