def _default_hparams(self):
# Data Dimensions
default_dict = AttrDict({
'batch_size': -1,
'max_seq_len': -1,
'n_actions': -1,
'state_dim': -1,
'input_nc': 3, # number of input feature maps
'device': None,
'data_conf': None,
'img_sz': None,
'goal_cond': True
})
# Network params
default_dict.update({
'use_convs': True,
'use_batchnorm': True, # TODO deprecate
'normalization': 'batch',
})
# add new params to parent params
parent_params = HParams()
for k in default_dict.keys():
parent_params.add_hparam(k, default_dict[k])
return parent_params
def _default_hparams(self):
default_dict = AttrDict({
'num_bins': 10,
})
# add new params to parent params
parent_params = super()._default_hparams()
for k in default_dict.keys():
parent_params.add_hparam(k, default_dict[k])
return parent_params
def _default_hparams(self):
default_dict = AttrDict({
'ndist_max': 10, # maximum temporal distance to classify
'use_skips': False, #todo try resnet architecture!
'ngf': 8,
'nz_enc': 64,
'classifier_restore_path': None # not really needed here.
})
# add new params to parent params
parent_params = super()._default_hparams()
for k in default_dict.keys():
parent_params.add_hparam(k, default_dict[k])
return parent_params
def _default_hparams(self):
default_dict = AttrDict({
'tmax_label':
10, # the highest label for temporal distance, values are clamped after that
'use_skips': False, #todo try resnet architecture!
'ngf': 8,
'nz_enc': 64,
'classifier_restore_path': None # not really needed here.
})
# add new params to parent params
parent_params = super()._default_hparams()
for k in default_dict.keys():
parent_params.add_hparam(k, default_dict[k])
return parent_params
def _default_hparams(self):
default_dict = AttrDict({
'use_skips': False, #todo try resnet architecture!
'ngf': 8,
'action_size': 2,
'nz_enc': 64,
'classifier_restore_path': None, # not really needed here.,
'low_dim': False,
'gamma': 0.0
})
# add new params to parent params
parent_params = super()._default_hparams()
for k in default_dict.keys():
parent_params.add_hparam(k, default_dict[k])
return parent_params
def _default_hparams(self):
default_dict = AttrDict({
'use_skips': False,
'ngf': 8,
'action_size': 2,
'state_size': 30,
'nz_enc': 64,
'linear_layer_size': 128,
'classifier_restore_path': None, # not really needed here.,
'low_dim': False,
'gamma': 0.0,
'terminal': True,
'update_target_rate': 1,
'action_range': [-1.0, 1.0],
'action_stds': [0.6, 0.6, 0.3, 0.3],
'est_max_samples': 100,
'binary_reward': [0, 1],
'n_step': 1,
'min_q': False,
'min_q_weight': 1.0,
'min_q_lagrange': False,
'min_q_eps': 0.1,
'sigmoid': False,
'optimize_actions': 'random_shooting',
'target_network_update': 'replace',
'polyak': 0.995,
'sg_sample': 'half_unif_half_first',
'geom_sample_p': 0.5,
'bellman_weight': 1.0,
'td_loss': 'mse',
'add_negative_sample': False,
'negative_sample_type': 'copy_arm', # also rand_arm, batch_goal
'gaussian_blur': False,
'twin_critics': False,
'add_action_noise': False,
'action_scaling': 1.0,
'eval_target_nets': True,
})
# add new params to parent params
parent_params = super()._default_hparams()
for k in default_dict.keys():
parent_params.add_hparam(k, default_dict[k])
return parent_params