The data that is fed into models.

[Jeriousman]

For example, for BERT4Rec, in its init, we should feed dataset. But as there is a flow of data (raw -> atomic -> dataframe -> dadaloader), I am a bit confused so I am asking to clarify.

What should the dataset that is fed to the init(config, dataset)? (below you can see the example of BERT4REC) Is it the dataset form of data I get when I do config 'save_dataloaders': True?

As far as I know, the dataloaders we get from data_preparation function is the same with data from 'save_dataloaders': True.

class BERT4Rec(SequentialRecommender):

    def __init__(self, config, dataset):
        super(BERT4Rec, self).__init__(config, dataset)

[leoleojie]

@Jeriousman Hello, thanks for your attention to RecBole! As you said, the overall data flow can be described as raw->atomic->dataset(dataframe)->dataloader(interaction). We use create_dataset() to convert atomic files into Dataset(That is what we fed to init.), during the transformation we make a series of preprocess. And then we use data_preparation to convert Dataset into Dataloader, where the data will be splitted. Here is an example:

    # dataset filtering
    dataset = create_dataset(config)
    logger.info(dataset)

    # dataset splitting
    train_data, valid_data, test_data = data_preparation(config, dataset)

    # model loading and initialization
    model = NewModel(config, train_data.dataset).to(config['device'])

[Jeriousman]

So what we feed into the init is the Dataset from create_dataset() and then model(config, Dataloader Dataset)? (Just to double check). Thank you!

[leoleojie]

Sorry, I don't know Dataloader Dataset means. Maybe what I said above is not clear. In the example above, after we get the dataset from create_dataset(), we use data_preparation(config, dataset) to get train_data, valid data, test_data(they are all dataloader ). but during data_preparation, the interaction features in dataset will be split and all the other attributes the same:

built_datasets = dataset.build()
train_dataset, valid_dataset, test_dataset = built_datasets
# they are all class:`~Dataset`

And it set train_data.dataset = train_dataset, valid_data.dataset =valid_dataset... So in the example above, model = NewModel(config, train_data.dataset) , train_data.dataset is just class:~Dataset, whose interaction features has been split.

Maybe not clear enough, but if you have any questions please ask. You can also read the code in this section to understand better

[Jeriousman]

I know what you mean. but I just asked is even though the difference is the whole data and the split of the data, data from create_dataset() is in atomic format, and data from data_preparation(config, dataset) is in dataloader format so what I asked is

data from create_dataset() goes to init of BERT4Rec for example, and, data from data_preparation(config, dataset) goes to below (let's call the data -> train_data which is the dataloader data from data_preparation)

iter_data = (
    tqdm(
        train_data,
        total=len(train_data),
        ncols=100,

    ) if False else train_data
)

for batch_idx, interaction in enumerate(iter_data):
    interaction = interaction.to(model.device)

for calculate_loss as dataloader format goes in for mini-batch training. I just wanted to be crystal clear if I am right. Your feedback has been awesome by the way. So, I guess I am right on this?

Thank you in advance.

[leoleojie]

Yeah, you are right.

[Jeriousman]

@leoleojie Thank you very much. Xiexie