def deb_archive_with_cartridge(cartridge_cmd, tmpdir, project_with_cartridge):
project = project_with_cartridge
pre_install_filepath = os.path.join(tmpdir, "pre.sh")
with open(pre_install_filepath, "w") as f:
f.write("""
/bin/sh -c 'touch $HOME/hello-bin-sh.txt'
/bin/touch $HOME/hello-absolute.txt
""")
post_install_filepath = os.path.join(tmpdir, "post.sh")
with open(post_install_filepath, "w") as f:
f.write("""
/bin/sh -c 'touch $HOME/bye-bin-sh.txt'
/bin/touch $HOME/bye-absolute.txt
""")
deps_filepath = os.path.join(tmpdir, "deps.txt")
with open(deps_filepath, "w") as f:
f.write("unzip>1,<=7\n" +
"stress\n" +
"neofetch < 25")
net_msg_max = 1024
user_param = 'user_data'
systemd_unit_params = os.path.join(tmpdir, "systemd-unit-params.yml")
with open(systemd_unit_params, "w") as f:
yaml.dump({
"instance-env": {"net-msg-max": net_msg_max, "user-param": user_param}
}, f)
replace_project_file(project, 'init.lua', INIT_CHECK_PASSED_PARAMS)
replace_project_file(project, 'stateboard.init.lua', INIT_CHECK_PASSED_PARAMS)
cmd = [
cartridge_cmd,
"pack", "deb",
"--deps-file", deps_filepath,
"--preinst", pre_install_filepath,
"--postinst", post_install_filepath,
"--unit-params-file", systemd_unit_params,
project.path,
"--use-docker",
]
process = subprocess.run(cmd, cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of deb archive with project"
filepath = find_archive(tmpdir, project.name, 'deb')
assert filepath is not None, "DEB archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def tgz_archive_with_cartridge(cartridge_cmd, tmpdir,
original_project_with_cartridge, request):
project = original_project_with_cartridge
cmd = [cartridge_cmd, "pack", "tgz", "--use-docker", project.path]
process = subprocess.run(cmd, cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of tgz archive with project"
filepath = find_archive(tmpdir, project.name, 'tar.gz')
assert filepath is not None, "TGZ archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def tgz_archive(cartridge_cmd, tmpdir, light_project, request):
project = light_project
cmd = [cartridge_cmd, "pack", "tgz", project.path]
if request.param == 'docker':
cmd.append('--use-docker')
process = subprocess.run(cmd, cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of tgz archive with project"
filepath = find_archive(tmpdir, project.name, 'tar.gz')
assert filepath is not None, "TGZ archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def deb_archive(tmpdir, light_project, request):
project = light_project
cmd = [os.path.join(basepath, "cartridge"), "pack", "deb", project.path]
if request.param == 'docker':
cmd.append('--use-docker')
process = subprocess.run(cmd, cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of deb archive with project"
filepath = find_archive(tmpdir, project.name, 'deb')
assert filepath is not None, "DEB archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def deb_archive_with_cartridge(cartridge_cmd, tmpdir,
original_project_with_cartridge, request):
project = original_project_with_cartridge
cmd = [cartridge_cmd, "pack", "deb", project.path]
if platform.system() == 'Darwin':
cmd.append("--use-docker")
process = subprocess.run(cmd, cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of deb archive with project"
filepath = find_archive(tmpdir, project.name, 'deb')
assert filepath is not None, "DEB archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def deb_archive(cartridge_cmd, tmpdir, light_project, request):
project = light_project
cmd = [cartridge_cmd, "pack", "deb", project.path]
if request.param == 'docker':
if project.deprecated_flow_is_used:
pytest.skip()
cmd.append('--use-docker')
process = subprocess.run(cmd, cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of deb archive with project"
filepath = find_archive(tmpdir, project.name, 'deb')
assert filepath is not None, "DEB archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def rpm_archive(tmpdir, light_project, request):
project = light_project
cmd = [os.path.join(basepath, "cartridge"), "pack", "rpm", project.path]
if request.param == 'docker':
if project.deprecated_flow_is_used:
pytest.skip()
cmd.append('--use-docker')
process = subprocess.run(cmd, cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of rpm archive with project"
filepath = find_archive(tmpdir, project.name, 'rpm')
assert filepath is not None, "RPM archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def rpm_archive_with_cartridge(cartridge_cmd, tmpdir, project_with_cartridge):
project = project_with_cartridge
pre_install_filepath = os.path.join(tmpdir, "pre.sh")
with open(pre_install_filepath, "w") as f:
f.write("""
/bin/sh -c 'touch $HOME/hello-bin-sh.txt'
/bin/touch $HOME/hello-absolute.txt
""")
post_install_filepath = os.path.join(tmpdir, "post.sh")
with open(post_install_filepath, "w") as f:
f.write("""
/bin/sh -c 'touch $HOME/bye-bin-sh.txt'
/bin/touch $HOME/bye-absolute.txt
""")
cmd = [
cartridge_cmd,
"pack",
"rpm",
"--deps",
"unzip>1,unzip<=7",
"--deps",
"wget",
"--deps",
"make>0.1.0",
"--preinst",
pre_install_filepath,
"--postinst",
post_install_filepath,
project.path,
"--use-docker",
]
process = subprocess.run(cmd, cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of rpm archive with project"
filepath = find_archive(tmpdir, project.name, 'rpm')
assert filepath is not None, "RPM archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def deb_archive(cartridge_cmd, tmpdir, light_project, request):
project = light_project
cmd = [cartridge_cmd, "pack", "deb", project.path]
if request.param == 'docker':
cmd.append('--use-docker')
rc, output = run_command_and_get_output(cmd, cwd=tmpdir)
if request.param == 'local' and platform.system() == 'Darwin':
assert rc == 1
assert "It's not possible to pack application into RPM or DEB on non-linux OS" in output
pytest.skip("Packing RPM and DEB locally should fail for Darwin")
assert rc == 0
filepath = find_archive(tmpdir, project.name, 'deb')
assert filepath is not None, "DEB archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def tgz_archive_with_cartridge(cartridge_cmd, tmpdir, project_with_cartridge):
project = project_with_cartridge
replace_project_file(project, 'app/roles/custom.lua',
ROUTER_WITH_EVAL_FILEPATH)
replace_project_file(project, 'init.lua',
INIT_ROLES_RELOAD_ALLOWED_FILEPATH)
cmd = [
cartridge_cmd,
"pack",
"tgz",
project.path,
"--use-docker",
]
process = subprocess.run(cmd, cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of tgz archive with project"
filepath = find_archive(tmpdir, project.name, 'tar.gz')
assert filepath is not None, "TGZ archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def rpm_archive_with_custom_units(cartridge_cmd, tmpdir, light_project):
project = light_project
unit_template = '''
[Unit]
Description=Tarantool service: ${app_name}
SIMPLE_UNIT_TEMPLATE
[Service]
Type=simple
ExecStart=${bindir}/tarantool ${app_dir}/init.lua
Environment=TARANTOOL_WORK_DIR=${workdir}
Environment=TARANTOOL_CONSOLE_SOCK=/var/run/tarantool/${app_name}.control
Environment=TARANTOOL_PID_FILE=/var/run/tarantool/${app_name}.pid
Environment=TARANTOOL_INSTANCE_NAME=${app_name}
[Install]
WantedBy=multi-user.target
Alias=${app_name}
'''
instantiated_unit_template = '''
[Unit]
Description=Tarantool service: ${app_name} %i
INSTANTIATED_UNIT_TEMPLATE
[Service]
Type=simple
ExecStartPre=mkdir -p ${workdir}.%i
ExecStart=${bindir}/tarantool ${app_dir}/init.lua
Environment=TARANTOOL_WORK_DIR=${workdir}.%i
Environment=TARANTOOL_CONSOLE_SOCK=/var/run/tarantool/${app_name}.%i.control
Environment=TARANTOOL_PID_FILE=/var/run/tarantool/${app_name}.%i.pid
Environment=TARANTOOL_INSTANCE_NAME=${app_name}@%i
[Install]
WantedBy=multi-user.target
Alias=${app_name}
'''
unit_template_filepath = os.path.join(tmpdir, "unit_template.tmpl")
with open(unit_template_filepath, 'w') as f:
f.write(unit_template)
inst_unit_template_filepath = os.path.join(
tmpdir, "instantiated_unit_template.tmpl")
with open(inst_unit_template_filepath, 'w') as f:
f.write(instantiated_unit_template)
process = subprocess.run([
cartridge_cmd, "pack", "rpm", "--unit-template", "unit_template.tmpl",
"--instantiated-unit-template", "instantiated_unit_template.tmpl",
project.path
],
cwd=tmpdir)
assert process.returncode == 0, \
"Error during creating of rpm archive with project"
filepath = find_archive(tmpdir, project.name, 'rpm')
assert filepath is not None, "RPM archive isn't found in work directory"
return Archive(filepath=filepath, project=project)
def __init__(self, args, device, env):
"""
Building the QDRL Algorithm
Args:
args: the arguments
device: the device
env: the environment
"""
self.device = device
self.args = args
self.env = env
print('err', args)
self.random_state = np.random.RandomState(args['--train_seed'])
action_dim = self.env.action_space.shape[0]
state_dim = self.env.observation_space.shape[0]
# We create the value and policy networks as well as their target
self.critic_net_diversity1, self.critic_net_diversity2, self.target_critic_net_diversity1, self.target_critic_net_diversity2, \
self.critic_net_quality1, self.critic_net_quality2, self.target_critic_net_quality1, self.target_critic_net_quality2 = [
ActionValueNetwork(state_dim, action_dim, args['--hidden']).to(device)
for _ in range(8)
]
self.selected_actors = []
# TODO check the copy system of selected actors
for _ in range(args['--population_size']):
actor = {
'net':
PolicyNetwork(state_dim,
action_dim,
args['--hidden'],
init_w=0.1).to(device)
}
# We create the optimizers
actor['optimizer'] = torch.optim.Adam(actor['net'].parameters(),
lr=args['--policy_lr'])
self.selected_actors.append(actor)
self.critic_quality1_optimizer = torch.optim.Adam(
self.critic_net_quality1.parameters(), lr=args['--value_lr'])
self.critic_quality2_optimizer = torch.optim.Adam(
self.critic_net_quality2.parameters(), lr=args['--value_lr'])
self.critic_diversity1_optimizer = torch.optim.Adam(
self.critic_net_diversity1.parameters(), lr=args['--value_lr'])
self.critic_diversity2_optimizer = torch.optim.Adam(
self.critic_net_diversity2.parameters(), lr=args['--value_lr'])
# We initialize the target models to be identical to the other models
soft_update(self.critic_net_quality1,
self.target_critic_net_quality1,
soft_tau=1.)
soft_update(self.critic_net_quality2,
self.target_critic_net_quality2,
soft_tau=1.)
soft_update(self.critic_net_diversity1,
self.target_critic_net_diversity1,
soft_tau=1.)
soft_update(self.critic_net_diversity2,
self.target_critic_net_diversity2,
soft_tau=1.)
# We create the replay buffer
self.replay_buffer = QDRLReplayBuffer(args['--buffer'])
# We create the criterion
# TODO: figure out why not two TD3
self.qdrl_criterion = QDRLCriterion(self.critic_net_diversity1,
self.critic_net_diversity2,
self.target_critic_net_diversity1,
self.target_critic_net_diversity2,
self.critic_net_quality1,
self.critic_net_quality2,
self.target_critic_net_quality1,
self.target_critic_net_quality2,
gamma=args['--gamma'],
soft_tau=args['--soft_tau'],
noise_std=args['--g_smooth_sigma'],
noise_clip=args['--g_smooth_clip'],
device=device)
# We prepare the experiment
exp_options = {
'critic_loss_diversity_1': {
'plot': 'line',
'yscale': 'log'
},
'critic_loss_diversity_2': {
'plot': 'line',
'yscale': 'log'
},
'critic_loss_quality_1': {
'plot': 'line',
'yscale': 'log'
},
'critic_loss_quality_2': {
'plot': 'line',
'yscale': 'log'
},
'score': {
'plot': 'line',
'yscale': 'linear'
},
'return_test': {
'plot': 'line',
'yscale': 'linear'
},
'return_test_sparse': {
'plot': 'line',
'yscale': 'linear'
},
'actor_loss': {
'plot': 'line',
'yscale': 'linear'
},
}
agent_id = 0
description = 'TD3: {} with {} frames for training'.format(
args['--env_name'], args['--budget'])
self.exp_id = create_experiment(args['--exp'], description, './',
exp_options)
my_logger.info('exp_id: {}'.format(self.exp_id))
# QDRL init
self.archive = Archive()
knn = KNN(k=args['--k_knn'], distance_func=distance_test)
self.novelty_score_evaluator = NoveltyScore(knn=knn)
self.storage_critic = ExpLoggerAgent(
self.exp_id, agent_id,
os.path.join(args['--exp'], 'agent_' + str(agent_id)), {
'critic_model_quality1': self.critic_net_quality1,
'critic_model_quality2': self.critic_net_quality2,
'critic_model_diversity1': self.critic_net_diversity1,
'critic_model_diversity2': self.critic_net_diversity2
}, {
'critic1_optimizer_quality': self.critic_quality1_optimizer,
'critic2_optimizer_quality': self.critic_quality2_optimizer,
'critic1_optimizer_diversity':
self.critic_diversity1_optimizer,
'critic2_optimizer_diversity': self.critic_diversity2_optimizer
}, {'hyperp': agent_id}) # no real use
self.storage_actor_ids = {}
self.step_idx = 0
self.algo_iter = 0
class QDRLAlgo:
"""
The QDRL Algorithm
Attributes
args: the arguments
device: the device
env: the environment
"""
def __init__(self, args, device, env):
"""
Building the QDRL Algorithm
Args:
args: the arguments
device: the device
env: the environment
"""
self.device = device
self.args = args
self.env = env
print('err', args)
self.random_state = np.random.RandomState(args['--train_seed'])
action_dim = self.env.action_space.shape[0]
state_dim = self.env.observation_space.shape[0]
# We create the value and policy networks as well as their target
self.critic_net_diversity1, self.critic_net_diversity2, self.target_critic_net_diversity1, self.target_critic_net_diversity2, \
self.critic_net_quality1, self.critic_net_quality2, self.target_critic_net_quality1, self.target_critic_net_quality2 = [
ActionValueNetwork(state_dim, action_dim, args['--hidden']).to(device)
for _ in range(8)
]
self.selected_actors = []
# TODO check the copy system of selected actors
for _ in range(args['--population_size']):
actor = {
'net':
PolicyNetwork(state_dim,
action_dim,
args['--hidden'],
init_w=0.1).to(device)
}
# We create the optimizers
actor['optimizer'] = torch.optim.Adam(actor['net'].parameters(),
lr=args['--policy_lr'])
self.selected_actors.append(actor)
self.critic_quality1_optimizer = torch.optim.Adam(
self.critic_net_quality1.parameters(), lr=args['--value_lr'])
self.critic_quality2_optimizer = torch.optim.Adam(
self.critic_net_quality2.parameters(), lr=args['--value_lr'])
self.critic_diversity1_optimizer = torch.optim.Adam(
self.critic_net_diversity1.parameters(), lr=args['--value_lr'])
self.critic_diversity2_optimizer = torch.optim.Adam(
self.critic_net_diversity2.parameters(), lr=args['--value_lr'])
# We initialize the target models to be identical to the other models
soft_update(self.critic_net_quality1,
self.target_critic_net_quality1,
soft_tau=1.)
soft_update(self.critic_net_quality2,
self.target_critic_net_quality2,
soft_tau=1.)
soft_update(self.critic_net_diversity1,
self.target_critic_net_diversity1,
soft_tau=1.)
soft_update(self.critic_net_diversity2,
self.target_critic_net_diversity2,
soft_tau=1.)
# We create the replay buffer
self.replay_buffer = QDRLReplayBuffer(args['--buffer'])
# We create the criterion
# TODO: figure out why not two TD3
self.qdrl_criterion = QDRLCriterion(self.critic_net_diversity1,
self.critic_net_diversity2,
self.target_critic_net_diversity1,
self.target_critic_net_diversity2,
self.critic_net_quality1,
self.critic_net_quality2,
self.target_critic_net_quality1,
self.target_critic_net_quality2,
gamma=args['--gamma'],
soft_tau=args['--soft_tau'],
noise_std=args['--g_smooth_sigma'],
noise_clip=args['--g_smooth_clip'],
device=device)
# We prepare the experiment
exp_options = {
'critic_loss_diversity_1': {
'plot': 'line',
'yscale': 'log'
},
'critic_loss_diversity_2': {
'plot': 'line',
'yscale': 'log'
},
'critic_loss_quality_1': {
'plot': 'line',
'yscale': 'log'
},
'critic_loss_quality_2': {
'plot': 'line',
'yscale': 'log'
},
'score': {
'plot': 'line',
'yscale': 'linear'
},
'return_test': {
'plot': 'line',
'yscale': 'linear'
},
'return_test_sparse': {
'plot': 'line',
'yscale': 'linear'
},
'actor_loss': {
'plot': 'line',
'yscale': 'linear'
},
}
agent_id = 0
description = 'TD3: {} with {} frames for training'.format(
args['--env_name'], args['--budget'])
self.exp_id = create_experiment(args['--exp'], description, './',
exp_options)
my_logger.info('exp_id: {}'.format(self.exp_id))
# QDRL init
self.archive = Archive()
knn = KNN(k=args['--k_knn'], distance_func=distance_test)
self.novelty_score_evaluator = NoveltyScore(knn=knn)
self.storage_critic = ExpLoggerAgent(
self.exp_id, agent_id,
os.path.join(args['--exp'], 'agent_' + str(agent_id)), {
'critic_model_quality1': self.critic_net_quality1,
'critic_model_quality2': self.critic_net_quality2,
'critic_model_diversity1': self.critic_net_diversity1,
'critic_model_diversity2': self.critic_net_diversity2
}, {
'critic1_optimizer_quality': self.critic_quality1_optimizer,
'critic2_optimizer_quality': self.critic_quality2_optimizer,
'critic1_optimizer_diversity':
self.critic_diversity1_optimizer,
'critic2_optimizer_diversity': self.critic_diversity2_optimizer
}, {'hyperp': agent_id}) # no real use
self.storage_actor_ids = {}
self.step_idx = 0
self.algo_iter = 0
def compute_pareto_front(self):
"""
Computes the Pareto front for QDRL
Returns:
(list) a list of ids of the actors in the pareto front
"""
pareto_front = []
final_size = min(self.args['--population_size'],
len(self.archive.elements))
# initial points
remaining_pts = []
for element in self.archive.elements:
remaining_pts.append({
'values': [element.quality, element.diversity],
'actor_id': element.id
})
iteration = 0
while len(pareto_front) < final_size:
efficient_pts = keep_efficient(remaining_pts)
efficient_pts_truncated = efficient_pts[0:final_size -
len(pareto_front)]
pareto_front += efficient_pts_truncated
new_remaining_pts = []
pareto_front_ids = [p['actor_id'] for p in pareto_front]
for p in remaining_pts:
if p['actor_id'] not in pareto_front_ids:
new_remaining_pts.append(p)
remaining_pts = new_remaining_pts
iteration += 1
if iteration == 1:
my_logger.info('true pareto_front {}'.format(pareto_front))
my_logger.info('final pareto_front {}'.format(pareto_front))
assert (len(set(pareto_front_ids)) == len(pareto_front_ids))
return pareto_front_ids
def qd_split_pareto_front(self, pareto_front_ids):
"""
Divides the pareto front into quality and diversity
Args:
pareto_front_ids: the pareto_front
Returns:
(list): the selected actors
"""
half = int(len(pareto_front_ids) / 2)
selected_actors = []
for pareto_idx, ac_id in enumerate(pareto_front_ids):
element = self.archive.elements[ac_id]
actor_ = {
'net': element.actor['net'],
'optimizer': element.actor['optimizer'],
'pareto_id': pareto_idx,
'id_copy': element.id
}
selected_actors.append(actor_)
selected_actors_copy = [copy.deepcopy(ac) for ac in selected_actors]
self.random_state.shuffle(selected_actors_copy)
self.qdrl_criterion.actors_diversity = selected_actors_copy[:half]
self.qdrl_criterion.actors_quality = selected_actors_copy[half:]
my_logger.info('quality actors {}'.format(
[ac['id_copy'] for ac in self.qdrl_criterion.actors_diversity]))
my_logger.info('diversity actors {}'.format(
[ac['id_copy'] for ac in self.qdrl_criterion.actors_quality]))
return selected_actors_copy
def update_actors_diversity(self):
"""
Updates the diversity of all the actors
"""
for element in self.archive.elements:
element.diversity = self.novelty_score_evaluator.compute_ns_actor(
self.archive, element)
my_logger.info('diversity {} of element {}'.format(
element.diversity, element.id))
def one_episode_test_and_store(self, actor):
"""
Run an episode from the actor on the env and decide whether or not to add to the archive and replay buffer if diverse enough
Args:
actor: the actor for which to execute an episode
Returns:
(int) the number of steps run in the environment
"""
replay_buffer_temp = []
actor['net'].eval()
state = self.env.reset()
done = False
episode_reward = 0
rollout = [state]
step_idx_in_episode = 0
detected = 0
idx = 0
while not done:
idx += 1
model_state = torch.FloatTensor(state).to(self.device)
action = actor['net'](model_state).detach().cpu().numpy()
next_state, reward, done, info_ = self.env.step(action)
detected += info_['detections']
episode_reward += reward
rollout.append(next_state)
if not done or step_idx_in_episode != 0:
replay_buffer_temp.append({
'state': state,
'action': action,
'reward': reward,
'next_state': next_state,
'done': done
})
# do not forget to update time and state
step_idx_in_episode += 1
state = next_state
success = (detected == 0) and (info_['arrived'] == 1)
element = create_element(episode_reward, success, rollout, actor)
distance_to_archive, nn = self.archive.distance(element, distance_test)
my_logger.info('new element distance to archive is {}'.format(
distance_to_archive))
if (distance_to_archive > self.args["--min_archive_inter_distance"]
or nn.quality < episode_reward or nn.success < success):
if nn is not None and nn.quality < episode_reward and distance_to_archive < .001:
print(
'CASE when close but better, replacing id {}, my parent {}',
nn.id, element.actor['id_copy']
if 'id_copy' in element.actor else None)
self.archive.replace_element(nn.id, element)
else:
actor_id = self.archive.add_element(element)
my_logger.info(
'+ADD+ to archive actor with id {} and episode reward {} success {} with parent {}'
.format(
actor_id, episode_reward, success,
element.actor['id_copy']
if 'id_copy' in element.actor else None))
for trans in replay_buffer_temp:
self.replay_buffer.push(trans['state'], trans['action'],
trans['reward'],
trans['next_state'], trans['done'],
actor_id)
else:
my_logger.info(
'-NOT- adding to archive actor with episode reward {} success {} because close to id {}, my parents {}'
.format(
episode_reward, success, nn.id, element.actor['id_copy']
if 'id_copy' in element.actor else None))
return len(rollout)
def sample_replay_buffer_and_gradient_descent(self):
"""
Routine that does the sampling from the replay buffer and does the gradient descent steps
"""
for actor in self.selected_actors:
actor['net'].train()
self.critic_net_diversity1.train()
self.critic_net_diversity2.train()
self.critic_net_quality1.train()
self.critic_net_quality2.train()
if len(self.replay_buffer) > self.args['--batch_size']:
# Sample from the relay buffer
state_replay, action_replay, reward_replay, next_state_replay, done_replay, actor_id_replay = self.replay_buffer.sample(
self.args['--batch_size'])
reward_diversity_replay = np.array([
self.archive.elements[ac_id].diversity
for ac_id in actor_id_replay
])
# Compute, store and optimize the losses
self.critic_loss_diversity1, self.critic_loss_diversity2, self.critic_loss_quality1, self.critic_loss_quality2, self.actor_losses = self.qdrl_criterion.loss(
state_replay, action_replay, reward_replay,
reward_diversity_replay, next_state_replay, done_replay)
self.critic_quality1_optimizer.zero_grad()
self.critic_diversity1_optimizer.zero_grad()
self.critic_loss_diversity1.backward(
retain_graph=True,
inputs=list(self.critic_net_diversity1.parameters()))
self.critic_loss_quality1.backward(
retain_graph=True,
inputs=list(self.critic_net_quality1.parameters()))
self.critic_quality2_optimizer.zero_grad()
self.critic_diversity2_optimizer.zero_grad()
self.critic_loss_diversity2.backward(
retain_graph=True,
inputs=list(self.critic_net_diversity2.parameters()))
self.critic_loss_quality2.backward(
retain_graph=True,
inputs=list(self.critic_net_quality2.parameters()))
if self.algo_iter % self.args['--delay_policy_update'] == 0:
for actor in self.selected_actors:
actor['optimizer'].zero_grad()
self.actor_losses[actor['pareto_id']].backward(
inputs=list(actor['net'].parameters()))
self.critic_quality1_optimizer.step()
self.critic_quality2_optimizer.step()
self.critic_diversity1_optimizer.step()
self.critic_diversity2_optimizer.step()
if self.algo_iter % self.args['--delay_policy_update'] == 0:
for actor in self.selected_actors:
actor['optimizer'].step()
soft_update(self.critic_net_quality1,
self.target_critic_net_quality1,
self.args['--soft_tau'])
soft_update(self.critic_net_quality2,
self.target_critic_net_quality2,
self.args['--soft_tau'])
soft_update(self.critic_net_diversity1,
self.target_critic_net_diversity1,
self.args['--soft_tau'])
soft_update(self.critic_net_diversity2,
self.target_critic_net_diversity2,
self.args['--soft_tau'])
def train_model(self):
"""
Train the QDRL
"""
my_logger.info('-------first evaluation of the initial actors')
for actor in self.selected_actors:
self.step_idx += self.one_episode_test_and_store(actor)
self.update_actors_diversity()
if self.args["--display_video_archive"]:
actors = [
transs(el.actor, self.device) for el in self.archive.elements
]
rollouts = [el.outcome for el in self.archive.elements]
plot_rol(self.env, rollouts, self.device)
my_logger.info('--------Starting the loop')
last_save_idx = 0
# We train the networks
while self.step_idx < self.args['--budget']:
if self.algo_iter % self.args['--delay_policy_update'] == 0:
self.archive.print_info(my_logger)
pareto_front_ids = self.compute_pareto_front()
my_logger.info('pareto front ids {}'.format(pareto_front_ids))
self.selected_actors = self.qd_split_pareto_front(
pareto_front_ids)
# Train/Update the actor and critic based on resampling transitions from the replay buffer
gradient_steps = int(self.step_idx *
self.args['--gradient_steps_ratio'] + 1)
my_logger.info('Algorithm Iteration {} -- steps done {}'.format(
self.algo_iter, self.step_idx))
my_logger.info('gradient_steps = {} {}'.format(
gradient_steps, self.args['--gradient_steps_ratio']))
for _ in range(gradient_steps):
self.sample_replay_buffer_and_gradient_descent()
if self.algo_iter % self.args['--delay_policy_update'] == 0:
my_logger.info('-------new evaluation of the actors')
for selected_actor in self.selected_actors:
self.step_idx += self.one_episode_test_and_store(
selected_actor)
self.update_actors_diversity()
if self.args[
"--display_video_archive"] and self.algo_iter % self.args[
"--display_video_archive_frequency"] == 0:
rollouts = [el.outcome for el in self.archive.elements]
plot_rol(self.env, rollouts, self.device)
#with open('archive' + str(self.algo_iter), 'wb') as f:
# import pickle
# pickle.dump(self.archive, f)
rollouts = [
el.outcome for el in self.archive.elements
if el.success
]
plot_rol(self.env, rollouts, self.device)
# Save and print performance information
if self.step_idx - last_save_idx > self.args[
"--save_interval_performance"] and self.step_idx > 1:
last_save_idx = self.step_idx
stor = {
'critic_loss_quality_1': self.critic_loss_quality1.item(),
'critic_loss_quality_2': self.critic_loss_quality2.item(),
'critic_loss_diversity_1':
self.critic_loss_diversity1.item(),
'critic_loss_diversity_2':
self.critic_loss_diversity2.item()
}
self.storage_critic.performance(self.step_idx, stor)
self.storage_critic.write()
my_logger.info(
'Loss at {}/{}: value1={:.4}, value2={:.4}'.format(
self.step_idx,
self.args['--budget'],
self.critic_loss_quality1.item(),
self.critic_loss_quality2.item(),
))
# save the weights of the model
if True: # step_idx % args['--save_interval'] == 0:
for idx, ac in enumerate(self.selected_actors):
actor_archive_idx = pareto_front_ids[ac['pareto_id']]
if actor_archive_idx in self.storage_actor_ids:
storage_ac = self.storage_actor_ids[actor_archive_idx]
else:
storage_ac = ExpLoggerAgent(
self.exp_id, actor_archive_idx + 1,
os.path.join(self.args['--exp'], 'agent_' +
str(actor_archive_idx + 1)),
{'actor_model': ac['net']},
{'actor_optimizer': ac['optimizer']},
{'hyperp': 1})
self.storage_actor_ids[actor_archive_idx] = storage_ac
stor = {}
stor['actor_loss'] = self.actor_losses[
ac['pareto_id']].item()
stor['return_test'] = self.archive.elements[
actor_archive_idx].quality
stor['return_test_sparse'] = self.archive.elements[
actor_archive_idx].success
storage_ac.performance(self.step_idx, stor)
storage_ac.write()
storage_ac.state(self.step_idx)
self.algo_iter += 1
self.env.close()
self.storage.close()
stop_experiment(self.exp_id)