def __init__(self, observation_dim: int, action_dim: int, ae_noise: float = 0.2, hidden_dim: int = 64,

hypo_dim: int = 64, eps: float = 0.03, **kwargs, ):

super(MemFFW, self).__init__()

self.ae_noise = ae_noise

self.state_dim = hypo_dim

self.eps = eps

self.encoder = nn.Sequential(nn.Linear(observation_dim + action_dim + 1, hidden_dim), nn.Sigmoid(),

nn.Linear(hidden_dim, hidden_dim), nn.Sigmoid(), nn.Linear(hidden_dim, hypo_dim), )

self.decoder = nn.Sequential(nn.Linear(observation_dim + action_dim + 1 + hypo_dim, hidden_dim), nn.Sigmoid(),

nn.Linear(hidden_dim, hidden_dim), nn.Sigmoid(),

nn.Linear(hidden_dim, observation_dim + action_dim + 1), )

def _mk_input(self, observation, action, reward):

x = torch.cat((observation, action, reward.reshape(-1, 1)), dim=-1)

return x

def encode(self, observation, action, reward, prev_vec=None):

x = self._mk_input(observation, action, reward)

encoded = F.normalize(self.encoder(x), eps=self.eps, dim=-1)

encoded = encoded.mean(dim=-2, keepdim=False).reshape(-1)

if prev_vec is not None:

prev_vec = prev_vec.detach().reshape(-1)

encoded = F.normalize(encoded + prev_vec, eps=self.eps, dim=-1)

return encoded

def sample_encode(self, observation, action, reward, mini_batch_size: int, prev_vec=None):

bsz = observation.size(0)

if mini_batch_size is None:

mini_batch_size = bsz // 2

x = self._mk_input(observation, action, reward)

indices = torch.randint(0, bsz, size=(bsz, mini_batch_size), device=x.device)

mini_batch_indices = (indices - torch.min(indices, dim=-1)[0][..., None] + torch.arange(bsz, device=x.device)[

..., None]) % bsz

mini_batches = x[mini_batch_indices]

encoded = F.normalize(self.encoder(mini_batches), eps=self.eps, dim=-1)

encoded = encoded.mean(dim=-2, keepdim=False).reshape(bsz, self.state_dim)

if prev_vec is not None:

assert (prev_vec.shape == (bsz, self.state_dim) or prev_vec.shape == (

self.state_dim,)), f"prev_vec shape not match!: {prev_vec.shape}"

encoded = F.normalize(encoded + prev_vec, eps=self.eps, dim=-1)

return encoded

def decode(self, observation, action, reward, prev_vec):

observations = self._mk_input(observation, action, reward)

bsz = observation.size(0)

if prev_vec.shape != (bsz, self.state_dim):

prev_vec = prev_vec.reshape(1, -1).expand(bsz, -1)

original_x = observations

noise_range = original_x.max(dim=0).values - original_x.min(dim=0).values

noise = torch.randn_like(original_x) * noise_range * self.ae_noise

x = original_x + noise

x = torch.cat((x, prev_vec), dim=-1)

denoised_x = self.decoder(x)

def __init__(self, state_dim, action_dim, max_action, hidden_dim: int = 64, hypo_dim: int = 64, ):

super(Actor, self).__init__()

self.hypo_dim = hypo_dim

self.l1 = nn.Linear(state_dim + hypo_dim, hidden_dim)

self.l2 = nn.Linear(hidden_dim, hidden_dim)

self.l3 = nn.Linear(hidden_dim, action_dim)

self.max_action = max_action

def forward(self, state, prev_vec):

bsz = state.shape[0]

if prev_vec.shape != (bsz, self.hypo_dim):

prev_vec = prev_vec.reshape(1, -1).expand(bsz, -1)

to_cat = [state, prev_vec]

state = torch.cat(to_cat, 1)

a = F.relu(self.l1(state))

a = F.relu(self.l2(a))

def __init__(self, state_dim, action_dim, hidden_dim: int = 64, hypo_dim: int = 64, ):

super(Critic, self).__init__()

self.hypo_dim = hypo_dim

# Q1 architecture

self.l1 = nn.Linear(state_dim + action_dim + hypo_dim, hidden_dim)

self.l2 = nn.Linear(hidden_dim, hidden_dim)

self.l3 = nn.Linear(hidden_dim, 1)

# Q2 architecture

self.l4 = nn.Linear(state_dim + action_dim + hypo_dim, hidden_dim)

self.l5 = nn.Linear(hidden_dim, hidden_dim)

self.l6 = nn.Linear(hidden_dim, 1)

def forward(self, state, action, prev_vec):

bsz = state.shape[0]

if prev_vec.shape != (bsz, self.hypo_dim):

prev_vec = prev_vec.reshape(1, -1).expand(bsz, -1)

to_cat = [state, action, prev_vec]

sa = torch.cat(to_cat, 1)

q1 = F.relu(self.l1(sa))

q1 = F.relu(self.l2(q1))

q1 = self.l3(q1)

q2 = F.relu(self.l4(sa))

q2 = F.relu(self.l5(q2))

q2 = self.l6(q2)

return q1, q2

def Q1(self, state, action, prev_vec):

bsz = state.shape[0]

if prev_vec.shape != (bsz, self.hypo_dim):

prev_vec = prev_vec.reshape(1, -1).expand(bsz, -1)

to_cat = [state, action, prev_vec]

sa = torch.cat(to_cat, 1)

q1 = F.relu(self.l1(sa))

q1 = F.relu(self.l2(q1))

q1 = self.l3(q1)

def __init__(self, state_dim, action_dim, max_action, discount=0.99, tau=0.005, policy_noise=0.2, noise_clip=0.5,

policy_freq=2, device: str = 'cpu', hidden_dim: int = 64, hypo_dim: int = 64, mini_batch_size=None,

**kwargs):

self.device = device

self.mini_batch_size = mini_batch_size

self.state_dim = state_dim

self.action_dim = action_dim

# M and \phi network

self.mem = MemFFW(state_dim, action_dim, hidden_dim=hidden_dim, hypo_dim=hypo_dim).to(device)

self.mem_optimizer = torch.optim.Adam(self.mem.parameters(), lr=3e-4)

# hypothesis-augmented actor/critic

self.actor = Actor(state_dim, action_dim, max_action, hidden_dim=hidden_dim, hypo_dim=hypo_dim).to(device)

self.actor_target = copy.deepcopy(self.actor)

self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=3e-4)

self.critic = Critic(state_dim, action_dim, hidden_dim=hidden_dim, hypo_dim=hypo_dim).to(device)

self.critic_target = copy.deepcopy(self.critic)

self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=3e-4)

# h_0

self.initial_state = torch.nn.Parameter(torch.zeros((hypo_dim,), dtype=torch.float32, device=device))

self.initial_state_optimizer = torch.optim.Adam([self.initial_state], lr=3e-4)

# other hyper params

self.max_action = max_action

self.discount = discount

self.tau = tau

self.policy_noise = policy_noise

self.noise_clip = noise_clip

self.policy_freq = policy_freq

self.total_it = 0

self._test_state = None

def train_mem_step(self, observation, action, reward) -> dict:

""" Optimize weak model"""

o_1_size = np.random.randint(1, observation.shape[0] - 1)

metrics = {}

batch = {'observation': observation, 'action': action, 'reward': reward, }

o_1 = {'observation': batch['observation'][:o_1_size], 'action': batch['action'][:o_1_size],

'reward': batch['reward'][:o_1_size], }

h_1 = self.mem.encode(**o_1, prev_vec=None)

_, loss_h_1 = self.mem.decode(**o_1, prev_vec=h_1)

o_2 = {'observation': batch['observation'][o_1_size:], 'action': batch['action'][o_1_size:],

'reward': batch['reward'][o_1_size:], }

h = self.mem.encode(**o_2, prev_vec=h_1)

_, loss_h = self.mem.decode(**batch, prev_vec=h)

diversity_loss = (-F.mse_loss(h_1, self.initial_state.detach()) - F.mse_loss(h, self.initial_state.detach()))

loss = (loss_h_1 + loss_h) + diversity_loss

metrics.update(

{'internal_mem_loss': loss, 'loss_h_1': loss_h_1, 'loss_h': loss_h, 'diversity_loss': diversity_loss})

# optimize D (\phi) and M (incremental fashion)

self.mem_optimizer.zero_grad()

loss.backward()

self.mem_optimizer.step()

# optimize h_0

_, D_train_mem_loss = self.mem.decode(**batch, prev_vec=self.initial_state)

self.initial_state_optimizer.zero_grad()

D_train_mem_loss.backward()

self.initial_state_optimizer.step()

metrics.update({'D_train_mem_loss': D_train_mem_loss, })

return metrics

def forget(self):

self._test_state = None

@property

def prev_state(self):

if self._test_state is None:

# initialize h'_0 as h_0

self._test_state = self.initial_state.detach()

return self._test_state

@prev_state.setter

def prev_state(self, value):

self._test_state = value

def watch(self, observation, action, reward) -> dict:

if not isinstance(observation, torch.Tensor):

observation = torch.tensor(observation, device=self.device, dtype=torch.float32)

observation = observation.reshape(-1, self.state_dim)

if not isinstance(action, torch.Tensor):

action = torch.tensor(action, device=self.device, dtype=torch.float32)

action = action.reshape(-1, self.action_dim)

if not isinstance(reward, torch.Tensor):

reward = torch.tensor(reward, device=self.device, dtype=torch.float32)

reward = reward.reshape(-1)

with torch.no_grad():

# update h'_0 with o_test

self.prev_state = self.mem.encode(observation, action, reward, self.prev_state)

return {}

def select_action(self, observation, return_batch=False):

if not isinstance(observation, torch.Tensor):

observation = torch.tensor(observation, device=self.device, dtype=torch.float32)

observation = observation.reshape(-1, self.state_dim)

out = self.actor(observation, getattr(self, 'prev_state', None)).cpu().data.numpy()

if return_batch:

return out

return out.flatten()

def train(self, replay_buffer, batch_size=256):

self.total_it += 1

# Sample replay buffer

state, action, next_state, reward, not_done = replay_buffer.sample(batch_size)

if self.total_it % 10 == 0:

self.train_mem_step(state, action, reward)

with torch.no_grad():

mini_batch_size = self.mini_batch_size if self.mini_batch_size is not None else batch_size // 2

hypothesis = self.mem.sample_encode(state, action, reward, mini_batch_size=mini_batch_size,

prev_vec=None)

with torch.no_grad():

# Select action according to policy and add clipped noise

noise = (torch.randn_like(action) * self.policy_noise).clamp(-self.noise_clip, self.noise_clip)

next_action = (self.actor_target(next_state, prev_vec=hypothesis) + noise).clamp(-self.max_action,

self.max_action)

# Compute the target Q value

target_Q1, target_Q2 = self.critic_target(next_state, next_action, prev_vec=hypothesis)

target_Q = torch.min(target_Q1, target_Q2)

target_Q = reward + not_done * self.discount * target_Q

# Get current Q estimates

current_Q1, current_Q2 = self.critic(state, action, prev_vec=hypothesis)

# Compute critic loss

critic_loss = F.mse_loss(current_Q1, target_Q) + F.mse_loss(current_Q2, target_Q)

# Optimize the critic

self.critic_optimizer.zero_grad()

critic_loss.backward()

self.critic_optimizer.step()

# Delayed policy updates

if self.total_it % self.policy_freq == 0:

# Compute actor loss

actor_loss = -self.critic.Q1(state, self.actor(state, prev_vec=hypothesis), prev_vec=hypothesis).mean()

# Optimize the actor

self.actor_optimizer.zero_grad()

actor_loss.backward()

self.actor_optimizer.step()

# Update the frozen target models

for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()):

target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)

for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()):

target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)

def save(self, filename):

torch.save(self.critic.state_dict(), filename + "_critic")

torch.save(self.critic_optimizer.state_dict(), filename + "_critic_optimizer")

torch.save(self.actor.state_dict(), filename + "_actor")

torch.save(self.actor_optimizer.state_dict(), filename + "_actor_optimizer")

torch.save(self.mem.state_dict(), filename + "_mem")

torch.save(self.mem_optimizer.state_dict(), filename + "_mem_optimizer")

torch.save(self.initial_state, filename + "_initial_state")

torch.save(self.initial_state_optimizer.state_dict(), filename + "_initial_state_optimizer")

def load(self, filename):

self.critic.load_state_dict(torch.load(filename + "_critic"))

self.critic_optimizer.load_state_dict(torch.load(filename + "_critic_optimizer"))

self.critic_target = copy.deepcopy(self.critic)

self.actor.load_state_dict(torch.load(filename + "_actor"))

self.actor_optimizer.load_state_dict(torch.load(filename + "_actor_optimizer"))

self.actor_target = copy.deepcopy(self.actor)

self.mem.load_state_dict(torch.load(filename + "_mem"))

self.mem_optimizer.load_state_dict(torch.load(filename + "_mem_optimizer"))

self.initial_state = torch.load(filename + "_initial_state").to(self.device)