LazyTensorStorage¶
- class torchrl.data.replay_buffers.LazyTensorStorage(max_size: int, *, device: device = 'cpu', ndim: int = 1)[source]¶
用于张量和 TensorDict 的预分配张量存储。
- 参数:
max_size (int) – 存储的大小,即缓冲区中存储的最大元素数量。
- 关键字参数:
device (torch.device, 可选) – 采样张量将被存储和发送到的设备。默认为
torch.device("cpu")
。如果传递 “auto”,则设备会自动从传递的第一批数据中收集。默认情况下不启用此功能,以避免数据被错误地放置在 GPU 上,从而导致 OOM 问题。ndim (int, 可选) – 测量存储大小时要考虑的维度数。例如,如果
ndim=1
,则形状为[3, 4]
的存储容量为3
;如果ndim=2
,则容量为12
。默认为1
。
示例
>>> data = TensorDict({ ... "some data": torch.randn(10, 11), ... ("some", "nested", "data"): torch.randn(10, 11, 12), ... }, batch_size=[10, 11]) >>> storage = LazyTensorStorage(100) >>> storage.set(range(10), data) >>> len(storage) # only the first dimension is considered as indexable 10 >>> storage.get(0) TensorDict( fields={ some data: Tensor(shape=torch.Size([11]), device=cpu, dtype=torch.float32, is_shared=False), some: TensorDict( fields={ nested: TensorDict( fields={ data: Tensor(shape=torch.Size([11, 12]), device=cpu, dtype=torch.float32, is_shared=False)}, batch_size=torch.Size([11]), device=cpu, is_shared=False)}, batch_size=torch.Size([11]), device=cpu, is_shared=False)}, batch_size=torch.Size([11]), device=cpu, is_shared=False) >>> storage.set(0, storage.get(0).zero_()) # zeros the data along index ``0``
此类也支持 tensorclass 数据。
示例
>>> from tensordict import tensorclass >>> @tensorclass ... class MyClass: ... foo: torch.Tensor ... bar: torch.Tensor >>> data = MyClass(foo=torch.randn(10, 11), bar=torch.randn(10, 11, 12), batch_size=[10, 11]) >>> storage = LazyTensorStorage(10) >>> storage.set(range(10), data) >>> storage.get(0) MyClass( bar=Tensor(shape=torch.Size([11, 12]), device=cpu, dtype=torch.float32, is_shared=False), foo=Tensor(shape=torch.Size([11]), device=cpu, dtype=torch.float32, is_shared=False), batch_size=torch.Size([11]), device=cpu, is_shared=False)
- attach(buffer: Any) None ¶
此函数将采样器附加到此存储。
从此存储读取数据的缓冲区必须通过调用此方法作为附加实体包含在内。这保证了即使存储与其他缓冲区(例如优先级采样器)共享,组件也能意识到存储中的数据更改。
- 参数:
buffer – 从此存储读取数据的对象。