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)
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