__init__.py

from torchvision import transforms as transforms
from torch.utils.data import DataLoader
from dataset.RGBDataset import Div2KTrain, Banchmark, CropDataset 
from dataset.YUVDataset import *
from utils.transforms import DescreteRandomRotation
from torch.utils.data import random_split
import os
import torch
import numpy as np



def RGBDataset(args, useBGR=False, test_only=False):
    # Dataset Set Up ###########################################

    ############ Image data augmentation for retrain ###########
    transform = transforms.Compose([
        DescreteRandomRotation(angles=[0, 90, 180, -90]),
        transforms.RandomVerticalFlip(),
        transforms.RandomHorizontalFlip(),
    ])

    # Select the Data load er for the datasets (biggest difference is the folder structure and the loading process)
    dataset = Div2KTrain(scaling=args.scale, useBGR=useBGR)
    dataset_test = Banchmark(scaling=args.scale, useBGR=useBGR)

    
    if args.test_mode:
        use_only = 0.2
        
        kept_num = round(len(dataset) * use_only)
        igonored_num = round(len(dataset) * (1-use_only))
        dataset, _ = random_split(dataset, [kept_num, igonored_num], generator=torch.Generator().manual_seed(args.seed))

        use_only = 0.1
        kept_num = round(len(dataset_test) * use_only)
        igonored_num = round(len(dataset_test) * (1-use_only))
        dataset_test, _ = random_split(dataset_test, [kept_num, igonored_num], generator=torch.Generator().manual_seed(args.seed))

    if not test_only :
        #Split the dataset in train and validation
        train_percentual = 0.8
        train_num = round(len(dataset) * train_percentual)
        val_num = round(len(dataset) * (1-train_percentual))
        train_data, val_data = random_split(dataset, [train_num, val_num], generator=torch.Generator().manual_seed(args.seed))
        train_data = CropDataset(train_data, scaling=args.scale, transform=transform, seed=args.seed, crop=args.crop, cropMode="random", noiseInensity=args.noise)
        val_data = CropDataset(val_data, scaling=args.scale, transform=None, seed=None, crop=args.crop, cropMode="center")
        
        # create the loader for the training set
        train_loader = DataLoader(train_data, shuffle=True,  batch_size=args.batch, num_workers=0, pin_memory=True, worker_init_fn = lambda id: np.random.seed(id))
        # create the loader for the validation set (to select the model after prune)
        val_loader = DataLoader(val_data, shuffle=False, batch_size=args.batch, num_workers=0, pin_memory=True, worker_init_fn = lambda id: np.random.seed(id))
        # create the loader for the test set (to evaluate the model performaces after selection)
        test_loader = DataLoader(dataset_test, shuffle=False, batch_size=1, num_workers=0, pin_memory=True, worker_init_fn = lambda id: np.random.seed(id))
    
        return (train_loader, val_loader, test_loader)

    else:
        test_loader = DataLoader(dataset_test, shuffle=False, batch_size=1, num_workers=0, pin_memory=True, worker_init_fn = lambda id: np.random.seed(id))
        return (train_loader, val_loader, test_loader)
    
def YUVDataset(args, test_only=False):
   # Dataset Set Up ###########################################

    ############ Image data augmentation for retrain ###########
    transform = transforms.Compose([
        DescreteRandomRotation(angles=[0, 90, 180, -90]),
        transforms.RandomVerticalFlip(),
        transforms.RandomHorizontalFlip(),
    ])

    if not test_only:
        dataset = DatasetFromFolderYUV(
                scale=args.scale,
                size=args.yuv_size,
                yuvFormat=args.yuv_format,
                y_only=args.only_y_channel,
                filter_only=args.filter,
                low_res_folder = os.path.normpath(args.low_res_data),
                prefix_low_res = args.prefix_lowres,
                high_res_folder = os.path.normpath(args.high_res_data),   
                prefix_high_res = args.prefix_highres,
                transform=transform,
                seed=args.seed,
            )

    if(args.yuv_testset.lower() == 'evcintra'):
        dataset_test = TestSequenciesEVCIntra(yOnly=args.only_y_channel)
    elif(args.yuv_testset.lower() == 'evcsdhd'):
        dataset_test = TestSequenciesEVCSDHD(yOnly=args.only_y_channel)
    elif(args.yuv_testset.lower() == 'evchd4k'):
        dataset_test = TestSequenciesEVCHD4K(yOnly=args.only_y_channel)
    elif(args.yuv_testset.lower() == 'vvcsdhd'):
        dataset_test = TestSequenciesVVCSDHD(yOnly=args.only_y_channel)
    elif(args.yuv_testset.lower() == 'vvchd4k'):
        dataset_test = TestSequenciesVVCHD4K(yOnly=args.only_y_channel)
    else:
        raise Exception(f"Unsupported Raw Testset:{args.yuv_testset}. see comand info for detrails.")

    if not test_only:
        #Split the dataset in train and validation
        train_percentual = 0.8
        train_num = round(len(dataset) * train_percentual)
        val_num = round(len(dataset) * (1-train_percentual))
        train_data, val_data = random_split(dataset, [train_num, val_num], generator=torch.Generator().manual_seed(args.seed))
        val_data.dataset.transform = None # Deactivate Transform in validation
        
        #create the loader for the training set
        train_loader = DataLoader(train_data, shuffle=True,  batch_size=args.batch, num_workers=1, pin_memory=True, worker_init_fn = lambda id: np.random.seed(id))
        #create the loader for the validation set
        val_loader = DataLoader(val_data, shuffle=False, batch_size=args.batch, num_workers=1, pin_memory=True, worker_init_fn = lambda id: np.random.seed(id))
        test_loader = DataLoader(dataset_test, shuffle=False, batch_size=1, num_workers=1, pin_memory=True, worker_init_fn = lambda id: np.random.seed(id))
        
        return(train_loader, val_loader, test_loader)
    else:
        test_loader = DataLoader(dataset_test, shuffle=False, batch_size=1, num_workers=1, pin_memory=True, worker_init_fn = lambda id: np.random.seed(id))
        return test_loader