dense_stack_tds¶
- class tensordict.dense_stack_tds(td_list: Union[Sequence[TensorDictBase], LazyStackedTensorDict], dim: Optional[int] = None)¶
密集地堆疊一個
TensorDictBase物件列表(或一個LazyStackedTensorDict),前提是它們具有相同的結構。此函式接受一個
TensorDictBase物件列表(可以直接傳入,或從LazyStackedTensorDict獲取)。與呼叫torch.stack(td_list)會返回一個LazyStackedTensorDict不同,此函式會展開輸入列表的第一個元素,然後將輸入列表堆疊到該元素上。這僅適用於輸入列表的所有元素具有相同結構的情況。返回的TensorDictBase將具有與輸入列表元素相同的型別。當某些需要堆疊的
TensorDictBase物件是LazyStackedTensorDict或在其條目(或巢狀條目)中包含LazyStackedTensorDict時,此函式非常有用。在這些情況下,呼叫torch.stack(td_list).to_tensordict()是不可行的。因此,此函式提供了一種密集堆疊所提供列表的替代方法。- 引數:
td_list (TensorDictBase 列表 或 LazyStackedTensorDict) – 要堆疊的 tds。
dim (int, 可選) – 堆疊它們的維度。如果 td_list 是 LazyStackedTensorDict,則將自動檢索。
示例
>>> import torch >>> from tensordict import TensorDict >>> from tensordict import dense_stack_tds >>> from tensordict.tensordict import assert_allclose_td >>> td0 = TensorDict({"a": torch.zeros(3)},[]) >>> td1 = TensorDict({"a": torch.zeros(4), "b": torch.zeros(2)},[]) >>> td_lazy = torch.stack([td0, td1], dim=0) >>> td_container = TensorDict({"lazy": td_lazy}, []) >>> td_container_clone = td_container.clone() >>> td_stack = torch.stack([td_container, td_container_clone], dim=0) >>> td_stack LazyStackedTensorDict( fields={ lazy: LazyStackedTensorDict( fields={ a: Tensor(shape=torch.Size([2, 2, -1]), device=cpu, dtype=torch.float32, is_shared=False)}, exclusive_fields={ }, batch_size=torch.Size([2, 2]), device=None, is_shared=False, stack_dim=0)}, exclusive_fields={ }, batch_size=torch.Size([2]), device=None, is_shared=False, stack_dim=0) >>> td_stack = dense_stack_tds(td_stack) # Automatically use the LazyStackedTensorDict stack_dim TensorDict( fields={ lazy: LazyStackedTensorDict( fields={ a: Tensor(shape=torch.Size([2, 2, -1]), device=cpu, dtype=torch.float32, is_shared=False)}, exclusive_fields={ 1 -> b: Tensor(shape=torch.Size([2, 2]), device=cpu, dtype=torch.float32, is_shared=False)}, batch_size=torch.Size([2, 2]), device=None, is_shared=False, stack_dim=1)}, batch_size=torch.Size([2]), device=None, is_shared=False) # Note that # (1) td_stack is now a TensorDict # (2) this has pushed the stack_dim of "lazy" (0 -> 1) # (3) this has revealed the exclusive keys. >>> assert_allclose_td(td_stack, dense_stack_tds([td_container, td_container_clone], dim=0)) # This shows it is the same to pass a list or a LazyStackedTensorDict