def agent(self, writer=DummyWriter(), train_steps=float('inf')):
n_updates = train_steps / self.hyperparameters["min_batch_size"]
feature_optimizer = Adam(self.feature_model.parameters(), lr=self.hyperparameters["lr_pi"], eps=self.hyperparameters["eps"])
value_optimizer = Adam(self.value_model.parameters(), lr=self.hyperparameters["lr_v"], eps=self.hyperparameters["eps"])
policy_optimizer = Adam(self.policy_model.parameters(), lr=self.hyperparameters["lr_pi"], eps=self.hyperparameters["eps"])
features = FeatureNetwork(
self.feature_model,
feature_optimizer,
scheduler=CosineAnnealingLR(feature_optimizer, n_updates),
clip_grad=self.hyperparameters["clip_grad"],
writer=writer
)
v = VNetwork(
self.value_model,
value_optimizer,
scheduler=CosineAnnealingLR(value_optimizer, n_updates),
loss_scaling=self.hyperparameters["value_loss_scaling"],
clip_grad=self.hyperparameters["clip_grad"],
writer=writer
)
policy = SoftmaxPolicy(
self.policy_model,
policy_optimizer,
scheduler=CosineAnnealingLR(policy_optimizer, n_updates),
clip_grad=self.hyperparameters["clip_grad"],
writer=writer
)
return DeepmindAtariBody(
VPG(features, v, policy, discount_factor=self.hyperparameters["discount_factor"], min_batch_size=self.hyperparameters["min_batch_size"]),
)
def agent(self, writer=DummyWriter(), train_steps=float('inf')):
n_updates = train_steps / self.hyperparameters['n_envs']
optimizer = Adam(self.model.parameters(),
lr=self.hyperparameters['lr'],
eps=self.hyperparameters['eps'])
q = QNetwork(self.model,
optimizer,
scheduler=CosineAnnealingLR(optimizer, n_updates),
writer=writer)
policy = ParallelGreedyPolicy(
q,
self.n_actions,
epsilon=LinearScheduler(
self.hyperparameters['initial_exploration'],
self.hyperparameters['final_exploration'],
0,
self.hyperparameters["final_exploration_step"] /
self.hyperparameters["n_envs"],
name="exploration",
writer=writer))
return DeepmindAtariBody(
VQN(q,
policy,
discount_factor=self.hyperparameters['discount_factor']), )
def test_agent(self):
q = QNetwork(copy.deepcopy(self.model))
return DeepmindAtariBody(
DDQNTestAgent(
q,
self.n_actions,
exploration=self.hyperparameters['test_exploration']))
def _vac(envs, writer=DummyWriter()):
value_model = value_model_constructor().to(device)
policy_model = policy_model_constructor(envs[0]).to(device)
feature_model = feature_model_constructor().to(device)
value_optimizer = Adam(value_model.parameters(), lr=lr_v, eps=eps)
policy_optimizer = Adam(policy_model.parameters(), lr=lr_pi, eps=eps)
feature_optimizer = Adam(feature_model.parameters(), lr=lr_pi, eps=eps)
v = VNetwork(
value_model,
value_optimizer,
loss_scaling=value_loss_scaling,
clip_grad=clip_grad,
writer=writer,
)
policy = SoftmaxPolicy(
policy_model,
policy_optimizer,
clip_grad=clip_grad,
writer=writer,
)
features = FeatureNetwork(feature_model,
feature_optimizer,
clip_grad=clip_grad,
writer=writer)
return DeepmindAtariBody(
VAC(features, v, policy, discount_factor=discount_factor), )
def _model_predictive_dqn(env, writer=None):
# models
feature_model = shared_feature_layers().to(device)
value_model = value_head().to(device)
reward_model = reward_head(env).to(device)
generator_model = Generator(env).to(device)
# optimizers
feature_optimizer = Adam(feature_model.parameters(), lr=lr, eps=eps)
value_optimizer = Adam(value_model.parameters(), lr=lr, eps=eps)
reward_optimizer = Adam(reward_model.parameters(), lr=lr, eps=eps)
generator_optimizer = Adam(generator_model.parameters(), lr=lr, eps=eps)
# approximators
f = FeatureNetwork(feature_model, feature_optimizer, writer=writer)
v = VNetwork(value_model, value_optimizer, writer=writer)
r = QNetwork(reward_model, reward_optimizer, name='reward', writer=writer)
g = Approximation(generator_model, generator_optimizer, name='generator', writer=writer)
# replay buffer
replay_buffer = ExperienceReplayBuffer(replay_buffer_size, device=device)
# create agent
agent = ModelPredictiveDQN(f, v, r, g, replay_buffer,
minibatch_size=minibatch_size,
replay_start_size=replay_start_size
)
# apply agent wrappers for better atari performance
return DeepmindAtariBody(agent, lazy_frames=True)
def _vqn(envs, writer=DummyWriter()):
env = envs[0]
model = nature_ddqn(env).to(device)
optimizer = RMSprop(model.parameters(), lr=lr, alpha=alpha, eps=eps)
q = QNetwork(
model,
optimizer,
env.action_space.n,
loss=smooth_l1_loss,
writer=writer
)
policy = GreedyPolicy(
q,
env.action_space.n,
epsilon=LinearScheduler(
initial_exploration,
final_exploration,
0,
final_exploration_frame,
name="epsilon",
writer=writer
)
)
return DeepmindAtariBody(
VQN(q, policy, gamma=discount_factor),
)
def _vqn(envs, writer=DummyWriter()):
action_repeat = 4
final_exploration_timestep = final_exploration_frame / action_repeat
env = envs[0]
model = model_constructor(env).to(device)
optimizer = Adam(model.parameters(), lr=lr, eps=eps)
q = QNetwork(
model,
optimizer,
writer=writer
)
policy = ParallelGreedyPolicy(
q,
env.action_space.n,
epsilon=LinearScheduler(
initial_exploration,
final_exploration,
0,
final_exploration_timestep,
name="epsilon",
writer=writer
)
)
return DeepmindAtariBody(
VQN(q, policy, discount_factor=discount_factor),
)
def agent_constructor(writer):
policy = GreedyPolicy(
q,
n_actions,
epsilon=LinearScheduler(
initial_exploration,
final_exploration,
replay_start_size,
final_exploration_step - replay_start_size,
name="epsilon",
writer=writer
)
)
return DeepmindAtariBody(
DQN(
q,
policy,
replay_buffer,
discount_factor=discount_factor,
loss=smooth_l1_loss,
minibatch_size=minibatch_size,
replay_start_size=replay_start_size,
update_frequency=update_frequency,
),
lazy_frames=True
)
def _dqn(env, writer=DummyWriter()):
_model = model
_optimizer = optimizer
if _model is None:
_model = conv_net(env, frames=agent_history_length).to(device)
if _optimizer is None:
_optimizer = Adam(_model.parameters(), lr=lr, eps=eps)
q = QNetwork(_model,
_optimizer,
env.action_space.n,
target_update_frequency=target_update_frequency,
loss=smooth_l1_loss,
writer=writer)
policy = GreedyPolicy(q,
env.action_space.n,
annealing_start=replay_start_size,
annealing_time=final_exploration_frame -
replay_start_size,
initial_epsilon=initial_exploration,
final_epsilon=final_exploration)
replay_buffer = ExperienceReplayBuffer(replay_buffer_size,
device=device)
return DeepmindAtariBody(DQN(
q,
policy,
replay_buffer,
discount_factor=discount_factor,
minibatch_size=minibatch_size,
replay_start_size=replay_start_size,
update_frequency=update_frequency,
),
env,
action_repeat=action_repeat,
frame_stack=agent_history_length,
noop_max=noop_max)
def _dqn(env, writer=DummyWriter()):
_model = nature_dqn(env).to(device)
_optimizer = Adam(_model.parameters(), lr=lr, eps=eps)
q = QNetwork(_model,
_optimizer,
env.action_space.n,
target=FixedTarget(target_update_frequency),
loss=smooth_l1_loss,
writer=writer)
policy = GreedyPolicy(q,
env.action_space.n,
epsilon=LinearScheduler(initial_exploration,
final_exploration,
replay_start_size,
final_exploration_frame,
name="epsilon",
writer=writer))
replay_buffer = ExperienceReplayBuffer(replay_buffer_size,
device=device)
return DeepmindAtariBody(
DQN(
q,
policy,
replay_buffer,
discount_factor=discount_factor,
minibatch_size=minibatch_size,
replay_start_size=replay_start_size,
update_frequency=update_frequency,
), )
def parallel_test_agent(self):
q = QNetwork(copy.deepcopy(self.model))
policy = ParallelGreedyPolicy(
q,
self.n_actions,
epsilon=self.hyperparameters['test_exploration'])
return DeepmindAtariBody(VQNTestAgent(policy))
def _ppo(envs, writer=DummyWriter()):
env = envs[0]
# Update epoch * minibatches times per update,
# but we only update once per n_steps,
# with n_envs and 4 frames per step
final_anneal_step = last_frame * epochs * minibatches / (n_steps *
n_envs * 4)
value_model = value_model_constructor().to(device)
policy_model = policy_model_constructor(env).to(device)
feature_model = feature_model_constructor().to(device)
feature_optimizer = Adam(feature_model.parameters(), lr=lr, eps=eps)
value_optimizer = Adam(value_model.parameters(), lr=lr, eps=eps)
policy_optimizer = Adam(policy_model.parameters(), lr=lr, eps=eps)
features = FeatureNetwork(feature_model,
feature_optimizer,
clip_grad=clip_grad,
scheduler=CosineAnnealingLR(
feature_optimizer, final_anneal_step),
writer=writer)
v = VNetwork(
value_model,
value_optimizer,
loss_scaling=value_loss_scaling,
clip_grad=clip_grad,
writer=writer,
scheduler=CosineAnnealingLR(value_optimizer, final_anneal_step),
)
policy = SoftmaxPolicy(
policy_model,
policy_optimizer,
clip_grad=clip_grad,
writer=writer,
scheduler=CosineAnnealingLR(policy_optimizer, final_anneal_step),
)
return DeepmindAtariBody(
PPO(
features,
v,
policy,
epsilon=LinearScheduler(clip_initial,
clip_final,
0,
final_anneal_step,
name='clip',
writer=writer),
epochs=epochs,
minibatches=minibatches,
n_envs=n_envs,
n_steps=n_steps,
discount_factor=discount_factor,
lam=lam,
entropy_loss_scaling=entropy_loss_scaling,
writer=writer,
))
def agent(self, writer=DummyWriter(), train_steps=float('inf')):
n_updates = train_steps * self.hyperparameters[
'epochs'] * self.hyperparameters['minibatches'] / (
self.hyperparameters['n_steps'] *
self.hyperparameters['n_envs'])
feature_optimizer = Adam(self.feature_model.parameters(),
lr=self.hyperparameters["lr"],
eps=self.hyperparameters["eps"])
value_optimizer = Adam(self.value_model.parameters(),
lr=self.hyperparameters["lr"],
eps=self.hyperparameters["eps"])
policy_optimizer = Adam(self.policy_model.parameters(),
lr=self.hyperparameters["lr"],
eps=self.hyperparameters["eps"])
features = FeatureNetwork(self.feature_model,
feature_optimizer,
scheduler=CosineAnnealingLR(
feature_optimizer, n_updates),
clip_grad=self.hyperparameters["clip_grad"],
writer=writer)
v = VNetwork(self.value_model,
value_optimizer,
scheduler=CosineAnnealingLR(value_optimizer, n_updates),
loss_scaling=self.hyperparameters["value_loss_scaling"],
clip_grad=self.hyperparameters["clip_grad"],
writer=writer)
policy = SoftmaxPolicy(self.policy_model,
policy_optimizer,
scheduler=CosineAnnealingLR(
policy_optimizer, n_updates),
clip_grad=self.hyperparameters["clip_grad"],
writer=writer)
return DeepmindAtariBody(
PPO(
features,
v,
policy,
epsilon=LinearScheduler(self.hyperparameters["clip_initial"],
self.hyperparameters["clip_final"],
0,
n_updates,
name='clip',
writer=writer),
epochs=self.hyperparameters["epochs"],
minibatches=self.hyperparameters["minibatches"],
n_envs=self.hyperparameters["n_envs"],
n_steps=self.hyperparameters["n_steps"],
discount_factor=self.hyperparameters["discount_factor"],
lam=self.hyperparameters["lam"],
entropy_loss_scaling=self.
hyperparameters["entropy_loss_scaling"],
writer=writer,
))
def _a2c(envs, writer=DummyWriter()):
env = envs[0]
final_anneal_step = last_frame / (n_steps * n_envs * 4)
value_model = nature_value_head().to(device)
policy_model = nature_policy_head(env).to(device)
feature_model = nature_features().to(device)
feature_optimizer = Adam(feature_model.parameters(), lr=lr, eps=eps)
value_optimizer = Adam(value_model.parameters(), lr=lr, eps=eps)
policy_optimizer = Adam(policy_model.parameters(), lr=lr, eps=eps)
features = FeatureNetwork(
feature_model,
feature_optimizer,
scheduler=CosineAnnealingLR(
feature_optimizer,
final_anneal_step,
),
clip_grad=clip_grad,
writer=writer
)
v = VNetwork(
value_model,
value_optimizer,
scheduler=CosineAnnealingLR(
value_optimizer,
final_anneal_step,
),
loss_scaling=value_loss_scaling,
clip_grad=clip_grad,
writer=writer
)
policy = SoftmaxPolicy(
policy_model,
policy_optimizer,
scheduler=CosineAnnealingLR(
policy_optimizer,
final_anneal_step,
),
clip_grad=clip_grad,
writer=writer
)
return DeepmindAtariBody(
A2C(
features,
v,
policy,
n_envs=n_envs,
n_steps=n_steps,
discount_factor=discount_factor,
entropy_loss_scaling=entropy_loss_scaling,
writer=writer
),
)
def agent_constructor():
q_dist = QDist(
self.model,
None,
self.n_actions,
self.hyperparameters['atoms'],
v_min=self.hyperparameters['v_min'],
v_max=self.hyperparameters['v_max'],
)
return DeepmindAtariBody(RainbowTestAgent(q_dist, self.n_actions, self.hyperparameters["test_exploration"]))
def test_agent(self):
q_dist = QDist(
copy.deepcopy(self.model),
None,
self.n_actions,
self.hyperparameters['atoms'],
v_min=self.hyperparameters['v_min'],
v_max=self.hyperparameters['v_max'],
)
return DeepmindAtariBody(C51TestAgent(q_dist, self.n_actions, self.hyperparameters["test_exploration"]))
def _dqn(env, writer=DummyWriter()):
action_repeat = 4
last_timestep = last_frame / action_repeat
last_update = (last_timestep - replay_start_size) / update_frequency
final_exploration_step = final_exploration_frame / action_repeat
model = nature_dqn(env).to(device)
optimizer = Adam(
model.parameters(),
lr=lr,
eps=eps
)
q = QNetwork(
model,
optimizer,
scheduler=CosineAnnealingLR(optimizer, last_update),
target=FixedTarget(target_update_frequency),
writer=writer
)
policy = GreedyPolicy(
q,
env.action_space.n,
epsilon=LinearScheduler(
initial_exploration,
final_exploration,
replay_start_size,
final_exploration_step - replay_start_size,
name="epsilon",
writer=writer
)
)
replay_buffer = ExperienceReplayBuffer(
replay_buffer_size,
device=device
)
return DeepmindAtariBody(
DQN(
q,
policy,
replay_buffer,
discount_factor=discount_factor,
loss=smooth_l1_loss,
minibatch_size=minibatch_size,
replay_start_size=replay_start_size,
update_frequency=update_frequency,
),
lazy_frames=True
)
def agent(self, writer=DummyWriter(), train_steps=float('inf')):
n_updates = (train_steps - self.hyperparameters['replay_start_size']) / self.hyperparameters['update_frequency']
optimizer = Adam(
self.model.parameters(),
lr=self.hyperparameters['lr'],
eps=self.hyperparameters['eps']
)
q = QNetwork(
self.model,
optimizer,
scheduler=CosineAnnealingLR(optimizer, n_updates),
target=FixedTarget(self.hyperparameters['target_update_frequency']),
writer=writer
)
policy = GreedyPolicy(
q,
self.n_actions,
epsilon=LinearScheduler(
self.hyperparameters['initial_exploration'],
self.hyperparameters['final_exploration'],
self.hyperparameters['replay_start_size'],
self.hyperparameters['final_exploration_step'] - self.hyperparameters['replay_start_size'],
name="exploration",
writer=writer
)
)
replay_buffer = ExperienceReplayBuffer(
self.hyperparameters['replay_buffer_size'],
device=self.device
)
return DeepmindAtariBody(
DQN(
q,
policy,
replay_buffer,
discount_factor=self.hyperparameters['discount_factor'],
loss=smooth_l1_loss,
minibatch_size=self.hyperparameters['minibatch_size'],
replay_start_size=self.hyperparameters['replay_start_size'],
update_frequency=self.hyperparameters['update_frequency'],
),
lazy_frames=True
)
def agent(self, writer=DummyWriter(), train_steps=float('inf')):
n_updates = (train_steps - self.hyperparameters['replay_start_size']
) / self.hyperparameters['update_frequency']
optimizer = Adam(self.model.parameters(),
lr=self.hyperparameters['lr'],
eps=self.hyperparameters['eps'])
q_dist = QDist(
self.model,
optimizer,
self.n_actions,
self.hyperparameters['atoms'],
scheduler=CosineAnnealingLR(optimizer, n_updates),
v_min=self.hyperparameters['v_min'],
v_max=self.hyperparameters['v_max'],
target=FixedTarget(
self.hyperparameters['target_update_frequency']),
writer=writer,
)
replay_buffer = NStepReplayBuffer(
self.hyperparameters['n_steps'],
self.hyperparameters['discount_factor'],
PrioritizedReplayBuffer(self.hyperparameters['replay_buffer_size'],
alpha=self.hyperparameters['alpha'],
beta=self.hyperparameters['beta'],
device=self.device))
return DeepmindAtariBody(Rainbow(
q_dist,
replay_buffer,
exploration=LinearScheduler(
self.hyperparameters['initial_exploration'],
self.hyperparameters['final_exploration'],
0,
train_steps - self.hyperparameters['replay_start_size'],
name="exploration",
writer=writer),
discount_factor=self.hyperparameters['discount_factor']**
self.hyperparameters["n_steps"],
minibatch_size=self.hyperparameters['minibatch_size'],
replay_start_size=self.hyperparameters['replay_start_size'],
update_frequency=self.hyperparameters['update_frequency'],
writer=writer,
),
lazy_frames=True,
episodic_lives=True)
def agent(self, writer=DummyWriter(), train_steps=float('inf')):
n_updates = (train_steps - self.hyperparameters['replay_start_size']) / self.hyperparameters['update_frequency']
optimizer = Adam(
self.model.parameters(),
lr=self.hyperparameters['lr'],
eps=self.hyperparameters['eps']
)
q = QDist(
self.model,
optimizer,
self.n_actions,
self.hyperparameters['atoms'],
v_min=self.hyperparameters['v_min'],
v_max=self.hyperparameters['v_max'],
target=FixedTarget(self.hyperparameters['target_update_frequency']),
scheduler=CosineAnnealingLR(optimizer, n_updates),
writer=writer,
)
replay_buffer = ExperienceReplayBuffer(
self.hyperparameters['replay_buffer_size'],
device=self.device
)
return DeepmindAtariBody(
C51(
q,
replay_buffer,
exploration=LinearScheduler(
self.hyperparameters['initial_exploration'],
self.hyperparameters['final_exploration'],
0,
self.hyperparameters["final_exploration_step"] - self.hyperparameters["replay_start_size"],
name="epsilon",
writer=writer,
),
discount_factor=self.hyperparameters["discount_factor"],
minibatch_size=self.hyperparameters["minibatch_size"],
replay_start_size=self.hyperparameters["replay_start_size"],
update_frequency=self.hyperparameters["update_frequency"],
writer=writer
),
lazy_frames=True,
episodic_lives=True
)
def _a2c(envs, writer=DummyWriter()):
env = envs[0]
value_model = nature_value_head().to(device)
policy_model = nature_policy_head(envs[0]).to(device)
feature_model = nature_features().to(device)
feature_optimizer = RMSprop(
feature_model.parameters(), alpha=alpha, lr=lr, eps=eps
)
value_optimizer = RMSprop(value_model.parameters(), alpha=alpha, lr=lr, eps=eps)
policy_optimizer = RMSprop(
policy_model.parameters(), alpha=alpha, lr=lr, eps=eps
)
features = FeatureNetwork(
feature_model,
feature_optimizer,
clip_grad=clip_grad,
writer=writer
)
v = VNetwork(
value_model,
value_optimizer,
loss_scaling=value_loss_scaling,
clip_grad=clip_grad,
writer=writer
)
policy = SoftmaxPolicy(
policy_model,
policy_optimizer,
env.action_space.n,
entropy_loss_scaling=entropy_loss_scaling,
clip_grad=clip_grad,
writer=writer
)
return DeepmindAtariBody(
A2C(
features,
v,
policy,
n_envs=n_envs,
n_steps=n_steps,
discount_factor=discount_factor,
),
)
def _c51(env, writer=DummyWriter()):
action_repeat = 4
last_timestep = last_frame / action_repeat
last_update = (last_timestep - replay_start_size) / update_frequency
model = nature_c51(env, atoms=atoms).to(device)
optimizer = Adam(
model.parameters(),
lr=lr,
eps=eps
)
q = QDist(
model,
optimizer,
env.action_space.n,
atoms,
v_min=v_min,
v_max=v_max,
target=FixedTarget(target_update_frequency),
scheduler=CosineAnnealingLR(optimizer, last_update),
writer=writer,
)
replay_buffer = ExperienceReplayBuffer(
replay_buffer_size,
device=device
)
return DeepmindAtariBody(
C51(
q,
replay_buffer,
exploration=LinearScheduler(
initial_exploration,
final_exploration,
0,
last_timestep,
name="epsilon",
writer=writer,
),
discount_factor=discount_factor,
minibatch_size=minibatch_size,
replay_start_size=replay_start_size,
update_frequency=update_frequency,
writer=writer
),
lazy_frames=True
)
def main():
parser = argparse.ArgumentParser(description="Run an Atari benchmark.")
parser.add_argument("env", help="Name of the Atari game (e.g. Pong)")
parser.add_argument("dir", help="Directory where the agent's model was saved.")
parser.add_argument(
"--device",
default="cpu",
help="The name of the device to run the agent on (e.g. cpu, cuda, cuda:0)",
)
parser.add_argument(
"--fps",
default=60,
help="Playback speed",
)
args = parser.parse_args()
env = AtariEnvironment(args.env, device=args.device)
agent = DeepmindAtariBody(GreedyAgent.load(args.dir, env))
watch(agent, env, fps=args.fps)
def _vpg_atari(env, writer=DummyWriter()):
feature_model = nature_features().to(device)
value_model = nature_value_head().to(device)
policy_model = nature_policy_head(env).to(device)
feature_optimizer = RMSprop(feature_model.parameters(),
alpha=alpha,
lr=lr * feature_lr_scaling,
eps=eps)
value_optimizer = RMSprop(value_model.parameters(),
alpha=alpha,
lr=lr,
eps=eps)
policy_optimizer = RMSprop(policy_model.parameters(),
alpha=alpha,
lr=lr,
eps=eps)
features = FeatureNetwork(feature_model,
feature_optimizer,
clip_grad=clip_grad,
writer=writer)
v = VNetwork(value_model,
value_optimizer,
loss_scaling=value_loss_scaling,
clip_grad=clip_grad,
writer=writer)
policy = SoftmaxPolicy(
policy_model,
policy_optimizer,
env.action_space.n,
entropy_loss_scaling=entropy_loss_scaling,
clip_grad=clip_grad,
writer=writer,
)
return DeepmindAtariBody(
VPG(features,
v,
policy,
gamma=discount_factor,
min_batch_size=min_batch_size), )
def _vpg_atari(env, writer=DummyWriter()):
value_model = nature_value_head().to(device)
policy_model = nature_policy_head(env).to(device)
feature_model = nature_features().to(device)
feature_optimizer = Adam(feature_model.parameters(), lr=lr, eps=eps)
value_optimizer = Adam(value_model.parameters(), lr=lr, eps=eps)
policy_optimizer = Adam(policy_model.parameters(), lr=lr, eps=eps)
features = FeatureNetwork(feature_model,
feature_optimizer,
scheduler=CosineAnnealingLR(
feature_optimizer,
final_anneal_step,
),
clip_grad=clip_grad,
writer=writer)
v = VNetwork(value_model,
value_optimizer,
scheduler=CosineAnnealingLR(
value_optimizer,
final_anneal_step,
),
loss_scaling=value_loss_scaling,
clip_grad=clip_grad,
writer=writer)
policy = SoftmaxPolicy(policy_model,
policy_optimizer,
scheduler=CosineAnnealingLR(
policy_optimizer,
final_anneal_step,
),
clip_grad=clip_grad,
writer=writer)
return DeepmindAtariBody(VPG(features,
v,
policy,
discount_factor=discount_factor,
min_batch_size=min_batch_size),
episodic_lives=True)
def _rainbow(env, writer=DummyWriter()):
action_repeat = 4
last_timestep = last_frame / action_repeat
last_update = (last_timestep - replay_start_size) / update_frequency
model = model_constructor(env, atoms=atoms, sigma=sigma).to(device)
optimizer = Adam(model.parameters(), lr=lr, eps=eps)
q = QDist(
model,
optimizer,
env.action_space.n,
atoms,
scheduler=CosineAnnealingLR(optimizer, last_update),
v_min=v_min,
v_max=v_max,
target=FixedTarget(target_update_frequency),
writer=writer,
)
replay_buffer = PrioritizedReplayBuffer(replay_buffer_size,
alpha=alpha,
beta=beta,
device=device)
replay_buffer = NStepReplayBuffer(n_steps, discount_factor,
replay_buffer)
agent = Rainbow(
q,
replay_buffer,
exploration=LinearScheduler(initial_exploration,
final_exploration,
0,
last_timestep,
name='exploration',
writer=writer),
discount_factor=discount_factor**n_steps,
minibatch_size=minibatch_size,
replay_start_size=replay_start_size,
update_frequency=update_frequency,
writer=writer,
)
return DeepmindAtariBody(agent, lazy_frames=True, episodic_lives=True)
def _c51(env, writer=DummyWriter()):
model = nature_c51(env, atoms=51).to(device)
optimizer = Adam(
model.parameters(),
lr=lr,
eps=eps
)
q = QDist(
model,
optimizer,
env.action_space.n,
atoms,
v_min=v_min,
v_max=v_max,
target=FixedTarget(target_update_frequency),
writer=writer,
)
replay_buffer = ExperienceReplayBuffer(
replay_buffer_size,
device=device
)
return DeepmindAtariBody(
C51(
q,
replay_buffer,
exploration=LinearScheduler(
initial_exploration,
final_exploration,
replay_start_size,
final_exploration_frame,
name="epsilon",
writer=writer,
),
discount_factor=discount_factor,
minibatch_size=minibatch_size,
replay_start_size=replay_start_size,
update_frequency=update_frequency,
writer=writer
)
)
def agent_constructor(writer):
return DeepmindAtariBody(
Rainbow(
q_dist,
replay_buffer,
exploration=LinearScheduler(
self.hyperparameters['initial_exploration'],
self.hyperparameters['final_exploration'],
0,
train_steps - self.hyperparameters['replay_start_size'],
name="exploration",
writer=writer
),
discount_factor=self.hyperparameters['discount_factor'] ** self.hyperparameters["n_steps"],
minibatch_size=self.hyperparameters['minibatch_size'],
replay_start_size=self.hyperparameters['replay_start_size'],
update_frequency=self.hyperparameters['update_frequency'],
writer=writer,
),
lazy_frames=True,
episodic_lives=True
)
def agent(self, writer=DummyWriter(), train_steps=float("inf")):
# optimizers
feature_optimizer = Adam(self.feature_model.parameters(), lr=self.hyperparameters["lr"], eps=self.hyperparameters["eps"])
value_optimizer = Adam(self.value_model.parameters(), lr=self.hyperparameters["lr"], eps=self.hyperparameters["eps"])
reward_optimizer = Adam(self.reward_model.parameters(), lr=self.hyperparameters["lr"], eps=self.hyperparameters["eps"])
generator_optimizer = Adam(self.generator_model.parameters(), lr=self.hyperparameters["lr"], eps=self.hyperparameters["eps"])
# approximators
f = FeatureNetwork(self.feature_model, feature_optimizer, writer=writer)
v = VNetwork(self.value_model, value_optimizer, writer=writer)
r = QNetwork(self.reward_model, reward_optimizer, name="reward", writer=writer)
g = Approximation(self.generator_model, generator_optimizer, name="generator", writer=writer)
# replay buffer
replay_buffer = ExperienceReplayBuffer(self.hyperparameters["replay_buffer_size"], device=self.device)
# create agent
agent = ModelBasedDQN(f, v, r, g, replay_buffer,
minibatch_size=self.hyperparameters["minibatch_size"],
replay_start_size=self.hyperparameters["replay_start_size"]
)
# apply atari wrappers for better performance
return DeepmindAtariBody(agent, lazy_frames=True)
def test_agent(self):
features = FeatureNetwork(copy.deepcopy(self.feature_model))
policy = SoftmaxPolicy(copy.deepcopy(self.policy_model))
return DeepmindAtariBody(VACTestAgent(features, policy))