From 5c990d51a87edd88b65744ffeb8b9256aec8cbfb Mon Sep 17 00:00:00 2001 From: nightcityblade Date: Mon, 22 Jun 2026 23:20:06 +0800 Subject: [PATCH] docs: clarify FID default feature dimension --- ignite/metrics/gan/fid.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/ignite/metrics/gan/fid.py b/ignite/metrics/gan/fid.py index 4a1aee350577..9dba75798a5d 100644 --- a/ignite/metrics/gan/fid.py +++ b/ignite/metrics/gan/fid.py @@ -83,6 +83,9 @@ class FID(_BaseInceptionMetric): Args: num_features: number of features predicted by the model or the reduced feature vector of the image. Default value is 1000. + The default Inception model uses its 1000-dimensional output logits as features. To match + `pytorch_fid` results, provide a feature extractor that returns the 2048-dimensional pool3 + features instead; see the example below. feature_extractor: 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