def launch_after_training(params, net_state_dict, device, episodes, opt_steps):
env = Assault(23)
net = DQN(env.state_sz,
env.action_sz,
"vae",
params["image_input"] == "True",
device=device).to(device)
net.load_state_dict(net_state_dict)
controller = FixedController(
lambda state, explore: net(state.to(device)).max(1)[1].item())
agent = Agent(env, controller)
plot_name = "AfterTraining"
log().add_plot(plot_name,
columns=("train_episode", "train_steps", "reward"))
pbar = tqdm(range(episodes))
total_steps = 0
for episode in pbar:
pbar.set_description("Episode [{}/{}] Step[{}/{}] Exploit".format(
episode + 1, episodes, total_steps, opt_steps))
reward, steps = agent.rollout(train=False)
total_steps += steps
log().add_plot_point(plot_name, (episode, total_steps, reward))
if total_steps >= opt_steps:
break
log().save_logs()
def __init__(self, id, env):
Agent.__init__(self, id, env)
self._energy = Parameters.initEnergy
self._reproductionMinEnergy = Parameters.reproductionMinEnergy
self._genotype = []
self._rand = Random()
getattr(self, "_initialization"+Parameters.initialization)()
self._fitness=None
self._updated=None
self._recalculateFitness()
self._fitnessCalls = 0
def __init__(self, id, env):
Agent.__init__(self, id, env)
self._energy = Parameters.initEnergy
self._reproductionMinEnergy = Parameters.reproductionMinEnergy
self._genotype = []
self._rand = Random()
rand = self._rand
for i in xrange(Parameters.genotypeLength):
if rand.randint(0, 100) > 0:
self._genotype.append(Parameters.cubeSize * rand.random())
else:
self._genotype.append(-1 * Parameters.cubeSize * rand.random())
def train(experiment):
env = experiment.env(random_state=experiment.random_seed)
memory = ReplayMemory(experiment.hyperparams.memory_config.memory_size)
controller = ControllerDQN(env=env,
memory=memory,
params=experiment.hyperparams,
prune_percent=experiment.prune_percent,
pruner=experiment.pruner,
stop_criterion=experiment.stop_criterion,
device=experiment.device)
agent = Agent(env, controller)
EXPLORE_ITERS = 1
EXPLOIT_ITERS = 1
episodes, prune_iters, opt_steps = experiment.episodes, experiment.prune_iters, experiment.opt_steps
for iter in range(prune_iters):
pbar = tqdm(range(episodes))
cur_percent = (1 - experiment.prune_percent / 100)**iter
explore_plot = "Explore_iter" + str(iter) + "_prune" + str(cur_percent)
exploit_plot = "Exploit_iter" + str(iter) + "_prune" + str(cur_percent)
log().add_plot(explore_plot,
columns=("train_episode", "train_steps", "reward"))
log().add_plot(exploit_plot,
columns=("train_episode", "train_steps", "reward"))
for episode in pbar:
# once in EXPLORE_ITERS train rollouts, do EXPLOIT_ITERS exploit rollouts
if episode % EXPLORE_ITERS == EXPLORE_ITERS - 1:
for _ in range(EXPLOIT_ITERS):
pbar.set_description(
"Iter[{}/{}] Episode [{}/{}] Step[{}/{}] Exploit".
format(iter + 1, prune_iters, episode + 1, episodes,
controller.steps_done, opt_steps))
exploit(agent, episode, exploit_plot)
pbar.set_description(
"Iter[{}/{}] Episode [{}/{}] Step[{}/{}] Explore".format(
iter + 1, prune_iters, episode + 1, episodes,
controller.steps_done, opt_steps))
explore(agent, episode, explore_plot)
if controller.steps_done >= opt_steps:
break
if controller.optimization_completed(
) and not iter + 1 == prune_iters: # no stop on last iteration
break
torch.cuda.empty_cache()
log().save_logs()
log().save_model(controller.get_state(),
"model:iter{}:{}".format(iter, cur_percent))
controller.prune()
controller.reinit()
agent = Agent(name="agent_" +
str(np.random.randint(low=1000000, high=9999999)),
actionScaling=1.0,
policyNetworkSize=[256, 256],
qNetworkSize=[256, 256],
numQuantiles=16,
policyNetworkLearningRate=3e-4,
qNetworkLearningRate=3e-4,
entropyCoefficient="auto",
tau=0.005,
gamma=0.99,
kappa=1.0,
maxMemoryLength=int(1e5),
priorityExponent=0.0,
batchSize=64,
nStep=3,
frameSkip=2,
maxEpisodes=4096,
trainSteps=1024,
maxTrainSteps=6000000,
minStepsBeforeTraining=10000,
rewardScaling=(10.0**-0.75),
actionShift=0.0,
stepsPerUpdate=1,
render=False,
showGraphs=True,
saveModel=True,
saveModelToS3=False,
restoreModel=False,
train=True,
testSteps=1024,
maxMinutes=360,
targetEntropy=-4.0,
maxGradientNorm=5.0,
meanRegularizationConstant=0.0,
varianceRegularizationConstant=0.0,
randomStartSteps=10000,
gradientSteps=1,
initialExtraNoise=0,
extraNoiseDecay=0,
evaluationEvery=25,
numFinalEvaluations=10)
def optimize_joint(system_nn, pol_nn, log_writer, **kwargs):
# unpack kwargs
horizon = kwargs.get("horizon")
nb_iterations = kwargs.get("nb_iterations")
batch_size = kwargs.get("batch_size")
policy_fit = kwargs.get("policy", False)
system_fit = kwargs.get("system", False)
mc_samples = kwargs.get("mc_samples", 128)
env = Environment(system_nn)
agent = Agent(pol_nn, env, horizon)
# Optimizers
parameters_list = []
if policy_fit:
parameters_list = parameters_list + list(pol_nn.parameters())
if system_fit:
parameters_list = parameters_list + list(system_nn.parameters())
if parameters_list:
lr = kwargs.get("learning_rate", .001)
optimizer = Adam(parameters_list, lr=lr)
for it in range(nb_iterations):
loss = {}
params = {}
# set gradient to zero
optimizer.zero_grad()
# generate the batch
_, states_batch, dist_batch, _, oha_batch, rew_batch = agent.sample_trajectory(
batch_size)
# Loss #
system_loss = system_error(system_nn, pol_nn, states_batch,
dist_batch, oha_batch, rew_batch)
system_loss.backward(retain_graph=policy_fit)
optimizer.step()
system_nn.project_parameters()
pol_nn.project_parameters()
if system_fit and log_writer is not None:
params['system'] = system_nn.unwrapped.named_parameters()
log_writer.add_system_parameters(system_nn.parameters_dict(),
step=it)
if policy_fit and log_writer is not None:
params['policy'] = pol_nn.named_parameters()
actions = pol_nn(
states_batch) # (B, H, A), need to stack along the B dim
log_writer.add_policy_histograms(actions.view(
-1, actions.shape[2]),
step=it)
if log_writer is not None:
loss['loss'] = system_loss.item()
log_writer.add_grad_histograms(params, step=it)
log_writer.add_loss(loss, step=it)
# performance of the agent on the epoch
ep_perf, return_estimate = agent.avg_performance(mc_samples)
log_writer.add_expected_return(ep_perf, step=it)
log_writer.add_return(return_estimate, step=it)
return env, agent
import torch
import torch.optim as optim
from agent.Agent import Agent
from environment.Environment import Environment
from policy.Policy import Policy
from utils.utils import plot_training_evolution
learning_rate = 0.01
discount_factor = 0.99
episodes = 5000
env = Environment('LunarLander-v2')
policy: Policy = Policy(env.observation_space(), env.action_space())
optimizer = optim.Adam(policy.parameters(), lr=learning_rate)
agent = Agent(env, policy, optimizer)
for episode in range(episodes):
agent.run_episode()
agent.update_policy(discount_factor=discount_factor)
if episode % 50 == 0:
print('Episode {}\tAverage reward: {}'.format(
episode,
np.array(policy.reward_history[-50:]).mean()))
if env.is_solved(np.array(policy.reward_history[-50:]).mean()):
break
torch.save(policy.state_dict(), 'saved_policy/policy.pt')
def __init__(self, env, childrenEnv=None):
addr = AddressManager.getAddress(None)
Agent.__init__(self, addr, env, childrenEnv)
self._reprCount = 0
self._rand = Random()
agent = Agent(name="agent_2943367",
actionScaling=1.0,
policyNetworkSize=[256, 256],
qNetworkSize=[256, 256],
policyNetworkLearningRate=3e-4,
qNetworkLearningRate=3e-4,
entropyCoefficient="auto",
tau=0.005,
gamma=0.99,
maxMemoryLength=int(5e6),
priorityExponent=0.0,
batchSize=256,
maxEpisodes=0,
trainSteps=1024,
minStepsBeforeTraining=4096,
rewardScaling=rewardScaling,
actionShift=0.0,
stepsPerUpdate=1,
render=True,
showGraphs=False,
saveModel=False,
restoreModel=True,
train=False,
testSteps=1024,
maxMinutes=60,
targetEntropy=-4.0,
maxGradientNorm=5.0,
meanRegularizationConstant=0.0,
varianceRegularizationConstant=0.0,
randomStartSteps=0,
gradientSteps=1,
initialExtraNoise=0,
extraNoiseDecay=0,
evaluationEvery=250,
numFinalEvaluations=10)
result = -20000
try:
agent = Agent(name="agent_" +
str(np.random.randint(low=1000000, high=9999999)),
policyNetworkSize=[256, 256],
qNetworkSize=[256, 256],
valueNetworkSize=[256, 256],
entropyCoefficient=entropyCoefficient,
valueNetworkLearningRate=learningRate,
policyNetworkLearningRate=learningRate,
qNetworkLearningRate=learningRate,
tau=0.005,
gamma=0.99,
maxMemoryLength=int(1e6),
priorityExponent=0,
batchSize=256,
maxGradientNorm=5,
maxEpisodes=1024,
trainSteps=1024,
minStepsBeforeTraining=4096,
rewardScaling=rewardScaling,
actionScaling=actionScaling,
actionShift=0.0,
stepsPerUpdate=1,
render=True,
showGraphs=True,
meanRegularizationConstant=weightRegularizationConstant,
varianceRegularizationConstant=weightRegularizationConstant,
testSteps=1024,
maxMinutes=600)
result = agent.execute()
def __init__(self, name, owner):
self.agent = Agent(name, owner)
self.name = name
self.owner = owner
parser.add_argument('--mode', type=str, help='Mode', default="train", choices=["train", "infer"])
args = parser.parse_args()
ray.init()
config = {}
with open(args.agent_config_path) as config_file:
config.update(json.load(config_file))
config["explore"] = True if args.mode == "train" else False
with open(args.environment_config_path) as config_file:
config.update(json.load(config_file))
env = Environment(config)
agent = Agent(config)
if args.mode == "train":
config.update({
"num_gpus": 0,
"num_workers": 1,
"monitor": False
})
tune.run(
agent.__class__,
name=env.__class__.__name__+"_"+agent.__class__.__name__,
stop={"episode_reward_mean": -100},
import queue
import time
import datetime
from data.DataSettings import DataSettings
from data_streamer.DataStreamer import DataStreamer
from data_logger.DataLogger import DataLogger
from agent.Agent import Agent
from sensors.PowerButton import PowerButton
from sensors.DataButton import DataButton
from sensors.RelaySensor import RelaySensor
data_settings = DataSettings()
data_streamer = DataStreamer()
data_logger = DataLogger(data_streamer)
relay = RelaySensor()
agent = Agent(data_settings, relay)
power_button_pin, data_button_pin = 18, 19 # Grove Base Hat connector D18
power_button = PowerButton(power_button_pin)
data_button = DataButton(data_button_pin)
command_queue = queue.Queue()
def setup():
data_settings.load_settings()
agent.update()
t1 = threading.Thread(target=read_incoming_serial)
t1.start()
def read_incoming_serial():
previous_serial_data = ""