class EnKFRunner:
def __init__(self, data, targets, connections, model_out, n_ensembles,
rng):
self.data = data
self.targets = targets
self.connections = connections
self.model_out = model_out
self.n_ensembles = n_ensembles
self.rng = rng
self.enkf = EnKF(maxit=1, online=False, n_batches=1)
self.connections = self._shape_connections(connections)
@staticmethod
def _shape_connections(connections):
df = pd.read_pickle(connections)
weights = df['weight'].values
return weights
def run(self):
gamma = np.eye(10) * 0.01
for i in range(self.rng):
self.enkf.fit(data=self.data,
ensemble=self.connections,
ensemble_size=self.n_ensembles,
moments1=self.connections.mean(0),
observations=self.targets,
model_output=self.model_out,
noise=0.0,
gamma=gamma,
)
ens = self.enkf.ensemble
return ens
def __init__(self, data, targets, connections, model_out, n_ensembles,
rng):
self.data = data
self.targets = targets
self.connections = connections
self.model_out = model_out
self.n_ensembles = n_ensembles
self.rng = rng
self.enkf = EnKF(maxit=1, online=False, n_batches=1)
self.connections = self._shape_connections(connections)
conv_loss_mnist = []
# average test losses
test_losses = []
np.random.seed(0)
torch.manual_seed(0)
batch_size = 64
model = MnistOptimizee(root=root,
batch_size=batch_size,
seed=0,
n_ensembles=n_ensembles).to(device)
conv_ens = None
gamma = np.eye(10) * 0.01
enkf = EnKF(tol=1e-5,
maxit=1,
stopping_crit='',
online=False,
shuffle=False,
n_batches=1,
converge=False)
rng = int(60000 / batch_size * 8)
for i in range(1):
model.generation = i + 1
if i == 0:
try:
out = model.load_model('')
# replace cov matrix with cov from weights (ensembles)
# m = torch.distributions.Normal(out['conv_params'].mean(),
# out['conv_params'].std())
# model.cov = m.sample((n_ensembles, model.length))
except FileNotFoundError as fe:
print(fe)