def _default_hparams(self):
default_dict = ParamDict({
'policy': ClosedLoopBlockStackDemoPolicy, # policy class
'env_params':
None, # parameters containing info about env -> set automatically
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'action_dim': 1, # dimensionality of the action space
'normalization': 'none', # normalization used in policy network ['none', 'batch']
'action_input': True, # forward takes actions as second argument if set to True
})
return default_dict
def _default_hparams(self):
default_dict = ParamDict({
'device': None, # pytorch device
'discount_factor': 0.99, # discount factor for RL update
'optimizer':
'adam', # supported: 'adam', 'radam', 'rmsprop', 'sgd'
'gradient_clip': None, # max grad norm, if None no clipping
'momentum': 0, # momentum in RMSProp / SGD optimizer
'adam_beta': 0.9, # beta1 param in Adam
'update_iterations':
1, # number of iteration steps per one call to 'update(...)'
'target_network_update_factor':
5e-3, # percentage of new weights that are carried over
'batch_size': 64, # size of the experience batch used for updates
'obs_normalizer':
DummyNormalizer, # observation normalization class
'obs_normalizer_params':
{}, # parameters for optimization norm class
'obs_norm_log_groups':
{}, # (optional) dict defining separation of state space for obsNormLog
'log_videos': True, # whether to log videos during logging
'log_video_caption': False, # whether to add captions to video
'num_workers':
None, # number of independent workers --> whether grads need sync
})
return default_dict
def _default_hparams(self):
default_dict = ParamDict({
'policy': None, # policy class
'policy_params': None, # parameters for the policy class
'policy_lr': 3e-4, # learning rate for policy update
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'unused_obs_size':
None, # dimensionality of split off observation part
'discard_part':
'back', # which part of observation to discard ['front', 'back']
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'prior_input_res': 32, # input resolution of prior images
'encoder_ngf': 8, # number of feature maps in shallowest level of encoder
'n_input_frames': 1, # number of prior input frames
})
# add new params to parent params
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'input_dim': 32, # dimensionality of the observation input
'n_layers': 3, # number of policy network layers
'nz_mid': 64, # size of the intermediate network layers
'output_dim': 1, # number of outputs, can be >1 for discrete action spaces
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'input_res': 32, # resolution of the image input
'input_nc': 3, # number of input channels
'ngf': 8, # number of channels in shallowest layer of image encoder
'nz_enc': 64, # number of dimensions in encoder-latent space
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'omega_schedule': ConstantSchedule, # schedule used for omega param
'omega_schedule_params': AttrDict( # parameters for omega schedule
p = 0.1,
),
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'device':
None, # device that all tensors should get transferred to
'screen_width': 400, # width of rendered images
'screen_height': 400, # height of rendered images
})
return default_dict
def _default_hparams(self):
default_dict = ParamDict({
'action_dim': 1, # dimensionality of the action space
'max_action_range':
1., # for cont. actions this defines a symmetric action range [-x, x]
'squash_output_dist':
True, # do not tanh adjust log prob if set to False
})
return default_dict
def _default_hparams(self):
default_dict = ParamDict({
'alpha_min': None, # minimum value alpha is clipped to, no clipping if None
'td_schedule': ConstantSchedule, # schedule used for target divergence param
'td_schedule_params': AttrDict( # parameters for target divergence schedule
p = 1.,
),
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'prior_model': None, # prior model class
'prior_model_params': None, # parameters for the prior model
'prior_model_checkpoint': None, # checkpoint path of the prior model
'prior_model_epoch': 'latest', # epoch that checkpoint should be loaded for (defaults to latest)
'load_weights': True, # optionally allows to *not* load the weights (ie train from scratch)
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'model': None, # policy class
'model_params': None, # parameters for the policy class
'model_checkpoint': None, # checkpoint path of the model
'model_epoch':
'latest', # epoch that checkpoint should be loaded for (defaults to latest)
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'input_dim': 32, # dimensionality of the observation input
'n_layers': 3, # number of policy network layers
'nz_mid': 64, # size of the intermediate network layers
'normalization':
'none', # normalization used in policy network ['none', 'batch']
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'clip_raw_obs': np.array(
float("Inf")), # symmetric value maximum for raw observation
'clip_norm_obs': np.array(float(
"Inf")), # symmetric value maximum for normalized observation
'update_horizon':
1e7, # number of values for which statistics get updated
})
return default_dict
def _default_hparams(self):
default_dict = ParamDict({
'name': None, # name of openai/gym environment
'reward_norm': 1., # reward normalization factor
'punish_reward':
-100, # reward used when action leads to simulation crash
'unwrap_time':
True, # removes time limit wrapper from envs so that done is not set on timeout
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'prior_model': None, # prior model class
'prior_model_params': None, # parameters for the prior model
'prior_model_checkpoint': None, # checkpoint path of the prior model
'prior_model_epoch': 'latest', # epoch that checkpoint should be loaded for (defaults to latest)
'prior_batch_size': -1, # optional: use separate batch size for prior network
'reverse_KL': False, # if True, computes KL[q||p] instead of KL[p||q] (can be more stable to opt)
'analytic_KL': False, # if True, computes KL divergence analytically, otherwise sampling based
'num_mc_samples': 10, # number of samples for monte-carlo KL estimate
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
default_dict = ParamDict({
'hl_agent': None, # high-level agent class
'hl_agent_params': None, # parameters of the high-level agent
'll_agent': None, # low-level agent class
'll_agent_params': None, # parameters of the low-level agent(s)
'update_hl': True, # whether to update high-level agent
'update_ll': True, # whether to update low-level agent(s)
'll_subgoal_reaching_reward': False, # whether to count ll subgoal reaching reward in training
'll_subgoal_reaching_reward_weight': 1e3, # weight for the subgoal reaching reward
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
# put new parameters in here:
return super()._default_hparams().overwrite(ParamDict({
'use_convs': False,
'device': None,
'state_dim': 1, # dimensionality of the state space
'action_dim': 1, # dimensionality of the action space
'nz_mid': 128, # number of dimensions for internal feature spaces
'n_processing_layers': 5, # number of layers in MLPs
'output_type': 'gauss', # distribution type for learned prior, ['gauss', 'gmm', 'flow']
'n_gmm_prior_components': 5, # number of Gaussian components for GMM learned prior
}))
def _default_hparams(self):
default_dict = ParamDict({
'critic': None, # critic class
'critic_params': None, # parameters for the critic class
'replay': None, # replay buffer class
'replay_params': None, # parameters for replay buffer
'critic_lr': 3e-4, # learning rate for critic update
'reward_scale': 1.0, # SAC reward scale
'clip_q_target': False, # if True, clips Q target
'target_entropy':
None, # target value for automatic entropy tuning, if None uses -action_dim
})
return super()._default_hparams().overwrite(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
def _default_hparams(self):
default_dict = ParamDict({
'policy_model': None, # policy model class
'policy_model_params': None, # parameters for the policy model
'policy_model_checkpoint':
None, # checkpoint path of the policy model
'policy_model_epoch':
'latest', # epoch that checkpoint should be loaded for (defaults to latest)
'load_weights':
True, # optionally allows to *not* load the weights (ie train from scratch)
'initial_log_sigma':
-50, # initial log sigma of policy dist (since model is deterministic)
})
return super()._default_hparams().overwrite(default_dict)
def _default_hparams(self):
return ParamDict({
'input_dim': None, # dimensionality of the vector input
'input_res': None, # resolution of image input
'output_dim': None, # dimensionality of output tensor
'input_nc': 3, # number of input channels
'ngf': 8, # number of channels in shallowest layer of image encoder
'nz_enc': 32, # number of dimensions in encoder-latent space
'nz_mid': 32, # number of dimensions for internal feature spaces
'n_layers': 3, # number of layers in MLPs
'normalization': 'none', # normalization used in encoder network ['none', 'batch']
'use_convs': False,
'device': None,
})
def _default_hparams(self):
default_dict = ParamDict({
'seed': None,
'agent': None,
'data_dir': None, # directory where dataset is in
'environment': None,
'sampler': Sampler, # sampler type used
'exp_path': None, # Path to the folder with experiments
'num_epochs': 200,
'max_rollout_len': 1000, # maximum length of the performed rollout
'n_steps_per_update': 1, # number of env steps collected per policy update
'n_steps_per_epoch': 20000, # number of env steps per epoch
'log_output_per_epoch': 100, # log the non-image/video outputs N times per epoch
'log_images_per_epoch': 4, # log images/videos N times per epoch
'logging_target': 'wandb', # where to log results to
'n_warmup_steps': 0, # steps of warmup experience collection before training
})
return default_dict
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):
default_dict = ParamDict({
'model': None,
'model_test': None,
'logger': None,
'logger_test': None,
'evaluator': None,
'data_dir': None, # directory where dataset is in
'batch_size': 16,
'exp_path': None, # Path to the folder with experiments
'num_epochs': 200,
'epoch_cycles_train': 1,
'optimizer':
'radam', # supported: 'adam', 'radam', 'rmsprop', 'sgd'
'lr': 1e-3,
'gradient_clip': None,
'momentum': 0, # momentum in RMSProp / SGD optimizer
'adam_beta': 0.9, # beta1 param in Adam
'top_of_n_eval': 1, # number of samples used at eval time
'top_comp_metric':
None, # metric that is used for comparison at eval time (e.g. 'mse')
})
return default_dict
def _default_hparams(self):
return BaseEnvironment._default_hparams(self).overwrite(ParamDict({
'n_blocks': 5, # number of blocks in env
'block_size': 0.04, # size of a block
'block_color': 'white', # color of the block
'rotate_blocks': False, # no block rotation if set to False
'allow_rotate': False, # if False, disallow gripper rotation
'table_size': (1.2, 1.2, 0.8), # size of table
'dimension': 2, # dimensionality for the task
'camera_name': 'frontview', # name of camera to render
'gripper_width': 0.02, # thickness of gripper to consider during placement
'task_generator': None, # task generator for generating HL plans
'task_params': AttrDict({}), # parameters for task generator
'perturb_actions': False, # if True, perturb action and init block placement
'perturb_prob': 0.3, # action perturb probability
'perturb_scale': 0.03, # action perturb scale
'n_steps': None, # number of steps in the task, default n_blocks - 1
'friction': 1, # friction for the boxes
'rand_task': False, # if True, randomizes the task in every reset (i.e. multi-task env)
'rand_init_pos': False, # if False, keeps initial position of blocks constant
'rand_init_gripper': False, # if True, randomizes gripper xy position at each episode reset
'include_quat': False, # if True, include quaternions in observation
'include_vel': True, # if True, include velocity of the gripper
'include_2d_rotation': False, # if True, adds 2D rotation representation for blocks to obs (sin+cos)
'clip_obs': 2.0, # if not None, clip observation values
'seed': None, # seed for generating block placements
'relative_shaped_reward': False, # if True, computes shaping reward as relative change towards the goal
'action_penalty_weight': 0., # penalty for action magnitude
'reward_density': 'dense', # integer defining how dense the reward is ['dense', 'sparse']
'number_blocks': False, # if True, print number on blocks
'fixed_task': None, # (optional) if provided is used as fixed task
'fixed_block_pos': None, # (optional) if provided is used as fixed block position
'add_boundary_walls': True, # if True, adds invisible walls that constrain movement
'reset_with_boundary': False, # if True, resets episode once agent leaves allowed region
'reward_scale': 1.0, # scale of the reward
}))
def _default_hparams(self):
return super()._default_hparams().overwrite(
ParamDict({
'obs_res': 64, # resolution of image observation
}))
def _default_hparams(self):
return super()._default_hparams().overwrite(
ParamDict({
'name': "Widow250OfficeFixed-v0",
}))
