resnet_classifier.py

import warnings
from argparse import ArgumentParser
from pathlib import Path

warnings.filterwarnings("ignore")

import pytorch_lightning as pl
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models as models
from torch.optim import SGD, Adam
from torch.utils.data import DataLoader
from torchmetrics import Accuracy
from torchvision.datasets import ImageFolder


# Here we define a new class to turn the ResNet model that we want to use as a feature extractor
# into a pytorch-lightning module so that we can take advantage of lightning's Trainer object.
# We aim to make it a little more general by allowing users to define the number of prediction classes.
class ResNetClassifier(pl.LightningModule):
    resnets = {
        18: models.resnet18,
        34: models.resnet34,
        50: models.resnet50,
        101: models.resnet101,
        152: models.resnet152,
    }
    weights = {
        18: models.ResNet18_Weights.IMAGENET1K_V1,
        34: models.ResNet34_Weights.IMAGENET1K_V1,
        50: models.ResNet50_Weights.IMAGENET1K_V2,
        101: models.ResNet101_Weights.IMAGENET1K_V2,
        152: models.ResNet152_Weights.IMAGENET1K_V2,
    }
    optimizers = {"adam": Adam, "sgd": SGD}

    def __init__(
        self,
        num_classes,
        resnet_version,
        train_path,
        val_path,
        test_path=None,
        optimizer="adam",
        lr=1e-3,
        batch_size=16,
        transfer=True,
        tune_fc_only=True,
    ):
        super().__init__()

        self.num_classes = num_classes
        self.train_path = train_path
        self.val_path = val_path
        self.test_path = test_path
        self.lr = lr
        self.batch_size = batch_size

        self.optimizer = self.optimizers[optimizer]
        # instantiate loss criterion
        self.loss_fn = (
            nn.BCEWithLogitsLoss() if num_classes == 1 else nn.CrossEntropyLoss()
        )
        # create accuracy metric
        self.acc = Accuracy(
            task="binary" if num_classes == 1 else "multiclass", num_classes=num_classes
        )
        weights = None
        if transfer:
            # Using a pretrained ResNet backbone
            weights=self.weights[resnet_version]
        self.resnet_model = self.resnets[resnet_version](weights=weights)
        # Inference transforms. Potential improvement: use train transforms from pytorch recipes (flip, ...)
        self.transforms = self.weights[resnet_version].transforms()
        # Replace old FC layer with Identity so we can train our own
        linear_size = list(self.resnet_model.children())[-1].in_features
        # replace final layer for fine tuning
        self.resnet_model.fc = nn.Linear(linear_size, num_classes)

        if tune_fc_only:  # option to only tune the fully-connected layers
            for child in list(self.resnet_model.children())[:-1]:
                for param in child.parameters():
                    param.requires_grad = False

    def forward(self, X):
        return self.resnet_model(X)

    def configure_optimizers(self):
        return self.optimizer(self.parameters(), lr=self.lr)

    def _step(self, batch):
        x, y = batch
        preds = self(x)

        if self.num_classes == 1:
            preds = preds.flatten()
            y = y.float()

        loss = self.loss_fn(preds, y)
        acc = self.acc(preds, y)
        return loss, acc

    def _dataloader(self, data_path, shuffle=False):
        # values here are specific to pneumonia dataset and should be updated for custom data
        img_folder = ImageFolder(data_path, transform=self.transforms)

        return DataLoader(img_folder, batch_size=self.batch_size, shuffle=shuffle)

    def train_dataloader(self):
        return self._dataloader(self.train_path, shuffle=True)

    def training_step(self, batch, batch_idx):
        loss, acc = self._step(batch)
        # perform logging
        self.log(
            "train_loss", loss, on_step=True, on_epoch=True, prog_bar=True, logger=True
        )
        self.log(
            "train_acc", acc, on_step=True, on_epoch=True, prog_bar=True, logger=True
        )
        return loss

    def val_dataloader(self):
        return self._dataloader(self.val_path)

    def validation_step(self, batch, batch_idx):
        loss, acc = self._step(batch)
        # perform logging
        self.log("val_loss", loss, on_epoch=True, prog_bar=False, logger=True)
        self.log("val_acc", acc, on_epoch=True, prog_bar=True, logger=True)

    def test_dataloader(self):
        return self._dataloader(self.test_path)

    def test_step(self, batch, batch_idx):
        loss, acc = self._step(batch)
        # perform logging
        self.log("test_loss", loss, on_step=True, prog_bar=True, logger=True)
        self.log("test_acc", acc, on_step=True, prog_bar=True, logger=True)


if __name__ == "__main__":
    parser = ArgumentParser()
    # Required arguments
    parser.add_argument(
        "model",
        help="""Choose one of the predefined ResNet models provided by torchvision. e.g. 50""",
        type=int,
    )
    parser.add_argument(
        "num_classes", help="""Number of classes to be learned.""", type=int
    )
    parser.add_argument("num_epochs", help="""Number of Epochs to Run.""", type=int)
    parser.add_argument(
        "train_set", help="""Path to training data folder.""", type=Path
    )
    parser.add_argument("val_set", help="""Path to validation set folder.""", type=Path)
    # Optional arguments
    parser.add_argument(
        "-amp",
        "--mixed_precision",
        help="""Use mixed precision during training. Defaults to False.""",
        action="store_true",
    )
    parser.add_argument(
        "-ts", "--test_set", help="""Optional test set path.""", type=Path
    )
    parser.add_argument(
        "-o",
        "--optimizer",
        help="""PyTorch optimizer to use. Defaults to adam.""",
        default="adam",
    )
    parser.add_argument(
        "-lr",
        "--learning_rate",
        help="Adjust learning rate of optimizer.",
        type=float,
        default=1e-3,
    )
    parser.add_argument(
        "-b",
        "--batch_size",
        help="""Manually determine batch size. Defaults to 16.""",
        type=int,
        default=16,
    )
    parser.add_argument(
        "-tr",
        "--transfer",
        help="""Determine whether to use pretrained model or train from scratch. Defaults to True.""",
        action="store_true",
    )
    parser.add_argument(
        "-to",
        "--tune_fc_only",
        help="Tune only the final, fully connected layers.",
        action="store_true",
    )
    parser.add_argument(
        "-s", "--save_path", help="""Path to save model trained model checkpoint."""
    )
    parser.add_argument(
        "-g", "--gpus", help="""Enables GPU acceleration.""", type=int, default=None
    )
    parser.add_argument('--from-pretrained', type=str, default=None, help='Load pretrained weights')
    args = parser.parse_args()

    if args.from_pretrained is not None:
        model = ResNetClassifier.load_from_checkpoint(
            args.from_pretrained,
            num_classes=args.num_classes,
            resnet_version=args.model,
            train_path=args.train_set,
            val_path=args.val_set,
            test_path=args.test_set,
            optimizer=args.optimizer,
            lr=args.learning_rate,
            batch_size=args.batch_size,
            transfer=args.transfer,
            tune_fc_only=args.tune_fc_only,
        )
    else:
        model = ResNetClassifier(
            num_classes=args.num_classes,
            resnet_version=args.model,
            train_path=args.train_set,
            val_path=args.val_set,
            test_path=args.test_set,
            optimizer=args.optimizer,
            lr=args.learning_rate,
            batch_size=args.batch_size,
            transfer=args.transfer,
            tune_fc_only=args.tune_fc_only,
        )

    save_path = args.save_path if args.save_path is not None else "./models"
    checkpoint_callback = pl.callbacks.ModelCheckpoint(
        dirpath=save_path,
        filename="resnet-model-{epoch}-{val_loss:.2f}-{val_acc:0.2f}",
        monitor="val_loss",
        save_top_k=3,
        mode="min",
        save_last=True,
    )

    stopping_callback = pl.callbacks.EarlyStopping(monitor="val_loss", mode="min")

    # Instantiate lightning trainer and train model
    trainer_args = {
        "accelerator": "gpu" if args.gpus else None,
        "devices": [0],
        "strategy": "ddp" if args.gpus > 1 else "auto",
        "max_epochs": args.num_epochs,
        "callbacks": [checkpoint_callback, stopping_callback],
        "precision": 16 if args.mixed_precision else 32,
    }
    trainer = pl.Trainer(**trainer_args)

    trainer.fit(model)

    if args.test_set:
        trainer.test(model)
    # Save trained model weights
    torch.save(trainer.model.resnet_model.state_dict(), save_path + "/trained_model.pt")