def main():
inputs = tf.placeholder(
dtype=tf.float32,
shape=[None, config.img_h, config.img_w, config.img_ch])
y_true_l = tf.placeholder(dtype=tf.float32,
shape=[
None, config.img_h // 8, config.img_w // 8,
3, 5 + len(cls_dict.label_num_dict)
])
y_true_m = tf.placeholder(dtype=tf.float32,
shape=[
None, config.img_h // 16, config.img_w // 16,
3, 5 + len(cls_dict.label_num_dict)
])
y_true_s = tf.placeholder(dtype=tf.float32,
shape=[
None, config.img_h // 32, config.img_w // 32,
3, 5 + len(cls_dict.label_num_dict)
])
y_true = [y_true_l, y_true_m, y_true_s]
fm1, fm2, fm3 = yolo3_net(inputs=inputs)
# anchors = [[20, 30], [40, 50], [60, 70]]
y_pred = [fm3, fm2, fm1]
calc_loss.build_loss(y_pred=y_pred, y_true=y_true)
def yolo3_model(mode, feature, label, batch_size):
# is_training = (mode == tf.estimator.ModeKeys.TRAIN)
fm1, fm2, fm3 = model_design_nn.yolo3_net(inputs=feature)
y_pred = [fm3, fm2, fm1]
if mode == tf.estimator.ModeKeys.TRAIN or mode == tf.estimator.ModeKeys.EVAL:
# calc ctc loss
loss = calc_loss.build_loss(y_pred=y_pred, y_true=label)
# # calc l2 loss
# l2_loss = tf.Variable(initial_value=0, dtype=tf.float32, trainable=False)
# for scope_name in ['CNN_Net', 'RNN_Net' 'FC_Net']:
# net_tv = tf.trainable_variables(scope=scope_name)
# r_lambda = 0.001
# regularization_cost = r_lambda * tf.reduce_sum([tf.nn.l2_loss(v) for v in net_tv])
# loss += regularization_cost
# l2_loss += regularization_cost
return y_pred, loss # , l2_loss, net_out, tensor_dict, seq_len
else:
# inference
loss = None
# l2_loss = None
return y_pred, loss # , l2_loss, net_out, tensor_dict, seq_len
def main():
# init directory
if not os.path.exists(config.model_save_dir):
os.mkdir(config.model_save_dir)
if not os.path.exists(config.summary_save_dir):
os.mkdir(config.summary_save_dir)
# data_load
Data_Gen = data_loader.DataGenerator(img_dir=config.img_dir,
img_h=config.img_h,
img_w=config.img_w,
img_ch=config.img_ch,
batch_size=config.batch_size)
Data_Gen.build_data()
generator = Data_Gen.next_batch()
# define train model
images = tf.placeholder(
dtype=tf.float32,
shape=[None, config.img_h, config.img_w, config.img_ch])
labels = tf.placeholder(dtype=tf.int32, shape=[None])
logits = model_design_nn.cls_net(inputs=images, is_training=True)
cls_loss = calc_loss.build_loss(logits=logits, labels=labels)
# l2_loss
l2_loss = tf.Variable(initial_value=0, dtype=tf.float32, trainable=False)
for scope_name in ['CNN_Module', 'FC_Module']:
module_train_vars = tf.trainable_variables(scope=scope_name)
regularization_cost = tf.reduce_sum(
[tf.nn.l2_loss(var) for var in module_train_vars])
l2_loss += regularization_cost * config.l2_loss_lambda
loss = cls_loss + l2_loss
tf.summary.scalar("Loss/0_total_loss", loss)
tf.summary.scalar("Loss/1_cls_loss", cls_loss)
tf.summary.scalar("Loss/2_l2_loss", l2_loss)
# summary_op
summary_op = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(config.summary_save_dir)
# train_op
global_step = tf.train.get_or_create_global_step()
train_op = tf.train.AdamOptimizer(
learning_rate=config.learning_rate).minimize(loss=loss,
global_step=global_step)
sess = tf.Session()
saver = tf.train.Saver(max_to_keep=3)
ckpt_path = tf.train.latest_checkpoint(config.model_save_dir)
print('latest_checkpoint_path: ', ckpt_path)
if ckpt_path is not None:
saver.restore(sess, ckpt_path)
prev_step = int(ckpt_path.split('-')[-1])
else:
init_op = tf.initialize_all_variables()
sess.run(init_op)
prev_step = -1
# train
with sess.as_default():
for i in range(config.train_steps):
_inputs, _outputs = next(generator) # .__next__()
_img_tensor = _inputs['images']
_label_tensor = _outputs['labels']
# print(_ant_tensor.shape)
_loss, _, _summary_op = sess.run([loss, train_op, summary_op],
feed_dict={
images: _img_tensor,
labels: _label_tensor
})
print('step: ', prev_step + 1 + i, 'loss: ', _loss)
train_writer.add_summary(_summary_op, prev_step + 1 + i)
train_writer.flush()
if i % config.save_n_iters == 0:
saver.save(sess=sess,
save_path=os.path.join(config.model_save_dir,
'model.ckpt'),
global_step=global_step)
def train_ver1():
anchors = [[12, 10], [18, 15], [22, 22], [29, 20], [29, 29], [37, 24],
[36, 34], [40, 43], [44, 55]]
DataSet = load_dataset.Read_Tfrecord(
tfrecord_dir='/data/data/weche_tfrecords', anchors=anchors)
config.batch_size = 2
imgs, y_true = DataSet.make_batch(batch_size=config.batch_size)
fm1, fm2, fm3 = model_design_nn.yolo3_net(inputs=imgs)
# anchors = [[20, 30], [40, 50], [60, 70]]
y_pred = [fm3, fm2, fm1]
loss = calc_loss.build_loss(y_pred=y_pred, y_true=y_true)
tf.summary.scalar('train_batch_statistics/total_loss', loss[0])
tf.summary.scalar('train_batch_statistics/loss_xy', loss[1])
tf.summary.scalar('train_batch_statistics/loss_wh', loss[2])
tf.summary.scalar('train_batch_statistics/loss_conf', loss[3])
tf.summary.scalar('train_batch_statistics/loss_class', loss[4])
# summary_op
summary_op = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('./train_summary')
# train_op
global_step = tf.train.get_or_create_global_step()
# add update_op for slim.bn
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(
loss=loss[0], global_step=global_step)
sess = tf.Session()
saver = tf.train.Saver(max_to_keep=3)
model_save_dir = './run_output'
if not os.path.exists(model_save_dir):
os.mkdir(model_save_dir)
ckpt_path = tf.train.latest_checkpoint(model_save_dir)
print('latest_checkpoint_path: ', ckpt_path)
if ckpt_path is not None:
saver.restore(sess, ckpt_path)
prev_step = int(ckpt_path.split('-')[-1])
else:
init_op = tf.global_variables_initializer()
sess.run(init_op)
prev_step = -1
with sess.as_default():
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for i in range(100):
_loss, _, _summary_op = sess.run([loss[0], train_op, summary_op])
print('step: ', prev_step + 1 + i, 'loss: ', _loss)
# train_writer.add_summary(_summary_op, prev_step + 1 + i)
# train_writer.flush()
if i % 10 == 0:
saver.save(sess=sess,
save_path=os.path.join(model_save_dir,
'model.ckpt'),
global_step=global_step)
coord.request_stop()
coord.join(threads=threads)
sess.close()