关于Lorenz在aistudio上的断点续训练

[xiaoniv]

你好。我在aistudio上跑通了那个Lorenz例子，想试下断点继续训练。我跑了10个epoch，按了停止退出codelab。然后重新进入codelab,安装paddlescience，再运行训练之前的代码。我想用初始化直接读取checkpoint，就运行以下代码。

OUTPUT_DIR = "./output/lorenz_transformer"
cp_path = f"{OUTPUT_DIR}/checkpoints"

new_solver = ppsci.solver.Solver(
    model=model,
    constraint=constraint,
    output_dir=OUTPUT_DIR,  # Added output_dir
    optimizer=optimizer,
    lr_scheduler=lr_scheduler,
    eval_during_train=True,
    eval_freq=50,
    validator=validator,
    visualizer=visualizer,
    checkpoint_path=cp_path,
) 

得到报错AttributeError: 'Solver' object has no attribute 'loss_aggregator'，显式给loss_aggregator赋值也不行。然后我换成用save_load把model和optimizer读出来，再尝试初始化。运行以下代码

import ppsci.utils.save_load as save_load

OUTPUT_DIR = "./output/lorenz_transformer"
checkpoint_dir = f"{OUTPUT_DIR}/checkpoints/latest"

save_load.load_checkpoint(
    checkpoint_dir,
    model,  # Your model
    optimizer  # Your optimizer
)

solver = ppsci.solver.Solver(
    model=model,  # Will use loaded weights
    constraint=constraint,
    output_dir=OUTPUT_DIR,
    optimizer=optimizer,  # Will use loaded state
    lr_scheduler=lr_scheduler,
    eval_during_train=True,
    eval_freq=50,
    validator=validator,
    visualizer=visualizer,
)

solver.train() 

得到报错AssertionError: Optimizer set error, layer_norm_1.w_0_moment1_0 should in state dict。
我看了latest.pdopt,layer_norm是从21开始的，layer_norm_21.w_0_moment1_0。
请帮我分析一下，谢谢