def run_maze():
step = 0
render_time = 0
max_episodes = 5000
episode_step_holder = []
success_holder = []
base_path = './logs/dqn/model/'
for i_episode in range(max_episodes):
episode_step = 0
s = env.reset().ravel()
while True:
env.render(render_time)
action = RL.choose_action(s)
s_, reward, done, info = env.step(action)
s_ = s_.ravel()
# print('action:{0} | reward:{1} | done: {2}'.format(action, reward, done))
RL.store_transition(s, action, reward, s_)
if step > 200:
RL.learn(done)
s = s_
step += 1
episode_step += 1
if episode_step > 299:
done = True
if done:
print('{0} -- {1} -- {2} -- {3}'.format(
i_episode, info, episode_step, RL.epsilon))
if info == 'running':
episode_step = 300
success_holder.append(0)
elif info == 'terminal':
if episode_step < 50:
success_holder.append(1)
else:
success_holder.append(1)
else:
raise Exception("Invalid info code.")
env.render(render_time)
episode_step_holder.append(episode_step)
break
# end of game
print('game over')
save_path = RL.saver.save(RL.sess, base_path + 'model_dqn.ckpt')
print("Model saved in path: {}".format(save_path))
RL.sess.close()
env.destroy()
# plot_cost(episode_step_holder, base_path + 'episode_steps.png')
plot_rate(success_holder, base_path, index=0)
def run_maze():
step = 0
render_time = 0
episode_step_holder = []
success_holder = []
base_path = './logs/dqn/'
for i_episode in range(400):
episode_step = 0
s = env.reset().ravel()
while True:
env.render(render_time)
action = RL.choose_action(s)
s_, reward, done, info = env.step(action)
s_ = s_.ravel()
# print('action:{0} | reward:{1} | done: {2}'.format(action, reward, done))
RL.store_transition(s, action, reward, s_)
if step > 200:
RL.learn()
s = s_
step += 1
episode_step += 1
if episode_step > 500:
done = True
if done:
print('{0} -- {1} -- {2}'.format(i_episode, info,
episode_step))
if info != 'success':
episode_step = 500
reward = -1
success_holder.append(0)
else:
success_holder.append(1)
env.render(render_time)
episode_step_holder.append(episode_step)
break
# end of game
print('game over')
save_path = RL.saver.save(RL.sess, base_path + 'model_dqn.ckpt')
print("Model saved in path: {}".format(save_path))
RL.sess.close()
env.destroy()
# plot_cost(episode_step_holder, base_path + 'episode_steps.png')
plot_rate(success_holder, base_path, index=15)
def run_maze():
# expert_counter = 0 # Record how many times we refer the expert action.
n_features = env.height * env.width
n_actions = 4
restore_path = None
dagger_itr = 5 # how many times we do dataset aggregation.
episode_step_holder = []
success_holder = []
obser_list = []
action_list = []
obsers_all = np.zeros((1, n_features))
actions_all = np.zeros((1, 1))
# # Collecting data.
for j in range(RUN_STEPS):
s = env.reset()
while True:
env.render(RENDER_TIME)
a = get_expert_action(s)
obser_list.append(s.ravel())
action_list.append(a)
s_, r, done, info = env.step(a)
if done:
env.render(RENDER_TIME)
break
s = s_
assert len(obser_list) == len(action_list)
for obser, act in zip(obser_list, action_list):
obsers_all = np.concatenate([obsers_all, obser[np.newaxis, :]], axis=0)
actions_all = np.concatenate([actions_all, np.array([act])[np.newaxis, :]], axis=0)
obsers_all = obsers_all[1:, :]
actions_all = actions_all[1:, :].astype(int)
actions_all = one_hot_encoding_numpy(actions_all.ravel().tolist(), 4)
# # Training an initialized policy.
net_s = tf.placeholder(tf.float32, [None, n_features], name='net_s') # observations
net_a = tf.placeholder(tf.int32, [None, n_actions], name='net_a') # expert actions
net_l1 = tf.contrib.layers.fully_connected(net_s, 32, activation_fn=tf.nn.relu)
# net_l2 = tf.contrib.layers.fully_connected(net_l1, 128, activation_fn=tf.nn.relu)
net_out = tf.contrib.layers.fully_connected(net_l1, n_actions, activation_fn=tf.nn.softmax)
net_loss = tf.losses.softmax_cross_entropy(onehot_labels=net_a, logits=net_out) # compute cost
train_op = tf.train.AdamOptimizer(0.001).minimize(net_loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
if restore_path is not None:
saver.restore(sess, restore_path)
print("Model restore in path: {}".format(restore_path))
else:
print("No pretrained model found.")
for step in range(TRAIN_STEPS):
b_s, b_a = get_batch(obsers_all, actions_all, batch_size=32)
_, loss_ = sess.run([train_op, net_loss], {net_s: b_s, net_a: b_a})
if step % 5 == 0:
output = sess.run(net_out, {net_s: obsers_all, net_a: actions_all})
item_a = np.argmax(output, axis=1)
item_b = np.argmax(actions_all, axis=1)
n_accuracy = np.where(np.equal(np.argmax(output, axis=1), np.argmax(actions_all, axis=1)))[0]
accuracy_ = n_accuracy.shape[0] / actions_all.shape[0]
print('Step:', step, '| train loss: %.4f' % loss_, '| test accuracy: %.2f' % accuracy_)
output = sess.run(net_out, {net_s: obsers_all, net_a: actions_all})
n_accuracy = np.where(np.equal(np.argmax(output, axis=1), np.argmax(actions_all, axis=1)))[0]
accuracy_ = n_accuracy.shape[0] / actions_all.shape[0]
print('Step:', step, '| train loss: %.4f' % loss_, '| test accuracy: %.2f' % accuracy_)
save_path = saver.save(sess, BASE_LOGS + 'model_init.ckpt')
print("Model saved in path: {}".format(save_path))
# # Dataset Aggregation and Retraining the policy.
for i in range(dagger_itr):
restore_path = save_path
# restore_path = BASE_LOGS + 'model_init.ckpt'
obser_list = []
action_list = []
obsers_new = np.zeros((1, n_features))
actions_new = np.zeros((1, 1))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
if restore_path is not None:
saver.restore(sess, restore_path)
print("Model restore in path: {}".format(restore_path))
else:
raise Exception("No pretrained model found.")
for j in range(RUN_STEPS):
episode_step = 0
s = env.reset()
while True:
env.render(RENDER_TIME)
a = sess.run(net_out, {net_s: s.ravel()[np.newaxis, :]})
a = np.argmax(a, axis=1)[0]
obser_list.append(s.ravel())
action_list.append(get_expert_action(s)) # add action what the expert teaches.
s_, r, done, info = env.step(a) # act action what the model output.
s = s_
episode_step += 1
if episode_step > 299:
done = True
if done:
if info == 'running':
episode_step = 300
success_holder.append(0)
elif info == 'terminal':
if episode_step < 50:
success_holder.append(1)
else:
success_holder.append(1)
else:
raise Exception("Invalid info code.")
env.render(RENDER_TIME)
episode_step_holder.append(episode_step)
break
assert len(obser_list) == len(action_list)
for obser, act in zip(obser_list, action_list):
obsers_new = np.concatenate([obsers_new, obser[np.newaxis, :]], axis=0)
actions_new = np.concatenate([actions_new, np.array([act])[np.newaxis, :]], axis=0)
obsers_new = obsers_new[1:, :]
actions_new = actions_new[1:, :].astype(int)
actions_new = one_hot_encoding_numpy(actions_new.ravel().tolist(), 4)
# Dataset Aggregation
obsers_all = np.concatenate([obsers_all, obsers_new], axis=0)
actions_all = np.concatenate([actions_all, actions_new], axis=0)
# Retraining the policy
saver = tf.train.Saver()
saver.restore(sess, restore_path)
print("Model restore in path: {}".format(restore_path))
for step in range(TRAIN_STEPS):
b_s, b_a = get_batch(obsers_all, actions_all, batch_size=32)
_, loss_ = sess.run([train_op, net_loss], {net_s: b_s, net_a: b_a})
if step % 5 == 0:
output = sess.run(net_out, {net_s: obsers_all, net_a: actions_all})
n_accuracy = np.where(np.equal(np.argmax(output, axis=1), np.argmax(actions_all, axis=1)))[0]
accuracy_ = n_accuracy.shape[0] / actions_all.shape[0]
print('Step:', step, '| train loss: %.4f' % loss_, '| test accuracy: %.2f' % accuracy_)
output = sess.run(net_out, {net_s: obsers_all, net_a: actions_all})
n_accuracy = np.where(np.equal(np.argmax(output, axis=1), np.argmax(actions_all, axis=1)))[0]
accuracy_ = n_accuracy.shape[0] / actions_all.shape[0]
print('Step:', step, '| train loss: %.4f' % loss_, '| test accuracy: %.2f' % accuracy_)
print("Size of the dataset {0}".format(obsers_all.shape[0]))
save_path = saver.save(sess, BASE_LOGS + 'model_{}.ckpt'.format(i))
print("Model saved in path: {}".format(save_path))
plot_rate(success_holder, './logs/dagger/model/', index=0)
# # Testing the final policy.
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
if restore_path is not None:
saver.restore(sess, restore_path)
print("Model restore in path: {}".format(restore_path))
else:
raise Exception("No pretrained model found.")
for epi in range(100):
s = env.reset()
while True:
env.render(0)
a = sess.run(net_out, {net_s: s.ravel()[np.newaxis, :]})
a = np.argmax(a, axis=1)[0]
s_, r, done, info = env.step(a) # act action what the model output.
if done:
env.render(0)
break
s = s_
# # destroy the env.
env.destroy()
"""
return super().append_input(data)
# The file path of packets' log
log_packet_file = "output/packet_log/packet-0.log"
# Use the object you created above
my_solution = MySolution()
# Create the emulator using your solution
# Specify USE_CWND to decide whether or not use crowded windows. USE_CWND=True by default.
# Specify ENABLE_LOG to decide whether or not output the log of packets. ENABLE_LOG=True by default.
# You can get more information about parameters at https://github.com/Azson/DTP-emulator/tree/pcc-emulator#constant
emulator = PccEmulator(solution=my_solution, USE_CWND=False, ENABLE_LOG=True)
# Run the emulator and you can specify the time for the emualtor's running.
# It will run until there is no packet can sent by default.
emulator.run_for_dur()
# print the debug information of links and senders
emulator.print_debug()
# Output the picture of pcc_emulator-analysis.png
# You can get more information from https://github.com/Azson/DTP-emulator/tree/pcc-emulator#pcc_emulator-analysispng.
analyze_pcc_emulator(log_packet_file, file_range="all")
# Output the picture of cwnd_changing.png
# You can get more information from https://github.com/Azson/DTP-emulator/tree/pcc-emulator#cwnd_changingpng
plot_rate(log_packet_file, file_range="all")
def run_maze():
step = 0
render_time = 0
episode_step_holder = []
success_holder = []
base_path = './logs/a2c/'
sess = tf.Session()
actor = Actor(sess, n_features=N_F, n_actions=N_A, lr=LR_A)
critic = Critic(sess, n_features=N_F, lr=LR_C)
sess.run(tf.global_variables_initializer())
# Add ops to save and restore all the variables.
saver = tf.train.Saver()
if RESTORE:
saver.restore(sess, base_path + 'model_a2c.ckpt')
print("Model restore in path: {}".format(base_path + 'model_a2c.ckpt'))
if OUTPUT_GRAPH:
tf.summary.FileWriter('.logs/', sess.graph)
for i_episode in range(MAX_EPISODE):
episode_step = 0
s = env.reset().ravel()
while True:
env.render(render_time)
a = actor.choose_action(s)
s_, r, done, info = env.step(a)
s_ = s_.ravel()
td_error = critic.learn(
s, r, s_) # gradient = grad[r + gamma * V(s_) - V(s)]
actor.learn(
s, a,
td_error) # true_gradient = grad[log(Pi(s,a)) * td_error]
s = s_
step += 1
episode_step += 1
if episode_step > 100:
done = True
if done:
print('{0} -- {1} -- {2}'.format(i_episode, info,
episode_step))
if info != 'success':
episode_step = 500
reward = -1
success_holder.append(0)
else:
success_holder.append(1)
env.render(render_time)
episode_step_holder.append(episode_step)
break
# end of game
print('game over')
if RESTORE:
save_path = saver.save(sess, base_path + 'model_a2c.ckpt')
print("Model saved in path: {}".format(save_path))
sess.close()
env.destroy()
# plot_cost(episode_step_holder, base_path + 'episode_steps.png')
plot_rate(success_holder, base_path + 'success_rate.png')
