cost=dict(
cost_type='nn',
network=(
nn.Relu(ds, 200) >> nn.Relu(200) >> nn.Gaussian(1)
),
),
),
train=dict(
num_epochs=1000,
learning_rate=1e-3,
model_learning_rate=1e-3 * horizon,
beta_start=1e-4,
beta_end=10.0,
beta_rate=5e-5,
beta_increase=0,
batch_size=2,
dump_every=100,
summary_every=50,
),
data=dict(
num_rollouts=100,
init_std=0.5,
smooth_noise=True,
),
dump_data=True,
seed=0,
out_dir='out/vae/reacher',
)
run(experiment, remote=False, gpu=False, num_threads=1, instance_type='m5.4xlarge')
from parasol.experiment import run, sweep
experiment = dict(
experiment_type='solar',
experiment_name='reacher-mpc',
env={
"environment_name": "Reacher",
},
control={
'control_type': 'mpc',
'horizon': 20,
},
# model='s3://parasol-experiments/vae/reacher-image/reacher-image_model{prior{prior_type}}-blds/weights/model-1700.pkl',
model='out/reacher_mpc/nnds/weights/model-final.pkl',
horizon=50,
seed=0,
rollouts_per_iter=2,
num_iters=20,
num_videos=2,
out_dir='out/solar/reacher_mpc',
model_train={'num_epochs': 100})
run(experiment, remote=False)
'random_target': False,
'pd_cost': True,
'easy_cost': False,
},
control=dict(
control_type='lqrflm',
data_strength=50,
prior_type='model',
horizon=50,
init_std=0.5,
kl_step=2.0,
),
model='data/vae/reacher-image/weights/model-final.pkl',
horizon=50,
seed=0,
rollouts_per_iter=40,
num_iters=10,
buffer_size=None,
smooth_noise=False,
num_videos=2,
out_dir='data/solar/reacher-image/',
model_train={
'num_epochs': 0
}
)
run(
experiment,
remote=False,
num_threads=1,
)