# Configure experiment
experiment_name = args.experiment_name
dataset_args = {
'use_proprioception': True,
'use_haptics': True,
'use_vision': True,
'vision_interval': 2,
'image_blackout_ratio': args.blackout,
'sequential_image_rate': args.sequential_image,
'start_timestep': args.start_timestep,
}
# Create models & training buddy
pf_image_model = panda_models.PandaParticleFilterNetwork(
panda_models.PandaDynamicsModel(),
panda_models.PandaMeasurementModel(units=args.hidden_units,
missing_modalities=['gripper_sensors']))
pf_force_model = panda_models.PandaParticleFilterNetwork(
panda_models.PandaDynamicsModel(),
panda_models.PandaMeasurementModel(units=args.hidden_units,
missing_modalities=['image']),
)
weight_model = fusion.CrossModalWeights(state_dim=1,
use_softmax=True,
use_log_softmax=True)
pf_fusion_model = fusion_pf.ParticleFusionModel(pf_image_model, pf_force_model,
weight_model)
buddy = fannypack.utils.Buddy(experiment_name,
pf_fusion_model,
optimizer_names=[
type=str,
choices=["mujoco", "omnipush"],
default="mujoco")
parser.add_argument("--hidden_units", type=int, default=64)
args = parser.parse_args()
# Some constants
E2E_EPOCHS = 10
# Configure experiment
experiment_name = args.experiment_name
# Create models & training buddy
dynamics_model = panda_models.PandaDynamicsModel(units=32)
measurement_model = panda_models.PandaMeasurementModel(units=args.hidden_units)
pf_model = panda_models.PandaParticleFilterNetwork(dynamics_model,
measurement_model)
buddy = fannypack.utils.Buddy(experiment_name + "_unfrozen",
pf_model,
optimizer_names=[
"e2e_fusion",
"e2e_image",
"e2e_force",
"dynamics_image",
"dynamics_force",
"dynamics_recurrent_image",
"dynamics_recurrent_force",
"measurement_image",
"measurement_force",
])
buddy.load_metadata(experiment_name=experiment_name)
