def get_updater(env):
policy = Policy(GoToPoint(), Deterministic(2), env.obs_shape)
# controller = FeedforwardCell(lambda inp, output_size: MLP([128, 128])(inp, output_size), 1)
controller = FeedforwardCell(
lambda inp, output_size: fully_connected(
inp, output_size, activation_fn=None), 1)
estimator = NeuralValueEstimator(controller, env.obs_shape)
alg = cfg.alg_class(estimator, name="critic")
updater = RLUpdater(env, policy, alg)
return updater
def build_object_network_controller(output_size, name):
from dps.utils.tf import ObjectNetwork
ff = ObjectNetwork(n_repeats=1, scope="collection_controller")
return FeedforwardCell(ff, output_size, name=name)
def build_attentional_relation_network(output_size, name):
from dps.utils.tf import AttentionalRelationNetwork
ff = AttentionalRelationNetwork(n_repeats=2, scope="collection_controller")
return FeedforwardCell(ff, output_size, name=name)
"Exponential(start=10000.0, end=0.000000001, decay_rate=0.1, decay_steps=3000, log=True)",
z_pres_temperature=1.0,
run_all_time_steps=False,
stopping_threshold=0.99,
per_process_gpu_memory_fraction=0.3,
training_wheels=0.0,
scale_prior_mean=-1.0,
scale_prior_std=np.sqrt(0.05),
shift_prior_mean=0.0,
shift_prior_std=3.0,
complete_rnn_input=False,
)
dair_config = air_config.copy(
difference_air=True,
build_cell=lambda scope: FeedforwardCell(MLP(n_units=[256, 256, 256, 256]),
cfg.rnn_n_units),
)
nem_config = alg_config.copy(
alg_name="nem",
build_network=nem.NEM_Network,
batch_size=16,
lr_schedule=0.001,
max_grad_norm=None,
threshold=-np.inf,
max_experiments=None,
render_hook=nem.NeuralEM_RenderHook(4),
render_step=5000,
# ------- from nem.py --------
noise_prob=0.2, # probability of annihilating the pixel
def __call__(self, param_shape, name=None):
return FeedforwardCell(MLP(), param_shape, name=name)
def __call__(self, param_shape, name=None):
return FeedforwardCell(MLP(*self.args, **self.kwargs),
param_shape,
name=name)
from dps import cfg
from dps.config import DEFAULT_CONFIG
from dps.env import BatchGymEnv
from dps.utils.tf import MLP, FeedforwardCell
from dps.rl import rl_render_hook, BuildSoftmaxPolicy, BuildMlpController
def build_env():
return BatchGymEnv(gym_env='MountainCar-v0')
controller = lambda params_dim, name: FeedforwardCell(
lambda inp, output_size: MLP(
[cfg.n_controller_units, cfg.n_controller_units])(inp, output_size),
params_dim, name=name)
config = DEFAULT_CONFIG.copy()
# So far, have not been able to solve this with a policy gradient method, the exploration problem is quite hard.
config.update(
env_name="mountain_car",
build_env=build_env,
build_controller=BuildMlpController(),
build_policy=BuildSoftmaxPolicy(one_hot=False),
exploration_schedule="1.0",