# 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,
batch_size=16,
dump_every=1000,
summary_every=50 / 2,
beta_start=1.0,
beta_rate=0,
),
data=dict(
num_rollouts=100,
init_std=.2,
),
def train(out_dir, dim, seed, load_model=None):
out_path = out_dir / 'models'
if not (out_path).exists():
out_path.mkdir()
boards = {}
for round in DATA:
for location in DATA[round]:
for board_id in DATA[round][location]:
if board_id not in boards:
boards[board_id] = set()
boards[board_id].add((round, location))
if load_model is None:
sensor_models = {
board_id: nn.Relu(100) >> nn.Relu(100) >> nn.Linear(dim)
for board_id in boards
# board_id: nn.Linear(3, dim) for board_id in boards
}
calibration_model = nn.Relu(dim + 3, args.hidden_size) >> nn.Relu(
args.hidden_size) >> nn.Linear(2)
split_model = SplitModel(sensor_models,
calibration_model,
log_dir=out_dir,
lr=args.lr,
batch_size=args.batch_size)
else:
split_model = joblib.load(load_model)
data = {}
print("Filtering: %s" % ignore)
for board_id in boards:
board_train = []
for round, location in boards[board_id]:
if (round, location, board_id) in ignore:
print("Removing: ", round, location, board_id)
continue
board_train.append(load(*(round, location, board_id))[0])
if len(board_train) > 0:
print("Loaded board[%u]: %u" % (board_id, len(board_train)))
board_train = pd.concat(board_train)
board_train['board'] = board_id
if board_id not in data:
data[board_id] = []
data[board_id].append(board_train)
data = [pd.concat(ds) for ds in data.values()]
max_size = max([d.shape[0] for d in data])
for i in range(len(data)):
d = data[i]
if d.shape[0] < max_size:
data[i] = d.append(d.sample(max_size - d.shape[0], replace=True))
data = pd.concat(data)
def cb(model):
with open(str(out_path / 'model_latest.pkl'), 'wb') as fp:
joblib.dump(split_model, fp)
print("Total data size:", data.shape)
split_model.fit(data[sensor_features],
data[env_features],
data['board'],
data[Y_features],
dump_every=(out_dir / 'models' / 'model_latest.pkl', 1000),
n_iters=args.num_iters,
cb=cb)
with open(str(out_path / 'model.pkl'), 'wb') as fp:
joblib.dump(split_model, fp)
joblib.dump(
(
board_id,
Model(X_features).fit(train_data[X_features], train_data[Y_features])
), fp
)
if __name__ == "__main__":
args = parse_args()
fs = s3fs.S3FileSystem(anon=False)
experiment_dir = Path(BUCKET_NAME) / args.experiment
out_dir = experiment_dir / args.name
features = X_features[:]
for feature in args.ignore_feature:
features.remove(feature)
if args.model == 'subu':
Model = SubuForest
elif args.model == 'linear':
Model = Linear
elif args.model == 'nn-2':
Model = lambda features, : NeuralNetwork(features, nn.Relu(len(features), 200) >> nn.Relu(200) >> nn.Linear(2))
elif args.model == 'nn-4':
Model = lambda features: NeuralNetwork(features, nn.Relu(len(features), 500) >> nn.Relu(500) >> nn.Relu(500) >> nn.Relu(500) >> nn.Linear(2))
if args.level1:
level1(out_dir, features)
if args.level2:
level2(out_dir, features)
if args.level3:
level3(out_dir, features, args.seed)
ds = 10
du = da = env.get_action_dim()
horizon = 50
experiment = dict(
experiment_name='vae',
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=sweep([
dict(prior_type='blds', smooth=True, time_varying=True),
dict(prior_type='normal'),
], ['blds-tv-smooth', 'normal']),
cost=dict(cost_type='quadratic', learn_stdev=sweep([False, True])),
),
if __name__ == "__main__":
args = parse_args()
out_dir = Path('results') / args.name
out_dir.mkdir(exist_ok=True, parents=True)
features = X_features[:]
for feature in args.ignore_feature:
features.remove(feature)
if args.model == 'subu':
Model = SubuForest
elif args.model == 'linear':
Model = Linear
elif args.model == 'nn-2':
Model = lambda features, : NeuralNetwork(
features,
nn.Relu(len(features), 200) >> nn.Relu(200) >> nn.Linear(2))
elif args.model == 'nn-4':
Model = lambda features: NeuralNetwork(
features,
nn.Relu(len(features), 500) >> nn.Relu(500) >> nn.Relu(500) >> nn.
Relu(500) >> nn.Linear(2))
if args.level1:
level1(out_dir, features)
if args.level2:
level2(out_dir, features)
if args.level3:
level3(out_dir, features, args.seed)
# >> 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-4,
batch_size=2,
dump_every=50,
summary_every=50 / 2,
),
data=dict(
num_rollouts=100,
init_std=0.1,
),
out_dir='s3://parasol-experiments/vae/reacher-noimage',
# out_dir='temp2/pm-noimage',
)
