How can I use the model for prediction without loading the whole dataset?

[arseny239]

Hi,

First of all let me thank you for such a great library. I really like it!

But I have a question. I use my own dataset; I created the .item and .iter files, it loads and trains ok.
But it is quite big (882 000 users, 1.4 mln items) so it takes quite a lot of memory and time to train the model.
(I try to use the sequential recommender - SASRec and SHAN)
So I need to use the trained model for prediction (to make some recomendations) without loading the full dataset.
How can I do it?

I've read this topic: #1288 but I still can not understand, how to load the model at all?
For now, I train the model using
run_recbole(model=my_model_name, dataset=DATASET_NAME, config_dict=parameter_dict)
and then
config, model, dataset, train_data, valid_data, test_data = load_data_and_model(model_file='saved/my_model.pth')
but it still takes a lot of memory

I checked the code of the load_data_and_model function and tried to do:

checkpoint = torch.load(model_file)
config = checkpoint["config"]
...
dataset = create_dataset(config)
logger.info(dataset)
train_data, valid_data, test_data = data_preparation(config, dataset)

model = get_model(config["model"])(config, train_data._dataset).to(config["device"])
model.load_state_dict(checkpoint["state_dict"])
model.load_other_parameter(checkpoint.get("other_parameter"))

but it is almost the same thing. As I can understand, it still loads the train_data, so it takes a lot of memory

Is there any way to avoid this?

And then, how should I create the interaction object for prediction ? I hardly can understand it. Could you explain it, please.
For today I use
scores, topk_iid_list = full_sort_topk([user], model, test_data, k=N_PREDICTIONS, device=config['device'])
but it requires the test_data, so I have to use data_preparation, and therefore load all the data (am I right?)
Or maybe there are some code examples somewhere?

Sorry for the naive questions, I'm new to working with the Recbole.

Sincerely yours,
Arseny

[throw-away-qq]

I wrote a blogpost for using a trained Sequential model, Please take a look https://throw-away-qq.github.io/recbole_to_production/
let me know if it is helpful!

Best,
Rajat