Memorization-Informed Frechet Inception Distance (MiFID)

Module Interface

class torchmetrics.image.mifid.MemorizationInformedFrechetInceptionDistance(feature=2048, reset_real_features=True, normalize=False, cosine_distance_eps=0.1, **kwargs)

Calculate Memorization-Informed Frechet Inception Distance (MIFID).

MIFID is a improved variation of the Frechet Inception Distance (FID) that penalizes memorization of the training set by the generator. It is calculated as

\[MIFID = \frac{FID(F_{real}, F_{fake})}{M(F_{real}, F_{fake})}\]

where \(FID\) is the normal FID score and \(M\) is the memorization penalty. The memorization penalty essentially corresponds to the average minimum cosine distance between the features of the real and fake distribution.

Using the default feature extraction (Inception v3 using the original weights from fid ref2), the input is expected to be mini-batches of 3-channel RGB images of shape (3 x H x W). If argument normalize is True images are expected to be dtype float and have values in the [0, 1] range, else if normalize is set to False images are expected to have dtype uint8 and take values in the [0, 255] range. All images will be resized to 299 x 299 which is the size of the original training data. The boolian flag real determines if the images should update the statistics of the real distribution or the fake distribution.

Hint

Using this metrics requires you to have scipy install. Either install as pip install torchmetrics[image] or pip install scipy

Hint

Using this metric with the default feature extractor requires that torch-fidelity is installed. Either install as pip install torchmetrics[image] or pip install torch-fidelity

As input to forward and update the metric accepts the following input

• imgs (Tensor): tensor with images feed to the feature extractor with

• real (bool): bool indicating if imgs belong to the real or the fake distribution

As output of forward and compute the metric returns the following output

• mifid (Tensor): float scalar tensor with mean MIFID value over samples

Parameters:

• feature (Union[int, Module]) –

  Either an integer or nn.Module:

  • an integer will indicate the inceptionv3 feature layer to choose. Can be one of the following: 64, 192, 768, 2048

  • an nn.Module for using a custom feature extractor. Expects that its forward method returns an (N,d) matrix where N is the batch size and d is the feature size.

• reset_real_features (bool) – Whether to also reset the real features. Since in many cases the real dataset does not change, the features can be cached them to avoid recomputing them which is costly. Set this to False if your dataset does not change.

• normalize (bool) – Whether to normalize the input images. If True the input is expected to be in the range [0, 1] and converted to uint8. If False the input is expected to already be in the range [0, 255] and of type uint8. If a custom feature extractor is used, this argument is ignored.

• cosine_distance_eps (float) – Epsilon value for the cosine distance. If the cosine distance is larger than this value it is set to 1 and thus ignored in the MIFID calculation.

• kwargs (Any) – Additional keyword arguments, see Advanced metric settings for more info.

Raises:

• RuntimeError – If torch is version less than 1.10

• ValueError – If feature is set to an int and torch-fidelity is not installed

• ValueError – If feature is set to an int not in [64, 192, 768, 2048]

• TypeError – If feature is not an str, int or torch.nn.Module

• ValueError – If reset_real_features is not an bool

Example::

>>> from torch import randint
>>> from torchmetrics.image.mifid import MemorizationInformedFrechetInceptionDistance
>>> mifid = MemorizationInformedFrechetInceptionDistance(feature=64)
>>> # generate two slightly overlapping image intensity distributions
>>> imgs_dist1 = randint(0, 200, (100, 3, 299, 299), dtype=torch.uint8)
>>> imgs_dist2 = randint(100, 255, (100, 3, 299, 299), dtype=torch.uint8)
>>> mifid.update(imgs_dist1, real=True)
>>> mifid.update(imgs_dist2, real=False)
>>> mifid.compute()
tensor(3003.3691)

plot(val=None, ax=None)

Plot a single or multiple values from the metric.

Parameters:

• val (Union[Tensor, Sequence[Tensor], None]) – Either a single result from calling metric.forward or metric.compute or a list of these results. If no value is provided, will automatically call metric.compute and plot that result.

• ax (Optional[Axes]) – An matplotlib axis object. If provided will add plot to that axis

Return type:

tuple[Figure, Union[Axes, ndarray]]

Returns:

Figure and Axes object

Raises:

ModuleNotFoundError – If matplotlib is not installed

>>> # Example plotting a single value
>>> import torch
>>> from torchmetrics.image.mifid import MemorizationInformedFrechetInceptionDistance
>>> imgs_dist1 = torch.randint(0, 200, (100, 3, 299, 299), dtype=torch.uint8)
>>> imgs_dist2 = torch.randint(100, 255, (100, 3, 299, 299), dtype=torch.uint8)
>>> metric = MemorizationInformedFrechetInceptionDistance(feature=64)
>>> metric.update(imgs_dist1, real=True)
>>> metric.update(imgs_dist2, real=False)
>>> fig_, ax_ = metric.plot()

>>> # Example plotting multiple values
>>> import torch
>>> from torchmetrics.image.mifid import MemorizationInformedFrechetInceptionDistance
>>> imgs_dist1 = lambda: torch.randint(0, 200, (100, 3, 299, 299), dtype=torch.uint8)
>>> imgs_dist2 = lambda: torch.randint(100, 255, (100, 3, 299, 299), dtype=torch.uint8)
>>> metric = MemorizationInformedFrechetInceptionDistance(feature=64)
>>> values = [ ]
>>> for _ in range(3):
...     metric.update(imgs_dist1(), real=True)
...     metric.update(imgs_dist2(), real=False)
...     values.append(metric.compute())
...     metric.reset()
>>> fig_, ax_ = metric.plot(values)

reset()

Reset metric states.

Return type:

None