use_linear_dynamics=False,
is_auto_encoder=False,
batch_size=64,
lr=1e-3,
),
save_period=50,
),
num_exps_per_instance=3,
)
search_space = {
'train_vae_variant.beta':[.5, 2.5],
'env_id':['SawyerPushAndReacherXYEnv-v0'],
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant,
)
variants = []
for exp_id, variant in enumerate(sweeper.iterate_hyperparameters()):
# if variant['env_id'] == 'SawyerPushAndReachXYEnv-No-Arena-v0':
# variant['train_vae_variant']['generate_vae_dataset_kwargs']['dataset_path'] = \
# 'datasets/SawyerPushAndReachXYEnv-No-Arena-v0_N5000_sawyer_pusher_camera_upright_v3_imsize84_random_oracle_split_0.npy'
# else:
# variant['train_vae_variant']['generate_vae_dataset_kwargs']['dataset_path'] = \
# 'datasets/SawyerPushAndReachXYEnv-No-Arena-v1_N5000_sawyer_pusher_camera_upright_v3_imsize84_random_oracle_split_0.npy'
variants.append(variant)
run_variants(grill_her_td3_full_experiment, variants, run_id=3)
# tensorboard=True,
# ),
load_demos=True,
pretrain_policy=True,
pretrain_rl=True,
# save_pretrained_algorithm=True,
# snapshot_mode="all",
env='pen-sparse-v0',
)
experiment(variant)
if __name__ == "__main__":
variant = dict()
search_space = {
'seedid': range(3),
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(local_exp, variants, run_id=0)
'seedid': range(5),
'algo_kwargs.base_kwargs.num_updates_per_env_step': [
16,
],
'replay_buffer_kwargs.fraction_goals_rollout_goals': [
0.1,
],
'replay_buffer_kwargs.fraction_goals_env_goals': [
0.5,
],
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
# n_seeds = 1
# mode = 'local'
# exp_prefix = 'test'
n_seeds = 1
mode = 'ec2'
exp_prefix = 'sawyer_pusher_state_final'
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(her_td3_experiment, variants, run_id=0)
distance_weight=0,
skew_dataset=False,
priority_function_kwargs=dict(
decoder_distribution='gaussian_identity_variance',
sampling_method='importance_sampling',
# sampling_method='true_prior_sampling',
num_latents_to_sample=10,
),
use_parallel_dataloading=False,
),
save_period=25,
),
)
search_space = {
'seedid': range(1),
'train_vae_variant.representation_size': [(16, 16)], #(3 * objects, 3 * colors)
'train_vae_variant.beta': [50], #THIS IS A LINEAR INTERPOLATION CURRENTLY
# 'train_vae_variant.generate_vae_dataset_kwargs.n_random_steps': [2],#, 50, 100],
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(experiment, variants, run_id=101)
wrap_mujoco_env=True,
track_qpos_goal=5,
do_state_based_exp=False,
exploration_noise=0.1,
reward_params=dict(
type="sparse",
epsilon=20.0,
),
save_video=False,
)
n_seeds = 3
search_space = {
'exploration_type': ['epsilon'],
'env_kwargs.arm_range': [1.0],
'algo_kwargs.reward_scale': [1],
'algo_kwargs.discount': [0.99],
'exploration_noise': [0.2],
'reward_params.epsilon': [20.0, 100.0, 5.0],
'replay_kwargs.fraction_goals_are_env_goals': [0.0, 0.5, 1.0],
'replay_kwargs.fraction_goals_are_rollout_goals': [0.2, 1.0],
# 'rdim': [2, 4, 8, 16],
'seedid': range(n_seeds),
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
run_variants(experiment, sweeper.iterate_hyperparameters(), run_id=1)
sampling_method='importance_sampling',
# sampling_method='true_prior_sampling',
num_latents_to_sample=10,
),
use_parallel_dataloading=False,
),
save_period=25,
),
)
search_space = {
'seedid': range(3),
'grill_variant.reward_params.type': ['latent_distance'],
'train_vae_variant.representation_size': [
(4, 4),
], #(3 * objects, 3 * colors)
'train_vae_variant.beta': [50],
'train_vae_variant.generate_vae_dataset_kwargs.n_random_steps': [100]
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(grill_her_twin_sac_online_vae_full_experiment,
variants,
run_id=10)
'env_id': [
'SawyerPushDebugLEAP-v3',
'SawyerPushDebugLEAPPuckRew-v3',
],
'rl_variant.use_masks': [True, False],
'rl_variant.max_path_length': [200],
'init_camera': [sawyer_xyz_reacher_camera_v0],
'rl_variant.vis_kwargs.vis_list': [
[],
],
# 'plt',
# ]],
'rl_variant.algo_kwargs.num_trains_per_train_loop': [
4000,
],
'rl_variant.es_kwargs.max_sigma': [0.2, 0.5, 0.8],
'seedid': range(3),
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
if not (variant['env_id'] == 'SawyerPushDebugLEAPPuckRew-v3'
and variant['rl_variant']['use_masks'] == True):
variants.append(variant)
run_variants(rl_experiment, variants, run_id=0)
trainer_kwargs=dict(
lr=1e-3,
loss_weights=dict(
mse=1.0,
)
),
save_period=100,
slurm_variant=dict(
timeout_min=48 * 60,
cpus_per_task=10,
gpus_per_node=1,
),
num_train_workers=8,
)
search_space = {
"dataset_kwargs.dataset_name": ["doors2", "pour1"],
"model_kwargs.delta_features": [True, ],
"model_kwargs.pretrained_features": [True, False, ],
"model_kwargs.normalize": [True, ],
"trainer_kwargs.loss_weights.mse": [0.0, 1.0],
"trainer_kwargs.loss_weights.classification_gradients": [False, True, ],
"trainer_class": [GeometricTimePredictorTrainer],
"seedid": range(1),
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant,
)
run_variants(train_rfeatures_model, sweeper.iterate_hyperparameters(), run_id=1)
4000,
],
'env_kwargs.reward_type': [
'hand_and_puck_distance',
],
'policy_kwargs.min_log_std': [
-6,
],
'expl_mask_distribution_kwargs.weights': [
(1, 1, 1),
(0, 0, 1),
(0, 1, 0),
],
'eval_mask_distribution_kwargs.weights': [
(1, 0, 0),
(0, 1, 0),
(0, 0, 1),
],
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(experiment, variants, process_args)
# # 'shaped',
# 'euclidean',
# ],
# 'exploration_type': [
# 'epsilon',
# 'ou',
# 'gaussian',
# ],
'seed_id':
range(n_seeds),
'replay_buffer_kwargs.fraction_resampled_goals_are_env_goals': [
0.0,
0.5,
],
'replay_buffer_kwargs.fraction_goals_are_rollout_goals': [
0.2,
1.0,
],
'env_kwargs.randomize_position_on_reset': [False, True],
'env_kwargs.wall_shape': ["u"],
'algo_kwargs.her_kwargs.rollout_goal_params.exploration_temperature':
[1.0, 100.0, 1e-2, 1e-4],
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
run_variants(her_td3_experiment,
sweeper.iterate_hyperparameters(),
run_id=0)
skew_config=dict(
method='vae_prob',
power=0,
),
skew_dataset=False,
priority_function_kwargs=dict(
decoder_distribution='gaussian_identity_variance',
sampling_method='importance_sampling',
# sampling_method='true_prior_sampling',
num_latents_to_sample=10,
),
use_parallel_dataloading=False,
),
save_period=25,
),
)
search_space = {
'seedid': range(5),
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(arl_full_experiment, variants, run_id=0)
)
search_space = {
'td3_bc_trainer_kwargs.use_awr': [False],
# 'td3_bc_trainer_kwargs.demo_beta':[1, 10],
'td3_bc_trainer_kwargs.bc_weight': [1, 0],
'td3_bc_trainer_kwargs.rl_weight': [1],
'algo_kwargs.num_epochs': [1000],
'algo_kwargs.num_eval_steps_per_epoch': [100],
'algo_kwargs.num_expl_steps_per_train_loop': [100],
'algo_kwargs.min_num_steps_before_training': [0],
# 'td3_bc_trainer_kwargs.add_demos_to_replay_buffer':[True, False],
# 'td3_bc_trainer_kwargs.num_trains_per_train_loop':[1000, 2000, 4000, 10000, 16000],
# 'exploration_noise':[0.1, .3, .5],
# 'pretrain_rl':[True],
# 'pretrain_policy':[False],
'pretrain_rl': [False],
'pretrain_policy': [False],
'seedid': range(5),
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(state_td3bc_experiment, variants, run_id=0)
skew_config=dict(
method='vae_prob',
power=0,
),
skew_dataset=False,
priority_function_kwargs=dict(
decoder_distribution='gaussian_identity_variance',
sampling_method='importance_sampling',
# sampling_method='true_prior_sampling',
num_latents_to_sample=10,
),
use_parallel_dataloading=False,
),
save_period=25,
),
)
search_space = {
'seedid': range(5),
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(grill_her_twin_sac_full_experiment, variants, run_id=0)
decoder_distribution='gaussian_identity_variance',
sampling_method='importance_sampling',
# sampling_method='true_prior_sampling',
num_latents_to_sample=10,
),
use_parallel_dataloading=False,
),
save_period=25,
),
region="us-west-2",
)
search_space = {
'seedid': range(2),
'grill_variant.algo_kwargs.rl_offpolicy_num_training_steps': [10, ],
# 'grill_variant.reward_params.type':['latent_bound'], #, 'latent_distance'
# 'train_vae_variant.representation_size': [(6, 8)], #(3 * objects, 3 * colors)
# 'train_vae_variant.beta': [1],
# 'train_vae_variant.generate_vae_dataset_kwargs.n_random_steps': [100]
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(grill_her_td3_offpolicy_online_vae_full_experiment, variants, run_id=0)
),
policy_kwargs=dict(
hidden_sizes=[32, 32],
),
save_video=True,
dump_video_kwargs=dict(
save_period=1,
# imsize=(3, 500, 300),
),
desired_trajectory="/home/anair/ros_ws/src/railrl-private/demos/door_demos_v3/demo_v3_0.pkl",
logger_variant=dict(
tensorboard=True,
),
model_path="/home/anair/data/s3doodad/facebook/models/rfeatures/multitask1/run2/id2/itr_4000.pt",
)
search_space = {
'seedid': range(1),
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(encoder_wrapped_td3bc_experiment, variants, run_id=3)
method='vae_prob',
power=0,
),
skew_dataset=False,
priority_function_kwargs=dict(
decoder_distribution='gaussian_identity_variance',
sampling_method='importance_sampling',
# sampling_method='true_prior_sampling',
num_latents_to_sample=10,
),
use_parallel_dataloading=False,
),
save_period=25,
),
)
search_space = {
'seedid': range(5),
'train_vae_variant.representation_size': [2, 4, 8, 16],
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(experiment, variants, run_id=2)
priority_function_kwargs=dict(
decoder_distribution='gaussian_identity_variance',
sampling_method='importance_sampling',
# sampling_method='true_prior_sampling',
num_latents_to_sample=10,
),
use_parallel_dataloading=False,
),
save_period=25,
),
region="us-west-1",
)
search_space = {
'seedid': range(5),
# 'grill_variant.reward_params.type':['latent_bound'], #, 'latent_distance'
# 'train_vae_variant.representation_size': [(6, 8)], #(3 * objects, 3 * colors)
# 'train_vae_variant.beta': [1],
# 'train_vae_variant.generate_vae_dataset_kwargs.n_random_steps': [100]
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(grill_her_td3_online_vae_full_experiment, variants, run_id=2)
