args.learn_start)
# Construct validation memory
val_mem = ReplayMemory(args, args.evaluation_size)
T, done = 0, True
while T < args.evaluation_size:
if done:
state, done = env.reset(), False
action = random.randint(0, action_space - 1)
if args.env == 'peg1-v0':
action = np.array([action // 16, action % 16 // 4, action % 4])
next_state, _, done, _ = env.step(action)
else:
next_state, _, done = env.step(action)
val_mem.append(state, None, None, done)
state = next_state
T += 1
if args.evaluate:
dqn.eval() # Set DQN (online network) to evaluation mode
avg_reward, avg_Q = test(args, 0, dqn, val_mem, evaluate=True) # Test
print('Avg. reward: ' + str(avg_reward) + ' | Avg. Q: ' + str(avg_Q))
else:
# Training loop
dqn.train()
T, done = 0, True
while T < args.T_max:
if done:
state, done = env.reset(), False
'online_net': dqn.online_net.state_dict()
}, ckptdir / 'last_ckpt.tar')
save_mem(mem, ckptdir)
log("Checkpoint successfully saved")
priority_weight_increase = (1 - args.priority_weight) / (args.T_max - args.learn_start)
# Construct validation memory
val_mem = ReplayMemory(args, args.evaluation_size)
T, done = 0, True
while T < args.evaluation_size:
if done:
state = env.reset()
next_state, _, done = env.step(np.random.randint(0, action_space))
val_mem.append(state, -1, 0.0, done)
state = next_state
T += 1
done = True
if args.evaluate:
dqn.eval() # Set DQN (online network) to evaluation mode
avg_reward, avg_Q = test(args, 0, dqn, val_mem, metrics, results_dir, evaluate=True) # Test
print('Avg. reward: ' + str(avg_reward) + ' | Avg. Q: ' + str(avg_Q))
else:
# Training loop
dqn.train()
T, done = 0, True
for T in trange(T_init, args.T_max + 1):
if done:
# Agent
dqn = Agent(args, env)
mem = ReplayMemory(args, args.memory_capacity)
priority_weight_increase = (1 - args.priority_weight) / (args.T_max -
args.learn_start)
# Construct validation memory
val_mem = ReplayMemory(args, args.evaluation_size)
T, done = 0, True
while T < args.evaluation_size - args.history_length + 1:
if done:
state, done = env.reset(), False
val_mem.preappend() # Set up memory for beginning of episode
val_mem.append(state, None, None)
state, _, done = env.step(random.randint(0, action_space - 1))
T += 1
# No need to postappend on done in validation memory
if args.evaluate:
dqn.eval() # Set DQN (policy network) to evaluation mode
avg_reward, avg_Q = test(args, 0, dqn, val_mem, evaluate=True) # Test
print('Avg. reward: ' + str(avg_reward) + ' | Avg. Q: ' + str(avg_Q))
else:
# Training loop
dqn.train()
T, done = 0, True
while T < args.T_max:
if done:
state, done = Variable(env.reset()), False
def train_agent(env, args, config):
"""
Args:
"""
# create CNN convert the [1,3,84,84] to [1, 200]
now = datetime.now()
dt_string = now.strftime("%d_%m_%Y_%H:%M:%S")
torch.manual_seed(config["seed"])
np.random.seed(config["seed"])
if torch.cuda.is_available() and not args.disable_cuda:
args.device = torch.device('cuda')
torch.cuda.manual_seed(np.random.randint(1, 10000))
torch.backends.cudnn.enabled = args.enable_cudnn
pathname = dt_string + "_seed" + str(config["seed"])
print("save tensorboard {}".format(config["locexp"]))
tensorboard_name = str(config["locexp"]) + '/runs/' + pathname
agent = Agent(args, env)
memory = ReplayMemory(args, args.memory_capacity)
#memory = ReplayBuffer((3, config["size"], config["size"]), (1,), config["expert_buffer_size"], int(config["image_pad"]), config["device"])
priority_weight_increase = (1 - args.priority_weight) / (args.T_max -
args.learn_start)
writer = SummaryWriter(tensorboard_name)
results_dir = os.path.join(str(config["locexp"]), args.id)
mkdir("", results_dir)
scores_window = deque(maxlen=100)
scores = []
t0 = time.time()
# Training loop
agent.train()
T, done = 0, True
print("result dir ", results_dir)
agent.save(results_dir, 'checkpoint-{}.pth'.format(T))
#eval_policy(env, agent, writer, T, config)
episode = -1
steps = 0
score = 0
print("save policy ", args.checkpoint_interval)
# eval_policy(env, agent, writer, 0, config)
for T in range(1, args.T_max + 1):
# print("\r {} of {}".format(T, args.T_max), end='')
if done:
episode += 1
scores_window.append(score) # save most recent scor
scores.append(score) # save most recent score
print(
'\rTime steps {} episode {} score {} Average Score: {:.2f} time: {}'
.format(T, episode, score, np.mean(scores_window),
time_format(time.time() - t0)),
end="")
writer.add_scalar('Episode_reward ', score, T)
average_reward = np.mean(scores_window)
writer.add_scalar('Average_reward ', average_reward, T)
state, done = env.reset(), False
steps = 0
score = 0
if T % args.replay_frequency == 0:
agent.reset_noise() # Draw a new set of noisy weights
action = agent.act(
state) # Choose an action greedily (with noisy weights)
next_state, reward, done, _ = env.step(action) # Step
score += reward
steps += 1
if steps == 30:
done = True
memory.append(state, action, reward,
done) # Append transition to memory
# Train and test
if T >= args.learn_start:
memory.priority_weight = min(
memory.priority_weight + priority_weight_increase,
1) # Anneal importance sampling weight β to 1
if T % args.replay_frequency == 0:
agent.learn(
memory
) # Train with n-step distributional double-Q learning
# Update target network
if T % args.target_update == 0:
agent.update_target_net()
# Checkpoint the network
if (args.checkpoint_interval !=
0) and (T % args.checkpoint_interval == 0):
print("Eval policy")
eval_policy(env, agent, writer, T, config)
agent.save(results_dir, 'checkpoint-{}.pth'.format(T))
state = next_state
def train(args, env):
action_space = env.action_space.n
print("show action space", action_space)
print("state space", env.observation_space)
# Agent
dqn_1 = Agent(args, env)
dqn_2 = Agent(args, env)
results_dir = os.path.join('results', args.id)
print("result dir", results_dir)
T, done = 0, True
# If a model is provided, and evaluate is fale, presumably we want to resume, so try to load memory
print(" ags training", args.continue_training)
args.continue_training = False
if args.continue_training:
print("Continue Training Load buffer 1 ...")
args.memory = results_dir + "/val_mem_1/memory.pkl"
mem_1 = load_memory(args.memory, args.disable_bzip_memory)
val_mem_1 = ReplayMemory(args, args.evaluation_size)
print("loaded memory buffer 1")
print("Continue Training Load buffer 2 ...")
args.memory = results_dir + "/val_mem_2/memory.pkl"
mem_2 = load_memory(args.memory, args.disable_bzip_memory)
val_mem_2 = ReplayMemory(args, args.evaluation_size)
print("loaded memory buffer 2")
else:
print("use empty Buffers")
args.memory = results_dir + "/val_mem_1/memory.pkl"
path = results_dir + "/val_mem_1"
print("save memory", args.memory)
os.makedirs(path, exist_ok=True)
val_mem_1 = ReplayMemory(args, args.evaluation_size)
mem_1 = ReplayMemory(args, args.memory_capacity)
args.memory = results_dir + "/val_mem_2/memory.pkl"
path = results_dir + "/val_mem_2"
print("save memory", args.memory)
os.makedirs(path, exist_ok=True)
val_mem_2 = ReplayMemory(args, args.evaluation_size)
mem_2 = ReplayMemory(args, args.memory_capacity)
priority_weight_increase = (1 - args.priority_weight) / (args.T_max -
args.learn_start)
metrics = {
'steps': [],
'rewards': [],
'Qs': [],
'step_rewards': [],
'train_rewards': [],
'best_avg_reward': -float('inf')
}
args.continue_training = True
def write_into_file(text, file_name='document.csv'):
"""
"""
with open(file_name, 'a', newline='\n') as fd:
fd.write(str(text) + "\n")
def log(s):
text = '[' + str(
datetime.now().strftime('%Y-%m-%dT%H:%M:%S')) + '] ' + s
write_into_file(text)
print(text)
if torch.cuda.is_available():
print("cuda")
def save_memory(memory, memory_path, disable_bzip):
if disable_bzip:
with open(memory_path, 'wb') as pickle_file:
pickle.dump(memory, pickle_file)
else:
with bz2.open(memory_path, 'wb') as zipped_pickle_file:
pickle.dump(memory, zipped_pickle_file)
("Create eval memory of size {} ".format(args.evaluation_size))
# Construct validation memory
size = 84
print("Fill eval memory")
# fill both memories at same time
# use the reward function for each
try:
while T < args.evaluation_size:
T += 1
print("steps ", T)
if done:
t = 0
done = False
state = env.reset()
state = torch.tensor(state,
dtype=torch.float32,
device=args.device).div_(255)
zeros = torch.zeros_like(state)
state_buffer = deque([], maxlen=args.history_length)
state_buffer.append(zeros)
state_buffer.append(zeros)
state_buffer.append(zeros)
state_buffer.append(state)
state = torch.stack(list(state_buffer), 0)
t += 1
if t == args.max_episode_length:
#if t == 5:
t = 0
done = True
next_state, _, _, _ = env.step(np.random.randint(0, action_space))
val_mem_1.append(state, None, None, done)
val_mem_2.append(state, None, None, done)
next_state = torch.tensor(next_state,
dtype=torch.float32,
device=args.device).div_(255)
state_buffer.append(next_state)
state = torch.stack(list(state_buffer), 0)
eps_1 = 1
eps_end_1 = 0.05
eps_decay_1 = 0.999978 # reaches 10% at 105000
eps_2 = 1
eps_end_2 = 0.05
eps_decay_2 = 0.999978 # reaches 10% at 10500
#args.evaluate = True
if args.evaluate:
print("Test")
dqn.eval() # Set DQN (online network) to evaluation mode
#avg_reward, avg_Q = test(args, 0, dqn, val_mem, metrics, results_dir, env, evaluate=True) # Test
avg_reward, avg_Q = test(args, T, dqn, val_mem, metrics,
results_dir, env) # Test
print('Avg. reward: ' + str(avg_reward) + ' | Avg. Q: ' +
str(avg_Q))
else:
if args.continue_training:
print("Start Training")
T = args.learn_start + 500
# Training loop
dqn_1.train()
dqn_2.train()
episode = 0
episode_reward = 0
mean_reward = deque(maxlen=100)
plot_rewards = []
print("Fill both memory buffers ")
while T < args.learn_start:
if T % args.max_episode_length == 0:
state, done = env.reset(), False
state = torch.tensor(state,
dtype=torch.float32,
device=args.device).div_(255)
zeros = torch.zeros_like(state)
state_buffer = deque([], maxlen=args.history_length)
state_buffer.append(zeros)
state_buffer.append(zeros)
state_buffer.append(zeros)
state_buffer.append(state)
state = torch.stack(list(state_buffer), 0)
# choose action at random
action = np.random.randint(0, action_space)
next_state, reward, done, reward_2 = env.step(action) # Step
text = "Step {} of {} ".format(T, args.learn_start)
print(text, end='\r', file=sys.stdout, flush=True)
# set done on the last transition
if (T + 1) % args.max_episode_length == 0:
done = True
mem_1.append(state, action, reward, done)
mem_2.append(state, action, reward_2, done)
next_state = torch.tensor(next_state,
dtype=torch.float32,
device=args.device).div_(255)
state_buffer.append(next_state)
state = torch.stack(list(state_buffer), 0)
T += 1
if T >= args.learn_start:
args.memory = results_dir + "/val_mem_1/memory.pkl"
print("save memory 1", args.memory)
save_memory(mem_1, args.memory, args.disable_bzip_memory)
args.memory = results_dir + "/val_mem_2/memory.pkl"
print("save memory 2", args.memory)
save_memory(mem_2, args.memory, args.disable_bzip_memory)
break
print("Start Training")
#for T in tqdm.trange(args.learn_start, args.T_max + 1):
for T in tqdm.trange(0, args.T_max + 1):
if T % args.max_episode_length == 0:
mean_reward.append(episode_reward)
print("Epiosde: {} Reward: {} Mean Reward: {} Goal1 {}".
format(episode, episode_reward, np.mean(mean_reward),
env.goal_counter_1))
plot_rewards.append(np.mean(mean_reward))
save_and_plot(T, plot_rewards)
episode_reward = 0
episode += 1
state, done = env.reset(), False
state = torch.tensor(state,
dtype=torch.float32,
device=args.device).div_(255)
zeros = torch.zeros_like(state)
state_buffer = deque([], maxlen=args.history_length)
state_buffer.append(zeros)
state_buffer.append(zeros)
state_buffer.append(zeros)
state_buffer.append(state)
state = torch.stack(list(state_buffer), 0)
g = 0
set_input = True
secondTask = False
if T % args.replay_frequency == 0:
pass
#dqn.reset_noise() # Draw a new set of noisy weights
"""
if env.task_one_complete or secondTask:
action = dqn_2.act_e_greedy(state, eps_2) # Choose an action greedily (with noisy weights)
secondTask = True
else:
action = dqn_1.act_e_greedy(state, eps_1) # Choose an action greedily (with noisy weights)
"""
if set_input:
set_input = False
g = input("Enter action : ")
action = int(g)
g = input("Enter steps : ")
g = int(g)
if g <= 0:
set_input = True
g -= 1
#print("step : {} action: {} eps: {}".format(T, action, eps))
next_state, reward, done, reward_2 = env.step(action) # Step
if args.reward_clip > 0:
reward = max(min(reward, args.reward_clip),
-args.reward_clip) # Clip rewards
reward_2 = max(min(reward_2, args.reward_clip),
-args.reward_clip) # Clip rewards
if env.task_one_complete or secondTask:
episode_reward += reward_2
eps_2 = max(eps_end_2, eps_decay_2 * eps_2)
mem_2.priority_weight = min(
mem_2.priority_weight + priority_weight_increase,
1) # Anneal importance sampling weight β to 1
else:
episode_reward += reward
eps_1 = max(eps_end_1, eps_decay_1 * eps_1)
mem_1.priority_weight = min(
mem_1.priority_weight + priority_weight_increase,
1) # Anneal importance sampling weight β to 1
#print(reward)
#print(reward_2)
# incase the last action set done to True
if T + 1 % args.max_episode_length == 0:
done = True
mem_1.append(state, action, reward,
done) # Append transition to memory
mem_2.append(state, action, reward_2,
done) # Append transition to memory
# Train and test
next_state = torch.tensor(next_state,
dtype=torch.float32,
device=args.device).div_(255)
# print("Main shape of next_state", next_state.shape)
state_buffer.append(next_state)
state = torch.stack(list(state_buffer), 0)
continue
# print("Main shape of state", state.shape)
if T % args.replay_frequency == 0:
dqn_1.learn(
mem_1
) # Train with n-step distributional double-Q learning
dqn_2.learn(
mem_2
) # Train with n-step distributional double-Q learning
if T % args.evaluation_interval == 0:
dqn_1.eval() # Set DQN (online network) to evaluation mode
print("Eval epsilon 1 {} epsilon 2 {} ".format(
eps_1, eps_2))
avg_reward, avg_Q = test(args, T, dqn_1, val_mem_1,
metrics, results_dir, env,
1) # Test
log('T = ' + str(T) + ' / ' + str(args.T_max) +
' | Avg. reward: ' + str(avg_reward) + ' | Avg. Q: ' +
str(avg_Q))
dqn_1.train(
) # Set DQN (online network) back to training mode
dqn_2.eval() # Set DQN (online network) to evaluation mode
avg_reward, avg_Q = test(args, T, dqn_2, val_mem_2,
metrics, results_dir, env,
2) # Test
log('T = ' + str(T) + ' / ' + str(args.T_max) +
' | Avg. reward: ' + str(avg_reward) + ' | Avg. Q: ' +
str(avg_Q))
dqn_2.train(
) # Set DQN (online network) back to training mode
# Update target network
if T % args.target_update == 0:
dqn_1.update_target_net()
dqn_2.update_target_net()
# checkpoint the network
if (args.checkpoint_interval !=
0) and (T % args.checkpoint_interval == 0):
#print("save memory", args.memory)
#save_memory(mem, args.memory, args.disable_bzip_memory)
print("epsilon 1: ", eps_1)
print("epsilon 2: ", eps_2)
print("Save model at ", results_dir)
dqn_1.save(results_dir, '{}-checkpoint.pth'.format(T))
dqn_2.save(results_dir, '{}-2-checkpoint.pth'.format(T))
except KeyboardInterrupt:
print("Keybaord error")
finally:
print("save state....")
print("Save model at ", results_dir)
dqn_1.save(results_dir, '{}-checkpoint.pth'.format(T))
dqn_2.save(results_dir, '{}-2-checkpoint.pth'.format(T))
args.memory = results_dir + "/val_mem_1/memory.pkl"
print("save memory 1 ...", args.memory)
save_memory(mem_1, args.memory, args.disable_bzip_memory)
args.memory = results_dir + "/val_mem_2/memory.pkl"
print("save memory 2 ...", args.memory)
save_memory(mem_2, args.memory, args.disable_bzip_memory)
print("Save model at ", results_dir)
dqn_1.save(results_dir, '{}-checkpoint.pth'.format(T))
dqn_2.save(results_dir, '{}-2-checkpoint.pth'.format(T))
print("... done Saving State")
sys.exit()
