experiment_name='reacher-image-mpc',
experiment_type='train_vae',
env=env_params,
model=dict(
do=do, du=du, ds=ds, da=da, horizon=horizon,
state_encoder=(nn.Reshape(do, [64, 64, 3])
>> nn.Convolution([7, 7, 64], strides=(1, 1)) >> nn.Relu()
>> nn.Convolution([5, 5, 32], strides=(2, 2))
>> nn.Convolution([3, 3, 8], strides=(2, 2))
>> nn.Flatten() >> nn.Relu(256) >> nn.Gaussian(ds)),
state_decoder=(nn.Relu(ds, 512) >> nn.Reshape([8, 8, 8])
>> nn.Deconvolution([3, 3, 32])
>> nn.Deconvolution([5, 5, 64])
>> nn.Deconvolution([7, 7, 3])
>> nn.Flatten() >> nn.Bernoulli()),
action_encoder=nn.IdentityVariance(variance=1e-4),
action_decoder=nn.IdentityVariance(variance=1e-4),
prior=dict(
prior_type='nnds',
network=(
nn.Relu(ds + da, 200) >> nn.Relu(200) >> nn.Gaussian(ds)
)
),
cost=dict(
cost_type='nn',
network=(
nn.Relu(ds, 200) >> nn.Relu(200) >> nn.Gaussian(1)
),
),
),
do=do,
du=du,
ds=ds,
da=da,
horizon=horizon,
# state_encoder=(nn.Reshape(do, [32, 32, 2])
# >> nn.Convolution([3, 3, 32]) >> nn.Relu()
# >> nn.Convolution([3, 3, 32]) >> nn.Relu()
# >> nn.Convolution([3, 3, 32]) >> nn.Relu()
# >> nn.Flatten() >> nn.Relu(200) >> nn.Gaussian(ds)),
# state_decoder=(nn.Relu(ds, 1024) >> nn.Reshape([16, 16, 4])
# >> nn.Convolution([3, 3, 32]) >> nn.Relu()
# >> nn.Deconvolution([2, 2, 32]) >> nn.Relu()
# >> nn.Convolution([3, 3, 32]) >> nn.Relu()
# >> nn.Convolution([2, 2, 2]) >> nn.Flatten() >> nn.Bernoulli()),
state_encoder=nn.IdentityVariance(),
state_decoder=nn.IdentityVariance(),
action_encoder=nn.IdentityVariance(),
action_decoder=nn.IdentityVariance(),
# prior={'prior_type': 'none'},
# prior={'prior_type': 'normal'},
# prior={'prior_type': 'lds'},
# prior={'prior_type': 'blds'},
prior={
'prior_type': 'nnds',
'network': nn.Relu(ds + da, 200) >> nn.Relu(200) >> nn.Gaussian(ds)
},
),
train=dict(
num_epochs=4000,
learning_rate=1e-3,