def next_frame_basic_recurrent():
"""Basic 2-frame recurrent model with stochastic tower."""
hparams = basic_stochastic.next_frame_basic_stochastic_discrete()
hparams.video_num_input_frames = 4
hparams.video_num_target_frames = 4
hparams.add_hparam("num_lstm_layers", 2)
hparams.add_hparam("num_lstm_filters", 256)
return hparams
def next_frame_basic_recurrent():
"""Basic 2-frame recurrent model with stochastic tower."""
hparams = basic_stochastic.next_frame_basic_stochastic_discrete()
hparams.filter_double_steps = 2
hparams.hidden_size = 64
hparams.video_num_input_frames = 4
hparams.video_num_target_frames = 4
hparams.concat_internal_states = False
hparams.add_hparam("num_lstm_layers", 2)
hparams.add_hparam("num_lstm_filters", 256)
return hparams
def next_frame_basic_recurrent():
"""Basic 2-frame recurrent model with stochastic tower."""
hparams = basic_stochastic.next_frame_basic_stochastic_discrete()
hparams.filter_double_steps = 2
hparams.hidden_size = 64
hparams.video_num_input_frames = 4
hparams.video_num_target_frames = 4
hparams.concat_internal_states = False
hparams.add_hparam("num_lstm_layers", 2)
hparams.add_hparam("num_lstm_filters", 256)
return hparams
def ppo_original_world_model_stochastic_discrete():
"""Atari parameters with stochastic discrete world model as policy."""
hparams = ppo_original_params()
hparams.policy_network = "next_frame_basic_stochastic_discrete"
hparams_keys = hparams.values().keys()
video_hparams = basic_stochastic.next_frame_basic_stochastic_discrete()
for (name, value) in six.iteritems(video_hparams.values()):
if name in hparams_keys:
hparams.set_hparam(name, value)
else:
hparams.add_hparam(name, value)
# To avoid OOM. Probably way to small.
hparams.optimization_batch_size = 1
return hparams
