How can change aggregation of a metric or get the total result of a metric

[Freed-Wu]

By default,

from torchmetrics.image.psnr import PeakSignalNoiseRatio
metric = PeakSignalNoiseRatio()
metric(img1, target1)
metric(img2, target2)
...
metric(img100, target100)
metric.compute()
# print the average of 100 pair of image's PSNRs

How can get

from torchmetrics.image.psnr import PeakSignalNoiseRatio
metric = PeakSignalNoiseRatio(aggregation = "concatenate")
metric(img1, target1)
metric(img2, target2)
...
metric(img100, target100)
metric.compute()
# print a list of 100 pair of image's PSNRs

TIA!

[viktor-ktorvi]

Hey, did you figure out how to do this?

[Borda]

As of v1.9.0, you can get per-sample metric values (instead of the default batch-averaged aggregation) using two approaches:

Approach 1 — Functional interface (simplest):

from torchmetrics.functional.image import peak_signal_noise_ratio

psnr_values = []
for img_pred, img_target in zip(preds, targets):
    psnr_values.append(
        peak_signal_noise_ratio(img_pred.unsqueeze(0), img_target.unsqueeze(0), data_range=1.0)
    )
psnr_per_image = torch.stack(psnr_values)

Approach 2 — Custom metric with list state:

import torch
from torchmetrics import Metric
from torchmetrics.functional.image import peak_signal_noise_ratio
from torchmetrics.utilities import dim_zero_cat


class PSNRPerSample(Metric):
    def __init__(self, data_range: float = 1.0, **kwargs):
        super().__init__(**kwargs)
        self.data_range = data_range
        self.add_state("values", default=[], dist_reduce_fx="cat")

    def update(self, preds: torch.Tensor, target: torch.Tensor) -> None:
        for i in range(preds.shape[0]):
            val = peak_signal_noise_ratio(
                preds[i].unsqueeze(0), target[i].unsqueeze(0),
                data_range=self.data_range,
            )
            self.values.append(val.unsqueeze(0))

    def compute(self) -> torch.Tensor:
        return dim_zero_cat(self.values)  # returns ALL per-sample values

Usage:

metric = PSNRPerSample(data_range=1.0)
for batch_pred, batch_target in dataloader:
    metric.update(batch_pred, batch_target)

all_psnrs = metric.compute()  # tensor of shape (N_total,)
print(f"Mean: {all_psnrs.mean()}, Std: {all_psnrs.std()}, Min: {all_psnrs.min()}")

This pattern works for any metric — just swap peak_signal_noise_ratio for your metric of choice.

Docs: PSNR | Custom Metrics