clean/video/dfd_fcg/src/model/base.py

import torch
import torch.nn as nn
import lightning.pytorch as pl

from torch import optim
from functools import partial
from torchmetrics import Metric
from torchmetrics.aggregation import MeanMetric
from torchmetrics.classification import AUROC, Accuracy, BinaryConfusionMatrix, AveragePrecision


class GenericStatistics(Metric):
    def __init__(self):
        super().__init__()
        self.add_state("values", default=torch.tensor([]), dist_reduce_fx="sum")

    def update(self, values: torch.Tensor):
        assert len(values.shape) == 1

        self.values = torch.cat([self.values, values])

    def compute(self):
        return torch.stack(
            [
                torch.mean(self.values),
                torch.std(self.values)
            ]
        )


class ODClassifier(pl.LightningModule):
    def __init__(self):
        super().__init__()
        params = dict(add_dataloader_idx=False, rank_zero_only=True)
        self.log = partial(self.log, **params)
        self.log_dict = partial(self.log_dict, **params)
        self.model = None

    def forward(self, *args, **kargs):
        return self.model(*args, **kargs)

    @property
    def transform(self):
        raise NotImplementedError()

    @property
    def n_px(self):
        raise NotImplementedError()

    def training_step(self, batch, batch_idx):
        results = [self.shared_step(batch[dts_name], 'train') for dts_name in batch]
        return sum([_results['loss'] for _results in results])

    def test_step(self, batch, batch_idx, dataloader_idx=0):
        result = self.shared_step(batch, 'test')
        return result['loss']

    def validation_step(self, batch, batch_idx, dataloader_idx=0):
        result = self.shared_step(batch, 'valid')
        return result['loss']

    def shared_step(self, batch, stage):
        x, y, z = batch["xyz"]
        indices = batch["indices"]
        dts_name = batch["dts_name"]
        logits = self(x, **z)
        loss = nn.functional.cross_entropy(logits, y)
        self.log(
            f"{stage}/{dts_name}/loss",
            loss,
            batch_size=x.shape[0]
        )
        return {
            "logits": logits,
            "labels": y,
            "loss": loss,
            "dts_name": dts_name,
            "indices": indices
        }

    def evaluate(self, x, **kargs):
        return self.model(x, **kargs)

    def on_validation_model_train(self):
        self.model.train()


class ODBinaryMetricClassifier(ODClassifier):
    def __init__(self):
        super().__init__()
        self.dts_metrics = {}
        self.metric_map = {
            "auc": partial(AUROC, task="BINARY", num_classes=2),
            "acc": partial(Accuracy, task="BINARY", num_classes=2),
            "cm": partial(BinaryConfusionMatrix, normalize="true"),
            "ap": partial(AveragePrecision, task="BINARY", num_classes=2),
            "loss": MeanMetric,
            # "stats/real": GenericStatistics,
            # "stats/fake": GenericStatistics
        }

    def get_metric(self, dts_name, metric_name, device):
        if (not dts_name in self.dts_metrics):
            self.dts_metrics[dts_name] = {}
        if (not metric_name in self.dts_metrics[dts_name]):
            self.dts_metrics[dts_name][metric_name] = self.metric_map[metric_name]().to(device)
        return self.dts_metrics[dts_name][metric_name]

    def reset_metrics(self):
        for dts_name, metrics in self.dts_metrics.items():
            for metric_name, metric_obj in metrics.items():
                metric_obj.reset()
        self.dts_metrics.clear()

    # shared procedures
    def shared_metric_update_procedure(self, result):
        # save metrics
        logits = result['logits'].detach().softmax(dim=-1)
        labels = result['labels'].detach()
        loss = result["loss"].detach()
        self.get_metric(result['dts_name'], 'auc', logits.device).update(logits[:, 1], labels)
        self.get_metric(result['dts_name'], 'acc', logits.device).update(logits[:, 1], labels)
        self.get_metric(result['dts_name'], 'ap', logits.device).update(logits[:, 1], labels)
        self.get_metric(result['dts_name'], 'cm', logits.device).update(logits[:, 1], labels)
        self.get_metric(result['dts_name'], 'loss', logits.device).update(loss)
        labels = labels.to(dtype=bool)
        # self.get_metric(result['dts_name'], 'stats/real', logits.device).update(logits[~labels, 0])
        # self.get_metric(result['dts_name'], 'stats/fake', logits.device).update(logits[labels, 1])

    def shared_beg_epoch_procedure(self, phase):
        self.reset_metrics()

    def shared_end_epoch_procedure(self, phase):
        log_datas = {}
        for dts_name, metrics in self.dts_metrics.items():
            for metric_name, metric_obj in metrics.items():
                values = metric_obj.compute().flatten()
                name = f'{phase}/{dts_name}/{metric_name}'
                if (len(values) == 1):
                    log_datas[name] = values
                else:
                    for i, v in enumerate(values):
                        log_datas[f"{name}/#{i}"] = v
        self.log_dict(log_datas)
        self.reset_metrics()

    # validation
    def on_validation_epoch_start(self) -> None:
        self.shared_beg_epoch_procedure('valid')

    def validation_step(self, batch, batch_idx, dataloader_idx=0):
        result = self.shared_step(batch, 'valid')
        self.shared_metric_update_procedure(result)
        return result['loss']

    def on_validation_epoch_end(self) -> None:
        self.shared_end_epoch_procedure('valid')

    # test
    def on_test_epoch_start(self) -> None:
        self.shared_beg_epoch_procedure('test')

    def test_step(self, batch, batch_idx, dataloader_idx=0):
        result = self.shared_step(batch, 'test')
        self.shared_metric_update_procedure(result)
        return result['loss']

    def on_test_epoch_end(self) -> None:
        self.shared_end_epoch_procedure('test')

    # predict
    def predict_step(self, batch, batch_idx, dataloader_idx=0):
        x, y, z = batch["xyz"]
        dts_name = batch['dts_name']
        names = batch["names"]
        z = {
            _k: z[_k]
            for _k in z
        }
        y = y.tolist()
        results = self.evaluate(
            x,
            **z
        )
        probs = results["logits"].softmax(dim=-1)[:, 1].flatten().cpu()

        return dict(
            y=y,
            probs=probs,
            names=names,
            dts_name=dts_name
        )