Creating and Training on Custom Dataset

[shahjaidev]

Hello!

I have a custom dataset that I have formatted according to recbole's Atomic Files guide (https://www.recbole.io/atomic_files.html) and created the custom_dataset.inter file

I place it in /dataset/custom_dataset/ , and then run:

python run_recbole.py
--model=DirectAU --dataset=custom_dataset
--learning_rate=1e-3 --weight_decay=1e-6
--gamma=1 --encoder=MF --train_batch_size=1024

However, while the true #users, # items ~ (2 million, 2 million), the terminal output says #Users are 360587 (and sets the batch size to that) . [ Pic Attached ]

Why this discrepancy?

I also tried with explicitly passing a config file as well.
See below

[THUwangcy]

RecBole will filter the dataset according to the number of user/item interactions. The default setting in sample.yaml is

min_user_inter_num: 5
min_item_inter_num: 5

You can change the setting to customize the filtering strategy.

[shahjaidev]

Thanks for the answer!

Where in the code, after run_recbole is called, is sample.yaml read?
I tried adding
min_user_inter_num: 0
min_item_inter_num: 0

to the original config file itself (passed as command line arg) but seems like these get overwritten.

[shahjaidev]

Similarly, settting embedding_size to 128 in my overall config file (pasted below for reference) also faces the same issue, embedding size remains 64

[THUwangcy]

I'm not very sure about the priority of config files in RecBole, but the command line args always work for me. You can try to set these hyperparameters by command line, and check whether the settings in the output log change accordingly.

[shahjaidev]

Thanks for your answer!