def decompress(self,
data: bytearray,
offset: Offset = None) -> torch.Tensor:
offset = offset or Offset()
torch.manual_seed(self.seed)
shape = unpack_tensor_shape(data, offset)
if len(shape) <= 1:
return unpack_tensor(data, offset)
else:
p = unpack_tensor(data, offset)
q = unpack_tensor(data, offset)
return torch.matmul(p, q.t()).view(shape)
def decompress(self, data: bytearray,
offset: Offset = None) -> torch.Tensor:
offset = offset or Offset()
# unpack tensor shape
shape = unpack_tensor_shape(data, offset)
# unpack sparse indices and value
tensor_indices = unpack_tensor(data, offset)
tensor_data = unpack_tensor(data, offset)
tensor = torch.zeros(int(np.prod(shape)))
tensor.index_copy_(0, tensor_indices, tensor_data)
tensor = tensor.reshape(list(shape))
return tensor
def decompress(self,
data: bytearray,
offset: Offset = None) -> torch.Tensor:
offset = offset or Offset()
shape = unpack_tensor_shape(data, offset)
if len(shape) <= 1:
return unpack_tensor(data, offset)
else:
mean = unpack_float32(data, offset)
S = unpack_tensor(data, offset)
U = unpack_tensor(data, offset)
V = unpack_tensor(data, offset)
return (torch.mm(torch.mm(U, torch.diag(S)), V.t()) +
mean).view(shape)
def decompress(self,
data: bytearray,
offset: Offset = None) -> torch.Tensor:
offset = offset or Offset()
# unpack original tensor shape
shape = unpack_tensor_shape(data, offset)
# unpack indices and data
indices = unpack_tensor(data, offset)
values = unpack_tensor(data, offset)
# re-construct tensor
tensor = torch.zeros(shape)
tensor.index_copy_(0, indices, values)
return tensor
def decompress(self,
data: bytearray,
offset: Offset = None) -> torch.Tensor:
offset = offset or Offset()
# unpack original tensor shape
shape = unpack_tensor_shape(data, offset)
# unpack compressed tensor
tensor_data = unpack_tensor(data, offset)
# restore random mask
k = math.ceil(self.ratio * np.prod(shape))
torch.manual_seed(self.seed)
mask = torch.randperm(int(np.prod(shape))).cuda() < k
idx = mask.int().sort(descending=True).indices
# restore tensor
tensor = torch.zeros(int(np.prod(shape)))
tensor.index_copy_(0, idx[:k].cpu(), tensor_data)
tensor = tensor.reshape(shape)
return tensor
def decompress(self,
data: bytearray,
offset: Offset = None) -> torch.Tensor:
offset = offset or Offset()
return unpack_tensor(data, offset)