torch_npu.utils.save_async

API接口

torch_npu.utils.save_async(obj, f, pickle_module=pickle, pickle_protocol=DEFAULT_PROTOCOL, _use_new_zipfile_serialization=True, model=None)

功能描述

异步保存一个对象到一个硬盘文件上。

参数说明

• obj (object) – 要保存的对象
• f (Union[str, PathLike, BinaryIO, IO[bytes]]) – 保存的目标文件路径或文件句柄
• pickle_module (Any,默认值为pickle) – 用于Pickle序列化的模块
• pickle_protocol (int,默认值为DEFAULT_PROTOCOL) – Pickle协议版本
• _use_new_zipfile_serialization (Bool,默认值为True) - 是否使用新的Zip文件序列化
• _disable_byteorder_record (Bool,默认值为False) - 是否禁用字节顺序记录
• model (torch.nn.Module,默认值为None) - 模型对象，需要对模型反向注册hook

示例

input = torch.tensor([1.,2.,3.,4.]).npu()
torch_npu.utils.save_async(input, "save_tensor.pt")
model = nn.Sequential(
	nn.Linear(100, 50),
	nn.ReLU(),
	nn.Linear(50, 20),
	nn.ReLU(),
	nn.Linear(20, 5),
	nn.ReLU()
)
model = model.npu()
criterion = nn.CrossEntropyLoss()
optimerizer = optim.SGD(model.parameters(), lr=0.01)
for epoch in range(3):
	for step in range(3):
		input_data = torch.ones(6400, 100).npu()
		labels = torch.randint(0, 5, (6400,)).npu()
		outputs = model(input_data)
		loss = criterion(outputs, labels)
		optimerizer.zero_grad()
		loss.backward()
		optimerizer.step()
	save_path = os.path.join(f"model_{epoch}_{step}.path")
	torch_npu.utils.save_async(model, save_path, model=model)