inputs = tf.placeholder(tf.float32, (None, model_config['doc_time_step'], model_config['input_dim']), name='input')
query = tf.placeholder(tf.float32, (None, model_config['query_time_step'], model_config['input_dim']), name='question')
labels = tf.placeholder(tf.float32, (None, model_config['n_entities']),name='labels')
learning_rate = tf.placeholder(tf.float32, shape=[], name='learning_rate')
keep_prob = tf.placeholder(tf.float32, name='dropout_prob')
doc_var_list = [
['d_attw',[2*model_config['ctx_lstm_size'], model_config['attention_mlp_hidden']]],
['q_attw',[2*model_config['question_lstm_size'], model_config['attention_mlp_hidden']]],
['wms',[model_config['attention_mlp_hidden'],1]],
['w_rg',[2*model_config['ctx_lstm_size'], model_config['n_entities'] ]],
['w_ug',[2*model_config['question_lstm_size'], model_config['n_entities'] ]]
]
doc_var = mut.create_var_xavier('Varibles',doc_var_list)#
x = tf.unstack(inputs, model_config['doc_time_step'], 1)
q = tf.unstack(query, model_config['query_time_step'], 1)
with tf.variable_scope("query"):
with tf.variable_scope("fw"):
qlstm_fw_cell = tf.contrib.rnn.LSTMCell(model_config['question_lstm_size'], forget_bias=1.0)
qlstm_fw_cell = tf.contrib.rnn.DropoutWrapper(qlstm_fw_cell, input_keep_prob=keep_prob)
with tf.variable_scope("bw"):
qlstm_bw_cell = tf.contrib.rnn.LSTMCell(model_config['question_lstm_size'], forget_bias=1.0)
qlstm_bw_cell = tf.contrib.rnn.DropoutWrapper(qlstm_bw_cell, input_keep_prob=keep_prob)
doc_net, fw, bw = rnn.static_bidirectional_rnn(qlstm_fw_cell, qlstm_bw_cell, q ,dtype=tf.float32)
y_q = tf.concat([fw[-1], bw[-1]],1)#
filename = "../../model/yolo_lsgan/fcann_v1.ckpt"
logfile = '../../log/yolo_lsgan'
graph_model = '../../model/yolo_lsgan/fcann_v1.ckpt-0.meta'
checkpoint_dir = '../../model/yolo_lsgan'
continue_training = 1
loop_num = 5500
d_loop_num = 3
batch_size = 64
keep_prob = tf.placeholder(tf.float32)
x = tf.placeholder(tf.float32, (None, 448, 448, 3))
label = tf.placeholder(tf.float32, (None, 1470))
yolo = YOLO_tiny_tf.YOLO_TF()
ds_yolo = mut.create_var_xavier('train', tmodel_var_list)
dis_var = mut.create_var_xavier('discriminator', discriminator_var)
theta_D = []
theta_G = []
for i in tmodel_var_list:
theta_G.append(ds_yolo[i[0]])
for i in discriminator_var:
theta_D.append(dis_var[i[0]])
##Train Phase
yolo_ds_train = nf.yolo_ds_all("yolo_train", x, ds_yolo, keep_prob, True)
d_real_logit, d_real_prob = nf.discriminator('discriminator', label, dis_var)
d_fake_logit, d_fake_prob = nf.discriminator('discriminator', yolo_ds_train,
dis_var)
save_epoch = 200
test_epoch = 500
modelTicket_G = {'root':'yolo_tiny', 'branch':'double_cut89'}
modelTicket_D = {'root':'discriminator', 'branch':'4layer'}
keep_prob = tf.placeholder(tf.float32)
x = tf.placeholder(tf.float32,(None,448,448,3))
test = tf.placeholder(tf.float32,(None,448,448,3))
label = tf.placeholder(tf.float32,(None,1470))
yolo = YOLO_tiny_tf.YOLO_TF()
gen_var = mut.create_var_xavier('train',mut.model_zoo(modelTicket_G))
dis_var = mut.create_var_xavier('discriminator', mut.model_zoo(modelTicket_D))
theta_D = []
theta_G = []
for i in mut.model_zoo(modelTicket_G):
theta_G.append(gen_var[i[0]])
for i in mut.model_zoo(modelTicket_D):
theta_D.append(dis_var[i[0]])
##Train Phase
yolo_ds_train = nf.yolo_dinception("yolo_train", x, gen_var, keep_prob, True)
lossTicket = {'loss':'L2norm'}
loss = mut.loss_zoo(lossTicket, yolo_ds_train, label)