deepscreen/models/dti.py

from functools import partial
from typing import Optional, Sequence, Dict

from torch import nn, optim, Tensor
from lightning import LightningModule
from torchmetrics import Metric, MetricCollection


class DTILightningModule(LightningModule):
    """
        Drug Target Interaction Prediction

        optimizer: a partially or fully initialized instance of class torch.optim.Optimizer
        drug_encoder: a fully initialized instance of class torch.nn.Module
        protein_encoder: a fully initialized instance of class torch.nn.Module
        classifier: a fully initialized instance of class torch.nn.Module
        model: a fully initialized instance of class torch.nn.Module
        metrics: a list of fully initialized instances of class torchmetrics.Metric
    """
    extra_return_keys = ['ID1', 'X1', 'ID2', 'X2', 'N']

    def __init__(
            self,
            optimizer: optim.Optimizer,
            scheduler: Optional[optim.lr_scheduler | Dict],
            predictor: nn.Module,
            metrics: Optional[Dict[str, Metric]] = (),
            out: nn.Module = None,
            loss: nn.Module = None,
            activation: nn.Module = None,
    ):
        super().__init__()

        self.predictor = predictor
        self.out = out
        self.loss = loss
        self.activation = activation

        # Automatically averaged over batches:
        # Separate metric instances for train, val and test step to ensure a proper reduction over the epoch
        metrics = MetricCollection(dict(metrics))
        self.train_metrics = metrics.clone(prefix="train/")
        self.val_metrics = metrics.clone(prefix="val/")
        self.test_metrics = metrics.clone(prefix="test/")

        # allows access to init params with 'self.hparams' attribute and ensures init params will be stored in ckpt
        self.save_hyperparameters(logger=False,
                                  ignore=['predictor', 'out', 'loss', 'activation', 'metrics'])

    def setup(self, stage):
        match stage:
            case 'fit':
                dataloader = self.trainer.datamodule.train_dataloader()
                dummy_batch = next(iter(dataloader))
                self.forward(dummy_batch)
            # case 'validate':
            #     dataloader = self.trainer.datamodule.val_dataloader()
            # case 'test':
            #     dataloader = self.trainer.datamodule.test_dataloader()
            # case 'predict':
            #     dataloader = self.trainer.datamodule.predict_dataloader()

        # for key, value in dummy_batch.items():
        #     if isinstance(value, Tensor):
        #         dummy_batch[key] = value.to(self.device)

    def forward(self, batch):
        output = self.predictor(batch['X1^'], batch['X2^'])
        target = batch.get('Y')
        indexes = batch.get('ID^')
        preds = None
        loss = None

        if isinstance(output, Tensor):
            output = self.out(output).squeeze(1)
            preds = self.activation(output)

        elif isinstance(output, Sequence):
            output = list(output)
            # If multi-objective, assume the zeroth element in `output` is main while the rest are auxiliary
            output[0] = self.out(output[0]).squeeze(1)
            # Downstream metrics evaluation only needs main-objective preds
            preds = self.activation(output[0])

        if target is not None:
            loss = self.loss(output, target.float())

        return preds, target, indexes, loss

    def training_step(self, batch, batch_idx):
        preds, target, indexes, loss = self.forward(batch)
        self.log('train/loss', loss, on_step=False, on_epoch=True, prog_bar=True, sync_dist=True)
        self.train_metrics(preds=preds, target=target, indexes=indexes.long())
        self.log_dict(self.train_metrics, on_step=False, on_epoch=True, prog_bar=True, sync_dist=True)

        return_dict = {
            'Y^': preds,
            'Y': target,
            'loss': loss
        }

        for key in self.extra_return_keys:
            if key in batch:
                return_dict[key] = batch[key]

        return return_dict

    def on_train_epoch_end(self):
        pass

    def validation_step(self, batch, batch_idx):
        preds, target, indexes, loss = self.forward(batch)

        self.log('val/loss', loss, on_step=False, on_epoch=True, prog_bar=True, sync_dist=True)
        self.val_metrics(preds=preds, target=target, indexes=indexes.long())
        self.log_dict(self.val_metrics, on_step=False, on_epoch=True, prog_bar=True, sync_dist=True)

        return_dict = {
            'Y^': preds,
            'Y': target,
            'loss': loss
        }

        for key in self.extra_return_keys:
            if key in batch:
                return_dict[key] = batch[key]

        return return_dict

    def on_validation_epoch_end(self):
        pass

    def test_step(self, batch, batch_idx):
        preds, target, indexes, loss = self.forward(batch)

        self.log('test/loss', loss, on_step=False, on_epoch=True, prog_bar=True, sync_dist=True)
        self.test_metrics(preds=preds, target=target, indexes=indexes.long())
        self.log_dict(self.test_metrics, on_step=False, on_epoch=True, prog_bar=True, sync_dist=True)

        return_dict = {
            'Y^': preds,
            'Y': target,
            'loss': loss
        }

        for key in self.extra_return_keys:
            if key in batch:
                return_dict[key] = batch[key]

        return return_dict

    def on_test_epoch_end(self):
        pass

    def predict_step(self, batch, batch_idx, dataloader_idx=0):
        preds, _, _, _ = self.forward(batch)
        # return a dictionary for callbacks like BasePredictionWriter
        return_dict = {
            'Y^': preds,
        }

        for key in self.extra_return_keys:
            if key in batch:
                return_dict[key] = batch[key]

        return return_dict

    def configure_optimizers(self):
        optimizers_config = {'optimizer': self.hparams.optimizer(params=self.parameters())}
        if self.hparams.get('scheduler'):
            if isinstance(self.hparams.scheduler, partial):
                optimizers_config['lr_scheduler'] = {
                    "scheduler": self.hparams.scheduler(optimizer=optimizers_config['optimizer']),
                    "monitor": "val/loss",
                    "interval": "epoch",
                    "frequency": 1,
                }
            else:
                self.hparams.scheduler['scheduler'] = self.hparams.scheduler['scheduler'](
                    optimizer=optimizers_config['optimizer']
                )
                optimizers_config['lr_scheduler'] = dict(self.hparams.scheduler)
        return optimizers_config