Python ExpConfig Exemples

Exemple #1

Fichier : train_planet_push.py Projet : wuphilipp/replay-overshooting

    pixel_res=256,
    raw=True,
    cond=False,
    half_image_size=False,
    kloss_dataset=True,  # true for mit push
)

model_config = PlaNetBaselineFilterConfig(
    latent_dim=hy_config.latent_dim,
    latent_obs_dim=32,
    hidden_units=64,
    ctrl_dim=6,
    dataset=dataset_config,
    overshoot=(OverShoot.LATENT, 2),
)

exp_config = ExpConfig(
    name=f"oo_planet_push_{model_config.overshoot[1]}" +
    f"{datetime.datetime.now().strftime('%a-%H-%M-%S')}",
    model=model_config,
    ramp_iters=200,
    batch_size=hy_config.batch_size,
    epochs=hy_config.epochs,
    log_iterations_simple=10,
    log_iterations_images=100,
    base_learning_rate=hy_config.learning_rate,
    learning_rate_function=lr5,
    gradient_clip_max_norm=100,
)
train(exp_config)

Exemple #2

Fichier : train_dt_ekf_kvae_pend_img_nop.py Projet : wuphilipp/replay-overshooting

    dkl_anneal_iter=1000,
    alpha=1.0,
    beta=2.0,
    atol=1e-9,  # default: 1e-9
    rtol=1e-7,  # default: 1e-7
    z_pred=False,
)
model_config = KVAEConfig(
    latent_dim=_model_config.latent_dim,
    ctrl_dim=_model_config.ctrl_dim,
    dataset=dataset_config,
    latent_obs_dim=2,
    kf_estimator_config=_model_config,
)

# experiment settings
exp_config = ExpConfig(
    name="dt_ekf_kvae_pend_img_nop",
    model=model_config,
    ramp_iters=(_model_config.ramp_iters
                if hasattr(_model_config, "ramp_iters") else 100),
    batch_size=hy_config.batch_size,
    epochs=hy_config.epochs,
    log_iterations_simple=10,
    log_iterations_images=model_config.kf_estimator_config.ramp_iters,
    base_learning_rate=hy_config.learning_rate,
    learning_rate_function=lr2,
    gradient_clip_max_norm=None,
)
train(exp_config)

Exemple #3

Fichier : train_le_ekf_push.py Projet : wuphilipp/replay-overshooting

    dyn_nonlinearity=nn.Softplus(beta=2, threshold=20),
    obs_hidden_units=64,
    obs_layers=3,
    obs_nonlinearity=nn.Softplus(beta=2, threshold=20),
    is_continuous=False,
    ramp_iters=200,
    burn_in=100,
    dkl_anneal_iter=1000,
    num_submodels=100,
    alpha=0.5,
    beta=1.0,
    atol=1e-9,  # default: 1e-9
    rtol=1e-7,  # default: 1e-7
    z_pred=True,
)

exp_config = ExpConfig(
    name="le_ekf_push",
    model=model_config,
    ramp_iters=(model_config.ramp_iters
                if hasattr(model_config, "ramp_iters") else 100),
    batch_size=hy_config.batch_size,
    epochs=hy_config.epochs,
    log_iterations_simple=10,
    log_iterations_images=100,
    base_learning_rate=hy_config.learning_rate,
    learning_rate_function=lr5,
    gradient_clip_max_norm=100,
)
train(exp_config)

Exemple #4

Fichier : train_planet_pend_img.py Projet : wuphilipp/replay-overshooting

dataset_config = ImageDynamicDatasetConfig(
    traj_len=pend_fixed.traj_len,
    num_trajectories=10000,
    num_viz_trajectories=pend_fixed.num_viz_trajectories,
    system=pend_fixed,
    policy=None,
)

model_config = PlaNetBaselineFilterConfig(
    latent_dim=hy_config.latent_dim,
    latent_obs_dim=8,
    hidden_units=64,
    ctrl_dim=1,
    dataset=dataset_config,
)

# experiment settings
exp_config = ExpConfig(
    name="planet_pend_img_try_reproduce",
    model=model_config,
    ramp_iters=100,
    batch_size=hy_config.batch_size,
    epochs=hy_config.epochs,
    log_iterations_simple=10,
    log_iterations_images=100,
    base_learning_rate=hy_config.learning_rate,
    learning_rate_function=lr2,
    gradient_clip_max_norm=500,
)
train(exp_config)

Exemple #5

    traj_len=50,
    num_viz_trajectories=20,
    pixel_res=256,
    raw=True,
    cond=False,
    half_image_size=False,
    kloss_dataset=True,  # true for mit push
)

model_config = GSSMBaselineConfig(
    latent_dim=hy_config.latent_dim,
    latent_obs_dim=16,
    hidden_units=64,
    ctrl_dim=6,
    dataset=dataset_config,
)

exp_config = ExpConfig(
    name="gssm_push",
    model=model_config,
    ramp_iters=200,
    batch_size=hy_config.batch_size,
    epochs=hy_config.epochs,
    log_iterations_simple=10,
    log_iterations_images=100,
    base_learning_rate=hy_config.learning_rate,
    learning_rate_function=lr5,
    gradient_clip_max_norm=500,
)
train(exp_config)

Exemple #6

Fichier : train_example.py Projet : wuphilipp/replay-overshooting

    dataset=dataset_config,
    dyn_hidden_units=32,
    dyn_layers=3,
    dyn_nonlinearity=nn.Softplus(beta=2, threshold=20),
    obs_hidden_units=32,
    obs_layers=3,
    obs_nonlinearity=nn.Softplus(beta=2, threshold=20),
    is_continuous=False,
    ramp_iters=100,
    burn_in=100,
    dkl_anneal_iter=1000,
    alpha=0.5,
    beta=1.0,
    atol=1e-9,  # default: 1e-9
    rtol=1e-7,  # default: 1e-7
    z_pred=False,
)

# experiment settings
exp_config = ExpConfig(
    model=model_config,
    ramp_iters=model_config.ramp_iters,
    batch_size=hy_config.batch_size,
    epochs=hy_config.epochs,
    log_iterations_simple=10,
    log_iterations_images=model_config.ramp_iters,
    base_learning_rate=hy_config.learning_rate,
    learning_rate_function=lr1,
)
train(exp_config)  # train the model