Statistics and data splitting scheme on PPI datasets
sduzxj opened this issue · 2 comments
Using the following code, I will get different statistics than the documentation[1, 2], how do I get the preprocessed dataset or the code about the dataset splitting scheme [2]？
code:
import torch
import torchdrug
from torchdrug import datasets
from torchdrug import core, datasets, tasks, models, layers
from torchdrug.datasets import HumanPPI,YeastPPI,PPIAffinity,Fold
from torchdrug import data, utils
from torchdrug import transforms as T
dataset = YeastPPI('./dataset/PPI', lazy =True)#, #transform=transforms)
train_set,valid_set,test_set =dataset.split(['train', 'valid', 'test'])#.split()
print(len(train_set))
print(len(valid_set))
print(len(test_set))
output statistics : 2421, 203, 326
[1] https://torchdrug.ai/docs/api/datasets.html
[2] https://torchprotein.ai/benchmark#leaderboard-for-yeast-ppi-prediction
I have the same issue with subcellular localization dataset. The number of samples in the training, validation and test sets are different compared to the PEER paper. What is the problem? Is there any additional post-prosessing step we need to do on the dataset?