FID

class ignite.metrics.FID(num_features=None, feature_extractor=None, output_transform=<function FID.<lambda>>, device=device(type='cpu'))

Calculates Frechet Inception Distance.

$$\text{FID} = |\mu_{1} - \mu_{2}| + \text{Tr}(\sigma_{1} + \sigma_{2} - {2}\sqrt{\sigma_1*\sigma_2})$$

where $\mu_1$ and $\sigma_1$ refer to the mean and covariance of the train data and $\mu_2$ and $\sigma_2$ refer to the mean and covariance of the test data.

More details can be found in Heusel et al. 2017

In addition, a faster and online computation approach can be found in Mathiasen et al. 2020

Remark:

This implementation is inspired by pytorch_fid package which can be found here

Note

The default Inception model requires the torchvision module to be installed. FID also requires scipy library for matrix square root calculations.

Parameters

• num_features (Optional[int]) – number of features predicted by the model or the reduced feature vector of the image. Default value is 1000.

• feature_extractor (Optional[Module]) – a torch Module for extracting the features from the input data. It returns a tensor of shape (batch_size, num_features). If neither num_features nor feature_extractor are defined, by default we use an ImageNet pretrained Inception Model and use model’s output logits as features. If only num_features is defined but feature_extractor is not defined, feature_extractor is assigned Identity Function. Please note that the model will be implicitly converted to device mentioned in the device argument.

• output_transform (Callable) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs. By default, metrics require the output as (y_pred, y) or {'y_pred': y_pred, 'y': y}.

• device (Union[str, device]) – specifies which device updates are accumulated on. Setting the metric’s device to be the same as your update arguments ensures the update method is non-blocking. By default, CPU.

Examples

For more information on how metric works with Engine, visit Attach Engine API.

from collections import OrderedDict

import torch
from torch import nn, optim

from ignite.engine import *
from ignite.handlers import *
from ignite.metrics import *
from ignite.metrics.clustering import *
from ignite.metrics.regression import *
from ignite.utils import *

# create default evaluator for doctests

def eval_step(engine, batch):
    return batch

default_evaluator = Engine(eval_step)

# create default optimizer for doctests

param_tensor = torch.zeros([1], requires_grad=True)
default_optimizer = torch.optim.SGD([param_tensor], lr=0.1)

# create default trainer for doctests
# as handlers could be attached to the trainer,
# each test must define his own trainer using `.. testsetup:`

def get_default_trainer():

    def train_step(engine, batch):
        return batch

    return Engine(train_step)

# create default model for doctests

default_model = nn.Sequential(OrderedDict([
    ('base', nn.Linear(4, 2)),
    ('fc', nn.Linear(2, 1))
]))

manual_seed(666)

metric = FID(num_features=1, feature_extractor=default_model)
metric.attach(default_evaluator, "fid")
y_true = torch.ones(10, 4)
y_pred = torch.ones(10, 4)
state = default_evaluator.run([[y_pred, y_true]])
print(state.metrics["fid"])

0.0

Note

The default torchvision model used is InceptionV3 pretrained on ImageNet. This can lead to differences in results with pytorch_fid. To find comparable results, the following model wrapper should be used:

import torch.nn as nn

# wrapper class as feature_extractor
class WrapperInceptionV3(nn.Module):

    def __init__(self, fid_incv3):
        super().__init__()
        self.fid_incv3 = fid_incv3

    @torch.no_grad()
    def forward(self, x):
        y = self.fid_incv3(x)
        y = y[0]
        y = y[:, :, 0, 0]
        return y

# use cpu rather than cuda to get comparable results
device = "cpu"

# pytorch_fid model
dims = 2048
block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
model = InceptionV3([block_idx]).to(device)

# wrapper model to pytorch_fid model
wrapper_model = WrapperInceptionV3(model)
wrapper_model.eval();

# comparable metric
pytorch_fid_metric = FID(num_features=dims, feature_extractor=wrapper_model)

Important, pytorch_fid results depend on the batch size if the device is cuda.

New in version 0.4.6.

Methods

compute	Computes the metric based on its accumulated state.
reset	Resets the metric to its initial state.
update	Updates the metric's state using the passed batch output.

compute()

Computes the metric based on its accumulated state.

By default, this is called at the end of each epoch.

Returns

the actual quantity of interest. However, if a Mapping is returned, it will be (shallow) flattened into engine.state.metrics when completed() is called.

Return type

Any

Raises

NotComputableError – raised when the metric cannot be computed.

reset()

Resets the metric to its initial state.

By default, this is called at the start of each epoch.

Return type

None

update(output)

Updates the metric’s state using the passed batch output.

By default, this is called once for each batch.

Parameters

output (Sequence[Tensor]) – the is the output from the engine’s process function.

Return type

None