with tf.Session() as sess: # start session
sess.run(tf.global_variables_initializer()) # Run initialize op
save_path = "save/"+dir_
predictions_length, predictions_length_w2opt, time_mmodel, time_l2opt = [], [], [], []
pred_all_2opt, time_all_2opt = [], []
for i in tqdm(range(1000)): # test instance
seed_ = 1+i
saver.restore(sess, save_path+"/actor.ckpt") # Restore variables from disk.
init_glob = [actor.global_step.assign(10000), actor.global_step2.assign(10000)]
sess.run(init_glob)
input_batch, dist_batch = dataset.test_batch(1, actor.max_length, actor.dimension, seed=seed_)
feed = {actor.input_: input_batch, actor.distances: dist_batch} # Get feed dict
#start
start_active_search = t()
for it_j in range(config.acive_search_steps):
reward, predictions, _, _ = sess.run([actor.reward, actor.predictions,
actor.trn_op1, actor.trn_op2], feed_dict=feed)
if it_j % 100 == 0:
print(i)
print('reward',np.min(reward))
print('predictions',predictions)
print(sess.run(actor.global_step))
print('time : ', t() - start_active_search)
variables_to_save = [v for v in tf.global_variables() if 'Adam' not in v.name] # Save & restore all the variables.
saver = tf.train.Saver(var_list=variables_to_save, keep_checkpoint_every_n_hours=1.0)
with tf.Session() as sess: # start session
sess.run(tf.global_variables_initializer()) # Run initialize op
save_path = "save/"+dir_
saver.restore(sess, save_path+"/actor.ckpt") # Restore variables from disk.
predictions_length, predictions_length_w2opt, time_mmodel, time_l2opt = [], [], [], []
pred_all_2opt, time_all_2opt = [], []
for i in tqdm(range(1000)): # test instance
seed_ = 1+i
input_batch, dist_batch = dataset.test_batch(1, config.max_length, config.dimension, seed=seed_)
feed = {actor.input_: input_batch, actor.distances: dist_batch} # Get feed dict
start = t()
tour, reward = sess.run([actor.from_, actor.reward], feed_dict=feed) # sample tours
model_time = (t() - start)
time_mmodel.append(model_time)
j = np.argmin(reward) # find best solution
best_permutation = tour[j]
predictions_length.append(reward[j])
#print('reward (before 2 opt)',reward[j])
#dataset.visualize_2D_trip(input_batch[0][best_permutation])
#dataset.visualize_sampling(tour)
start2 = t()
opt_tour, opt_length = dataset.loop2opt(best_permutation, dist_batch[0])
l2opt_time = t()- start + model_time