def init_imgnetwork(pc_trans_feat):
step = tf.Variable(0)
with tf.variable_scope("img_var"):
img_placeholder = tf.placeholder(
tf.float32, shape=[FEAT_BARCH_SIZE * BATCH_DATA_SIZE, 256, 256, 3])
img_feat, _ = resnet.endpoints(img_placeholder, None, is_training=True)
img_feat = tf.reshape(
img_feat, [FEAT_BARCH_SIZE, BATCH_DATA_SIZE, img_feat.shape[1]])
q_img_vec, pos_img_vec, neg_img_vec, oth_img_vec = tf.split(
img_feat, [1, POS_NUM, NEG_NUM, OTH_NUM], 1)
img_loss = lazy_quadruplet_loss(q_img_vec, pos_img_vec, neg_img_vec,
oth_img_vec, MARGIN1, MARGIN2)
tf.summary.scalar('img_loss', img_loss)
epoch_num_placeholder = tf.placeholder(tf.float32, shape=())
learning_rate = get_learning_rate(epoch_num_placeholder)
tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate)
img_train_op = optimizer.minimize(img_loss, global_step=step)
merged = tf.summary.merge_all()
ops = {
"img_placeholder": img_placeholder,
"epoch_num_placeholder": epoch_num_placeholder,
"img_loss": img_loss,
"img_train_op": img_train_op,
"merged": merged,
"step": step
}
return ops
def init_imgnetwork(pc_trans_feat):
with tf.variable_scope("img_var"):
img_placeholder = tf.placeholder(tf.float32,
shape=[BATCH_SIZE, 240, 320, 3])
img_feat, img_pc_feat = resnet.endpoints(img_placeholder,
pc_trans_feat,
is_training=False)
return img_placeholder, img_feat, img_pc_feat
def init_imgnetwork(pc_trans_feat):
with tf.variable_scope("img_var"):
img_placeholder = tf.placeholder(
tf.float32, shape=[FEAT_BARCH_SIZE * BATCH_DATA_SIZE, 256, 256, 3])
img_feat, img_pc_feat = resnet.endpoints(img_placeholder,
pc_trans_feat,
is_training=True)
return img_placeholder, img_feat, img_pc_feat
