#!/usr/bin/env python3
import torch
from ..utils.broadcasting import _matmul_broadcast_shape, _mul_broadcast_shape, _pad_with_singletons
from ..utils.getitem import _noop_index
from ..utils.memoize import cached
from .diag_lazy_tensor import DiagLazyTensor
from .lazy_tensor import LazyTensor
from .non_lazy_tensor import NonLazyTensor, lazify
def _inner_repeat(tensor, amt):
return tensor.unsqueeze(-1).repeat(amt, 1).squeeze(-1)
def _outer_repeat(tensor, amt):
return tensor.unsqueeze(-1).repeat(1, amt).view(-1)
[docs]class MatmulLazyTensor(LazyTensor):
def __init__(self, left_lazy_tensor, right_lazy_tensor):
left_lazy_tensor = lazify(left_lazy_tensor)
right_lazy_tensor = lazify(right_lazy_tensor)
# Match batch dimensions
batch_shape = _mul_broadcast_shape(left_lazy_tensor.batch_shape, right_lazy_tensor.batch_shape)
if left_lazy_tensor.batch_shape != batch_shape:
left_lazy_tensor = left_lazy_tensor._expand_batch(batch_shape)
if right_lazy_tensor.batch_shape != batch_shape:
right_lazy_tensor = right_lazy_tensor._expand_batch(batch_shape)
super().__init__(left_lazy_tensor, right_lazy_tensor)
batch_shape = _mul_broadcast_shape(left_lazy_tensor.batch_shape, right_lazy_tensor.batch_shape)
if left_lazy_tensor.batch_shape != batch_shape:
self.left_lazy_tensor = left_lazy_tensor._expand_batch(batch_shape)
else:
self.left_lazy_tensor = left_lazy_tensor
if right_lazy_tensor.batch_shape != batch_shape:
self.right_lazy_tensor = right_lazy_tensor._expand_batch(batch_shape)
else:
self.right_lazy_tensor = right_lazy_tensor
def _expand_batch(self, batch_shape):
return self.__class__(
self.left_lazy_tensor._expand_batch(batch_shape), self.right_lazy_tensor._expand_batch(batch_shape)
)
def _get_indices(self, row_index, col_index, *batch_indices):
row_index = row_index.unsqueeze(-1)
col_index = col_index.unsqueeze(-1)
batch_indices = tuple(batch_index.unsqueeze(-1) for batch_index in batch_indices)
inner_index = torch.arange(0, self.left_lazy_tensor.size(-1), device=self.device)
inner_index = _pad_with_singletons(inner_index, row_index.dim() - 1, 0)
left_tensor = self.left_lazy_tensor._get_indices(
row_index, inner_index, *batch_indices[-len(self.left_lazy_tensor.batch_shape) :]
)
right_tensor = self.right_lazy_tensor._get_indices(
inner_index, col_index, *batch_indices[-len(self.right_lazy_tensor.batch_shape) :]
)
res = (left_tensor * right_tensor).sum(-1)
return res
def _getitem(self, row_index, col_index, *batch_indices):
# Make sure we're not generating more memory with our "efficient" method
if torch.is_tensor(row_index) and torch.is_tensor(col_index):
num_indices = row_index.numel()
if num_indices > self.matrix_shape.numel():
return lazify(self.evaluate())._getitem(row_index, col_index, *batch_indices)
left_tensor = self.left_lazy_tensor._getitem(row_index, _noop_index, *batch_indices)
right_tensor = self.right_lazy_tensor._getitem(_noop_index, col_index, *batch_indices)
res = MatmulLazyTensor(left_tensor, right_tensor)
return res
def _matmul(self, right_lazy_tensor):
return self.left_lazy_tensor._matmul(self.right_lazy_tensor._matmul(right_lazy_tensor))
def _t_matmul(self, right_lazy_tensor):
return self.right_lazy_tensor._t_matmul(self.left_lazy_tensor._t_matmul(right_lazy_tensor))
def _quad_form_derivative(self, left_vecs, right_vecs):
if left_vecs.ndimension() == 1:
left_vecs = left_vecs.unsqueeze(1)
right_vecs = right_vecs.unsqueeze(1)
right_vecs_times_right_lazy_tensor = self.right_lazy_tensor._matmul(right_vecs)
left_vecs_times_left_lazy_tensor_t = self.left_lazy_tensor._t_matmul(left_vecs)
left_grad = self.left_lazy_tensor._quad_form_derivative(left_vecs, right_vecs_times_right_lazy_tensor)
right_grad = self.right_lazy_tensor._quad_form_derivative(left_vecs_times_left_lazy_tensor_t, right_vecs)
left_grad = (left_grad,) if not isinstance(left_grad, tuple) else left_grad
right_grad = (right_grad,) if not isinstance(right_grad, tuple) else right_grad
return left_grad + right_grad
def _permute_batch(self, *dims):
return self.__class__(self.left_lazy_tensor._permute_batch(*dims), self.right_lazy_tensor._permute_batch(*dims))
def _size(self):
return _matmul_broadcast_shape(self.left_lazy_tensor.shape, self.right_lazy_tensor.shape)
def _transpose_nonbatch(self, *args):
return self.__class__(self.right_lazy_tensor._transpose_nonbatch(), self.left_lazy_tensor._transpose_nonbatch())
def diag(self):
if isinstance(self.left_lazy_tensor, NonLazyTensor) and isinstance(self.right_lazy_tensor, NonLazyTensor):
return (self.left_lazy_tensor.tensor * self.right_lazy_tensor.tensor.transpose(-1, -2)).sum(-1)
elif isinstance(self.left_lazy_tensor, DiagLazyTensor) or isinstance(self.right_lazy_tensor, DiagLazyTensor):
return self.left_lazy_tensor.diag() * self.right_lazy_tensor.diag()
else:
return super().diag()
@cached
def evaluate(self):
return torch.matmul(self.left_lazy_tensor.evaluate(), self.right_lazy_tensor.evaluate())