Problem in load_from_checkpoint

1

I have a lightning module:

class MyClassifier(pl.LightningModule):

    def __init__(self, config):
        super().__init__()
        self.save_hyperparameters()
        self.num_classes = config["arch"]["args"]["num_classes"]
        self.model_args = config["arch"]["args"]
        self.model, self.tokenizer = get_model_and_tokenizer(**self.model_args)
        self.bias_loss = False

I train the model using this:

model = MyClassifier(config)

checkpoint_callback = ModelCheckpoint(
    save_top_k=5,
    verbose=True,
    monitor="val_f1",
    mode="max",
)

trainer = pl.Trainer(
    accelerator='auto',
    strategy="ddp",
    max_epochs=args.n_epochs,
    accumulate_grad_batches=config["accumulate_grad_batches"],
    callbacks=[checkpoint_callback],
    default_root_dir="saved/" + config["name"],
    deterministic=True,
    precision=16
)
trainer.fit(model, data_loader, valid_data_loader)

I try to load a checkpoint using this:

model = MyClassifier.load_from_checkpoint(PATH)

but I get this error:

RuntimeError: Error(s) in loading state_dict for MyClassifier:
        Missing key(s) in state_dict: "model.roberta.embeddings.position_ids".

I have also tried this:

model = MyClassifier.load_from_checkpoint(PATH,config=config)

but I get the same error. I also tried loading state_dict using this:

ckpoint = torch.load(PATH)
model = MyClassifier.load_state_dict(ckpoint['state_dict'])

but I get this error:

TypeError: Module.load_state_dict() missing 1 required positional argument: 'state_dict'

My Pytorch Lightning version is 2.0.7. I run training on 4 GPUs.

2

Hi, I also encountered this problem. Did you work this out now?