[Q]: What's the native way to split datasets into train, validation and test? about disent HOT 4 CLOSED

import math

import pytorch_lightning as pl
import torch
from torch.utils.data import DataLoader

from disent.dataset import DisentDataset
from disent.dataset.data import Cars3dData
from disent.dataset.transform import ToImgTensorF32
from disent.frameworks.ae import Ae
from disent.metrics import metric_dci
from disent.metrics import metric_mig
from disent.model import AutoEncoder
from disent.model.ae import DecoderConv64
from disent.model.ae import EncoderConv64


# normalise the data
vis_mean = [0.8976676149976628, 0.8891658020067508, 0.885147515814868]
vis_std = [0.22503195531503034, 0.2399461278981261, 0.24792106319684404]
data = Cars3dData(transform=ToImgTensorF32(size=64, mean=vis_mean, std=vis_std))

# SOLUTION:
# -- split the data using built-in functions (no longer ground-truth datasets, but subsets)
data_train, data_val, data_test = torch.utils.data.random_split(data, [
    int(math.floor(len(data)*0.6)),
    int(math.ceil(len(data)*0.2)),
    int(math.ceil(len(data)*0.2)),
])
# -- create multiple disent datasets
dataset_train = DisentDataset(data_train)
dataset_val   = DisentDataset(data_val)
dataset_test  = DisentDataset(data_test)
# -- create dataloaders
dataloader_train = DataLoader(dataset=dataset_train, batch_size=4, shuffle=True, num_workers=0)
dataloader_val   = DataLoader(dataset=dataset_val, batch_size=4, shuffle=True, num_workers=0)
dataloader_test  = DataLoader(dataset=dataset_test, batch_size=4, shuffle=True, num_workers=0)

# create the pytorch lightning system
module: pl.LightningModule = Ae(
    model=AutoEncoder(
        encoder=EncoderConv64(x_shape=(3, 64, 64), z_size=6),
        decoder=DecoderConv64(x_shape=(3, 64, 64), z_size=6),
    ),
    cfg=Ae.cfg(
        optimizer="adam", optimizer_kwargs=dict(lr=1e-3)
    ),
)

# PROBLEM: unfortunately the framework does not yet implement the pytorch-lightning validation step
#          I'll add this to the roadmap and this should work in future.
trainer = pl.Trainer(max_steps=10000, checkpoint_callback=False, gpus=1 if torch.cuda.is_available() else None)
trainer.fit(module, dataloader_train, dataloader_val)

# PROBLEM: unfortunately the metrics will no longer work with the subsets
#          of data. You could instead pass the original full dataset to the
#          metrics, but this may be considered an information leak?
#          -- This will crash!
get_repr = lambda x: module.encode(x.to(module.device))
metrics = {
    **metric_dci(dataset_test, get_repr, num_train=1000, num_test=500, show_progress=True),
    **metric_mig(dataset_test, get_repr, num_train=2000),
}
print("metrics:", metrics)

from disent.dataset.sampling import RandomSampler

from disent.dataset.sampling import GroundTruthPairOrigSampler
from disent.dataset.sampling import GroundTruthPairSampler