PDGrapher/data/scripts/splits/create_standard_splits.py

##Create splits to be used for training across all models

import os
import os.path as osp
from sklearn.model_selection import KFold, train_test_split
import torch
from sklearn.metrics import jaccard_score
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt



def create_splits_n_fold(dataset, splits_type, nfolds, outdir):
	kf = KFold(nfolds, shuffle=True, random_state=42)
	os.makedirs(outdir, exist_ok = True)

	#datasets forward and backward
	dataset_forward = dataset[0]; dataset_backward = dataset[1]
	splits = {}


	if splits_type =='random':
		i = 1
		if len(dataset_forward)> 0:
			for train_test_index_forward, train_test_index_backward in zip(kf.split(dataset_forward), kf.split(dataset_backward)):
				#Forward
				train_index_forward = train_test_index_forward[0]; test_index_forward = train_test_index_forward[1]
				train_index_backward = train_test_index_backward[0]; test_index_backward = train_test_index_backward[1]
				#Backward
				train_index_forward, val_index_forward = train_test_split(train_index_forward, test_size=0.2, random_state=42)
				train_index_backward, val_index_backward = train_test_split(train_index_backward, test_size=0.2, random_state=42)

				assert len(dataset_backward) == len(train_index_backward) +  len(val_index_backward) +  len(test_index_backward), 'Splitted datasets should have the same number of samples as full dataset'
				assert  len(set(test_index_forward).intersection(train_index_forward)) ==0, "Overlap between train and test indices should be zero"
				assert  len(set(test_index_forward).intersection(val_index_forward)) ==0, "Overlap between val and test indices should be zero"
				assert  len(set(train_index_forward).intersection(val_index_forward)) ==0, "Overlap between train and val indices should be zero"
				assert  len(set(test_index_backward).intersection(train_index_backward)) ==0, "Overlap between train and test indices should be zero"
				assert  len(set(test_index_backward).intersection(val_index_backward)) ==0, "Overlap between val and test indices should be zero"
				assert  len(set(train_index_backward).intersection(val_index_backward)) ==0, "Overlap between train and val indices should be zero"

				splits[i] = {'train_index_forward': train_index_forward, 
								'val_index_forward': val_index_forward,
								'test_index_forward': test_index_forward,
								'train_index_backward': train_index_backward,
								'val_index_backward': val_index_backward,
								'test_index_backward': test_index_backward}
				i += 1
		else:
			for train_test_index_backward in kf.split(dataset_backward):
				#Backward
				train_index_backward = train_test_index_backward[0]; test_index_backward = train_test_index_backward[1]
				train_index_backward, val_index_backward = train_test_split(train_index_backward, test_size=0.2, random_state=42)

				assert len(dataset_backward) == len(train_index_backward) +  len(val_index_backward) +  len(test_index_backward), 'Splitted datasets should have the same number of samples as full dataset'
				assert  len(set(test_index_backward).intersection(train_index_backward)) ==0, "Overlap between train and test indices should be zero"
				assert  len(set(test_index_backward).intersection(val_index_backward)) ==0, "Overlap between val and test indices should be zero"
				assert  len(set(train_index_backward).intersection(val_index_backward)) ==0, "Overlap between train and val indices should be zero"

				splits[i] = {'train_index_forward': None, 
								'val_index_forward': None,
								'test_index_forward': None,
								'train_index_backward': train_index_backward,
								'val_index_backward': val_index_backward,
								'test_index_backward': test_index_backward}
				i += 1
		
	torch.save(splits, osp.join(outdir,'splits.pt'))
	return




###Generate splits
os.makedirs('../../processed/splits/', exist_ok=True)
nfolds = 5
for dataset_type in ['genetic', 'chemical']:
	for dataset_name in ['A375', 'A549', 'MCF7', 'PC3', 'HT29', 'ES2', 'BICR6', 'YAPC', 'AGS', 'U251MG', 'VCAP', 'MDAMB231', 'BT20', 'HA1E', 'HELA']:
		for splits_type in ['random']:
			splits_setting = '{}fold'.format(nfolds)
			outdir = '../../processed/splits/{}/{}/{}/{}'.format(dataset_type, dataset_name, splits_type, splits_setting)

			if dataset_type == 'genetic':
				base_path = "../../processed/torch_data/real_lognorm"
			elif dataset_type =='chemical':
				base_path = "../../processed/torch_data/chemical/real_lognorm"

			try:
				path = osp.join(base_path, 'data_forward_{}.pt'.format(dataset_name))
				dataset_forward = torch.load(path)
				path = osp.join(base_path, 'data_backward_{}.pt'.format(dataset_name))
				dataset_backward = torch.load(path)
				dataset = [dataset_forward, dataset_backward]

				create_splits_n_fold(dataset, splits_type, nfolds, outdir)
			except:
				continue