def __init__(self,
optimizer=None,
optimizer_args=None,
observation_permutation=None,
action_permutation=None,
sym_loss_weight=0.0001,
clip_param=0.2,
adam_batchsize=128,
adam_epochs=10,
**kwargs):
if optimizer is None:
if optimizer_args is None:
optimizer_args = dict()
optimizer = LbfgsOptimizer(**optimizer_args)
super(PPO_Clip_Sym, self).__init__(optimizer=optimizer, **kwargs)
self.observation_permutation = observation_permutation
self.action_permutation = action_permutation
self.sym_loss_weight = sym_loss_weight
self.clip_param = clip_param
self.adam_batchsize = adam_batchsize
self.adam_epochs = adam_epochs
self.obs_perm_mat = np.zeros(
(len(observation_permutation), len(observation_permutation)))
self.act_per_mat = np.zeros(
(len(action_permutation), len(action_permutation)))
for i, perm in enumerate(self.observation_permutation):
self.obs_perm_mat[i][int(np.abs(perm))] = np.sign(perm)
for i, perm in enumerate(self.action_permutation):
self.act_per_mat[i][int(np.abs(perm))] = np.sign(perm)
def __init__(self,
optimizer=None,
optimizer_args=None,
positive_adv=None,
**kwargs):
Serializable.quick_init(self, locals())
if optimizer is None:
if optimizer_args is None:
optimizer_args = dict()
optimizer = LbfgsOptimizer(**optimizer_args)
super(ERWR, self).__init__(
optimizer=optimizer,
positive_adv=True if positive_adv is None else positive_adv,
**kwargs)
def __init__(
self,
input_shape,
output_dim,
mean_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
optimizer=None,
use_trust_region=True,
step_size=0.01,
learn_std=True,
init_std=1.0,
adaptive_std=False,
std_share_network=False,
std_hidden_sizes=(32, 32),
std_nonlinearity=None,
normalize_inputs=True,
normalize_outputs=True,
name=None,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean network.
:param hidden_nonlinearity: Non-linearity used for each layer of the mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param step_size: KL divergence constraint for each iteration
:param learn_std: Whether to learn the standard deviations. Only effective if adaptive_std is False. If
adaptive_std is True, this parameter is ignored, and the weights for the std network are always learned.
:param adaptive_std: Whether to make the std a function of the states.
:param std_share_network: Whether to use the same network as the mean.
:param std_hidden_sizes: Number of hidden units of each layer of the std network. Only used if
`std_share_network` is False. It defaults to the same architecture as the mean.
:param std_nonlinearity: Non-linearity used for each layer of the std network. Only used if `std_share_network`
is False. It defaults to the same non-linearity as the mean.
"""
Serializable.quick_init(self, locals())
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer()
else:
optimizer = LbfgsOptimizer()
self._optimizer = optimizer
if mean_network is None:
mean_network = MLP(
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=None,
)
l_mean = mean_network.output_layer
if adaptive_std:
l_log_std = MLP(
input_shape=input_shape,
input_var=mean_network.input_layer.input_var,
output_dim=output_dim,
hidden_sizes=std_hidden_sizes,
hidden_nonlinearity=std_nonlinearity,
output_nonlinearity=None,
).output_layer
else:
l_log_std = ParamLayer(
mean_network.input_layer,
num_units=output_dim,
param=lasagne.init.Constant(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
LasagnePowered.__init__(self, [l_mean, l_log_std])
xs_var = mean_network.input_layer.input_var
ys_var = TT.matrix("ys")
old_means_var = TT.matrix("old_means")
old_log_stds_var = TT.matrix("old_log_stds")
x_mean_var = theano.shared(np.zeros((1, ) + input_shape),
name="x_mean",
broadcastable=(True, ) +
(False, ) * len(input_shape))
x_std_var = theano.shared(np.ones((1, ) + input_shape),
name="x_std",
broadcastable=(True, ) +
(False, ) * len(input_shape))
y_mean_var = theano.shared(np.zeros((1, output_dim)),
name="y_mean",
broadcastable=(True, False))
y_std_var = theano.shared(np.ones((1, output_dim)),
name="y_std",
broadcastable=(True, False))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
normalized_ys_var = (ys_var - y_mean_var) / y_std_var
normalized_means_var = L.get_output(
l_mean, {mean_network.input_layer: normalized_xs_var})
normalized_log_stds_var = L.get_output(
l_log_std, {mean_network.input_layer: normalized_xs_var})
means_var = normalized_means_var * y_std_var + y_mean_var
log_stds_var = normalized_log_stds_var + TT.log(y_std_var)
normalized_old_means_var = (old_means_var - y_mean_var) / y_std_var
normalized_old_log_stds_var = old_log_stds_var - TT.log(y_std_var)
dist = self._dist = DiagonalGaussian()
normalized_dist_info_vars = dict(mean=normalized_means_var,
log_std=normalized_log_stds_var)
mean_kl = TT.mean(
dist.kl_sym(
dict(mean=normalized_old_means_var,
log_std=normalized_old_log_stds_var),
normalized_dist_info_vars,
))
loss = -TT.mean(
dist.log_likelihood_sym(normalized_ys_var,
normalized_dist_info_vars))
self._f_predict = compile_function([xs_var], means_var)
self._f_pdists = compile_function([xs_var], [means_var, log_stds_var])
self._l_mean = l_mean
self._l_log_std = l_log_std
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[normalized_means_var, normalized_log_stds_var],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [
xs_var, ys_var, old_means_var, old_log_stds_var
]
else:
optimizer_args["inputs"] = [xs_var, ys_var]
self._optimizer.update_opt(**optimizer_args)
self._use_trust_region = use_trust_region
self._name = name
self._normalize_inputs = normalize_inputs
self._normalize_outputs = normalize_outputs
self._x_mean_var = x_mean_var
self._x_std_var = x_std_var
self._y_mean_var = y_mean_var
self._y_std_var = y_std_var
def __init__(
self,
input_shape,
output_dim,
predict_all=False, # CF
prob_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
optimizer=None,
use_trust_region=True,
step_size=0.01,
normalize_inputs=True,
name=None,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean network.
:param hidden_nonlinearity: Non-linearity used for each layer of the mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer()
else:
optimizer = LbfgsOptimizer()
self.output_dim = output_dim
self._optimizer = optimizer
if prob_network is None:
prob_network = GRUNetwork(
input_shape=input_shape,
output_dim=output_dim,
hidden_dim=hidden_sizes[0], # this gives 32 by default
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=NL.softmax,
)
l_prob = prob_network.output_layer
LasagnePowered.__init__(self, [l_prob])
xs_var = prob_network.input_layer.input_var
ys_var = TT.itensor3("ys")
old_prob_var = TT.tensor3("old_prob")
x_mean_var = theano.shared(
np.zeros(
(
1,
1,
) + input_shape
), # this syntax makes the shape (1,1,*input_shape,). The first is traj
name="x_mean",
broadcastable=(
True,
True,
) + (False, ) * len(input_shape))
x_std_var = theano.shared(np.ones((
1,
1,
) + input_shape),
name="x_std",
broadcastable=(
True,
True,
) + (False, ) * len(input_shape))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
prob_var_all = L.get_output(
l_prob, {prob_network.input_layer: normalized_xs_var})
if predict_all:
prob_var = prob_var_all
else:
# take only last dim but keep the shape
prob_var_last = TT.reshape(
prob_var_all[:, -1, :],
(TT.shape(prob_var_all)[0], 1, TT.shape(prob_var_all)[2]))
# padd along the time dimension to obtain the same shape as before
padded_prob_var = TT.tile(prob_var_last,
(1, TT.shape(prob_var_all)[1], 1))
# give it the standard name
prob_var = padded_prob_var
old_info_vars = dict(prob=old_prob_var)
info_vars = dict(prob=prob_var)
dist = self._dist = Categorical(output_dim)
mean_kl = TT.mean(dist.kl_sym(old_info_vars, info_vars))
loss = -TT.mean(dist.log_likelihood_sym(ys_var, info_vars))
predicted_flat = special.to_onehot_sym(
TT.flatten(TT.argmax(prob_var, axis=-1)), output_dim)
predicted = TT.reshape(predicted_flat, TT.shape(prob_var))
self._f_predict = ext.compile_function([xs_var], predicted)
self._f_prob = ext.compile_function([xs_var], prob_var)
self._prob_network = prob_network
self._l_prob = l_prob
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[prob_var],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [xs_var, ys_var, old_prob_var]
else:
optimizer_args["inputs"] = [xs_var, ys_var]
self._optimizer.update_opt(**optimizer_args)
self._use_trust_region = use_trust_region
self._name = name
self._normalize_inputs = normalize_inputs
self._x_mean_var = x_mean_var
self._x_std_var = x_std_var
def __init__(
self,
input_shape,
output_dim,
prob_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
optimizer=None,
use_trust_region=True,
step_size=0.01,
normalize_inputs=True,
name=None,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean network.
:param hidden_nonlinearity: Non-linearity used for each layer of the mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer()
else:
optimizer = LbfgsOptimizer()
self.output_dim = output_dim
self._optimizer = optimizer
if prob_network is None:
prob_network = MLP(
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=NL.softmax,
)
l_prob = prob_network.output_layer
LasagnePowered.__init__(self, [l_prob])
xs_var = prob_network.input_layer.input_var
ys_var = TT.imatrix("ys")
old_prob_var = TT.matrix("old_prob")
x_mean_var = theano.shared(
np.zeros((1,) + input_shape),
name="x_mean",
broadcastable=(True,) + (False,) * len(input_shape)
)
x_std_var = theano.shared(
np.ones((1,) + input_shape),
name="x_std",
broadcastable=(True,) + (False,) * len(input_shape)
)
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
prob_var = L.get_output(l_prob, {prob_network.input_layer: normalized_xs_var})
old_info_vars = dict(prob=old_prob_var)
info_vars = dict(prob=prob_var)
dist = self._dist = Categorical(output_dim)
mean_kl = TT.mean(dist.kl_sym(old_info_vars, info_vars))
loss = - TT.mean(dist.log_likelihood_sym(ys_var, info_vars))
predicted = special.to_onehot_sym(TT.argmax(prob_var, axis=1), output_dim)
self._f_predict = ext.compile_function([xs_var], predicted)
self._f_prob = ext.compile_function([xs_var], prob_var)
self._prob_network = prob_network
self._l_prob = l_prob
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[prob_var],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [xs_var, ys_var, old_prob_var]
else:
optimizer_args["inputs"] = [xs_var, ys_var]
self._optimizer.update_opt(**optimizer_args)
self._use_trust_region = use_trust_region
self._name = name
self._normalize_inputs = normalize_inputs
self._x_mean_var = x_mean_var
self._x_std_var = x_std_var
raise ValueError('No values provided for param %s' % param_name)
else:
try:
param_values = [[float(y) for y in x.split(',')]
for x in args.param_val.split(',,')]
except:
param_values = [[y for y in x.split(',')]
for x in args.param_val.split(',,')]
############################################################################
## POSTPROCESSING OF PARAMETERS
if args.logdir[-1] != '/':
args.logdir += '/'
if params['hide_baseline_net_params']['optimizer'] == 'LbfgsOptimizer':
params['hide_baseline_net_params']['optimizer'] = LbfgsOptimizer(
max_opt_itr=params['hide_baseline_net_params']['max_opt_itr'])
params['hide_baseline_net_params'].pop('max_opt_itr', None)
else:
raise ValueError('Unknown optimizer: %s',
params['hide_baseline_net_params']['optimizer'])
if params['seek_baseline_net_params']['optimizer'] == 'LbfgsOptimizer':
params['seek_baseline_net_params']['optimizer'] = LbfgsOptimizer(
max_opt_itr=params['seek_baseline_net_params']['max_opt_itr'])
params['seek_baseline_net_params'].pop('max_opt_itr', None)
else:
raise ValueError('Unknown optimizer: %s',
params['seek_baseline_net_params']['optimizer'])
## All possible combinations of hyperparameters
param_values.append(seeds)
def __init__(
self,
input_shape,
output_dim,
predict_all=False,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
optimizer=None,
use_trust_region=True,
step_size=0.01,
normalize_inputs=True,
name=None,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param predict_all: use the prediction made at every step about the latent variables (not only the last step)
:param hidden_sizes: Number of hidden units of each layer of the mean network.
:param hidden_nonlinearity: Non-linearity used for each layer of the mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer()
else:
optimizer = LbfgsOptimizer()
self.output_dim = output_dim
self._optimizer = optimizer
p_network = GRUNetwork(
input_shape=input_shape,
output_dim=output_dim,
hidden_dim=hidden_sizes[0],
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=NL.sigmoid,
)
l_p = p_network.output_layer # this is every intermediate latent state! but I only care about last
LasagnePowered.__init__(self, [l_p])
xs_var = p_network.input_layer.input_var
ys_var = TT.itensor3("ys") # this is 3D: (traj, time, lat_dim)
old_p_var = TT.tensor3("old_p")
x_mean_var = theano.shared(np.zeros((
1,
1,
) + input_shape),
name="x_mean",
broadcastable=(
True,
True,
) + (False, ) * len(input_shape))
x_std_var = theano.shared(np.ones((
1,
1,
) + input_shape),
name="x_std",
broadcastable=(
True,
True,
) + (False, ) * len(input_shape))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
# this is the previous p_var, from which I only want the last time-step padded along all time-steps
p_var_all = L.get_output(l_p,
{p_network.input_layer: normalized_xs_var})
# take only last dim but keep the shape
p_var_last = TT.reshape(
p_var_all[:, -1, :],
(TT.shape(p_var_all)[0], 1, TT.shape(p_var_all)[2]))
# padd along the time dimension to obtain the same shape as before
padded_p = TT.tile(p_var_last, (1, TT.shape(p_var_all)[1], 1))
# give it the standard name
if predict_all:
p_var = p_var_all
else:
p_var = padded_p
old_info_vars = dict(p=old_p_var)
info_vars = dict(
p=p_var
) # posterior of the latent at every step, wrt obs-act. Same along batch if recurrent
dist = self._dist = Bernoulli(output_dim)
mean_kl = TT.mean(dist.kl_sym(old_info_vars, info_vars))
loss = -TT.mean(dist.log_likelihood_sym(
ys_var,
info_vars)) # regressor just wants to min -loglik of data ys
predicted = p_var >= 0.5
self._f_predict = ext.compile_function([xs_var], predicted)
self._f_p = ext.compile_function(
[xs_var], p_var
) # for consistency with gauss_mlp_reg this should be ._f_pdists
self._l_p = l_p
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[p_var],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [xs_var, ys_var, old_p_var]
else:
optimizer_args["inputs"] = [xs_var, ys_var]
self._optimizer.update_opt(**optimizer_args)
self._use_trust_region = use_trust_region
self._name = name
self._normalize_inputs = normalize_inputs
self._x_mean_var = x_mean_var
self._x_std_var = x_std_var
def __init__(
self,
input_shape,
output_dim,
predict_all=True,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
optimizer=None,
use_trust_region=True,
step_size=0.01,
normalize_inputs=True,
name=None,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean network.
:param hidden_nonlinearity: Non-linearity used for each layer of the mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer()
else:
optimizer = LbfgsOptimizer()
self.output_dim = output_dim
self._optimizer = optimizer
p_network = MLP(
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=NL.sigmoid,
)
l_p = p_network.output_layer
LasagnePowered.__init__(self, [l_p])
xs_var = p_network.input_layer.input_var
ys_var = TT.imatrix("ys")
old_p_var = TT.matrix("old_p")
x_mean_var = theano.shared(np.zeros((1, ) + input_shape),
name="x_mean",
broadcastable=(True, ) +
(False, ) * len(input_shape))
x_std_var = theano.shared(np.ones((1, ) + input_shape),
name="x_std",
broadcastable=(True, ) +
(False, ) * len(input_shape))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
p_var = L.get_output(l_p, {p_network.input_layer: normalized_xs_var})
old_info_vars = dict(p=old_p_var)
info_vars = dict(
p=p_var
) # posterior of the latent at every step, wrt obs-act. Same along batch if recurrent
dist = self._dist = Bernoulli(output_dim)
mean_kl = TT.mean(dist.kl_sym(old_info_vars, info_vars))
self._mean_kl = ext.compile_function(
[xs_var, old_p_var], mean_kl) # if not using TR, still log KL
loss = -TT.mean(dist.log_likelihood_sym(
ys_var,
info_vars)) # regressor just wants to min -loglik of data ys
predicted = p_var >= 0.5 # this gives 0 or 1, depending what is closer to the p_var
self._f_predict = ext.compile_function([xs_var], predicted)
self._f_p = ext.compile_function(
[xs_var], p_var
) # for consistency with gauss_mlp_reg this should be ._f_pdists
self._l_p = l_p
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[p_var],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [xs_var, ys_var, old_p_var]
else:
optimizer_args["inputs"] = [xs_var, ys_var]
self._optimizer.update_opt(**optimizer_args)
self._use_trust_region = use_trust_region
self._name = name
self._normalize_inputs = normalize_inputs
self._x_mean_var = x_mean_var
self._x_std_var = x_std_var