I also set continual_learning: true, and when I set it to false, the error disappeared，Can't both of them be enabled at the same time in backtesting? Or in backtest mode, will the previous model be automatically loaded when training a new model?

I found the reason, the save method of PyTorchModelTrainer saves self

torch.save({
             "model_state_dict": self.model.state_dict(),
             "optimizer_state_dict": self.optimizer.state_dict(),
             "model_meta_data": self.model_meta_data,
             "pytrainer": self
         }, path)

But I don't know how to fix it

Hello,

continual_learning for PyTorch models may only be suppprted in Dry Live modes.

If you need continual_learning in backtesting as well, you can use one of the other supported 18 models like XGBoost/Catboost.

You may have noticed in the documentation of continual_learning that this is a highly experimental feature, so if you would like to use it with PyTorch in backtesting, you may need to wait until someone has time to try to debug it for you. I will try my best, but it is not a top priority at the moment since it is an edge case on an experimental feature.

https://www.freqtrade.io/en/stable/freqai-running/#continual-learning

cheers,

rob

Hello,

continual_learning for PyTorch models may only be suppprted in Dry Live modes.

If you need continual_learning in backtesting as well, you can use one of the other supported 18 models like XGBoost/Catboost.

You may have noticed in the documentation of continual_learning that this is a highly experimental feature, so if you would like to use it with PyTorch in backtesting, you may need to wait until someone has time to try to debug it for you. I will try my best, but it is not a top priority at the moment since it is an edge case on an experimental feature.

https://www.freqtrade.io/en/stable/freqai-running/#continual-learning

cheers,

rob

Thank you for your work. Using continual_learning in Dry mode does not work. This problem may only exist with PyTorchTransformerRegressor. I have not tried continual_learning on other pytorch-based models, and I think turning on continual_learning on the Transformer model may have better results. Hope this issue can be resolved.

I solved this problem by modifying the following code，Modify the fit method of the file PyTorchTransformerRegressor.py

def fit(self, data_dictionary: Dict, dk: FreqaiDataKitchen, **kwargs) -> Any:
        """
        User sets up the training and test data to fit their desired model here
        :param data_dictionary: the dictionary holding all data for train, test,
            labels, weights
        :param dk: The datakitchen object for the current coin/model
        """

        n_features = data_dictionary["train_features"].shape[-1]
        n_labels = data_dictionary["train_labels"].shape[-1]
        model = PyTorchTransformerModel(
            input_dim=n_features,
            output_dim=n_labels,
            time_window=self.window_size,
            **self.model_kwargs
        )
        model.to(self.device)
        optimizer = torch.optim.AdamW(model.parameters(), lr=self.learning_rate)
        criterion = torch.nn.MSELoss()
        # check if continual_learning is activated, and retreive the model to continue training
        temp_trainer = self.get_init_model(dk.pair)
        if temp_trainer is None:
            trainer = PyTorchTransformerTrainer(
                model=model,
                optimizer=optimizer,
                criterion=criterion,
                device=self.device,
                data_convertor=self.data_convertor,
                window_size=self.window_size,
                tb_logger=self.tb_logger,
                **self.trainer_kwargs,
            )
        else:
            trainer = PyTorchTransformerTrainer(
                model=temp_trainer.model,
                optimizer=temp_trainer.optimizer,
                model_meta_data=temp_trainer.model_meta_data,
                criterion=criterion,
                device=self.device,
                data_convertor=self.data_convertor,
                window_size=self.window_size,
                tb_logger=self.tb_logger,
                **self.trainer_kwargs,
            )
        trainer.fit(data_dictionary, self.splits)
        return trainer

I don't know if this modification is correct, but it makes it run.