def main(argv):
# print_squidward()
# FileManager: load specs used in experiment
fm = FileManager()
try:
spec_summary = fm.load_spec_summary(FLAGS.specs)
print(spec_summary)
fm.change_cwd(spec_summary["ROOT_DIR"])
except:
print("Loading specs/model failed. Have you selected the right path?")
exit()
fm.create_test_file()
agent = setup_agent(spec_summary)
agent.DQN.load(FLAGS.model)
agent.set_testing_mode()
# setup environment in testing mode
env = gym.make("sc2-v0")
obs, reward, done, info = env.setup(spec_summary, "testing")
while(True):
# Action selection
action = agent.policy(obs, reward, done, info)
# print(action)
if (action is 'reset'): # Resetting the environment
obs, reward, done, info = env.reset()
# No saving of model in test_mode
else: # Peforming selected action
obs, reward, done, info = env.step(action)
test_report = agent.evaluate(obs, reward, done, info)
fm.log_test_reports(test_report)
if env.finished:
print("Finished testing.")
break
def main(argv):
try:
# FileManager: Save specs and create experiment
fm = FileManager()
try:
fm.create_experiment(agent_specs["EXP_NAME"]) # Automatic cwd switch
fm.save_specs(agent_specs, mv2beacon_specs)
except:
print("Creating experiment or saving specs failed.")
exit()
fm.create_train_file()
# Create HDRL agent
agent = setup_agent(agent_specs)
# Extract skills
skill_name_list = ['move2beacon','collectmineralshards']
skill_specs_list = fm.extract_skill_specs(skill_name_list)
agent_list = setup_multiple_agents(skill_specs_list)
move2beacon_agent = agent_list[0]
collectmineralshards_agent = agent_list[1]
move2beacon_agent.DQN.load(fm.main_dir + '/assets/skills/' + skill_name_list[0] + '/model.pt') # Load model into agent
collectmineralshards_agent.DQN.load(fm.main_dir + '/assets/skills/' + skill_name_list[1] + '/model.pt') # Load model into agent
move2beacon_skill = move2beacon_agent.DQN # Extract DQN skill network
collectmineralshards_skill = collectmineralshards_agent.DQN
agent.add_skill_list([move2beacon_skill, collectmineralshards_skill])
print_ts("Code ran through")
exit()
# Create environment
env = gym.make("gym-sc2-m2b-v0")
obs, reward, done, info = env.setup(mv2beacon_specs, "learning")
while(True):
# Action selection
action = agent.policy(obs, reward, done, info)
if (action is 'reset'): # Resetting the environment
obs, reward, done, info = env.reset()
if agent.episodes % agent.model_save_period == 0:
agent.save_model(fm.get_cwd())
else: # Peforming selected action
obs, reward, done, info = env.step(action)
dict_agent_report = agent.evaluate(obs, reward, done, info)
fm.log_training_reports(dict_agent_report)
if env.finished:
print("Finished learning.")
break
except KeyboardInterrupt:
agent.save_model(fm.get_cwd(), emergency=True)
exit()
def main(argv):
# print_squidward()
try:
# FileManager: Save specs and create experiment
fm = FileManager()
try:
fm.create_experiment(
agent_specs["EXP_NAME"]) # Automatic cwd switch
fm.save_specs(agent_specs, mv2beacon_specs)
except:
print("Creating eperiment or saving specs failed.")
exit()
fm.create_train_file()
agent = setup_agent(agent_specs)
# Loading of model possible fot training mode
agent.set_supervised_mode()
# setup environment in learning mode
env = gym.make("gym-sc2-m2b-v0")
obs, reward, done, info = env.setup(mv2beacon_specs, "learning")
while (True):
# Action selection
action = agent.policy(obs, reward, done, info)
if (action is 'reset'): # Resetting the environment
obs, reward, done, info = env.reset()
if agent.episodes % agent_specs['MODEL_SAVE_PERIOD'] == 0:
agent.save_model(fm.get_cwd())
else: # Peforming selected action
obs, reward, done, info = env.step(action)
dict_agent_report = agent.evaluate(obs, reward, done, info)
fm.log_training_reports(dict_agent_report)
if env.finished:
print("Finished learning.")
break
except KeyboardInterrupt:
agent.save_model(fm.get_cwd(), emergency=True)
exit()
def main(argv):
# print_squidward()
# FileManager: Save specs and create experiment
fm = FileManager()
try:
fm.create_experiment(agent_specs["EXP_NAME"]) # Automatic cwd switch
fm.save_specs(agent_specs, cartpole_specs)
except:
print("Creating eperiment or saving specs failed.")
exit()
fm.create_train_file()
# show_extracted_screen(get_screen(env))
plotter = Plotter()
# No FileManager yet
agent = CartPoleAgent(agent_specs)
agent.set_learning_mode()
env = gym.make('CartPole-v0').unwrapped
num_episodes = cartpole_specs["EPISODES"]
for e in range(num_episodes):
# Initialize the environment and state
state, reward, done, info = env.reset()
last_screen = get_screen(env)
current_screen = get_screen(env)
state = current_screen - last_screen
if e % 50 == 0 and e != 0:
agent.save_model(fm.get_cwd())
for t in count():
# Select and perform an action
action = agent.policy(state, reward, done, info)
_, reward, done, info = env.step(action)
last_screen = current_screen
current_screen = get_screen(env)
if not done:
next_state = current_screen - last_screen
else:
next_state = None
agent.episodes += 1
# agent.update_target_network()
plotter.episode_durations.append(t + 1)
plotter.plot_durations()
print(e, t, reward, action, len(agent.DQN.memory), agent.epsilon)
# print(agent.DQN.state_q_values)
# Store the transition in memory
train_report = agent.evaluate(next_state, reward, done, info)
fm.log_training_reports(train_report)
# Move to the next state
state = next_state
if done:
break
agent.update_target_network()
agent.save_model(fm.get_cwd())
print('Training complete')
env.close()
plotter.close()
def main(argv):
# print_squidward()
# FileManager: Save specs and create experiment
fm = FileManager()
try:
spec_summary = fm.load_spec_summary(FLAGS.specs)
fm.change_cwd(spec_summary["ROOT_DIR"])
except:
print("Loading specs/model failed. Have you selected the right path?")
exit()
fm.create_test_file()
# show_extracted_screen(get_screen(env))
plotter = Plotter()
# No FileManager yet
agent = CartPoleAgent(agent_specs)
agent.DQN.load(FLAGS.model)
agent.set_testing_mode()
list_reward_per_episode = []
env = gym.make('CartPole-v0').unwrapped
num_episodes = int(spec_summary["TEST_EPISODES"])
for e in range(num_episodes):
# Initialize the environment and state
state, reward, done, info = env.reset()
last_screen = get_screen(env)
current_screen = get_screen(env)
state = current_screen - last_screen
reward_cnt = 0
try:
print("Episode {} | Last reward: {}".format(
e, list_reward_per_episode[-1]))
except:
pass
for t in count():
# Select and perform an action
action = agent.policy(state, reward, done, info)
_, reward, done, info = env.step(action)
reward_cnt += reward
last_screen = current_screen
current_screen = get_screen(env)
if not done:
next_state = current_screen - last_screen
else:
next_state = None
agent.episodes += 1
plotter.episode_durations.append(t + 1)
plotter.plot_durations()
# Store the transition in memory
test_report = agent.evaluate(next_state, reward, done, info)
fm.log_test_reports(test_report)
# Move to the next state
state = next_state
if done:
list_reward_per_episode.append(reward_cnt)
dict_test_report = {
"RewardPerEpisode": list_reward_per_episode
}
fm.log_test_reports(dict_test_report)
break
print('Testing complete')
env.close()
plotter.close()