def init_from_tensor(self, tensor, **kwargs):
with torch.no_grad():
# TODO: deal properly with wrong kwargs
factors = tensor_train(tensor, self.rank)
self.factors = FactorList([nn.Parameter(f) for f in factors])
return self
def from_tensor(cls, tensor, tensorized_row_shape, tensorized_column_shape, rank='same', **kwargs):
full_shape = tensorized_row_shape + tensorized_column_shape
n_matrices = _ensure_tuple(tensor.shape[:-len(full_shape)])
rank = tl.tt_tensor.validate_tt_rank(n_matrices + full_shape, rank)
with torch.no_grad():
factors = tensor_train(tensor, rank, **kwargs)
return cls([nn.Parameter(f) for f in factors], tensorized_row_shape, tensorized_column_shape, rank=rank, n_matrices=n_matrices)
def from_tensor(cls, tensor, rank='same', **kwargs):
shape = tensor.shape
rank = tl.tt_tensor.validate_tt_rank(shape, rank)
with torch.no_grad():
# TODO: deal properly with wrong kwargs
factors = tensor_train(tensor, rank)
return cls([nn.Parameter(f) for f in factors])
def build_network(five_tuple, k, chi, omega):
s, a, P, R, gamma = five_tuple
H = []
H_core = []
data = softmax_by_state(torch.randn((s * a, 1)), s, a)
data.requires_grad = True
for i in range(k):
H.append(initialize_H(five_tuple, i + 1))
if i >= 1:
factors = tensor_train(H[i].get_tensor(), chi).factors
combined_cores, _ = put_mps(factors)
H_core.append(combined_cores)
else:
masked_H = H[i].get_tensor() * omega[i]
H_core.append([tn.Node(masked_H, backend=backend)])
return H, H_core, data
def init_from_tensor(self, tensor, **kwargs):
with torch.no_grad():
factors = tensor_train(tensor, self.rank, **kwargs)
self.factors = FactorList([nn.Parameter(f) for f in factors])
return self