jdit.assessment

FID

jdit.assessment.FID_score(source, target, sample_prop=1.0, gpu_ids=(), dim=2048, batchsize=128, verbose=True)[source]

Compute FID score from Tensor, DataLoader or a directory``path``.

Parameters:
  • source – source data.
  • target – target data.
  • sample_prop – If passing a Tensor source, set this rate to sample a part of data from source.
  • gpu_ids – gpu ids.
  • dim –

    The number of features. Three options available.

    • 64: The first max pooling features of Inception.
    • 192: The Second max pooling features of Inception.
    • 768: The Pre-aux classifier features of Inception.
    • 2048: The Final average pooling features of Inception.

    Default: 2048.

  • batchsize – Only using for passing paths of source and target.
  • verbose – If show processing log.
Returns:

fid score

Attention

If you are passing Tensor as source and target. Make sure you have enough memory to load these data in _InceptionV3. Otherwise, please passing path of DataLoader to compute them step by step.

Example:

>>> from jdit.dataset import Cifar10
>>> loader = Cifar10(root=r"../../datasets/cifar10", batch_shape=(32, 3, 32, 32))
>>> target_tensor = loader.samples_train[0]
>>> source_tensor = loader.samples_valid[0]
>>> # using Tensor to compute FID score
>>> fid_value = FID_score(source_tensor, target_tensor, sample_prop=0.01, depth=768)
>>> print('FID: ', fid_value)
>>> # using DataLoader to compute FID score
>>> fid_value = FID_score(loader.loader_test, loader.loader_valid, depth=768)
>>> print('FID: ', fid_value)