Convert the whole codes to tensorflow 2.x for enabling eager_mode for eclipse debugger

[sekigh]

I am a almost beginner for tensorflow.
I am trying to convert the whole codes, priority is seq_group.py, to tensorflow 2.4.1 for enabling eager_mode (define by run mode).
I followed instructions upto "Top-level behavioral changes" on tensorflow documents (https://www.tensorflow.org/guide/migrate) and successfully had them run on 2.4.1 in session.run() mode. I attach successfully running codes on 2.4.1 with session.run() here.
Now I am trying to do "Create code for TensorFlow 2.x" on that document and try to remove session.run() but do not know how to do for following three things;

1. How to build and treat multiple asynchronous training data threading without session.run() and feed those threadings into training flow. This question is associated with queue/dequeue parts and threading parts as follows;

q = tf.queue.PaddingFIFOQueue(50, [tf.float32, tf.float32, tf.float32, tf.float32, tf.float32, tf.float32, tf.float32, tf.float32, tf.int32], shapes=[[None, n_input], [None, n_input], [None, n_output], [None, n_output], [None, n_output], [None, n_output], [None, n_output2], [None, n_output2], [None]])
enqueue_op = q.enqueue([xr_batch, xi_batch, y1r_batch, y1i_batch, y2r_batch, y2i_batch, est1_batch, est2_batch, seq_len_batch])
queue_size = q.size()

Define dequeue operation without predefined batch size

xr_, xi_, y1r_, y1i_, y2r_, y2i_, est1_, est2_, seq_len_ = q.dequeue()

and

def load_and_enqueue(sess, enqueue_op, coord, queue_index, total_queues):

and

coord = tf.train.Coordinator()
num_threads = 5 # Use 5 threads for data loading, each thread touches different part of training data
t = [threading.Thread(target=load_and_enqueue, args=(sess,enqueue_op,coord,i,num_threads)) for i in range(num_threads)]
for tmp_t in t:
tmp_t.start()

2. How to implement optimizer and forward/backpropagation for training without session.run(). This question is associated with defining optimizer and forward/backpropagation in training as follows;

Define loss and optimizer

extra_update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
if FLAGS.is_adam == 1:
optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=lr,beta2=0.9).minimize(cost) # Adam Optimizer
else:
optimizer = tf.compat.v1.train.MomentumOptimizer(learning_rate=lr, momentum=0.9).minimize(cost) # Momentum Optimizer

and

Run one batch of training

    _, train_cost, train_cost_pit, q_size\
    = sess.run([optimizer, cost, cost_pit, queue_size], \
    feed_dict={lr: learning_rate, keep_prob: FLAGS.keep_prob, is_training: True})

3. How to replace model Saver save/restore without session.run() ? This question is associated with defining tf.train.Saver and its methods for save and restore as follows;

Model dir and model saver

model_dir = os.getcwd()+"/exp/"+FLAGS.exp_name+"/models/"+FLAGS.time_stamp
if not os.path.exists(model_dir):
os.makedirs(model_dir)
saver = tf.compat.v1.train.Saver(max_to_keep = None)

and

If there is a pickle file, then load to restore previous training states

with open(training_info_pickle, 'rb') as f_pickle:
step, best_cv_cost, best_cv_step, learning_rate = pickle.load(f_pickle)
saver.restore(sess, os.getcwd()+"/exp/"+FLAGS.exp_name+"/models/"+FLAGS.time_stamp+'/'+FLAGS.exp_name+"_model.ckpt" + "step"+ str(step))

and

Saving models and training specs after evaluation

save_path = saver.save(sess, os.getcwd()+"/exp/"+FLAGS.exp_name+"/models/"+FLAGS.time_stamp+'/'+FLAGS.exp_name+"_model.ckpt" + "step"+ str(step))

The end of text.
seq_group_and_utility_running on tf2.zip