def main(args):
hidden_units = args.hidden_units
msg_dim = args.msg_dim
model_path = os.getcwd() + "/" + args.model_dir
ray.init(log_to_driver=False)
env_test_instance = gym.make('BipedalWalker-v3')
if args.baseline:
from smp.baseline import TD3Net
action_dimension = copy(env_test_instance.action_space.shape[0])
else:
from smp.smp import TD3Net
action_dimension = 1
model_kwargs = {
# action dimension for modular actions
'action_dimension': action_dimension,
'min_action': copy(env_test_instance.action_space.low)[0],
'max_action': copy(env_test_instance.action_space.high)[0],
'msg_dimension': msg_dim,
'fix_sigma': True,
'hidden_units': hidden_units
}
del env_test_instance
manager = SampleManager(TD3Net,
'BipedalWalker-v3',
num_parallel=(os.cpu_count() - 1),
total_steps=150,
action_sampling_type="continuous_normal_diagonal",
is_tf=True,
model_kwargs=model_kwargs)
manager.load_model(model_path)
manager.test(200, test_episodes=5, render=True)
ray.shutdown()
optimizer.apply_gradients(
zip(gradients, agent.model.trainable_variables))
actor_losses.append(actor_loss)
critic_losses.append(critic_loss)
losses.append(total_loss)
# Set new weights
manager.set_agent(agent.get_weights())
print('TEST')
# Update aggregator
steps, current_rewards = manager.test(
max_steps=1000,
test_episodes=10,
render=False,
evaluation_measure="time_and_reward",
)
#if (e+1) % 5 == 0:
manager.test(max_steps=1000, test_episodes=1, render=True)
manager.update_aggregator(loss=losses,
reward=current_rewards,
time=steps)
# Collect all rewards
rewards.extend(current_rewards)
# Average reward over last 100 episodes
avg_reward = sum(rewards[-100:]) / min(len(rewards), 100)
# Print progress
saving_path = os.getcwd() + "/progress_cartpole"
# keys for replay buffer -> what you will need for optimization
optim_keys = ["state", "action", "reward", "state_new", "not_done"]
# initialize buffer
manager.initilize_buffer(buffer_size, optim_keys)
# initilize progress aggregator
manager.initialize_aggregator(path=saving_path,
saving_after=5,
aggregator_keys=["loss", "time_steps"])
# initial testing:
print("test before training: ")
manager.test(test_steps, test_episodes=10, do_print=True, render=True)
# get initial agent
agent = manager.get_agent()
for e in range(epochs):
# training core
# experience replay
print("collecting experience..")
data = manager.get_data(total_steps=100)
manager.store_in_buffer(data)
# sample data to optimize on from buffer
sample_dict = manager.sample(sample_size)
optim_keys = ["state", "action", "reward", "state_new", "not_done"]
# initialize buffer
manager.initilize_buffer(buffer_size, optim_keys)
# initilize progress aggregator
manager.initialize_aggregator(
path=saving_path, saving_after=5, aggregator_keys=["loss", "time_steps"]
)
# this was given
print("test before training: ")
manager.test(
max_steps=50,
test_episodes=5,
#render=True,
#do_print=True,
evaluation_measure="time_and_reward",
)
# get initial agent
# agent = manager.get_agent()
losses = []
for epoch in range(epochs):
# training core
print("collecting experience..")
data = manager.get_data()
manager.store_in_buffer(data)
# sample data to optimize on
sample_dict = manager.sample(sample_size)
# Initialize
ray.init(log_to_driver=False)
manager = SampleManager(**kwargs)
# Where to load your results from
loading_path = os.getcwd() + "/progress_CartPole"
# Load model
manager.load_model(loading_path)
print("done")
print("testing optimized agent")
manager.test(
1000,
test_episodes=10,
render=True,
do_print=True,
evaluation_measure="time_and_reward",
)
print('Prepare LunarLander')
env = gym.make("LunarLander-v2")
model_kwargs = {"layers": [32, 32, 32], "num_actions": env.action_space.n}
kwargs = {
"model": QNet,
"environment": "LunarLander-v2",
"num_parallel": 1,
"total_steps": 1000,
"model_kwargs": model_kwargs,
with tf.GradientTape() as tape:
# positive critic loss for gradient descent with MSE
critic_loss = tf.reduce_mean((mc - agent.v(state))**2)
critic_gradients = tape.gradient(
critic_loss, agent.model.critic.trainable_variables)
optimizer.apply_gradients(
zip(critic_gradients, agent.model.critic.trainable_variables))
total_loss += actor_loss + critic_loss
# Update the agent
manager.set_agent(agent.get_weights())
agent = manager.get_agent()
reward = manager.test(test_steps, evaluation_measure="reward")
manager.update_aggregator(loss=total_loss, reward=reward)
# print progress
print(
f"epoch ::: {e} loss ::: {total_loss} avg env steps ::: {np.mean(reward)}"
)
if e % saving_after == 0:
# you can save models
manager.save_model(saving_path, e)
# and load mmodels
manager.load_model(saving_path)
print("done")
print("testing optimized agent")
manager.test(test_steps, test_episodes=10, render=True)
# keys needed for tabular q
optim_keys = ['state', 'action', 'reward', 'state_new', 'not_done']
# initialize buffer
manager.initialize_buffer(buffer_size)
agent = manager.get_agent()
# initialize progress aggregator
manager.initialize_aggregator(path=saving_path,
saving_after=saving_after,
aggregator_keys=['time_steps', 'rewards'])
# initial testing:
print('test before training: ')
manager.test(test_steps,
test_episodes=10,
do_print=True,
evaluation_measure='time_and_reward')
for e in range(epochs):
print('collecting experience..')
data = manager.get_data()
manager.store_in_buffer(data)
# sample data to optimize on from buffer
experience_dict = manager.sample(sample_size)
print('optimizing...')
for states, actions, rewards, states_new, not_dones in zip(
*[experience_dict[k] for k in optim_keys]):
train_step(agent.model, states, actions, rewards, states_new,
def train_td3(args, model, action_dimension=None):
print(args)
tf.keras.backend.set_floatx('float32')
ray.init(log_to_driver=False)
# hyper parameters
buffer_size = args.buffer_size # 10e6 in their repo, not possible with our ram
epochs = args.epochs
saving_path = os.getcwd() + "/" + args.saving_dir
saving_after = 5
sample_size = args.sample_size
optim_batch_size = args.batch_size
gamma = args.gamma
test_steps = 100 # 1000 in their repo
policy_delay = 2
rho = .046
policy_noise = args.policy_noise
policy_noise_clip = .5
msg_dim = args.msg_dim # 32 in their repo
learning_rate = args.learning_rate
save_args(args, saving_path)
env_test_instance = gym.make('BipedalWalker-v3')
if action_dimension is None:
action_dimension = copy(env_test_instance.action_space.shape[0])
model_kwargs = {
# action dimension for modular actions
'action_dimension': action_dimension,
'min_action': copy(env_test_instance.action_space.low)[0],
'max_action': copy(env_test_instance.action_space.high)[0],
'msg_dimension': msg_dim,
'fix_sigma': True,
'hidden_units': args.hidden_units
}
del env_test_instance
manager = SampleManager(model,
'BipedalWalker-v3',
num_parallel=(os.cpu_count() - 1),
total_steps=150,
action_sampling_type="continuous_normal_diagonal",
is_tf=True,
model_kwargs=model_kwargs)
optim_keys = [
'state',
'action',
'reward',
'state_new',
'not_done',
]
manager.initialize_buffer(buffer_size, optim_keys)
manager.initialize_aggregator(path=saving_path,
saving_after=saving_after,
aggregator_keys=["loss", "reward"])
agent = manager.get_agent()
optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)
# fill buffer
print("Filling buffer before training..")
while len(manager.buffer.buffer[
manager.buffer.keys[0]]) < manager.buffer.size:
# Gives you state action reward trajectories
data = manager.get_data()
manager.store_in_buffer(data)
# track time while training
timer = time.time()
last_t = timer
target_agent = manager.get_agent()
for e in range(epochs):
# off policy
sample_dict = manager.sample(sample_size, from_buffer=True)
print(f"collected data for: {sample_dict.keys()}")
# cast values to float32 and create data dict
sample_dict['state'] = tf.cast(sample_dict['state'], tf.float32)
sample_dict['action'] = tf.cast(sample_dict['action'], tf.float32)
sample_dict['reward'] = tf.cast(sample_dict['reward'], tf.float32)
sample_dict['state_new'] = tf.cast(sample_dict['state_new'],
tf.float32)
sample_dict['not_done'] = tf.cast(sample_dict['not_done'], tf.float32)
data_dict = dict_to_dict_of_datasets(sample_dict,
batch_size=optim_batch_size)
total_loss = 0
for state, action, reward, state_new, not_done in \
zip(data_dict['state'],
data_dict['action'],
data_dict['reward'],
data_dict['state_new'],
data_dict['not_done']):
action_new = target_agent.act(state_new)
# add noise to action_new
action_new = action_new + tf.clip_by_value(
tf.random.normal(action_new.shape, 0., policy_noise),
-policy_noise_clip, policy_noise_clip)
# clip action_new to action space
action_new = tf.clip_by_value(
action_new, manager.env_instance.action_space.low,
manager.env_instance.action_space.high)
# calculate target with double-Q-learning
state_action_new = tf.concat([state_new, action_new], axis=-1)
q_values0 = target_agent.model.critic0(state_action_new)
q_values1 = target_agent.model.critic1(state_action_new)
q_values = tf.concat([q_values0, q_values1], axis=-1)
q_targets = tf.squeeze(tf.reduce_min(q_values, axis=-1))
critic_target = reward + gamma * not_done * q_targets
state_action = tf.concat([state, action], axis=-1)
# update critic 0
with tf.GradientTape() as tape:
q_output = agent.model.critic0(state_action)
loss = tf.keras.losses.MSE(tf.squeeze(critic_target),
tf.squeeze(q_output))
total_loss += loss
gradients = tape.gradient(loss,
agent.model.critic0.trainable_variables)
optimizer.apply_gradients(
zip(gradients, agent.model.critic0.trainable_variables))
# update critic 1
with tf.GradientTape() as tape:
q_output = agent.model.critic1(state_action)
loss = tf.keras.losses.MSE(tf.squeeze(critic_target),
tf.squeeze(q_output))
total_loss += loss
gradients = tape.gradient(loss,
agent.model.critic1.trainable_variables)
optimizer.apply_gradients(
zip(gradients, agent.model.critic1.trainable_variables))
# update actor with delayed policy update
if e % policy_delay == 0:
with tf.GradientTape() as tape:
actor_output = agent.model.actor(state)
action = reparam_action(actor_output,
agent.model.action_dimension,
agent.model.min_action,
agent.model.max_action)
state_action = tf.concat([state, action], axis=-1)
q_val = agent.model.critic0(state_action)
actor_loss = -tf.reduce_mean(q_val)
total_loss += actor_loss
actor_gradients = tape.gradient(
actor_loss, agent.model.actor.trainable_variables)
optimizer.apply_gradients(
zip(actor_gradients,
agent.model.actor.trainable_variables))
# Update agent
manager.set_agent(agent.get_weights())
agent = manager.get_agent()
if e % policy_delay == 0:
# Polyak averaging
new_weights = list(rho * np.array(target_agent.get_weights()) +
(1. - rho) * np.array(agent.get_weights()))
target_agent.set_weights(new_weights)
reward = manager.test(test_steps, evaluation_measure="reward")
manager.update_aggregator(loss=total_loss, reward=reward)
print(
f"epoch ::: {e} loss ::: {total_loss} avg reward ::: {np.mean(reward)}"
)
if e % saving_after == 0:
manager.save_model(saving_path, e)
# needed time and remaining time estimation
current_t = time.time()
time_needed = (current_t - last_t) / 60.
time_remaining = (current_t - timer) / 60. / (e + 1) * (epochs -
(e + 1))
print(
'Finished epoch %d of %d. Needed %1.f min for this epoch. Estimated time remaining: %.1f min'
% (e + 1, epochs, time_needed, time_remaining))
last_t = current_t
manager.load_model(saving_path)
print("done")
print("testing optimized agent")
manager.test(test_steps, test_episodes=10, render=True)
ray.shutdown()
alpha = 0.1
gamma = 0.95
optim_keys = ["state", "action", "reward", "state_new", "not_done"]
# initialize buffer
manager.initilize_buffer(buffer_size, optim_keys)
# initialize progress aggregator
manager.initialize_aggregator(path=saving_path,
saving_after=5,
aggregator_keys=["time_steps"])
print("Testing before training... ")
manager.test(max_steps=10, test_episodes=3, render=False, do_print=True)
# get initial agent
agent = manager.get_agent()
for e in range(epochs):
# experience replay
print("Collecting experience...")
data = manager.get_data(total_steps=1000)
manager.store_in_buffer(data)
# sample data to optimize on from buffer
sample_dict = manager.sample(sample_size)
print("Optimizing...")
training = True
optim_keys = [
'state', 'action', 'reward', 'state_new', 'not_done', 'value_estimate'
]
# initialize aggregator
aggregator_keys = ['actor_loss', 'time_steps', 'reward']
manager.initialize_aggregator(path=saving_path,
saving_after=10,
aggregator_keys=aggregator_keys)
# test before training
print('Test before training:')
manager.test(test_steps,
render=False,
do_print=True,
evaluation_measure="time_and_reward")
# get initial agent
agent = manager.get_agent()
# initialize the optimizer
adam = tf.keras.optimizers.Adam(learning_rate=0.0001)
# initialize mse loss for training critic
mse = tf.keras.losses.MeanSquaredError()
if training:
# main trainig loop
for e in range(epochs):
print("training epoch: ", e)
optim_keys = ["state", "action", "reward", "state_new", "not_done"]
# initialize buffer
manager.initilize_buffer(buffer_size, optim_keys)
# initilize progress aggregator
manager.initialize_aggregator(path=saving_path,
saving_after=5,
aggregator_keys=["loss", "time_steps"])
# initial testing:
print("test before training: ")
manager.test(test_steps,
test_episodes=10,
do_print=True,
render=True,
evaluation_measure="time_and_reward")
# get initial agent
agent = manager.get_agent()
for e in range(epochs):
# training core
# experience replay off (-> on policy)
print("collecting experience..")
data = manager.get_data()
manager.store_in_buffer(data)
# sample data to optimize on from buffer
# use samples from buffer to compute q_target and fit model
q_target = sample_dict["reward"] + GAMMA * agent.max_q(
sample_dict["state_new"])
sample_dict["q_target"] = q_target
agent.model.train(sample_dict, optim_batch_size)
# Get last lost
loss = agent.model.history.history['loss']
# Get optimized weights and update agent
new_weights = agent.model.get_weights()
manager.set_agent(new_weights)
agent = manager.get_agent()
# update aggregator
time_steps = manager.test(test_steps)
manager.update_aggregator(loss=loss, time_steps=time_steps)
print(
f"epoch ::: {e} loss ::: {np.mean([np.mean(l) for l in loss])} avg env steps ::: {np.mean(time_steps)}"
)
# update epsilon
new_epsilon = EPSILONSTART * (1 - EPSILONDECAYRATE)**(
(e - epsilon_decay_started_at_e) + 1)
if (new_epsilon < EPSILONMIN):
new_epsilon = EPSILONMIN
print("Updating epsilon to " + str(new_epsilon))
manager.set_epsilon(epsilon=new_epsilon)
if e % saving_after == 0:
manager.save_model(saving_path, e)
saving_after = 5
test_steps = 1000
optim_keys = ["state", "action", "reward", "state_new", "not_done"]
manager.initilize_buffer(buffer_size, optim_keys)
manager.initialize_aggregator(path=saving_path,
saving_after=5,
aggregator_keys=["loss", "time_steps"])
print("test before training: ")
manager.test(
max_steps=12, #100
test_episodes=1, #10,
render=True,
do_print=True,
evaluation_measure="time_and_reward",
)
agent = manager.get_agent()
epochs = 20
for e in range(epochs):
print("collecting experience..")
data = manager.get_data()
manager.store_in_buffer(data)
sample_dict = manager.sample(sample_size)
print(f"collected data for: {sample_dict.keys()}")
print("optimizing...")
saving_path = Path(os.getcwd() + SAVING_DIRECTORY)
saving_path.mkdir(parents=True, exist_ok=True)
saving_path_model = Path(os.getcwd() + SAVING_DIRECTORY + '/model')
saving_path_model.mkdir(parents=True, exist_ok=True)
# initialize manager
optim_keys = ['state', 'action', 'reward', 'state_new', 'not_done']
manager.initilize_buffer(BUFFER_SIZE, optim_keys)
aggregator_keys=['loss', 'time_steps', 'reward']
manager.initialize_aggregator(saving_path, 5, aggregator_keys)
# initialize the optimizers
optimizer = Adam(learning_rate=LEARNING_RATE)
print('# =============== INITIAL TESTING =============== #')
manager.test(MAX_TEST_STEPS, 5, evaluation_measure='time_and_reward', do_print=True, render=True)
# get the initial agent
agent = manager.get_agent()
print('# =============== START TRAINING ================ #')
for e in range(1, EPOCHS+1):
print(f'# ============== EPOCH {e}/{EPOCHS} ============== #')
print('# ============= collecting samples ============== #')
# collect experience and save it to ERP buffer
data = manager.get_data(do_print=False)
manager.store_in_buffer(data)
# get some samples from the ERP buffer and create a dataset
sample_dict = manager.sample(sample_size=SAMPLE_SIZE)
data_dict = dict_to_dict_of_datasets(sample_dict, batch_size=BATCH_SIZE)