def forward(
self,
obs,
reparameterize=True,
deterministic=False,
return_log_prob=False,
):
"""
:param obs: Observation
:param deterministic: If True, do not sample
:param return_log_prob: If True, return a sample and its log probability
"""
h = obs
# import pdb; pdb.set_trace()
for i, fc in enumerate(self.fcs):
h = self.hidden_activation(fc(h))
mean = self.last_fc(h)
if self.std is None:
log_std = self.last_fc_log_std(h)
log_std = torch.clamp(log_std, LOG_SIG_MIN, LOG_SIG_MAX)
std = torch.exp(log_std)
else:
std = self.std
log_std = self.log_std
log_prob = None
entropy = None
mean_action_log_prob = None
pre_tanh_value = None
if deterministic:
action = torch.tanh(mean)
else:
tanh_normal = TanhNormal(mean, std)
if return_log_prob:
if reparameterize is True:
action, pre_tanh_value = tanh_normal.rsample(
return_pretanh_value=True)
else:
action, pre_tanh_value = tanh_normal.sample(
return_pretanh_value=True)
log_prob = tanh_normal.log_prob(action,
pre_tanh_value=pre_tanh_value)
log_prob = log_prob.sum(dim=1, keepdim=True)
else:
if reparameterize is True:
action = tanh_normal.rsample()
else:
action = tanh_normal.sample()
return (
action,
mean,
log_std,
log_prob,
entropy,
std,
mean_action_log_prob,
pre_tanh_value,
)
def forward(
self,
obs,
reparameterize=False,
deterministic=False,
return_log_prob=False,
):
"""
:param obs: Observation
:param deterministic: If True, do not sample
:param return_log_prob: If True, return a sample and its log probability
"""
t, b, _ = obs.size()
h = obs
h = self.inner_forward(h)
mean = self.last_fc(h)
if self.std is None:
log_std = self.last_fc_log_std(h)
log_std = torch.clamp(log_std, LOG_SIG_MIN, LOG_SIG_MAX)
std = torch.exp(log_std)
else:
std = self.std
log_std = self.log_std
log_prob = None
expected_log_prob = None
mean_action_log_prob = None
pre_tanh_value = None
if deterministic:
action = torch.tanh(mean)
else:
tanh_normal = TanhNormal(mean, std)
if return_log_prob:
if reparameterize:
action, pre_tanh_value = tanh_normal.rsample(
return_pretanh_value=True)
else:
action, pre_tanh_value = tanh_normal.sample(
return_pretanh_value=True)
log_prob = tanh_normal.log_prob(action,
pre_tanh_value=pre_tanh_value)
log_prob = log_prob.sum(dim=1, keepdim=True)
else:
if reparameterize:
action = tanh_normal.rsample()
else:
action = tanh_normal.sample()
return (
action,
mean,
log_std,
log_prob,
expected_log_prob,
std,
mean_action_log_prob,
pre_tanh_value,
)
def forward(
self,
obs,
reparameterize=reparameterize,
deterministic=False,
):
"""
:param obs: Observation
:param deterministic: If True, do not sample
:param return_log_prob: If True, return a sample and its log probability
"""
h = obs
for i, fc in enumerate(self.fcs):
h = self.hidden_activation(fc(h))
mean = self.last_fc(h)
if self.std is None:
log_std = self.last_fc_log_std(h)
log_std = torch.clamp(log_std, LOGMIN, LOGMAX)
std = torch.exp(log_std)
else:
std = self.std
log_std = self.log_std
log_prob = None
expected_log_prob = None
mean_action_log_prob = None
pre_tanh_value = None
tanh_normal = TanhNormal(mean, std)
if reparameterize:
action, pre_tanh_value = tanh_normal.rsample(
return_pretanh_value=True)
else:
action, pre_tanh_value = tanh_normal.sample(
return_pretanh_value=True)
log_prob = tanh_normal.log_prob(action, pre_tanh_value=pre_tanh_value)
log_prob = log_prob.sum(dim=1, keepdim=True)
return action.cuda(), log_prob.cuda(), pre_tanh_value.cuda(
), mean, log_std
def forward(
self,
obs,
reparameterize=True,
deterministic=False,
return_log_prob=False,
):
"""
:param obs: Observation
:param deterministic: If True, do not sample
:param return_log_prob: If True, return a sample and its log probability
"""
h = obs
for i, fc in enumerate(self.fcs):
h = self.hidden_activation(fc(h))
mean = self.last_fc(h) # actions * heads
if self.std is None:
log_std = self.last_fc_log_std(h)
log_std = torch.clamp(log_std, LOG_SIG_MIN, LOG_SIG_MAX)
std = torch.exp(log_std)
else:
std = self.std
log_std = self.log_std
log_prob = None
entropy = None
mean_action_log_prob = None
pre_tanh_value = None
log_stds = None
log_probs = None
if deterministic:
means = mean.view(-1, self.action_dim)
actions = torch.tanh(means)
actions = actions.view(-1, self.heads, self.action_dim)
else:
all_actions = []
means = mean.view(-1, self.action_dim)
stds = std.view(-1, self.action_dim)
log_stds = log_std.view(-1, self.action_dim)
tanh_normal = TanhNormal(means, stds)
if return_log_prob:
if reparameterize is True:
action, pre_tanh_value = tanh_normal.rsample(
return_pretanh_value=True)
else:
action, pre_tanh_value = tanh_normal.sample(
return_pretanh_value=True)
log_prob = tanh_normal.log_prob(action,
pre_tanh_value=pre_tanh_value)
log_prob = log_prob.sum(dim=1, keepdim=True)
log_probs = log_prob.view(-1, self.heads, self.action_dim)
else:
if reparameterize is True:
action = tanh_normal.rsample()
else:
action = tanh_normal.sample()
actions = action.view(-1, self.heads, self.action_dim)
return (
actions,
means,
log_stds,
log_probs,
entropy,
std,
mean_action_log_prob,
pre_tanh_value,
)
def forward(
self,
obs,
reparameterize=True,
deterministic=False,
return_log_prob=False,
):
"""
:param obs: Observation
:param deterministic: If True, do not sample
:param return_log_prob: If True, return a sample and its log probability
"""
h = obs
for i, fc in enumerate(self.fcs):
h = self.hidden_activation(fc(h))
out = self.last_fc(h).view(-1, self.k, (2 * self.action_dim + 1))
log_w = out[..., 0]
mean = out[..., 1:1 + self.action_dim]
log_std = out[..., 1 + self.action_dim:]
log_std = torch.clamp(log_std, LOG_SIG_MIN, LOG_SIG_MAX)
log_w = torch.clamp(log_w, min=LOG_W_MIN)
self.log_w = log_w
self.mean = mean
self.log_std = log_std
std = torch.exp(log_std)
log_prob = None
entropy = None
mean_action_log_prob = None
pre_tanh_value = None
arange = torch.arange(out.shape[0])
if deterministic:
ks = log_w.view(-1, self.k).argmax(1)
action = torch.tanh(mean[arange, ks])
else:
sample_ks = Categorical(logits=log_w.view(-1, self.k)).sample()
tanh_normal = TanhNormal(mean[arange, sample_ks], std[arange,
sample_ks])
if return_log_prob:
action, pre_tanh_value = tanh_normal.sample(
return_pretanh_value=True)
# (NxKxA), (NxKxA), (Nx1xA) => (NxK)
log_p_xz_t = log_gaussian(mean, log_std,
pre_tanh_value[:, None, :].data)
log_p_x_t = torch.logsumexp(log_p_xz_t + log_w,
1) - torch.logsumexp(log_w, 1)
# squash correction
log_prob = log_p_x_t - torch.log(1 - action**2 + 1e-6).sum(1)
log_prob = log_prob[:, None]
else:
if reparameterize is True:
action = tanh_normal.rsample()
else:
action = tanh_normal.sample()
return (
action,
mean,
log_std,
log_prob,
entropy,
std,
mean_action_log_prob,
pre_tanh_value,
)
def forward(
self,
obs,
reparameterize=True,
deterministic=False,
return_log_prob=False,
):
"""
:param obs: Observation
:param deterministic: If True, do not sample
:param return_log_prob: If True, return a sample and its log probability
"""
# This is a bit messed up TODO clean it
if obs.shape[0] == 1: # if obs is single image: flatten
h = torch.flatten(obs)
h = h.view(1, -1)
else: # else if obs comes from replay buffer --> it is already flat and comes in a batch --> DO NOT flatten!
h = obs
h = super().forward(h, None, complete=False)
mean = self.last_fc(h)
if self.std is None:
log_std = self.last_fc_log_std(h)
log_std = torch.clamp(log_std, LOG_SIG_MIN, LOG_SIG_MAX)
std = torch.exp(log_std)
else:
std = self.std
log_std = self.log_std
log_prob = None
entropy = None
mean_action_log_prob = None
pre_tanh_value = None
if deterministic:
action = torch.tanh(mean)
else:
tanh_normal = TanhNormal(mean, std)
if return_log_prob:
if reparameterize is True:
action, pre_tanh_value = tanh_normal.rsample(
return_pretanh_value=True)
else:
action, pre_tanh_value = tanh_normal.sample(
return_pretanh_value=True)
log_prob = tanh_normal.log_prob(action,
pre_tanh_value=pre_tanh_value)
log_prob = log_prob.sum(dim=1, keepdim=True)
else:
if reparameterize is True:
action = tanh_normal.rsample()
else:
action = tanh_normal.sample()
return (
action,
mean,
log_std,
log_prob,
entropy,
std,
mean_action_log_prob,
pre_tanh_value,
)
def forward(
self,
meta_size,
batch_size,
obs,
reparameterize=False,
deterministic=False,
return_log_prob=False,
):
"""
:param obs: Observation
:param deterministic: If True, do not sample
:param return_log_prob: If True, return a sample and its log probability
"""
h = obs
for i, fc in enumerate(self.fcs):
h = fc(h)
if self.layer_norm and i < len(self.fcs) - 1:
h = self.layer_norms[i](h)
h = self.hidden_activation(h)
if self.use_dropout and i < len(self.fcs) - 1:
h = self.dropouts[i](h)
mean = self.last_fc(h)
if self.std is None:
log_std = self.last_fc_log_std(h)
log_std = torch.clamp(log_std, LOG_SIG_MIN, LOG_SIG_MAX)
std = torch.exp(log_std)
else:
std = self.std
log_std = self.log_std
log_prob = None
expected_log_prob = None
mean_action_log_prob = None
pre_tanh_value = None
if deterministic:
action = torch.tanh(mean)
else:
tanh_normal = TanhNormal(mean, std)
if return_log_prob:
if reparameterize:
action, pre_tanh_value = tanh_normal.rsample(
return_pretanh_value=True
)
else:
action, pre_tanh_value = tanh_normal.sample(
return_pretanh_value=True
)
log_prob = tanh_normal.log_prob(
action,
pre_tanh_value=pre_tanh_value
)
log_prob = log_prob.sum(dim=1, keepdim=True)
else:
if reparameterize:
action = tanh_normal.rsample()
else:
action = tanh_normal.sample()
return (
action, mean, log_std, log_prob, expected_log_prob, std,
mean_action_log_prob, pre_tanh_value,
)
def forward(
self,
obs,
reparameterize=True,
deterministic=False,
return_log_prob=False,
return_entropy=False,
return_log_prob_of_mean=False,
):
"""
:param obs: Observation
:param deterministic: If True, do not sample
:param return_log_prob: If True, return a sample and its log probability
:param return_entropy: If True, return the true expected log
prob. Will not need to be differentiated through, so this can be a
number.
:param return_log_prob_of_mean: If True, return the true expected log
prob. Will not need to be differentiated through, so this can be a
number.
"""
h = self.obs_processor(obs)
h = self.mean_and_log_std_net(h)
mean, log_std = torch.split(h, self.action_dim, dim=1)
log_std = torch.clamp(log_std, LOG_SIG_MIN, LOG_SIG_MAX)
std = torch.exp(log_std)
log_prob = None
entropy = None
mean_action_log_prob = None
pre_tanh_value = None
if deterministic:
action = torch.tanh(mean)
else:
tanh_normal = TanhNormal(mean, std)
if return_log_prob:
if reparameterize is True:
action, pre_tanh_value = tanh_normal.rsample(
return_pretanh_value=True)
else:
action, pre_tanh_value = tanh_normal.sample(
return_pretanh_value=True)
log_prob = tanh_normal.log_prob(action,
pre_tanh_value=pre_tanh_value)
log_prob = log_prob.sum(dim=1, keepdim=True)
else:
if reparameterize is True:
action = tanh_normal.rsample()
else:
action = tanh_normal.sample()
if return_entropy:
entropy = log_std + 0.5 + np.log(2 * np.pi) / 2
# I'm not sure how to compute the (differential) entropy for a
# tanh(Gaussian)
entropy = entropy.sum(dim=1, keepdim=True)
raise NotImplementedError()
if return_log_prob_of_mean:
tanh_normal = TanhNormal(mean, std)
mean_action_log_prob = tanh_normal.log_prob(
torch.tanh(mean),
pre_tanh_value=mean,
)
mean_action_log_prob = mean_action_log_prob.sum(dim=1,
keepdim=True)
return (
action,
mean,
log_std,
log_prob,
entropy,
std,
mean_action_log_prob,
pre_tanh_value,
)