def forward(self, obs):
h = self.pre_fc(obs)
mu, log_std = self.fc(h).chunk(2, dim=-1)
# constrain log_std inside [log_std_min, log_std_max]
log_std = torch.tanh(log_std)
log_std_min, log_std_max = self.log_std_bounds
log_std = log_std_min + 0.5 * (log_std_max - log_std_min) * (log_std +
1)
std = log_std.exp()
dist = utils.SquashedNormal(mu, std)
return dist
def forward(self, obs, detach_encoder=False):
obs = self.encoder(obs, detach=detach_encoder)
mu, log_std = self.trunk(obs).chunk(2, dim=-1)
# constrain log_std inside [log_std_min, log_std_max]
log_std = torch.tanh(log_std)
log_std_min, log_std_max = self.log_std_bounds
log_std = log_std_min + 0.5 * (log_std_max - log_std_min) * (log_std + 1)
std = log_std.exp()
self.outputs['mu'] = mu
self.outputs['std'] = std
dist = utils.SquashedNormal(mu, std)
return dist
def forward(self, observation, desired_goal, detach_encoder=False):
observation = self.encoder(observation,
desired_goal,
detach=detach_encoder)
mu, log_std = self.trunk(observation).chunk(2, dim=-1)
# constrain log_std inside [log_std_min, log_std_max]
log_std = torch.tanh(log_std)
log_std_min, log_std_max = self.log_std_bounds
log_std = log_std_min + 0.5 * (log_std_max - log_std_min) * (log_std +
1)
std = log_std.exp()
dist = utils.SquashedNormal(mu, std)
return dist
def noise(self, obs, n, detach_encoder=False):
obs = self.encoder(obs, detach=detach_encoder)
obs[0][np.random.choice(range(len(obs[0])), n, replace=False)] = 0
mu, log_std = self.trunk(obs).chunk(2, dim=-1)
# constrain log_std inside [log_std_min, log_std_max]
log_std = torch.tanh(log_std)
log_std_min, log_std_max = self.log_std_bounds
log_std = log_std_min + 0.5 * (log_std_max - log_std_min) * (log_std +
1)
std = log_std.exp()
self.outputs['mu'] = mu
self.outputs['std'] = std
dist = utils.SquashedNormal(mu, std)
return dist
def forward(self, obs, detach_encoder=False):
img, robot_state = split_obs(obs)
conv_features = self.encoder(img, detach=detach_encoder)
robot_state_out = self.state_input(robot_state)
h1 = torch.cat((conv_features, robot_state_out), 1)
mu, log_std = self.trunk(h1).chunk(2, dim=-1)
# constrain log_std inside [log_std_min, log_std_max]
log_std = torch.tanh(log_std)
log_std_min, log_std_max = self.log_std_bounds
log_std = log_std_min + 0.5 * (log_std_max - log_std_min) * (log_std +
1)
std = log_std.exp()
self.outputs['mu'] = mu
self.outputs['std'] = std
dist = utils.SquashedNormal(mu, std)
return dist
def forward(self, obs, state_goals=None, detach_encoder=False):
obs = self.encoder(obs, detach=detach_encoder)
if state_goals is not None:
if len(state_goals.shape) > len(obs.shape):
state_goals = torch.squeeze(state_goals, 1)
obs = torch.cat((obs, state_goals), axis=-1)
mu, log_std = self.trunk(obs).chunk(2, dim=-1)
# constrain log_std inside [log_std_min, log_std_max]
log_std = torch.tanh(log_std)
log_std_min, log_std_max = self.log_std_bounds
log_std = log_std_min + 0.5 * (log_std_max - log_std_min) * (log_std +
1)
std = log_std.exp()
self.outputs['mu'] = mu
self.outputs['std'] = std
dist = utils.SquashedNormal(mu, std)
return dist