"""Various methods for splitting chemical datasets. We mostly adapt them from deepchem (https://github.com/deepchem/deepchem/blob/master/deepchem/splits/splitters.py). """ # pylint: disable= no-member, arguments-differ, invalid-name # pylint: disable=E0611 from collections import defaultdict from functools import partial from itertools import accumulate, chain from rdkit import Chem from rdkit.Chem import rdMolDescriptors from rdkit.Chem.rdmolops import FastFindRings from rdkit.Chem.Scaffolds import MurckoScaffold import dgl.backend as F import numpy as np from dgl.data.utils import split_dataset, Subset __all__ = ['ConsecutiveSplitter', 'RandomSplitter', 'MolecularWeightSplitter', 'ScaffoldSplitter', 'SingleTaskStratifiedSplitter'] def base_k_fold_split(split_method, dataset, k, log): """Split dataset for k-fold cross validation. Parameters ---------- split_method : callable Arbitrary method for splitting the dataset into training, validation and test subsets. dataset We assume ``len(dataset)`` gives the size for the dataset and ``dataset[i]`` gives the ith datapoint. k : int Number of folds to use and should be no smaller than 2. log : bool Whether to print a message at the start of preparing each fold. Returns ------- all_folds : list of 2-tuples Each element of the list represents a fold and is a 2-tuple (train_set, val_set), which are all :class:`Subset` instances. """ assert k >= 2, 'Expect the number of folds to be no smaller than 2, got {:d}'.format(k) all_folds = [] frac_per_part = 1. / k for i in range(k): if log: print('Processing fold {:d}/{:d}'.format(i + 1, k)) # We are reusing the code for train-validation-test split. train_set1, val_set, train_set2 = split_method(dataset, frac_train=i * frac_per_part, frac_val=frac_per_part, frac_test=1. - (i + 1) * frac_per_part) # For cross validation, each fold consists of only a train subset and # a validation subset. train_set = Subset(dataset, np.concatenate( [train_set1.indices, train_set2.indices]).astype(np.int64)) all_folds.append((train_set, val_set)) return all_folds def train_val_test_sanity_check(frac_train, frac_val, frac_test): """Sanity check for train-val-test split Ensure that the fractions of the dataset to use for training, validation and test add up to 1. Parameters ---------- frac_train : float Fraction of the dataset to use for training. frac_val : float Fraction of the dataset to use for validation. frac_test : float Fraction of the dataset to use for test. """ total_fraction = frac_train + frac_val + frac_test assert np.allclose(total_fraction, 1.), \ 'Expect the sum of fractions for training, validation and ' \ 'test to be 1, got {:.4f}'.format(total_fraction) def indices_split(dataset, frac_train, frac_val, frac_test, indices): """Reorder datapoints based on the specified indices and then take consecutive chunks as subsets. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset and ``dataset[i]`` gives the ith datapoint. frac_train : float Fraction of data to use for training. frac_val : float Fraction of data to use for validation. frac_test : float Fraction of data to use for test. indices : list or ndarray Indices specifying the order of datapoints. Returns ------- list of length 3 Subsets for training, validation and test, which are all :class:`Subset` instances. """ frac_list = np.array([frac_train, frac_val, frac_test]) assert np.allclose(np.sum(frac_list), 1.), \ 'Expect frac_list sum to 1, got {:.4f}'.format(np.sum(frac_list)) num_data = len(dataset) lengths = (num_data * frac_list).astype(int) lengths[-1] = num_data - np.sum(lengths[:-1]) return [Subset(dataset, list(indices[offset - length:offset])) for offset, length in zip(accumulate(lengths), lengths)] def count_and_log(message, i, total, log_every_n): """Print a message to reflect the progress of processing once a while. Parameters ---------- message : str Message to print. i : int Current index. total : int Total count. log_every_n : None or int Molecule related computation can take a long time for a large dataset and we want to learn the progress of processing. This can be done by printing a message whenever a batch of ``log_every_n`` molecules have been processed. If None, no messages will be printed. """ if (log_every_n is not None) and ((i + 1) % log_every_n == 0): print('{} {:d}/{:d}'.format(message, i + 1, total)) def prepare_mols(dataset, mols, sanitize, log_every_n=1000): """Prepare RDKit molecule instances. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset, ``dataset[i]`` gives the ith datapoint and ``dataset.smiles[i]`` gives the SMILES for the ith datapoint. mols : None or list of rdkit.Chem.rdchem.Mol None or pre-computed RDKit molecule instances. If not None, we expect a one-on-one correspondence between ``dataset.smiles`` and ``mols``, i.e. ``mols[i]`` corresponds to ``dataset.smiles[i]``. sanitize : bool This argument only comes into effect when ``mols`` is None and decides whether sanitization is performed in initializing RDKit molecule instances. See https://www.rdkit.org/docs/RDKit_Book.html for details of the sanitization. log_every_n : None or int Molecule related computation can take a long time for a large dataset and we want to learn the progress of processing. This can be done by printing a message whenever a batch of ``log_every_n`` molecules have been processed. If None, no messages will be printed. Default to 1000. Returns ------- mols : list of rdkit.Chem.rdchem.Mol RDkit molecule instances where there is a one-on-one correspondence between ``dataset.smiles`` and ``mols``, i.e. ``mols[i]`` corresponds to ``dataset.smiles[i]``. """ if mols is not None: # Sanity check assert len(mols) == len(dataset), \ 'Expect mols to be of the same size as that of the dataset, ' \ 'got {:d} and {:d}'.format(len(mols), len(dataset)) else: if log_every_n is not None: print('Start initializing RDKit molecule instances...') mols = [] for i, s in enumerate(dataset.smiles): count_and_log('Creating RDKit molecule instance', i, len(dataset.smiles), log_every_n) mols.append(Chem.MolFromSmiles(s, sanitize=sanitize)) return mols class ConsecutiveSplitter(object): """Split datasets with the input order. The dataset is split without permutation, so the splitting is deterministic. """ @staticmethod def train_val_test_split(dataset, frac_train=0.8, frac_val=0.1, frac_test=0.1): """Split the dataset into three consecutive chunks for training, validation and test. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset and ``dataset[i]`` gives the ith datapoint. frac_train : float Fraction of data to use for training. By default, we set this to be 0.8, i.e. 80% of the dataset is used for training. frac_val : float Fraction of data to use for validation. By default, we set this to be 0.1, i.e. 10% of the dataset is used for validation. frac_test : float Fraction of data to use for test. By default, we set this to be 0.1, i.e. 10% of the dataset is used for test. Returns ------- list of length 3 Subsets for training, validation and test that also have ``len(dataset)`` and ``dataset[i]`` behaviors """ return split_dataset(dataset, frac_list=[frac_train, frac_val, frac_test], shuffle=False) @staticmethod def k_fold_split(dataset, k=5, log=True): """Split the dataset for k-fold cross validation by taking consecutive chunks. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset and ``dataset[i]`` gives the ith datapoint. k : int Number of folds to use and should be no smaller than 2. Default to be 5. log : bool Whether to print a message at the start of preparing each fold. Returns ------- list of 2-tuples Each element of the list represents a fold and is a 2-tuple ``(train_set, val_set)``. ``train_set`` and ``val_set`` also have ``len(dataset)`` and ``dataset[i]`` behaviors. """ return base_k_fold_split(ConsecutiveSplitter.train_val_test_split, dataset, k, log) class RandomSplitter(object): """Randomly reorder datasets and then split them. The dataset is split with permutation and the splitting is hence random. """ @staticmethod def train_val_test_split(dataset, frac_train=0.8, frac_val=0.1, frac_test=0.1, random_state=None): """Randomly permute the dataset and then split it into three consecutive chunks for training, validation and test. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset and ``dataset[i]`` gives the ith datapoint. frac_train : float Fraction of data to use for training. By default, we set this to be 0.8, i.e. 80% of the dataset is used for training. frac_val : float Fraction of data to use for validation. By default, we set this to be 0.1, i.e. 10% of the dataset is used for validation. frac_test : float Fraction of data to use for test. By default, we set this to be 0.1, i.e. 10% of the dataset is used for test. random_state : None, int or array_like, optional Random seed used to initialize the pseudo-random number generator. Can be any integer between 0 and 2**32 - 1 inclusive, an array (or other sequence) of such integers, or None (the default). If seed is None, then RandomState will try to read data from /dev/urandom (or the Windows analogue) if available or seed from the clock otherwise. Returns ------- list of length 3 Subsets for training, validation and test, which also have ``len(dataset)`` and ``dataset[i]`` behaviors. """ return split_dataset(dataset, frac_list=[frac_train, frac_val, frac_test], shuffle=True, random_state=random_state) @staticmethod def k_fold_split(dataset, k=5, random_state=None, log=True): """Randomly permute the dataset and then split it for k-fold cross validation by taking consecutive chunks. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset and ``dataset[i]`` gives the ith datapoint. k : int Number of folds to use and should be no smaller than 2. Default to be 5. random_state : None, int or array_like, optional Random seed used to initialize the pseudo-random number generator. Can be any integer between 0 and 2**32 - 1 inclusive, an array (or other sequence) of such integers, or None (the default). If seed is None, then RandomState will try to read data from /dev/urandom (or the Windows analogue) if available or seed from the clock otherwise. log : bool Whether to print a message at the start of preparing each fold. Default to True. Returns ------- list of 2-tuples Each element of the list represents a fold and is a 2-tuple ``(train_set, val_set)``. ``train_set`` and ``val_set`` also have ``len(dataset)`` and ``dataset[i]`` behaviors. """ # Permute the dataset only once so that each datapoint # will appear once in exactly one fold. indices = np.random.RandomState(seed=random_state).permutation(len(dataset)) return base_k_fold_split(partial(indices_split, indices=indices), dataset, k, log) # pylint: disable=I1101 class MolecularWeightSplitter(object): """Sort molecules based on their weights and then split them.""" @staticmethod def molecular_weight_indices(molecules, log_every_n): """Reorder molecules based on molecular weights. Parameters ---------- molecules : list of rdkit.Chem.rdchem.Mol Pre-computed RDKit molecule instances. We expect a one-on-one correspondence between ``dataset.smiles`` and ``mols``, i.e. ``mols[i]`` corresponds to ``dataset.smiles[i]``. log_every_n : None or int Molecule related computation can take a long time for a large dataset and we want to learn the progress of processing. This can be done by printing a message whenever a batch of ``log_every_n`` molecules have been processed. If None, no messages will be printed. Returns ------- indices : list or ndarray Indices specifying the order of datapoints, which are basically argsort of the molecular weights. """ if log_every_n is not None: print('Start computing molecular weights.') mws = [] for i, mol in enumerate(molecules): count_and_log('Computing molecular weight for compound', i, len(molecules), log_every_n) mws.append(rdMolDescriptors.CalcExactMolWt(mol)) return np.argsort(mws) @staticmethod def train_val_test_split(dataset, mols=None, sanitize=True, frac_train=0.8, frac_val=0.1, frac_test=0.1, log_every_n=1000): """Sort molecules based on their weights and then split them into three consecutive chunks for training, validation and test. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset, ``dataset[i]`` gives the ith datapoint and ``dataset.smiles[i]`` gives the SMILES for the ith datapoint. mols : None or list of rdkit.Chem.rdchem.Mol None or pre-computed RDKit molecule instances. If not None, we expect a one-on-one correspondence between ``dataset.smiles`` and ``mols``, i.e. ``mols[i]`` corresponds to ``dataset.smiles[i]``. Default to None. sanitize : bool This argument only comes into effect when ``mols`` is None and decides whether sanitization is performed in initializing RDKit molecule instances. See https://www.rdkit.org/docs/RDKit_Book.html for details of the sanitization. Default to be True. frac_train : float Fraction of data to use for training. By default, we set this to be 0.8, i.e. 80% of the dataset is used for training. frac_val : float Fraction of data to use for validation. By default, we set this to be 0.1, i.e. 10% of the dataset is used for validation. frac_test : float Fraction of data to use for test. By default, we set this to be 0.1, i.e. 10% of the dataset is used for test. log_every_n : None or int Molecule related computation can take a long time for a large dataset and we want to learn the progress of processing. This can be done by printing a message whenever a batch of ``log_every_n`` molecules have been processed. If None, no messages will be printed. Default to 1000. Returns ------- list of length 3 Subsets for training, validation and test, which also have ``len(dataset)`` and ``dataset[i]`` behaviors """ # Perform sanity check first as molecule instance initialization and descriptor # computation can take a long time. train_val_test_sanity_check(frac_train, frac_val, frac_test) molecules = prepare_mols(dataset, mols, sanitize, log_every_n) sorted_indices = MolecularWeightSplitter.molecular_weight_indices(molecules, log_every_n) return indices_split(dataset, frac_train, frac_val, frac_test, sorted_indices) @staticmethod def k_fold_split(dataset, mols=None, sanitize=True, k=5, log_every_n=1000): """Sort molecules based on their weights and then split them for k-fold cross validation by taking consecutive chunks. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset, ``dataset[i]`` gives the ith datapoint and ``dataset.smiles[i]`` gives the SMILES for the ith datapoint. mols : None or list of rdkit.Chem.rdchem.Mol None or pre-computed RDKit molecule instances. If not None, we expect a one-on-one correspondence between ``dataset.smiles`` and ``mols``, i.e. ``mols[i]`` corresponds to ``dataset.smiles[i]``. Default to None. sanitize : bool This argument only comes into effect when ``mols`` is None and decides whether sanitization is performed in initializing RDKit molecule instances. See https://www.rdkit.org/docs/RDKit_Book.html for details of the sanitization. Default to be True. k : int Number of folds to use and should be no smaller than 2. Default to be 5. log_every_n : None or int Molecule related computation can take a long time for a large dataset and we want to learn the progress of processing. This can be done by printing a message whenever a batch of ``log_every_n`` molecules have been processed. If None, no messages will be printed. Default to 1000. Returns ------- list of 2-tuples Each element of the list represents a fold and is a 2-tuple ``(train_set, val_set)``. ``train_set`` and ``val_set`` also have ``len(dataset)`` and ``dataset[i]`` behaviors. """ molecules = prepare_mols(dataset, mols, sanitize, log_every_n) sorted_indices = MolecularWeightSplitter.molecular_weight_indices(molecules, log_every_n) return base_k_fold_split(partial(indices_split, indices=sorted_indices), dataset, k, log=(log_every_n is not None)) # pylint: disable=W0702 class ScaffoldSplitter(object): """Group molecules based on their Bemis-Murcko scaffolds and then split the groups. Group molecules so that all molecules in a group have a same scaffold (see reference). The dataset is then split at the level of groups. References ---------- Bemis, G. W.; Murcko, M. A. “The Properties of Known Drugs. 1. Molecular Frameworks.” J. Med. Chem. 39:2887-93 (1996). """ @staticmethod def get_ordered_scaffold_sets(molecules, include_chirality, log_every_n): """Group molecules based on their Bemis-Murcko scaffolds and order these groups based on their sizes. The order is decided by comparing the size of groups, where groups with a larger size are placed before the ones with a smaller size. Parameters ---------- molecules : list of rdkit.Chem.rdchem.Mol Pre-computed RDKit molecule instances. We expect a one-on-one correspondence between ``dataset.smiles`` and ``mols``, i.e. ``mols[i]`` corresponds to ``dataset.smiles[i]``. include_chirality : bool Whether to consider chirality in computing scaffolds. log_every_n : None or int Molecule related computation can take a long time for a large dataset and we want to learn the progress of processing. This can be done by printing a message whenever a batch of ``log_every_n`` molecules have been processed. If None, no messages will be printed. Returns ------- scaffold_sets : list Each element of the list is a list of int, representing the indices of compounds with a same scaffold. """ if log_every_n is not None: print('Start computing Bemis-Murcko scaffolds.') scaffolds = defaultdict(list) for i, mol in enumerate(molecules): count_and_log('Computing Bemis-Murcko for compound', i, len(molecules), log_every_n) # For mols that have not been sanitized, we need to compute their ring information try: FastFindRings(mol) mol_scaffold = MurckoScaffold.MurckoScaffoldSmiles( mol=mol, includeChirality=include_chirality) # Group molecules that have the same scaffold scaffolds[mol_scaffold].append(i) except: print('Failed to compute the scaffold for molecule {:d} ' 'and it will be excluded.'.format(i + 1)) # Order groups of molecules by first comparing the size of groups # and then the index of the first compound in the group. scaffold_sets = [ scaffold_set for (scaffold, scaffold_set) in sorted( scaffolds.items(), key=lambda x: (len(x[1]), x[1][0]), reverse=True) ] return scaffold_sets @staticmethod def train_val_test_split(dataset, mols=None, sanitize=True, include_chirality=False, frac_train=0.8, frac_val=0.1, frac_test=0.1, log_every_n=1000): """Split the dataset into training, validation and test set based on molecular scaffolds. This spliting method ensures that molecules with a same scaffold will be collectively in only one of the training, validation or test set. As a result, the fraction of dataset to use for training and validation tend to be smaller than ``frac_train`` and ``frac_val``, while the fraction of dataset to use for test tends to be larger than ``frac_test``. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset, ``dataset[i]`` gives the ith datapoint and ``dataset.smiles[i]`` gives the SMILES for the ith datapoint. mols : None or list of rdkit.Chem.rdchem.Mol None or pre-computed RDKit molecule instances. If not None, we expect a one-on-one correspondence between ``dataset.smiles`` and ``mols``, i.e. ``mols[i]`` corresponds to ``dataset.smiles[i]``. Default to None. sanitize : bool This argument only comes into effect when ``mols`` is None and decides whether sanitization is performed in initializing RDKit molecule instances. See https://www.rdkit.org/docs/RDKit_Book.html for details of the sanitization. Default to True. include_chirality : bool Whether to consider chirality in computing scaffolds. Default to False. frac_train : float Fraction of data to use for training. By default, we set this to be 0.8, i.e. 80% of the dataset is used for training. frac_val : float Fraction of data to use for validation. By default, we set this to be 0.1, i.e. 10% of the dataset is used for validation. frac_test : float Fraction of data to use for test. By default, we set this to be 0.1, i.e. 10% of the dataset is used for test. log_every_n : None or int Molecule related computation can take a long time for a large dataset and we want to learn the progress of processing. This can be done by printing a message whenever a batch of ``log_every_n`` molecules have been processed. If None, no messages will be printed. Default to 1000. Returns ------- list of length 3 Subsets for training, validation and test, which also have ``len(dataset)`` and ``dataset[i]`` behaviors """ # Perform sanity check first as molecule related computation can take a long time. train_val_test_sanity_check(frac_train, frac_val, frac_test) molecules = prepare_mols(dataset, mols, sanitize) scaffold_sets = ScaffoldSplitter.get_ordered_scaffold_sets( molecules, include_chirality, log_every_n) train_indices, val_indices, test_indices = [], [], [] train_cutoff = int(frac_train * len(molecules)) val_cutoff = int((frac_train + frac_val) * len(molecules)) for group_indices in scaffold_sets: if len(train_indices) + len(group_indices) > train_cutoff: if len(train_indices) + len(val_indices) + len(group_indices) > val_cutoff: test_indices.extend(group_indices) else: val_indices.extend(group_indices) else: train_indices.extend(group_indices) return [Subset(dataset, train_indices), Subset(dataset, val_indices), Subset(dataset, test_indices)] @staticmethod def k_fold_split(dataset, mols=None, sanitize=True, include_chirality=False, k=5, log_every_n=1000): """Group molecules based on their scaffolds and sort groups based on their sizes. The groups are then split for k-fold cross validation. Same as usual k-fold splitting methods, each molecule will appear only once in the validation set among all folds. In addition, this method ensures that molecules with a same scaffold will be collectively in either the training set or the validation set for each fold. Note that the folds can be highly imbalanced depending on the scaffold distribution in the dataset. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset, ``dataset[i]`` gives the ith datapoint and ``dataset.smiles[i]`` gives the SMILES for the ith datapoint. mols : None or list of rdkit.Chem.rdchem.Mol None or pre-computed RDKit molecule instances. If not None, we expect a one-on-one correspondence between ``dataset.smiles`` and ``mols``, i.e. ``mols[i]`` corresponds to ``dataset.smiles[i]``. Default to None. sanitize : bool This argument only comes into effect when ``mols`` is None and decides whether sanitization is performed in initializing RDKit molecule instances. See https://www.rdkit.org/docs/RDKit_Book.html for details of the sanitization. Default to True. include_chirality : bool Whether to consider chirality in computing scaffolds. Default to False. k : int Number of folds to use and should be no smaller than 2. Default to be 5. log_every_n : None or int Molecule related computation can take a long time for a large dataset and we want to learn the progress of processing. This can be done by printing a message whenever a batch of ``log_every_n`` molecules have been processed. If None, no messages will be printed. Default to 1000. Returns ------- list of 2-tuples Each element of the list represents a fold and is a 2-tuple ``(train_set, val_set)``. ``train_set`` and ``val_set`` also have ``len(dataset)`` and ``dataset[i]`` behaviors. """ assert k >= 2, 'Expect the number of folds to be no smaller than 2, got {:d}'.format(k) molecules = prepare_mols(dataset, mols, sanitize) scaffold_sets = ScaffoldSplitter.get_ordered_scaffold_sets( molecules, include_chirality, log_every_n) # k buckets that form a relatively balanced partition of the dataset index_buckets = [[] for _ in range(k)] for group_indices in scaffold_sets: bucket_chosen = int(np.argmin([len(bucket) for bucket in index_buckets])) index_buckets[bucket_chosen].extend(group_indices) all_folds = [] for i in range(k): if log_every_n is not None: print('Processing fold {:d}/{:d}'.format(i + 1, k)) train_indices = list(chain.from_iterable(index_buckets[:i] + index_buckets[i + 1:])) val_indices = index_buckets[i] all_folds.append((Subset(dataset, train_indices), Subset(dataset, val_indices))) return all_folds class SingleTaskStratifiedSplitter(object): """Splits the dataset by stratification on a single task. We sort the molecules based on their label values for a task and then repeatedly take buckets of datapoints to augment the training, validation and test subsets. """ @staticmethod def train_val_test_split(dataset, labels, task_id, frac_train=0.8, frac_val=0.1, frac_test=0.1, bucket_size=10, random_state=None): """Split the dataset into training, validation and test subsets as stated above. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset, ``dataset[i]`` gives the ith datapoint and ``dataset.smiles[i]`` gives the SMILES for the ith datapoint. labels : tensor of shape (N, T) Dataset labels all tasks. N for the number of datapoints and T for the number of tasks. task_id : int Index for the task. frac_train : float Fraction of data to use for training. By default, we set this to be 0.8, i.e. 80% of the dataset is used for training. frac_val : float Fraction of data to use for validation. By default, we set this to be 0.1, i.e. 10% of the dataset is used for validation. frac_test : float Fraction of data to use for test. By default, we set this to be 0.1, i.e. 10% of the dataset is used for test. bucket_size : int Size of bucket of datapoints. Default to 10. random_state : None, int or array_like, optional Random seed used to initialize the pseudo-random number generator. Can be any integer between 0 and 2**32 - 1 inclusive, an array (or other sequence) of such integers, or None (the default). If seed is None, then RandomState will try to read data from /dev/urandom (or the Windows analogue) if available or seed from the clock otherwise. Returns ------- list of length 3 Subsets for training, validation and test, which also have ``len(dataset)`` and ``dataset[i]`` behaviors """ train_val_test_sanity_check(frac_train, frac_val, frac_test) if random_state is not None: np.random.seed(random_state) if not isinstance(labels, np.ndarray): labels = F.asnumpy(labels) task_labels = labels[:, task_id] sorted_indices = np.argsort(task_labels) train_bucket_cutoff = int(np.round(frac_train * bucket_size)) val_bucket_cutoff = int(np.round(frac_val * bucket_size)) + train_bucket_cutoff train_indices, val_indices, test_indices = [], [], [] while sorted_indices.shape[0] >= bucket_size: current_batch, sorted_indices = np.split(sorted_indices, [bucket_size]) shuffled = np.random.permutation(range(bucket_size)) train_indices.extend( current_batch[shuffled[:train_bucket_cutoff]].tolist()) val_indices.extend( current_batch[shuffled[train_bucket_cutoff:val_bucket_cutoff]].tolist()) test_indices.extend( current_batch[shuffled[val_bucket_cutoff:]].tolist()) # Place rest samples in the training set. train_indices.extend(sorted_indices.tolist()) return [Subset(dataset, train_indices), Subset(dataset, val_indices), Subset(dataset, test_indices)] @staticmethod def k_fold_split(dataset, labels, task_id, k=5, log=True): """Sort molecules based on their label values for a task and then split them for k-fold cross validation by taking consecutive chunks. Parameters ---------- dataset We assume ``len(dataset)`` gives the size for the dataset, ``dataset[i]`` gives the ith datapoint and ``dataset.smiles[i]`` gives the SMILES for the ith datapoint. labels : tensor of shape (N, T) Dataset labels all tasks. N for the number of datapoints and T for the number of tasks. task_id : int Index for the task. k : int Number of folds to use and should be no smaller than 2. Default to be 5. log : bool Whether to print a message at the start of preparing each fold. Returns ------- list of 2-tuples Each element of the list represents a fold and is a 2-tuple ``(train_set, val_set)``. ``train_set`` and ``val_set`` also have ``len(dataset)`` and ``dataset[i]`` behaviors. """ if not isinstance(labels, np.ndarray): labels = F.asnumpy(labels) task_labels = labels[:, task_id] sorted_indices = np.argsort(task_labels).tolist() return base_k_fold_split(partial(indices_split, indices=sorted_indices), dataset, k, log)