Source code for gt4sd.frameworks.gflownet.dataloader.dataset

#
# MIT License
#
# Copyright (c) 2022 GT4SD team
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
#
import ast
import os
import pickle
from typing import Any, Callable, Dict, List, NewType, Tuple, Union, Optional

import numpy as np
import pandas as pd
import rdkit.Chem as Chem
import torch
import torch.nn as nn
import torch_geometric.data as gd
from torch import Tensor
from torch.utils.data import Dataset

from ..envs.graph_building_env import GraphActionCategorical
from ..util import MPModelPlaceholder, wrap_model_mp

# This type represents an unprocessed list of reward signals/conditioning information
FlatRewards = NewType("FlatRewards", torch.tensor)  # type: ignore

# This type represents the outcome for a multi-objective task of
# converting FlatRewards to a scalar, e.g. (sum R_i omega_i) ** beta
RewardScalar = NewType("RewardScalar", torch.tensor)  # type: ignore


[docs]class GFlowNetDataset(Dataset): """A dataset for gflownet."""
[docs] def __init__( self, h5_file: Optional[str] = None, target: str = "gap", properties: List[str] = [], ) -> None: """Initialize a gflownet dataset. If the dataset is in a format compatible with h5 file, we can directly load it. If the dataset is in a format compatible with xyz file, we have to convert it to h5 file. Code adapted from: https://github.com/recursionpharma/gflownet/tree/trunk/src/gflownet/data. Args: h5_file: data file in h5 format. target: reward target. properties: relevant properties for the task. Raises: ValueError: if the dataset is not in a format compatible with h5 file or not present. """ if h5_file: try: # pickle on py3.8 uses prot 5 pickle.HIGHEST_PROTOCOL = 5 self.df = pd.HDFStore(h5_file, "r")["df"] except FileNotFoundError: ValueError("No h5 compatible database found in the path.") else: raise ValueError("The h5_file path is None. Please specify a h5 file path.") self.target = target self.properties = properties
[docs] def set_indexes(self, ixs: torch.Tensor): """Set the indexes of the dataset split (train/val/test). Args: ixs: indexes of the dataset split. """ self.idcs = ixs
[docs] def get_len(self): """Get the length of the full dataset (before splitting). Returns: len: the length of the full dataset. """ return len(self.df)
[docs] def get_stats(self, percentile: float = 0.95) -> Tuple[float, float, Any]: """Get the stats of the dataset. Args: percentile: percentile. Returns: min, max, percentile. """ y = self.df[self.target].astype(float) return y.min(), y.max(), np.sort(y)[int(y.shape[0] * percentile)]
[docs] @staticmethod def convert_xyz_to_h5( xyz_path: str = "data/xyz", h5_path: str = "data/qm9.h5", property_names: List[str] = [ "rA", "rB", "rC", "mu", "alpha", "homo", "lumo", "gap", "r2", "zpve", "U0", "U", "H", "G", "Cv", ], ) -> None: """Convert the data from xyz to h5. Assumes that the xyz files are extracted in the xyz_path folder. Args: xyz_path: path to the xyz file in input. h5_path: path to the h5 file in output. property_names: names of the properties we want to use/overwrite. """ # Reads the xyz files and return the properties, smiles and coordinates data = [] smiles = [] properties = [] i = 0 for file in os.listdir(xyz_path): try: path = os.path.join(xyz_path, file) atoms, coordinates, smile, prop = GFlowNetDataset._read_xyz(path) data.append( (atoms, coordinates) ) # A tuple with the atoms and its coordinates smiles.append(smile) # The SMILES representation properties.append(prop) # The molecules properties except ValueError: print(path) i += 1 # rename relevant properties to match GFN schema labels = property_names df = pd.DataFrame(properties, columns=labels) df["SMILES"] = smiles df.to_hdf(h5_path, key="df", mode="w")
[docs] @staticmethod def _read_xyz(path: str): """Reads the xyz files in the directory on 'path'. Code adapted from: https://www.kaggle.com/code/rmonge/predicting-molecule-properties-based-on-its-smiles/notebook Args: path: the path to the folder. Returns: atoms: list with the characters representing the atoms of a molecule. coordinates: list with the cartesian coordinates of each atom. smile: list with the SMILE representation of a molecule. prop: list with the scalar properties. """ atoms = [] coordinates = [] with open(path, "r") as file: lines = file.readlines() n_atoms = int(lines[0]) # the number of atoms smile = lines[n_atoms + 3].split()[0] # smiles string prop = lines[1].split()[2:] # scalar properties # to retrieve each atmos and its cartesian coordenates for atom in lines[2 : n_atoms + 2]: line = atom.split() # which atom atoms.append(line[0]) # its coordinate # Some properties have '*^' indicading exponentiation try: coordinates.append((float(line[1]), float(line[2]), float(line[3]))) except ValueError: coordinates.append( ( float(line[1].replace("*^", "e")), float(line[2].replace("*^", "e")), float(line[3].replace("*^", "e")), ) ) return atoms, coordinates, smile, prop
[docs] def __len__(self): """Dataset split (train/val/test) length. Returns: length of the dataset. """ return len(self.idcs)
[docs] def __getitem__(self, idx: torch.Tensor) -> Tuple[Any, float]: """Retrieve an item from the dataset by index. Args: index: index for the item. Returns: an tuple (item, reward). """ return ( Chem.MolFromSmiles(self.df["SMILES"][self.idcs[idx]]), self.df[self.target][self.idcs[idx]], )
[docs]class GFlowNetTask: """A task for gflownet."""
[docs] def __init__( self, configuration: Dict[str, Any], dataset: GFlowNetDataset, reward_model: Optional[nn.Module] = None, wrap_model: Optional[Callable[[nn.Module], nn.Module]] = None, ) -> None: """Initialize a generic gflownet task. The task specifies the reward model for the trajectory. We consider the task as part of the dataset. Code adapted from: https://github.com/recursionpharma/gflownet/tree/trunk/src/gflownet/tasks. Args: configuration: a dictionary with the task configuration. dataset: a dataset instance. reward_model: The model that is used to generate the conditional reward. wrap_model: a wrapper function that is applied to the model. """ hps = configuration self._wrap_model = wrap_model self.num_workers = hps["num_workers"] self.device = hps["device"] self.rng = hps["rng"] # if reward model provided, use it if reward_model: self.model = {"model_task": reward_model} else: self.model = self.load_task_models() self.dataset = dataset self.temperature_sample_dist = hps["temperature_sample_dist"] self.temperature_dist_params = ast.literal_eval(hps["temperature_dist_params"]) self._min, self._max, self._percentile_95 = self.dataset.get_stats(percentile=0.05) # type: ignore self._width = self._max - self._min self._rtrans = "unit+95p"
[docs] def load_task_models(self) -> Dict[str, nn.Module]: """Loads the task models. Returns: model: a dictionary with the task models. """ raise NotImplementedError()
[docs] def sample_conditional_information(self, n: int) -> Dict[str, Any]: """Samples conditional information for a minibatch. Args: n: number of samples. Returns: cond_info: a dictionary with the sampled conditional information. """ raise NotImplementedError()
[docs] def cond_info_to_reward( self, cond_info: Dict[str, Any], flat_reward: FlatRewards ) -> RewardScalar: """Combines a minibatch of reward signal vectors and conditional information into a scalar reward. Args: cond_info: a dictionary with various conditional informations (e.g. temperature). flat_reward: a 2d tensor where each row represents a series of flat rewards. Returns: reward: a 1d tensor, a scalar reward for each minibatch entry. """ raise NotImplementedError()
[docs] def compute_flat_rewards(self, x: List[Any]) -> Tuple[RewardScalar, Any]: """Compute the flat rewards of mols according the the tasks' proxies. Args: mols: a list of RDKit molecules. Returns: reward: a 1d tensor, a scalar reward for each molecule. is_valid: a 1d tensor, a boolean indicating whether the molecule is valid. """ raise NotImplementedError()
[docs] def flat_reward_transform(self, y: Union[float, Tensor]) -> FlatRewards: """Transforms a reward with a generic structure to a flat vector. Args: y: scalar reward for a trajectory. """ raise NotImplementedError()
[docs] def _wrap_model_mp(self, model: nn.Module) -> Union[nn.Module, MPModelPlaceholder]: """Wraps a nn.Module instance so that it can be shared to `DataLoader` workers. Args: model: a nn.Module instance. """ if self.num_workers > 0: placeholder = wrap_model_mp( model, self.num_workers, cast_types=(gd.Batch, GraphActionCategorical) ) return placeholder return model