The Role of ImageNet Classes in Fréchet Inception Distance
Authors: Tuomas Kynkäänniemi, Tero Karras, Miika Aittala, Timo Aila, Jaakko Lehtinen
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We investigate a root cause of these discrepancies, and visualize what FID looks at in generated images. We show that the feature space that FID is (typically) computed in is so close to the Image Net classifications that aligning the histograms of Top-N classifications between sets of generated and real images can reduce FID substantially without actually improving the quality of results. Thus, we conclude that FID is prone to intentional or accidental distortions. |
| Researcher Affiliation | Collaboration | Tuomas Kynk a anniemi Aalto University tuomas.kynkaanniemi@aalto.fi Tero Karras NVIDIA tkarras@nvidia.com Miika Aittala NVIDIA maittala@aalto.fi Timo Aila NVIDIA taila@aalto.fi Jaakko Lehtinen Aalto University & NVIDIA jlehtinen@nvidia.com |
| Pseudocode | Yes | Algorithm 1 shows the pseudocode for our resampling method. |
| Open Source Code | Yes | Code is available at https://github.com/kynkaat/role-of-imagenet-classes-in-fid. |
| Open Datasets | Yes | We study this in the same context as Sauer et al. (2021) by training a Projected Fast GAN (Liu et al., 2021; Sauer et al., 2021) that uses an Image Net pre-trained Efficient Net (Tan & Le, 2019) as a feature extractor of the discriminator, and compare it against Style GAN2 in FFHQ.6 |
| Dataset Splits | Yes | Following standard practice, we compute FID against the training set, using 50k randomly chosen real and generated images and the official Tensor Flow version of Inception-V3.2 |
| Hardware Specification | Yes | We use 32GB NVIDIA Tesla V100 GPU to run our resampling experiments. |
| Software Dependencies | No | The paper mentions using 'Tensor Flow' and 'Py Torch' frameworks and specific models/codebases like 'Inception-V3', 'Res Net-50', 'CLIP', 'Style GAN2', but it does not provide specific version numbers for these software dependencies (e.g., 'PyTorch 1.9'). |
| Experiment Setup | Yes | In our experiments, we use Style GAN2 auto-config trained in 256 x 256 resolution without adaptive discriminator augmentation (ADA). The only exception is AFHQ-V2 DOG, where we enable ADA and train in 512 x 512 resolution. [...] We use learning rate α = 10.0 when optimizing pre-logits features and α = 5.0 when optimizing logits or binarized class probabilities. We optimize the weights until convergence, which typically requires 100k iterations. |