def _default_hparams(self):
# Data Dimensions
default_dict = ParamDict({
'batch_size': -1,
})
# Network params
default_dict.update({
'normalization': 'batch',
})
# Misc params
default_dict.update({
})
return default_dict
def _default_hparams(self):
# put new parameters in here:
default_dict = ParamDict({
'use_convs': False,
'device': None,
'n_rollout_steps': 10, # number of decoding steps
'cond_decode':
False, # if True, conditions decoder on prior inputs
})
# Network size
default_dict.update({
'state_dim': 1, # dimensionality of the state space
'action_dim': 1, # dimensionality of the action space
'nz_enc': 32, # number of dimensions in encoder-latent space
'nz_vae': 10, # number of dimensions in vae-latent space
'nz_mid': 32, # number of dimensions for internal feature spaces
'nz_mid_lstm': 128, # size of middle LSTM layers
'n_lstm_layers': 1, # number of LSTM layers
'n_processing_layers': 3, # number of layers in MLPs
})
# Learned prior
default_dict.update({
'n_prior_nets': 1, # number of prior networks in ensemble
'num_prior_net_layers':
6, # number of layers of the learned prior MLP
'nz_mid_prior':
128, # dimensionality of internal feature spaces for prior net
'nll_prior_train':
True, # if True, trains learned prior by maximizing NLL
'learned_prior_type':
'gauss', # distribution type for learned prior, ['gauss', 'gmm', 'flow']
'n_gmm_prior_components':
5, # number of Gaussian components for GMM learned prior
})
# Loss weights
default_dict.update({
'reconstruction_mse_weight':
1., # weight of MSE reconstruction loss
'kl_div_weight': 1., # weight of KL divergence loss
})
# add new params to parent params
parent_params = super()._default_hparams()
parent_params.overwrite(default_dict)
return parent_params